a sociology essay
cesare07
1-s2.0-S0167629611000701-main.pdf
S
T a
b
a
A R R A A
J I Q Q
K R A C
1
p e i i u e h
a c S W r P a a
S
( e
0 d
Journal of Health Economics 30 (2011) 987– 999
Contents lists available at ScienceDirect
Journal of Health Economics
j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / e c o n b a s e
chool buses, diesel emissions, and respiratory health�
imothy K.M. Beatty a,∗, Jay P. Shimshack b,1
Department of Applied Economics, University of Minnesota, St Paul, MN 55108, United States Department of Economics, Tulane University, New Orleans, LA 70118, United States
r t i c l e i n f o
rticle history: eceived 7 July 2009 eceived in revised form 28 May 2011 ccepted 31 May 2011 vailable online 21 June 2011
EL classification: 18
a b s t r a c t
School buses contribute disproportionately to ambient air quality, pollute near schools and residential areas, and their emissions collect within passenger cabins. This paper examines the impact of school bus emissions reductions programs on health outcomes. A key contribution relative to the broader literature is that we examine localized pollution reduction programs at a fine level of aggregation. We find that school bus retrofits induced reductions in bronchitis, asthma, and pneumonia incidence for at-risk populations. Back of the envelope calculations suggest conservative benefit–cost ratios between 7:1 and 16:1.
© 2011 Elsevier B.V. All rights reserved.
58 53
eywords: espiratory health ir pollution lean school bus
l i m
o d W
. Introduction
Pollution regulations are controversial, and economists and olicy-makers debate their efficiency and cost effectiveness. Most conomic evaluations of environmental quality examine the mpact of ambient air pollution on health outcomes.2 These stud- es are important for understanding national policy, but they are
nlikely to shed light on programs targeting localized pollution xposure because widely dispersed ambient air quality monitors ide large local differences in pollution. Moreover, localized pol-
� Seniority of authorship is shared. We thank program managers Michael Boyer nd Charlie Stansel and installation contractor Steve Reihs for data and helpful dis- ussions. We are grateful to seminar participants at Tulane, Chicago, York, Guelph, askatchewan, Lausanne, the AERE Sessions of the AEA meetings, the Colorado orkshop on Environmental Economics, the AERE workshop on health and the envi-
onment, and the Canadian Health Economics Study Group. We also thank Matthew litch and Emily O’Connor for excellent research assistance. The views in this paper re solely those of the authors, and errors of omission or commission are solely the uthors’ responsibility. ∗ Corresponding author. Tel.: +1 612 208 9702.
E-mail addresses: [email protected] (T.K.M. Beatty), [email protected] (J.P. himshack). 1 Tel.: +1 504 862 8353. 2 Notable studies in this vein include Chay et al. (2003), Chay and Greenstone
2003), Neidell (2004), Currie and Neidell (2005), Currie et al. (2009a,b), and Janke t al. (2009).
u r s fi 2 r i d s t 2 c r t p y r t l
167-6296/$ – see front matter © 2011 Elsevier B.V. All rights reserved. oi:10.1016/j.jhealeco.2011.05.017
ution policies may be especially effective at the margin; the basic nsight is that abating pollution where people live, work, and study
ay return large benefits per dollar of cost. This paper studies the health impacts and cost effectiveness
f a new localized emissions reduction program that retrofits iesel school buses with aggressive pollution control technologies. e focus on school buses for several reasons. First, the partic-
late matter and air toxics common in diesel pollution may be esponsible for as many as 15,000 premature deaths annually. In ome regions, diesel mortality levels are similar to those of traf- c accidents and second-hand smoke (CA Air Resources Board, 002). Second, school buses are ubiquitous. In 2005, buses car- ied nearly 25 million children between 5 and 6 billion miles n the United States. Third, school buses are disproportionately irty. The national average bus age is over 9 years, and estimates uggest that the average school bus emits twice as many con- aminants per mile as the average tractor-trailer truck (Monahan, 006). Fourth, school bus pollution has important local effects. In ontrast to most diesel vehicles, buses primarily travel through esidential areas and so individuals who are sensitive to pollu- ion may be affected by bus emissions where they live. Diesel air ollutants also collect inside of passenger cabins and in school-
ards, so school-aged children may be further impacted. Recent esearch finds that within-bus concentrations of particulate mat- er and air toxics were 4–12 times higher than ambient pollution evels.
9 l of He
r d m O s t g a c n h d i e b e a p
c i t c c r w i a A o q
w a e T s a t c
d o c i “ o t o t a i
d r g m i a c
(
2
2
p p o A l
c t p t 2 a i s m o
b t v l i W i a
r b 1 l i m e A v c c t t 2
2
n t o l p 2 w l i i
88 T.K.M. Beatty, J.P. Shimshack / Journa
Despite the potentially large health benefits of school bus etrofit programs, we know very little about their impacts. The earth of empirical studies stems from at least two challenges. First, any of these programs are relatively new and data are scarce. ur study uses a comprehensive dataset on bus retrofits from the
tate of Washington, and detailed information includes retrofit ype, retrofit date, and retrofit cost. We combine the novel pro- ram data with comprehensive morbidity and demographic data t the school district level. Second, statistical identification can be hallenging. Health outcomes may drive program adoption or sig- ificant unobservables may influence both program adoption and ealth outcomes. We exploit a natural experiment and employ a ifferences-in-differences research design to help isolate causal
mpacts. Treatment school districts retrofitted eligible buses by the nd of our sample period and control school districts retrofitted no uses by the end of our sample period. Identification exploits differ- nces in adoption timing, rather than the adoption decision itself, s nearly all non-adopters began retrofits shortly after our sample eriod ends.
We find that school bus retrofits induced statistically signifi- ant and large reductions in bronchitis, asthma, and pneumonia ncidence for both children and adults with chronic respira- ory conditions. Empirical magnitudes are typically larger for hildren’s health outcomes than for chronically ill adult out- omes. Results, especially for asthma and bronchitis incidence, are obust to several falsification and sensitivity checks. Most notably, hile adopters and non-adopters experienced differential trends
n health outcomes over the retrofit period, adopters and non- dopters experienced comparable trends in the pre-retrofit period. dopters and non-adopters also experienced comparable trends ver the retrofit period for illnesses plausibly unrelated to air uality.
To put our results in context, we combine our empirical results ith the cost-of-treatment health valuation literature and perform
back of the envelope benefit–cost analysis. We conservatively stimate program benefits between 7 and 16 times program costs. his interpretation suggests that if the many states not aggres- ively pursuing school bus retrofits were to do so, social benefits re potentially large. Buses are an inexpensive and safe means of ransport (in an accident sense), but our results suggest that they ould be made safer (in the broadest sense) at modest cost.
We believe our analysis makes three contributions. First, our ata and methods permit the first empirical economic assessment f the impact of school bus retrofit programs on morbidity out- omes. Second, we show that local pollution policies can significantly mpact public health, and that these programs may produce a large bang per buck”. Third, our results may provide additional evidence n the broader effects of air pollution on health.3 We cannot directly race retrofit programs to lower ambient pollution levels, since ur spatial unit of analysis is significantly smaller than the spa- ial distribution of pollution monitors. However, we do show that
program targeting air pollution exposure significantly reduces llnesses plausibly related to air quality (and only those illnesses).
The paper proceeds as follows. Section 2 provides institutional etail on school buses, diesel emissions, retrofit programs, and espiratory health. Section 3 describes our unique retrofit, demo- raphic, weather, and health data. Section 4 presents our empirical ethods and Section 5 presents key results. Section 6 explores our
dentification and other empirical assumptions. Section 7 provides
conservative back of the envelope benefit–cost assessment and oncludes.
3 In this sense, our paper is in the spirit of recent work by Currie and Walker 2011), Schlenker and Walker (2010), and Moretti and Neidell (2011).
a o i
o 2
alth Economics 30 (2011) 987– 999
. Background
.1. School buses and diesel emissions
Diesel emissions make up a substantial portion of ambient air ollution. Particulate matter from diesel engines accounts for 26 ercent of total air pollution from fuel combustion and 66 percent f particulate air pollution from on-road sources (American Lung ssociation, 2008). On-road mobile sources emissions are often the
argest single source of air pollution in a region. School buses are common and polluting. In 2005, 25 million
hildren traveled between 5 and 6 billion miles on school buses in he United States. Median routes for many of the buses in our sam- le were approximately 6 miles each way. The average child riding hese buses spent nearly 45 min per day on the bus (Adar et al., 008). The average bus age in the United States is over 9 years, nd the average is substantially higher in many states. Research ndicates that, per mile, school buses are twice as polluting as emi trucks. The average bus emits nearly 15 pounds of particulate atter and approximately 400 pounds of smog-forming nitrogen
xides and hydrocarbons per year (Monahan, 2006). In addition to affecting background ambient air quality, school
us diesel pollution has important local effects. Since buses travel hrough residential areas, their emissions may impact at-risk indi- iduals at a neighborhood level. Research indicates that people iving near roads are exposed to pollution levels that are signif- cantly greater than broad ambient levels (Pearson et al., 2000;
ilhelm and Ritz, 2003). Further, emissions from groups of buses dling outside schools can concentrate pollution within schoolyards nd schools themselves.
Pollution exposure may be particularly high for children who ide buses. Air pollution concentrations inside mobile sources may e as much as 10 times background ambient levels (Shikiya et al., 989; Chan et al., 1991; Lawryk et al., 1996). Diesel emissions col-
ect through mechanisms such as direct flows from leaks or cracks n the crankcase or exhaust system. Such leaks or cracks may be
ore common in school buses than in other vehicles, as school bus ngines are often less regularly maintained (Behrentz et al., 2004). dar et al. (2008) installed pollution monitors in a subset of the ehicles in our study. Their estimates suggest that within-bus con- entrations of harmful particulates were more than twice roadway oncentrations and 4 times ambient levels. Related studies found hat within-school bus concentrations of particulate matter and air oxics were 4–12 times higher than ambient levels (Wargo et al., 002; Sabin et al., 2005).
.2. Diesel emissions and respiratory health
Diesel fumes contain high levels of particulate matter, air toxics, itrogen oxides, and hydrocarbons. Even at relatively low levels, hese contaminants are known to exacerbate or cause asthma and ther respiratory ailments. Daily changes in air pollution have been inked to daily changes in mortality, hospital admissions, and other ublic health indicators (Spix et al., 1998; Brunekreef and Holgate, 002; Dockery, 2009). Air toxics defined broadly are associated ith asthma, lung inflammation, coughing, wheezing, and reduced
ung function (Peden, 2002). The fine particulate matter common n diesel emissions is linked to reduced lung function and increased ncidences of pneumonia (Cohen and Nikula, 1999; McCreanor et l., 2007). Nitrogen oxides cause ground-level ozone, and high zone concentrations are associated with aggravated respiratory
llness and increased respiratory symptoms.
All children are potentially susceptible to the adverse effects f particulates and ozone (Committee on Environmental Health, 004). Impacts on children are due to ongoing physiological res-
l of He
p a h b g i i e i a
2
c a u b y r m y e L
r g A o c a e A a g
o d d a b 5
t a a t C d p t
2
n m p
d d t c
a p t
f t p a t g r r w g s o v
w t w D w y a r u t r
2
s l i b l T c S c p
r u s o b o e a d
T.K.M. Beatty, J.P. Shimshack / Journa
iratory development, smaller average lung size, and increased ctivity levels. The fine particulates contained in diesel exhaust ave been shown to contribute significantly to children’s mor- idity and mortality, especially reduced lung function and lung rowth (Gauderman et al., 2000, 2004). Diesel fumes can also ncrease the severity of children’s asthma and can induce asthma n otherwise healthy children (McConnell et al., 2002; Peters t al., 2004). In contrast to children’s morbidity, adult morbidity s pronounced only for individuals with pre-existing respiratory ilments.
.3. Clean school bus initiatives
This study examines the impacts of the Washington State lean school bus program. Washington was an aggressive early dopter of retrofits, but their program is otherwise similar to those nder consideration or under way in other states. State senate ill ESSB6072 provided $5 million in annual funding for the 5 ears spanning 2003–2008. The legislation’s primary goal was to etrofit older school buses with modern pollution control equip- ent. Legislative priorities included targeting buses with model
ears prior to 1994, since retrofitting older buses yields greater missions reductions than retrofitting newer buses (Boyer and yons, 2004).
Under ESSB6072, the state offered school districts complete etrofit rebates. Further, a small number of districts were eli- ible for federal funding from the US Environmental Protection gency’s Clean School Bus USA Program. Washington’s Department f Ecology (DOE) and the state’s seven air quality control agen- ies administered the program. Washington emphasized retrofits, nd approximately 88 percent of expenditures were devoted to quipment installations in the program’s first operational year. bout 7 percent of expenditures were devoted to administration nd less than 5 percent went to low sulfur diesel fuel pro- rams.
Adopting districts typically began retrofits by installing diesel xidation catalysts (DOCs). DOCs are add-on ceramic substrate evices that catalyze chemical reactions in emissions, breaking own harmful pollutants into less harmful substances. On aver- ge, DOCs are expected to reduce particulates from these vehicles y approximately 20–30 percent and hydrocarbons by as much as 0–70 percent.
More recently, the program coupled DOC retrofits with addi- ional crankcase ventilation filter (CCV) retrofits. The primary dvantage of CCVs is that they reduce within-bus pollution. CCVs re add-on devices that filter unburned fuel and blow-by gases from he crankcase, the largest chamber of most diesel engines. When CVs are installed, they are almost always coupled with diesel oxi- ation catalysts. CCVs are expected to reduce PM levels by 10–20 ercent more than DOCs alone and may reduce within-bus concen- rations by much greater margins.4
.4. Retrofit timing and scope
According to interviews with program staff members, the total umber of districts retrofitting at a given time was largely deter- ined by external budget considerations. While the state initially
romised $5 million in regular installments, actual funding arrived
4 After our sample period ends, the state moved towards greater use of low sulfur iesel fuel and greater use of diesel particulate filters. Both were extremely rare uring our sample period, and remain relatively rare. However, increased pene- ration of these pollution control methods suggests that the marginal benefits of onventional retrofits will decrease over time.
H
m e b i t m a d
alth Economics 30 (2011) 987– 999 989
t irregular and unpredictable intervals. Thus, the actual timing of rogram adoption across districts was exogenous to the districts hemselves.
The order in which districts participated was driven by many actors. Program managers approached school district adminis- rators early in our sample period. Most administrators offered ermission to proceed with retrofits provided their bus fleet man- gers also agreed. A few administrators were slow in responding o agency requests. Conversations with school administrators sug- ested that unrelated administrative burdens may explain slow esponse during the sample period. Some districts’ mechanics, once eached, immediately agreed to start retrofits. Many, however, anted to see evidence that other districts had adopted the pro-
ram without substantial disruptions to their buses or normal work chedules. These fleet managers eventually came on board, but ften required sustained persuasion over time including repeated isits by DOE staff.
Once a district began the retrofit process, implementation scope as largely determined by available technologies and bus charac-
eristics. Program managers and contractors maintained that buses ith model years prior to 1982 were largely unsuited for retrofits. iesel oxidation catalysts alone were most appropriate for buses ith model years from 1982 to 1987. For most buses with model
ears later than 1987, DOCs coupled with CCVs were the appropri- te technologies. Retrofitting newer buses with DOCs or CCVs was arely cost effective. Since technology and bus fleet characteristics sually determined the number of buses retrofitted and retrofit ype, program scope was essentially exogenous to the district. As a ule, adopters retrofitted all suitable buses.
.5. Linking retrofits and health
To summarize much of Section 2, school bus retrofits should ubstantially reduce emissions of diesel-related toxics, particu- ates, and other contaminants. These emissions reductions should n turn improve the respiratory health of at-risk individuals. First, us tailpipe emissions impact background air quality as well as
ocalized air quality in residential neighborhoods and near schools. he retrofits may therefore reduce adverse respiratory health out- omes for both adults with chronic respiratory illness and children. econd, since bus emissions may concentrate within passenger abins and in schoolyards, retrofits should further reduce children’s ollution exposure.
Two details bear noting. First, the relative health benefits of etrofits for children versus adults with chronic conditions are nknown a priori. While children may experience reduced expo- ure from retrofits through improvements in both inside-bus and utside-bus air quality, individuals with chronic conditions may e especially sensitive to changes in contaminant exposure. Sec- nd, health outcomes may respond quickly to changes in diesel missions exposure. The medical literature indicates that hospital dmissions and other public health outcomes respond to day-to- ay variations in air pollution (Spix et al., 1998; Brunekreef and olgate, 2002; Dockery, 2009).
Section 2 also raises an additional issue. While many ele- ents that influenced adoption timing and scope were plausibly
xogenous, some factors that influenced adoption status may have een non-random and correlated with other elements influenc-
ng health. A plausible research design must therefore be robust
o some non-random assignment. We adopt a natural experi-
ent estimation strategy that is robust to endogeneity in levels nd we explore the validity of our identifying assumptions in etail.
9 l of He
3
t W E d W f a
4 o b d h m ( B a u
f R o m d m s o d e
a a m c b C p p N a
w m a a o c m i
4
f
e w
r b
p s w c T f b t m
4
i a a t e a e o d d p
g t t r b p i
c a g s I c a t r t s r t c t c
t
�
90 T.K.M. Beatty, J.P. Shimshack / Journa
. Data
Our data allow the first large-scale empirical assessment of he health benefits stemming from school bus retrofit programs.
e use program data from the Washington State Department of cology and the Puget Sound Clean Air Agency. We use hospital ischarge data from the Washington State Department of Health. e also augment retrofit and health data with demographic data
rom the National Center of Educational Statistics and temperature nd precipitation data from the US Historical Climatology Network.
Our administrative retrofit database consists of approximately 000 buses in 53 school districts of the Puget Sound area. We focus n the state of Washington because it is an innovator in school us programs. We examine specifically the Puget Sound region ue to extensive administrative data collection and the relative omogeneity of districts. For each bus in the area, we observe infor- ation related to equipment installations such as the retrofit type
DOC/CCV), the date of the retrofit, and the cost of the retrofit. us-specific data are aggregated to the district level to construct
fleet profile consisting of the cumulative share of buses having ndergone each type of retrofit.
Data on health outcomes from 1996 to 2006 were extracted rom the Washington State Comprehensive Hospital Abstract eporting System (CHARS). CHARS data include a complete record f hospital inpatient discharges. Each observation consists of treat- ent date, patient age, patient sex, patient home zip code, and a
etailed diagnosis code. Illness-specific data is aggregated to the onth level by summing the number of illnesses in each diagno-
is code for each zip code over individual days. We obtain a profile f respiratory illness incidence for at-risk populations (which we efine to be children and adults with chronic conditions) within ach zip code each month.
Since we focus on an air pollution control program, our primary nalysis emphasizes the major respiratory ailments: bronchitis, sthma, pleurisy, and pneumonia.5 Bronchitis, asthma, and pneu- onia outcomes for at-risk populations have been linked to air
ontaminants. To the best of our knowledge, pleurisy has not een linked to the pollutants common in diesel exhaust. However, HARS diagnosis codes aggregate pneumonia with pleurisy.6 Any otential measurement error from including pleurisy biases our rogram impacts towards zero and reduces statistical precision. o related measurement error is present for the analysis of asthma nd bronchitis.
We match all data, including retrofit, health, demographic, and eather data, at the school-district by month level. Details of the atching procedure are provided in the data appendix. Bus fleets
nd retrofit programs are managed at the district level, so data nalysis at spatial scales finer than school districts (e.g. zip codes r schools) would artificially inflate precision. Some health out- omes are rare events, so data analysis at temporal scales finer than onths (e.g. days or weeks) would not ensure meaningful variation
n outcome measures.
. Methods
Our goal is to assess the impact of retrofits on health outcomes or those districts that retrofitted eligible buses. Since we have a
5 The relevant diagnosis-related group (DRG) codes are 089, 091, 096, and 098. For ach disease category, we consider the diagnosis group for children and for adults ith chronic conditions. 6 Pleurisy is an inflammation of the mucus membrane enveloping the lungs and
ib cage. Individuals suffering from pleurisy typically report difficult and painful reathing.
o t i w
y
w n
alth Economics 30 (2011) 987– 999
anel dataset, it is perhaps tempting to assess these impacts by imply comparing adopting districts’ pre-retrofit health outcomes ith post-retrofit health outcomes. However, we cannot attribute
hanges in health outcomes after the retrofits to the retrofits alone. he clearest way to isolate causal effects of the retrofits, accounting or confounding factors, would be to examine outcome differences etween randomly assigned adopter districts and non-adopter dis- ricts over time. While this is not possible ex-post, our methods
imic this general structure.
.1. Two-period, two-group difference-in-differences
Our first empirical design is a standard two-period difference- n-differences approach. Here, we examine differential trends for dopter districts and non-adopter districts over time. If non- dopters provide information on the expected health outcome rends for adopters had adoption not occurred, the quasi- xperiment afforded by the difference-in-differences in outcomes cross adopting and non-adopting districts should remove the ffect of confounding factors and isolate the effect of bus retrofits n health outcomes. Note that the analysis that follows exploits ifferences in adoption timing rather than differences in adoption ecisions, as most non-adopters began retrofits after our sample eriod ended.
Some notation is helpful in presenting the estimator. Two roups g ∈ [a,b] experience health outcomes y in two-periods
∈ [1,2]. YR is the health outcome in the presence of the retrofit reatment and YNR is the health outcome in the absence of the etrofit treatment. Group a is the non-adopter control group, group
is the adopter treatment group, t = 1 is the pre-treatment time eriod, and t = 2 is the post-treatment time period. The treatment
s observed only if g = b and t = 2. In our two-period difference-in-differences context, time t = 1
orresponds to 2002, the year before retrofits began in earnest, nd time t = 2 corresponds to 2006, the last sample year. Group
= b typically corresponds to the 34 districts that had completed ignificant retrofits by the beginning of the 2006/2007 school year. n practice, these treatment districts retrofitted more than 30 per- ent of their total bus fleet with diesel oxidation catalysts (DOCs) nd/or crankcase ventilation filters (CCVs) by the beginning of he 2006/2007 school year. Alternatively, group g = b may cor- espond to the 8 district subset of the broader treatment group hat had significantly augmented DOC retrofits with more aggres- ive CCV retrofits. In practice, these alternative treatment districts etrofitted more than 10 percent of their total bus fleet with CCVs by he beginning of the 2006/2007 school year. In all cases, group g = a orresponds to the 9 districts that retrofitted no buses by the start of he 2006/2007 school year. Districts that are neither treatment nor ontrol are dropped from the main two-period, two-group analysis.
In the most basic model, the average treatment effect on the reated in the presence of retrofits can be written as:
DID = E[Y R b2] − E[Y
NR b2 ] = E[Yb2] − E[Yb1] − (E[Ya2] − E[Ya1]) (1)
A regression analog to this model allows us to control for bservable differences in the distribution of characteristics of the reatment and control groups. This regression model is parameter- zed following the difference-in-differences literature, and can be
ritten as:
= ̨ + ıt + �g + ˇt · g + �X + ε, (2)
here y continues to represent health outcomes. In practice, we ormalize respiratory health outcomes as the ratio of illness cases
T.K.M. Beatty, J.P. Shimshack / Journal of Health Economics 30 (2011) 987– 999 991
Table 1 Summary statistics for the baseline (pre-adoption) year.
Characteristics The 53 full sample districts
The 9 non-adopter districts
The 34 adopter districts
Difference: adopter vs. non-adopter
p-Value for difference
Student population 11,239 2335 10,597 −8262 0.01 Per capita income 24,841 23,609 24,166 −557 0.79 Percent of pop. below poverty
line 0.074 0.086 0.075 0.011 0.48
School staff members per student
0.115 0.120 0.100 0.020 0.15
Percent white 0.799 0.777 0.737 0.040 0.42 School bus fleet age 8.92 8.94 9.59 −0.65 0.52 Children’s bronchitis and
asthma cases (per 100,000 children)
22.83 14.22 23.87 −9.64 0.01
Chronic adult bronchitis and asthma cases (per 100,000 adults)
2.38 0.68 2.73 −2.04 0.01
Children’s pneumonia and pleurisy cases (per 100,000 children)
9.28 3.29 11.11 −7.82 0.01
t v s i e � p e o
a a a s a p r a l t c r e t i i d c t a t
J s 8 d m
c t
t v w m (
a W o d t ( a
a a d i i p m p p m e s r W r
4
Chronic adult pneumonia and pleurisy cases (per 100,000 adults)
19.27 14.32
o affected population numbers.7 X represents a vector of control ariables including per capita income, poverty status, racial compo- ition, and school staff per student ratios. ε represents the standard diosyncratic disturbance term. The coefficient ı represents the ffect of time on health outcomes for the non-treated group and
represents the effect of the treatment on health outcomes in the re-treatment period. ̌ is the coefficient of interest, and it is our stimator for the difference-in-differences effect of the treatment n the treated (the regression analog of �DID).
The key difference-in-difference assumption is that non- dopters would experience the same trends in health outcomes s adopters absent the treatment, after conditioning on observ- bles. Table 1 shows that demographic characteristics are generally imilar across adoption classification groups. Adopters and non- dopters have statistically indistinguishable per capita income, overty status, racial composition, and school staff per student atios. Moreover, bus fleet average age is similar between adopter nd non-adopter districts. Treatment districts are systematically arger, but we analyze health outcome variables scaled by popula- ion size. While demographics are generally similar across adoption lassification groups, the final rows of Table 1 indicate that respi- atory health treatments are significantly more common among arly adopting districts. This suggests that it is possible that adop- ion timing across districts was not purely random. However, this s not necessarily a concern. Our difference-in-differences strategy s robustness to endogeneity in health levels; initial health may iffer across groups as long as expected trends (conditional on the ovariates) are expected to be similar across the groups absent the reatment. We test this key assumption for pre-treatment periods, nd our later sensitivity section devotes considerable attention to he plausibility of all identifying assumptions in the model.
In all two-period difference-in-differences analyses, we omit uly and August from the sample since schools were not in ses- ion during those months. The resulting general dataset consists of
60 observations: we observe 34 treatment districts and 9 control istricts over 10 months in 2002 (before the retrofits) and over 10 onths in 2006 (after the retrofits). A parallel dataset for examining
7 For children, the ratio is respiratory illness cases among children per 100,000 hildren. For adults, the ratio is respiratory cases among adults with chronic condi- ions per 100,000 adults.
d F a i v a l w r
20.38 −6.06 0.01
he impact of retrofits that emphasize more aggressive crankcase entilation filter (CCV) installations consists of 340 observations: e observe 8 treatment districts and 9 control districts over 10 onths in 2002 (before the retrofits) and over 10 months in 2006
after the retrofits). In all two-period analyses, we cluster standard errors to
llow for arbitrary within-group correlations at the district level. e cluster over individual districts as they represent the unit
f analysis. Clustered standard errors are then used to test ifference-in-difference hypotheses against one-sided alterna- ives. We hypothesize that our key difference-in-differences results the impacts of retrofits on illness outcomes) will be negative, so the ppropriate alternative hypothesis is a non-negative coefficient.
Results from two-period, two-group difference-in-differences pproaches are presented in Table 2. Clustered standard errors re reported in parentheses below the coefficient estimates. We efer interpretation of our key two-period, two-group difference-
n-difference coefficients to the next section, but we note here the mpact of controls on health outcomes. Districts with higher pro- ortions of their population below the poverty line experienced ore respiratory ailments on average, especially for pleurisy and
neumonia. Districts with greater staff-student ratios and higher ercentages of white students experienced fewer respiratory ail- ents on average. Interestingly, districts with older bus fleets
xperienced statistically more respiratory ailments on average, upporting the basic hypothesis that diesel school buses impact espiratory health for children and adults with chronic conditions.
armer and wetter districts, on average, experienced fewer respi- atory ailments on average.
.2. Multiple period approaches
We believe the two-period, two-group difference-in- ifferences approach outlined above offers several advantages. irst, it facilitates a transparent econometric analysis and gener- tes readily interpretable empirical estimates (i.e. ‘what happens f an average district retrofits its eligible buses?’). Second, the alidity of the identifying assumptions can be more directly
ssessed. Third, the two-period, two-group approach imposes very ittle parametric structure on the problem. Imposing linearity,
hich implies the relationship between a district’s share of buses etrofitted and health outcomes is exactly the same at every
9 9
2 T.K
.M .
B ea
tty ,
J.P .
Sh im
sh a ck
/ Jo
u rn
a l
o f
H ea
lth E
co n
o m
ics 3
0 (2
0 1
1 ) 9
8 7
– 9
9 9
Table 2 Results: two-period, two-group difference-in-differences.
All retrofits bronchitis & asthma cases All retrofits pleurisy & pneumonia cases CCV retrofits bronchitis & asthma cases CCV retrofits pleurisy & pneumonia cases
All at-risk groups
Children Adults with chronic illness
All at-risk groups
Children Adults with chronic illness
All at-risk groups
Children Adults with chronic illness
All at-risk groups
Children Adults with chronic illness
Diff-in-diff −2.88** −5.35 −1.80** −3.94 −4.14** −3.20 −4.33*** −10.2** −1.83** −5.44* −5.06* −5.13 (1.47) (5.22) (0.85) (3.30) (2.34) (4.34) (1.66) (5.96) (0.79) (3.64) (3.30) (4.41)
Treatment group 2.55 3.13 1.81** 4.21 6.78*** 3.16 3.26** 7.17 1.28* 0.76 6.46** −1.44 (1.64) (5.60) (0.87) (2.57) (2.27) (3.00) (1.35) (4.34) (0.69) (3.40) (3.00) (4.50)
Post-treatment 0.62 −1.47 1.29* 2.91 −0.37 3.76 0.78 −0.98 1.30 2.97 −0.41 3.92 (1.31) (5.02) (0.73) (3.15) (2.07) (4.01) (1.37) (5.24) (0.79) (3.62) (2.20) (4.21)
Per capita income −0.00 0.00 0.00 0.00 0.00 0.00 0.00** 0.00** 0.00 0.00 0.00*** 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)
% below poverty 12.2 46.8 4.18 95.3** 20.9 120** 33.5 180 4.90 35.9 2.68 43.0 (23.8) (49.7) (22.5) (45.5) (38.8) (55.6) (42.0) (139) (17.4) (115) (50.6) (144)
Staff per student −23.9 −80.8* −2.50 −69.9** −26.8 −88.8** −28.7 −123 −5.02 −39.1 −11.6 −50.0 (15.4) (42.2) (12.2) (29.1) (19.3) (34.5) (22.5) (73.7) (9.89) (52.9) (28.4) (66.3)
Percent white −5.76* −28.3*** −0.77 −1.98 1.95 4.17 −12.2* −39.6* −3.25 −9.46 −10.5 −7.92 (2.92) (9.32) (1.66) (6.44) (5.00) (7.97) (6.15) (19.0) (2.79) (15.4) (6.24) (20.2)
Average bus age 0.29** 0.90** 0.09** 1.11*** 0.40 1.37** 0.56 1.21 0.18 2.29** 0.73 2.94*
(0.11) (0.40) (0.04) (0.41) (0.25) (0.61) (0.36) (1.00) (0.18) (1.06) (0.46) (1.50) Temperature −0.02*** −0.06*** −0.01*** −0.04*** −0.04*** −0.03** −0.03*** −0.08*** −0.01** −0.04** −0.03*** −0.05
(0.01) (0.01) (0.00) (0.01) (0.01) (0.01) (0.01) (0.02) (0.00) (0.02) (0.01) (0.03) Precipitation −0.01* −0.01 −0.00*** −0.00 −0.00 0.00 −0.00** −0.00 −0.00 −0.00 −0.00 −0.00
(0.00) (0.01) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.01) Constant 21.8*** 68.8*** 6.05** 19.7** 23.7*** 17.4 18.5** 53.3** 5.64*** 21.3 12.9** 24.5
(4.16) (12.1) (2.42) (9.52) (6.65) (12.2) (6.61) (24.1) (1.90) (17.2) (5.39) (22.4)
Obs. 860 860 860 860 860 860 340 340 340 340 340 340 F-Statistic 14.2 12.0 10.1 10.2 10.7 8.69 13.6 10.4 8.87 8.61 21.7 6.19 Prob > F 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Notes: Clustered standard errors appear in parentheses. DID coefficients tested against one-sided alternatives. * Indicates significance at the 10 percent level.
** Indicates significance at the 5 percent level. *** Indicates significance at the 1 percent level.
l of Health Economics 30 (2011) 987– 999 993
r P d i f
i m a b o f a
y
w p t p b o v t
p a k w b v i C o t s t p r t
t T 5 ( t d
g d d a t t c
s p p d n o y F a
ti p
le -p
er io
d
re g
re ss
io n
sp ec
ifi ca
ti o
n s.
A ll
re tr
o fi
ts
b ro
n ch
it is
&
as th
m a
ca se
s am
o n
g
al l
at -r
is k
g ro
u p
s C
C V
re tr
o fi
ts
b ro
n ch
it is
&
as th
m a
ca se
s am
o n
g
al l
at -r
is k
g ro
u p
s A
ll
re tr
o fi
ts
p le
u ri
sy
&
p n
eu m
o n
ia ca
se s
am o
n g
al l
at -r
is k
g ro
u p
s C
C V
re tr
o fi
ts
p le
u ri
sy
&
p n
eu m
o n
ia ca
se s
am o
n g
al l
at -r
is k
g ro
u p
s
D is
cr et
e
sp ec
ifi ca
ti o
n
C o
n ti
n u
o u
s sp
ec ifi
ca ti
o n
D is
cr et
e
sp ec
ifi ca
ti o
n
C o
n ti
n u
o u
s sp
ec ifi
ca ti
o n
D is
cr et
e
sp ec
ifi ca
ti o
n
C o
n ti
n u
o u
s sp
ec ifi
ca ti
o n
D is
cr et
e
sp ec
ifi ca
ti o
n
C o
n ti
n u
o u
s sp
ec ifi
ca ti
o n
−0 .8
1 **
−1 .9
4 **
−0 .7
9
−1 .0
9
−0 .1
0
−0 .0
2
−0 .7
1
−3 .1
2 (0
.5 0
)
(1 .0
3 )
(0 .8
1 )
(1 .9
1 )
(0 .9
6 )
(1 .9
7 )
(1 .5
5 )
(3 .6
5 )
C o
n st
an t
in cl
u d
ed
C o
n st
an t
in cl
u d
ed
C o
n st
an t
in cl
u d
ed
C o
n st
an t
in cl
u d
ed d
ef fe
ct s
5 2
d is
tr ic
t
fi x
ed
ef fe
ct s
5 2
d is
tr ic
t
fi x
ed
ef fe
ct s
5 2
d is
tr ic
t
fi x
ed
ef fe
ct s
5 2
d is
tr ic
t
fi x
ed
ef fe
ct s
ef fe
ct s
1 0
y ea
r
fi x
ed
ef fe
ct s
1 0
y ea
r
fi x
ed
ef fe
ct s
1 0
y ea
r
fi x
ed
ef fe
ct s
1 0
y ea
r
fi x
ed
ef fe
ct s
d
ef fe
ct s
9
m o
n th
fi x
ed
ef fe
ct s
9
m o
n th
fi x
ed
ef fe
ct s
9
m o
n th
fi x
ed
ef fe
ct s
9
m o
n th
fi x
ed
ef fe
ct s
T em
p . a
n d
p re
ci p
. c o
n tr
o ls
T em
p . a
n d
p re
ci p
. c o
n tr
o ls
T em
p . a
n d
p re
ci p
. c o
n tr
o ls
T em
p . a
n d
p re
ci p
. c o
n tr
o ls
5 8
3 0
5 8
3 0
5 8
3 0
5 8
3 0
5 8
3 0
5 8
3 0
5 8
3 0
5 8
3 0
3 1
.4
3 1
.4
3 1
.3
3 1
.3
1 7
.3
1 7
.3
1 7
.3
1 7
.4 0
.0 0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
re d
st an
d ar
d
er ro
rs
ap p
ea r
in
p ar
en th
es es
. D ID
co ef
fi ci
en ts
te st
ed
ag ai
n st
o n
e- si
d ed
al te
rn at
iv es
.
si g
n ifi
ca n
ce
at
th e
1 0
p er
ce n
t
le v
el .
si g
n ifi
ca n
ce
at
th e
5
p er
ce n
t
le v
el .
si g
n ifi
ca n
ce
at
th e
1
p er
ce n
t
le v
el .
T.K.M. Beatty, J.P. Shimshack / Journa
etrofit level and in every time period, is a strong assumption. recise dose-response relationships in the real world are very ifficult to understand a priori, so we prefer a specification that
mposes little structure on relationships about which we have no unctional form guidance.
Nevertheless, the intuition of the difference-in-differences dentification strategy can be applied in more general settings with
ultiple time periods and multiple treatment classifications. This pproach imposes significantly more structure on the problem, ut exploits more information about the extent and exact timing f retrofit adoption. Our multiple time period regression model, or district i in month m of year t, can be most generally written s:
itm = ˛i + �t + ˇRitm + �Xitm + εitm, (3)
here y continues to represent health outcomes scaled by affected opulation numbers. ˛i are district-level fixed effects, which con- rol for average group status and time invariant factors like income, overty status, racial composition, school staff ratios, and relative us fleet age. �t are year (time) dummies and X represents a vector f time varying controls including month dummies and weather ariables. ε continues to represent an idiosyncratic disturbance erm.
In (3), R is a retrofit policy variable that takes one of four ossible forms. First, the policy variable may be an indicator vari- ble for a district that had completed significant retrofits of any ind by month m of year t. This variable takes a value of 1 hen a district retrofitted more than 30 percent of their total
us fleet with diesel oxidation catalysts (DOCs) and/or crankcase entilation filters (CCVs). Second, the policy variable may be an ndicator variable for a district that had completed significant CV retrofits by month m of year t. This variable takes a value f 1 when the district retrofitted more than 10 percent of their otal bus fleet with CCVs. Third, the policy variable may repre- ent a district’s continuous share of buses retrofitted with any echnology prior to (inclusive of) month m of year t. Finally, the olicy variable may represent a district’s continuous share of buses etrofitted with CCVs prior to (inclusive of) month m of year .
In all multiple period analyses, we omit July and August from he sample since schools were not in session during those months. he resulting dataset consists of 5830 observations: we observe all 3 districts over 10 months for each of the 11 years spanning 1996 the earliest year for which we were able to obtain health data) o 2006 (the last year we were able to obtain health and program ata).
We again cluster standard errors to allow for arbitrary within- roup correlations at the district level. We cluster over individual istricts as they represent the unit of analysis. Clustered stan- ard errors are then used to test program impact hypotheses gainst one-sided alternatives. We continue to hypothesize that he impacts of retrofits on illness outcomes will be negative, so he appropriate alternative hypothesis is a non-negative coeffi- ient.
Results from multiple period regression specifications are pre- ented in Table 3. Clustered standard errors are reported in arentheses below the coefficient estimates. We defer inter- retation of our key multiple group multiple time period ifference-in-difference coefficients to the next section, but we ote here the impact of several controls. There were no obvi-
us monotonic trends in respiratory ailments across years. Within ears, respiratory ailments tended to be lowest in the late fall. inally, on average, warmer districts experienced fewer respiratory ilments. Ta
b le
3 R
es u
lt s:
m u
l
R et
ro fi
ts
C o
n st
an t
D is
tr ic
t
fi x
e Y
ea r
fi x
ed
M o
n th
fi x
e W
ea th
er
O b
s.
F- St
at is
ti c
P ro
b
>
F
N o te
s:
C lu
st e
* In
d ic
at es
** In
d ic
at es
** *
In d
ic at
es
9 l of He
5
c a 1 o t A d 1 f a a 1 p p
n d a i o t a b M d 1 t m c (
i c 4 a t i e c e m c c r 5 r r p t
a t t 0 m e b d i r b o
p d fl l m n t i
i t i s s i H s i e t f a p a r
6
6
t i r t p f m m
s e a s a s r l n due to unobservables but adopters’ incidence was declining faster due to higher initial illness incidence and non-linear responses to unobservables. In all of these cases, difference-in-difference results
8 In continuous specifications, we observe 53 districts over 10 months (July/August omitted) for each of the 11 years spanning 1996 to 2006. Within and between variations are 8.14 and 2.41 for bronchitis and asthma cases among at-risk groups; 15.21 and 4.03 for pleurisy and pneumonia cases among at-risk groups; 0.29 and 0.08 for discrete adoption of all retrofits; 0.15 and 0.04 for continuous adoption of all retrofits; 0.14 and 0.06 for discrete adoption of CCV retrofits; and 0.06 and 0.02
94 T.K.M. Beatty, J.P. Shimshack / Journa
. Results
The two-period, two-group difference-in-differences coeffi- ients presented in row 1 of Table 2 are all large in magnitude nd for the most part statistically significant. Results in columns –3 provide the impact of significant retrofits for adopting districts n asthma and bronchitis outcomes when treatment classifica- ion is defined by the adoption of any type of retrofit technology. fter controlling for changes to a quasi-control group, adopter istricts experienced 2.9 fewer bronchitis and asthma cases per 00,000 individuals per month. More specifically, after controlling or confounders, adopter districts experienced 5.4 fewer asthma nd bronchitis cases per 100,000 children per month and 1.8 fewer sthma and bronchitis cases for those with chronic conditions per 00,000 adults per month. These are substantial changes: 5.4 cases er 100,000 children per month is a 23 percent drop relative to re-retrofit (2002) levels for the adopter districts.
Results in columns 7–9 of Table 2 provide the impact of sig- ificant retrofits on asthma and bronchitis outcomes for adopting istricts, when treatment classification is defined only by the doption of more aggressive crankcase ventilation filter (CCV) nstallations. Recall that CCV retrofits typically augment diesel xidation catalysts with add-on devices that further reduce con- aminants, especially within buses. After controlling for changes to
quasi-control group, CCV adopter districts experienced 4.3 fewer ronchitis and asthma cases per 100,000 individuals per month. ore specifically, after controlling for confounders, CCV adopter
istricts experienced 10.2 fewer asthma and bronchitis cases per 00,000 children per month and 1.8 fewer asthma and bronchi- is cases for those with chronic conditions per 100,000 adults per
onth. Again, these are substantial changes: 10.2 cases per 100,000 hildren per month is a 33 percent drop relative to pre-retrofit 2002) levels for the CCV adopter districts.
Results in columns 4–6 and 10–12 of Table 2 provide the mpact of significant retrofits on pneumonia and pleurisy out- omes for adopting districts. Treatment classification in columns –6 is defined by the adoption of any type of retrofit technology nd treatment classification in columns 10–12 is defined only by he adoption of more aggressive crankcase ventilation filter (CCV) nstallations. In general, the results suggest that retrofits influ- nce pneumonia outcomes, but perhaps only for children. After ontrolling for changes to a quasi-control group, adopter districts xperienced 4.1 fewer pneumonia cases per 100,000 children per onth. CCV adopter districts experienced 5.1 fewer pneumonia
ases per 100,000 children per month. For context, note that 4.1 ases per 100,000 children is a 37 percent drop relative to pre- etrofit (2002) levels for the ‘any technology adopter’ districts and .1 cases per 100,000 children is a 40 percent drop relative to pre- etrofit (2002) levels for the ‘CCV adopter’ districts. In contrast to esults for children, results presenting the impact of retrofits on neumonia outcomes for adults with chronic conditions are statis- ically insignificant and often small relative to pre-retrofit levels.
Multiple period regression results in Table 3 suggest that sthma and bronchitis results are reasonably robust to specifica- ion. Results in column 1 indicate that a district that retrofits more han 30 percent of its bus fleet with any technology experiences .81 fewer asthma and bronchitis cases per 100,000 individuals per onth. These results suggest that, on average, a retrofitting district
xperiences an approximately 10 percent reduction in asthma and ronchitis cases (across both children and adults with chronic con- itions) relative to its pre-adoption (2002) baseline levels. Results
n column 2 indicate that a 10 percent increase in the share of buses etrofitted with any technology is associated with 0.194 fewer ronchitis or asthma cases per month. These results suggest that, n average, a retrofitting district experiences an approximately 2.3
f t g s m
alth Economics 30 (2011) 987– 999
ercent reduction in asthma and bronchitis cases (across both chil- ren and adults with chronic conditions) for every 10% of its bus eet that is retrofitted. Bronchitis and asthma effects of CCV instal-
ations, presented in columns 3 and 4 of Table 3, are similar in agnitude to ‘any retrofit technology’ results. These CCV results are
ot, however, statistically significant. This may reflect smaller sta- istical variation in technology adoption for CCVs or non-linearity n dose-responses relationships for CCV installations.8
Multiple period regression results in columns 5–8 of Table 3 ndicate that pneumonia and pleurisy results are not robust to he multiple period specifications. While all of the difference- n-differences coefficients are negative, none are statistically ignificantly different from zero. This is not necessarily surprising, ince pleurisy has not been linked to pollutants commonly found n diesel exhaust. Measurement error induced by Department of ealth data aggregations can substantially reduce statistical preci-
ion. Note, however, that coefficient magnitudes for the CCV retrofit mpacts are broadly similar to asthma and bronchitis results. For xample, results in column 8 indicate that a 10 percent increase in he share of buses retrofitted with CCVs is associated with 0.312 ewer pneumonia cases per month. These results suggest that, on verage, a retrofitting district experiences an approximately 1.7 ercent reduction in pneumonia cases (across both children and dults with chronic conditions) for every 10% of its bus fleet that is etrofitted.
. Robustness
.1. Robustness: identifying assumptions
As is often the case in real world policy evaluations, our reatment status may not be randomly assigned. The difference- n-differences research design, however, still estimates the policy elevant average treatment effect on the treated as long as health rends are uncorrelated with adoption. This section explores the lausibility of our identifying assumptions. For transparency, we ocus all robustness checks on the simpler two-period, two-group
odel. Note that these falsification test results are not sensitive to odeling choices. What are the concerns? It is possible a priori that some group-
pecific trends are correlated with treatment status. Suppose, for xample, that some unobserved factor was associated with both
decrease in health incidences and an increased likelihood of chool bus retrofits. Active policy-makers might implement other ir pollution or public health policies in conjunction with clean chool bus programs. Or, suppose that school districts that expe- ienced or expected falling health outcomes over time were more ikely to adopt clean school bus programs. Finally, suppose that ill- ess incidence was declining for both adopter and non-adopters
or continuous adoption of CCV retrofits. Two issues bear noting. First, within varia- ion is always greater than between variation, suggesting that fixed effects models enerally have a source of variation. Second, variation for all retrofit variables is ubstantially greater than variation for CCV retrofit variables. This smaller variation ay affect statistical precision in CCV multiple period regression results.
l of Health Economics 30 (2011) 987– 999 995
m t
•
•
•
•
•
e p
ri o
d
b ro
n ch
it is
& P
re -r
et ro
fi t
p er
io d
p le
u ri
sy
& p
n eu
m o
n ia
ca se
s R
et ro
fi t
p er
io d
g as
tr o
in te
st in
al ca
se s
R et
ro fi
t
p er
io d
k id
n ey
& u
ri n
ar y
tr ac
t
ca se
s H
ea lt
h y
ad u
lt s
C h
il d
re n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n A
d u
lt s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
B ro
n ch
. & as
th m
a P
le u
ri sy
& p
n eu
m .
3 .0
3 2
.5 0
3 .8
9
0 .5
2
4 .7
3
0 .0
2
−0 .0
5
0 .2
8
0 .2
8
7 .9
8
−1 .4
9
−0 .1
9
0 .2
3 (3
.8 8
)
(2 .5
7 )
(3 .8
4 )
(1 .7
8 )
(4 .8
2 )
(4 .1
9 )
(1 .7
3 )
(5 .2
1 )
(3 .1
1 )
(7 .9
9 )
(2 .9
1 )
(2 .5
4 )
(0 .6
8 )
1 .3
3
−1 .4
3
2 .6
2
3 .2
5 *
2 .9
1
2 .4
9
−1 .5
2
3 .9
8
2 .6
6
−1 .3
8
3 .5
9 **
2 .5
3
1 .0
5 (2
.9 5
) (1
.4 2
) (2
.3 0
) (1
.8 0
) (2
.9 2
) (2
.3 7
) (1
.8 4
) (2
.8 3
) (1
.9 5
)
(4 .3
4 )
(1 .6
8 )
(2 .4
2 )
(1 .0
7 )
−0 .5
1
−2 .1
0
−0 .4
8
−0 .3
0
−0 .1
5
2 .8
0
0 .4
6
3 .5
1
0 .8
6
−8 .3
2
3 .3
1
−2 .3
9
−1 .0
8 (3
.6 1
) (2
.4 6
) (3
.7 0
)
(1 .1
9 )
(4 .7
0 )
(4 .2
5 )
(1 .5
9 )
(5 .2
9 )
(3 .1
0 )
(7 .8
2 )
(2 .8
9 )
(2 .0
5 )
(0 .6
5 )
d ed
C o
n st
an t
in cl
u d
ed
C o
n st
an t
in cl
u d
ed
C o
n st
an t
in cl
u d
ed
In cl
u d
ed h
ic s
in cl
u d
ed
Fu ll
d em
o g
ra p
h ic
s
in cl
u d
ed
Fu ll
d em
o g
ra p
h ic
s
in cl
u d
ed
Fu ll
d em
o g
ra p
h ic
s
in cl
u d
ed
In cl
u d
ed ci
p . i
n cl
u d
ed T
em p
. a n
d
p re
ci p
. i n
cl u
d ed
T em
p . a
n d
p re
ci p
. i n
cl u
d ed
T em
p . a
n d
p re
ci p
. i n
cl u
d ed
In cl
u d
ed
8 6
0 8
6 0
8 6
0 8
6 0
8 6
0 8
6 0
8 6
0
8 6
0
8 6
0
8 6
0
8 6
0
8 6
0
8 6
0 2
3 .4
2 .6
2
1 0
.3
1 4
.1
1 1
.6
5 .0
4
1 4
.0
4 .2
1
3 .4
8
2 .8
9
7 .8
6
7 .2
3
3 .7
2 0
.0 0
0 .0
1
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
0
0 .0
1
0 .0
0
0 .0
0
0 .0
0
rs
ap p
ea r
in
p ar
en th
es es
. D ID
co ef
fi ci
en ts
te st
ed
ag ai
n st
o n
e- si
d ed
al te
rn at
iv es
. e
1 0
p er
ce n
t
le v
el .
e
5
p er
ce n
t
le v
el .
e
1
p er
ce n
t
le v
el .
T.K.M. Beatty, J.P. Shimshack / Journa
ay be biased. However, some simple sensitivity analyses alleviate hese concerns:
Adopters and non-adopters experienced the same trends in health outcomes until retrofits were actually installed. We replicated our analysis in periods before adoption occurred. Columns 1 through 6 of Table 4 show that adopter districts had similar changes in respiratory ailments to non-adopter districts between 1996 and 2001, before retrofits began in earnest. Other pre-retrofit year comparisons yield similar results. As before, adopters sometimes had significantly higher initial levels of respiratory illness, but difference-in-difference coefficients were always insignificant. Standard errors were similar in magnitude or larger than the esti- mated coefficients themselves. In other words, improvements in health outcomes for adopters relative to non-adopters only appear over the sample period when retrofits were installed. For illnesses plausibly unrelated to air quality, adopters and non- adopters experienced similar health trends over the sample period. We replicated our difference-in-differences regressions for gas- trointestinal diseases and kidney/urinary tract illnesses.9 These are the most commonly treated non-respiratory ailments. Results in columns 7 through 12 of Table 4 show no significant differences in other illness changes between treated and quasi-control dis- tricts over our sample period, even though initial health outcome levels can differ. In other words, treatment and control districts experienced differential trends for diseases plausibly related to air quality, and only for diseases plausibly related to air quality. For individuals plausibly unaffected by air quality, adopter and non- adopter districts experienced similar health trends over the sample period. We replicated our difference-in-differences regressions for adults without chronic respiratory conditions. These individ- uals should not be sensitive to marginal changes in air quality. Results in columns 13 and 14 of Table 4 show no consistent differ- ence in major respiratory ailment differences for healthy adults between treatment and control groups. In other words, treatment and control districts experienced differential respiratory health trends for individuals plausibly sensitive to air quality, and only for individuals plausibly sensitive to air quality. Adopters and non-adopters experienced the same trends in health outcomes during summer months, when buses operate much less frequently. We replicated our analysis, but only for months when school was not in session. In other words, we performed a fal- sification test leveraging the fact that buses run infrequently during the summer. Results indicated that adopters have statisti- cal significantly higher initial levels of respiratory illness during summer months, but the key difference-in-difference coefficients were always statistically insignificant. In other words, signifi- cant improvements in health outcomes for adopters relative to non-adopters only appear in months where buses are active.10
Non-adopters do not systematically experience falling respira- tory health outcomes. Results on the post-treatment dummy in Tables 4 and 2 indicate that non-adopters experienced no sta- tistically significant trends in respiratory health outcomes over the pre-retrofit period (1996–2001) or over the retrofit sample period (2002–2006). As a result, it is not possible that illness inci-
dence was declining for all districts due to unobservables. Our results cannot be explained by non-linear responses to common factors causing illness declines in all districts.
9 The relevant diagnosis-related group (DRG) codes are 182, 184, 320, and 322. 10 A relatively quick change in the response of health outcomes to changes in xposure is consistent with epidemiological evidence on the short-run effects of air ollution (Spix et al., 1998; Brunekreef and Holgate, 2002; Dockery, 2009). T
a b
le
4 R
es u
lt s:
fa ls
ifi ca
ti o
n
te st
s.
P re
-r et
ro fi
t
p e
as th
m a
ca se
s
A ll
at -r
is k
g ro
u p
s
D if
f- in
-d if
f 2
.6 6
(2 .2
9 )
T re
at m
en t
−0 .5
8
G ro
u p
(1 .3
2 )
P o
st
−1 .6
3
T re
at m
en t
(2 .1
7 )
C o
n st
an t
C o
n st
an t
in cl
u C
o n
tr o
ls Fu
ll
d em
o g
ra p
W ea
th er
T em
p . a
n d
p re
O b
s. 8
6 0
F- St
at is
ti cs
2 2
.3
P ro
b
>
F 0
.0 0
N o te
s:
C lu
st er
ed
st an
d ar
d
er ro
* In
d ic
at es
si g
n ifi
ca n
ce
at
th **
In d
ic at
es
si g
n ifi
ca n
ce
at
th **
* In
d ic
at es
si g
n ifi
ca n
ce
at
th
9 l of Health Economics 30 (2011) 987– 999
•
•
6
n c c s o s a w
r a s o c t d v t
p a s w t p w a h p
d e F i a s p R d S n
7
t s
se n
si ti
v it
y
re su
lt s.
A n
n u
al
d at
a
C o
n tr
o l
g ro
u p
co n
ta in
s
re tr
o fi
ts
in
p ro
g re
ss
A ll
re tr
o fi
ts
b ro
n ch
it is
&
as th
m a
ca se
s A
ll
re tr
o fi
ts
p le
u ri
sy
&
p n
eu m
o n
ia
ca se
s A
ll
re tr
o fi
ts
b ro
n ch
it is
&
as th
m a
ca se
s A
ll
re tr
o fi
ts
p le
u ri
sy
&
p n
eu m
o n
ia
ca se
s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
−2 .8
3 **
−5 .1
8
−1 .7
5 **
−3 .8
8
−4 .0
1 **
−3 .1
5
−1 .6
0 *
−2 .2
9
−1 .0
3 *
−1 .7
5
−3 .3
7 **
−0 .6
5 (1
.6 0
) (5
.6 3
) (0
.9 3
)
(3 .5
9 )
(2 .4
9 )
(4 .7
4 )
(1 .0
0 )
(2 .9
9 )
(0 .6
4 )
(1 .8
5 )
(1 .5
2 )
(2 .6
1 )
g ro
u p
2 .6
7 3
.5 5
1 .9
4 **
4 .6
6 *
7 .3
4 **
* 3
.5 4
0 .6
8 −0
.9 0
0 .5
8 1
.0 5
3 .4
6 **
0 .0
4 (1
.6 3
)
(5 .5
1 )
(0 .9
2 )
(2 .6
6 )
(2 .3
1 )
(3 .2
4 )
(1 .1
0 )
(3 .3
4 )
(0 5
9 )
(1 .8
2 )
(1 .6
1 )
(2 .3
5 )
en t
0 .3
3 −2
.7 4
1 .0
2 2
.9 1
−0 .9
3
3 .9
2 −0
.6 8
−4 .5
9 *
0 .5
1
0 .8
1
−1 .1
2
1 .3
3 (1
.8 3
) (5
.6 1
) (0
.9 6
) (4
.3 1
) (2
.5 1
) (5
.7 1
) (0
.7 3
) (2
.6 2
) (0
.4 1
) (1
.6 1
) (1
.0 4
) (2
.0 3
) ls
In cl
u d
ed In
cl u
d ed
In cl
u d
ed
In cl
u d
ed In
cl u
d ed
In cl
u d
ed In
cl u
d ed
In cl
u d
ed
8 6
8 6
8 6
8 6
8 6
8 6
1 0
6 0
1 0
6 0
1 0
6 0
1 0
6 0
1 0
6 0
1 0
6 0
re d
st an
d ar
d
er ro
rs
ap p
ea r
in
p ar
en th
es es
. D ID
co ef
fi ci
en ts
te st
ed
ag ai
n st
o n
e- si
d ed
al te
rn at
iv es
.
si g
n ifi
ca n
ce
at
th e
1 0
p er
ce n
t
le v
el .
si g
n ifi
ca n
ce
at
th e
5
p er
ce n
t
le v
el .
si g
n ifi
ca n
ce
at
th e
1
p er
ce n
t
le v
el .
96 T.K.M. Beatty, J.P. Shimshack / Journa
Adopters and non-adopters do not systematically differ on most important characteristics. Results in Table 1 show that demo- graphic characteristics are similar across adoption classification. Adopters and non-adopters have statistically indistinguishable per capita income, poverty status, racial composition, and school staff per student ratios. Bus fleet average age is also similar between adopter and non-adopter districts. Treatment dis- tricts are systematically larger, but we analyze health outcome variables scaled by population size. Again note that most non- adopters adopt the program shortly after our sample period ends. School bus retrofit decisions and implementation occur at the school district level. In contrast, most air quality and public health pro- grams are instituted at the county, state, or national level.
.2. Robustness: specification
In the two-period, two-group analysis, treatment status does ot vary within the year. A possible concern is that district by month ells induce artificial statistical precision. To address this we use lustered standard errors, which allow for arbitrary within-district erial correlation. Nevertheless, as a sensitivity check, we repeated ur two-period analysis using annual data. Results are in the first ix columns of Table 5. Point estimates were very slightly smaller nd standard error magnitudes were very slightly larger, but results ere identical for all practical purposes.
In the two-period, two-group analysis, we omit districts with etrofits in progress from the treatment group. These districts re included in the multiple period analysis. Nevertheless, as a ensitivity check, we repeated the two-period analysis including mitted districts in the treatment group. Results are in the latter six olumns of Table 5. As expected with a less cleanly identified con- rol group (which now includes districts with retrofits in progress), ifference-in-differences point estimates were smaller in absolute alue. Nevertheless, signs and statistical significance were similar o presented results.
In the two-period, two-group analysis, as well as the multiple eriod, two-group analysis, our ‘adopter’ thresholds are admittedly d hoc. As a sensitivity check, we repeated all two-group analy- es using both lower and higher adoption thresholds. Specifically, e increased the adoption threshold from 30 to 40 percent and
hen subsequently decreased the adoption threshold from 30 to 20 ercent. Results are in Table 6. Patterns of statistical significance ere unchanged. As expected, lower adoption thresholds gener-
lly yielded somewhat smaller health impact point estimates and igher adoption thresholds yielded somewhat larger health impact oint estimates.
In the two-period, two-group analysis, we scale our depen- ent variable by population. This control function approach is an specially flexible way to account for the influence of population. urther, our multiple period analysis contains fixed effects which nherently condition on factors like population. Nevertheless, as
sensitivity check, we repeated the two-period analysis with a pecification that uses a dependent variable that is not scaled by opulation and includes an explanatory variable for population. esults are in Table 7. Note that coefficient magnitudes are not irectly comparable to coefficient magnitudes in the main results. ign patterns, however, were identical and results were more sig- ificant statistically.
. Discussion and interpretation
This paper analyzes the impact of school bus emissions reduc- ions on human health. We find that school bus retrofits induced tatistically significant and large reductions in bronchitis, asthma, Ta
b le
5 Sp
ec ifi
ca ti
o n
D if
f- in
-d if
f
T re
at m
en t
P o
st -t
re at
m
Fu ll
co n
tr o
C o
n st
an t
O b
s.
N o te
s:
C lu
st e
* In
d ic
at es
** In
d ic
at es
** *
In d
ic at
es
T.K.M. Beatty, J.P. Shimshack / Journal of He
T a
b le
6 Sp
ec ifi
ca ti
o n
se n
si ti
v it
y
re su
lt s.
H ig
h er
th re
sh o
ld
fo r
‘a d
o p
te r’
st at
u s
Lo w
er
th re
sh o
ld
fo r
‘a d
o p
te r’
st at
u s
A ll
re tr
o fi
ts
b ro
n ch
it is
&
as th
m a
ca se
s
A ll
re tr
o fi
ts
p le
u ri
sy
&
p n
eu m
o n
ia
ca se
s
A ll
re tr
o fi
ts
b ro
n ch
it is
&
as th
m a
ca se
s
A ll
re tr
o fi
ts
p le
u ri
sy
&
p n
eu m
o n
ia
ca se
s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
A ll
at -r
is k
g ro
u p
s C
h il
d re
n
A d
u lt
s
w it
h ch
ro n
ic
il ln
es s
D if
f- in
-d if
f −2
.9 1
** −4
.8 8
−1 .9
3 **
−3 .8
2
−5 .0
1 **
−2 .5
9
−2 .8
1 **
−5 .3
0
−1 .7
6 **
−3 .9
9
−3 .6
7 *
−3 .4
8 (1
.5 8
)
(5 .3
4 )
(0 .9
7 )
(3 .4
0 )
(2 .5
0 )
(4 .6
0 )
(1 .4
3 )
(5 .1
6 )
(0 .8
1 )
(3 .2
6 )
(2 .2
7 )
(4 .2
4 )
T re
at m
en t
g ro
u p
2 .4
9 2
.8 7
1 .8
0 **
3 .6
7
7 .4
3 **
* 2
.0 7
2 .5
0
2 .8
4
1 .9
1 **
4 .5
3 *
6 .4
5 **
* 3
.6 8
(1 .5
4 )
(5 .0
5 )
(0 .7
9 )
(2 .7
5 )
(2 .0
8 )
(3 .3
9 )
(1 .6
3 )
(5 .6
8 )
(0 .8
0 )
(2 .4
7 )
(2 .2
1 )
(2 .7
9 )
P o
st -t
re at
m en
t 0
.6 7
−1 .2
8
1 .2
9 *
2 .9
8 −0
.3 4
3 .8
5
0 .5
8
−1 .5
9
1 .2
7
2 .9
0
−0 .3
5
3 .7
4 (1
.3 2
)
(5 .0
5 )
(0 .7
4 )
(3 .1
7 )
(2 .1
0 )
(4 .0
5 )
(1 .3
0 )
(5 .0
1 )
(0 .7
1 )
(3 .1
5 )
(2 .0
6 )
(3 .9
9 )
Fu ll
co n
tr o
ls In
cl u
d ed
In cl
u d
ed In
cl u
d ed
In cl
u d
ed C
o n
st an
t In
cl u
d ed
In cl
u d
ed
In cl
u d
ed
In cl
u d
ed
O b
s. 6
8 0
6 8
0
6 8
0
6 8
0
6 8
0
6 8
0
9 8
0
9 8
0
9 8
0
9 8
0
9 8
0
9 8
0
N o te
s:
C lu
st er
ed
st an
d ar
d
er ro
rs
ap p
ea r
in
p ar
en th
es es
. D ID
co ef
fi ci
en ts
te st
ed
ag ai
n st
o n
e- si
d ed
al te
rn at
iv es
. *
In d
ic at
es
si g
n ifi
ca n
ce
at
th e
1 0
p er
ce n
t
le v
el .
** In
d ic
at es
si g
n ifi
ca n
ce
at
th e
5
p er
ce n
t
le v
el .
** *
In d
ic at
es
si g
n ifi
ca n
ce
at
th e
1
p er
ce n
t
le v
el .
a d i e m a l
d a d c r fi o
d s q r c r p
s e h n m S p e r t m P n F d s p i a r
r o a E o b a i
a s h i s a o s t w e
alth Economics 30 (2011) 987– 999 997
nd pneumonia incidence for children and adults with chronic con- itions. Broad findings are robust to a host of sensitivity checks,
ncluding falsification tests examining the pre-retrofit period, dis- ases not associated with pollution, and individuals not sensitive to arginal changes in air quality. Results for bronchitis and asthma
re very robust to specification as well; results for pneumonia are ess so.
Findings from our preferred specifications indicate that adopter istricts experienced 23 percent fewer children’s bronchitis and sthma cases per month, relative to a control group. These same istricts also experienced 37 percent fewer children’s pneumonia ases per month. We also typically find larger effects for the CCV etrofits, suggesting that the more modern crankcase ventilation lters may play a larger role in health improvements than diesel xidation catalysts alone.
Asthma and bronchitis illness reductions occurred for both chil- ren and adults with chronic conditions. This suggests that clean chool bus programs may impact background and localized air uality. However, when we detect significant pneumonia illness eductions, they occur for children but not for adults with chronic onditions. While there are several plausible explanations, these esults are consistent with school bus programs further impacting ublic health through reductions in within-bus exposure.
Our analysis permits the first empirical economic assessment of chool bus programs and illness outcomes. However, we note sev- ral limitations. First, we do not directly observe individual-level ealth and bus ridership data. We assume that children ride buses ear where they live and that the share of bus ridership across treat- ent and control districts does not differentially change over time.
econd, our analysis only considers acute, short-run responses to ollution exposure. Cumulative effects may also matter; to the xtent that these responses are important, our program impact esults are understated. Third, we are unable to directly examine he impact of bus programs on air quality. Only seven air pollution
onitors regularly measure air quality in the entire four-county uget Sound region, and most have incomplete data, so we can- ot separately match monitors to treatment and control areas. ourth, people with vulnerable children may move into retrofitting istricts. We do not see evidence for selective migration on a large- cale, as student populations at non-adopting districts grew by 6.3 ercent over the retrofit period while student populations at adopt-
ng districts grew by only 2.2 percent. To the extent that migration s averting behavior is important, however, our program impact esults are understated.
We also note caveats to external validity. First, Washington’s espiratory illness rates are among the highest in the nation. Sec- nd, the Puget Sound region of Washington is whiter, wealthier, nd less dense than most other urban areas in the United States. xtrapolating the numerical benefits of retrofits from our dataset to ther contexts may misrepresent the case. Nevertheless, we would e surprised if overall policy implications differed substantially cross the country since detected retrofit benefits were very large n our study area.
Several interesting directions for future research arise from this nalysis. First, Currie et al. (2009a,b) show that pollution influences chool absences, so absences might provide some evidence on the uman capital impacts of retrofits programs. Such an exploration
s beyond the scope of this paper, however, as the Washington tate schools in our sample were not required to track excused bsences for our sample period. Second, we would ideally compare ur results to a “cash for clunkers” school bus program or to a low
ulfur diesel school bus program. Unfortunately, we do not have he data to run the necessary empirical evaluations ourselves, and e are unaware of any other studies that provide the necessary
mpirical estimates for these comparisons.
998 T.K.M. Beatty, J.P. Shimshack / Journal of Health Economics 30 (2011) 987– 999
Table 7 Specification sensitivity results.
Dependent variables not scaled by population
All retrofits bronchitis & asthma cases All retrofits pleurisy & pneumonia cases
All at-risk groups
Children Adults with chronic illness
All at-risk groups
Children Adults with chronic illness
Diff-in-diff −1.02*** −0.90*** −0.12* −0.67* −0.43*** −0.24 (0.23) (0.22) (0.08) (0.42) (0.12) (0.37)
Treatment group 0.41 0.32 0.12 1.17** 0.38*** 0.77 (0.30) (0.28) (0.10) (0.50) (0.13) (0.40)
Post-Treatment 0.09 0.01 0.08 0.06 0.20 0.04 (0.10) (0.13) (0.05) (0.29) (0.07) (0.25)
Population 0.08*** – – 0.15*** – – (0.01) (0.02)
Children population – 0.21*** – – 0.07*** – (0.02) (0.01)
Adult population – – 0.02*** – – 0.18***
(0.01) (0.02) Full controls Included Included Constant Included Included
Obs. 860 860 860 860 860 860
Notes: Clustered standard errors appear in parentheses. DID coefficients tested against one-sided alternatives.
i t l t t w o o c c c a
e $ e d t c c c c d a t fi e c
e E l E p t
s
r f r e d t v T p a
r 4 t a l t r
A
g m
A
t t i o (i.e. GIS layers) cannot be used to assign an individual who appears
* Indicates significance at the 10 percent level. ** Indicates significance at the 5 percent level.
*** Indicates significance at the 1 percent level.
Our empirical analysis investigates the impact of retrofits on ncidences of bronchitis, asthma, and pneumonia severe enough o warrant hospital or clinical treatment. Retrofit impacts for ess severe respiratory illnesses are unobserved. Further, the rela- ionships between health outcomes and communicable disease ransmission, pain and suffering considerations, and long-term elfare effects are complex. Addressing the full benefits and costs
f bus retrofits are beyond the scope of this study. However, in rder to provide an approximate guide to the economic signifi- ance of our results, we combine our empirical point estimates with ost-of-treatment health valuation estimates and observed retrofit osts to compute a conservative back of the envelope benefit–cost ssessment of school bus retrofits.
The medical literature estimates health care costs per inpatient pisode of bronchitis, asthma, and pneumonia at approximately 3000–7000/visit. See, for example, Stanford et al. (1999) and Lave t al. (1999), for a more complete discussion. The average school istrict in our dataset serves approximately 10,000 children. For he average district, our crankcase ventilation filter (CCV) coeffi- ients translate approximately into 12.2 avoided hospital visits for hildren’s asthma and bronchitis per year.11 Similarly, CCV coeffi- ients translate approximately into 6.1 avoided hospital visits for hildren’s pneumonia and pleurisy per year. Estimates of a single istrict’s annual benefits for children’s health from observed CCV doption range from approximately $54,900–128,100 (18.3 visits imes $3000/visit and 18.3 times $7000/visit). Again, these bene- t calculations exclude non-respiratory illnesses, long-term health ffects, suffering considerations, and any impacts on adults with hronic respiratory conditions.
The average CCV adopting district retrofitted approximately 25 ligible buses of its 66 total buses with CCVs over the sample period. ach CCV retrofit cost approximately $1200 in total, including parts, abor, and testing. Most CCVs are coupled with pre-existing DOCs.
ach DOC retrofit cost approximately $1300 in total, including arts, labor, and testing. Therefore, the average adopter school dis- rict spent approximately $62,500 (25 buses times $2500) on CCV
11 The relevant DID coefficient in column 8 of Table 2 is −10.2 cases per 100,000 tudents: (10.2/100,000) × 10,000 students × 12 months = 12.2.
i a
b c
etrofits. Total CCV retrofit costs are likely less than annual benefits or children alone. Assuming a 5 percent discount rate and a useful etrofit life of 10 years, the net present value children’s health ben- fits are between 424,000 and 989,000 dollars per adopter school istrict.12 Even excluding benefits to adults with chronic condi- ions and omitting suffering considerations, the ratio of present alue benefits to present value costs ranges between 7:1 and 16:1. his interpretation suggests that if the many states not aggressively ursuing school bus retrofits were to do so, potential social benefits re likely to be large.
For perspective, best estimates suggest that the benefit–cost atios of the 1990 Clean Air Act amendments are between 1:1 and :1 (USEPA, 1999; Portney, 2000). Like most major pollution con- rol programs, the goals of the Clean Air Act are improving ambient ir quality. The difference between our back of the envelope calcu- ations and these estimates suggests that, on the margin, policies argeting localized air pollution may be particularly cost effective elative to ambient air pollution policies.
ppendix A. Data appendix
This data appendix describes how data observed at a finer geo- raphic level (zip code and latitude/longitude) is aggregated to atch our unit analysis: school districts.
.1. Health data
Health outcome data consists of hospital discharge records from he Washington State Department of Health. Because of confiden- iality concerns, records in the database only contain home zip code nformation, rather than a complete street address. As we do not bserve a patient’s exact street address, school district boundaries
n the health data to a specific school district. As a result, we need n algorithm to map health treatments observed at the five-digit
12 For a district, the range of the NPV of benefits is based on the upper and lower ounds of per incident health care costs (7000 and 3000), respectively. NPVs are alculated as (18.3 × 3000/0.05(1 − 1/1.0510 )) and (18.3 × 7000/0.05(1 − 1/1.0510 )).
l of He
z z w
1
2
3
A
t C t e f d t
R
A
A
B
B
B
C
C
C
C
C
C
C
C
C
C
D
G
G
J
L
L
M
M
M
M
N
P
P
P
P
S
S
S
S
S
U
T.K.M. Beatty, J.P. Shimshack / Journa
ip code level to school districts.Our algorithm uses schools within ip codes (as opposed to surface area, for example) as assignment eights. Specifically:
. We compile a list of all schools in the Puget Sound region, their address zip codes, and their school district.
. For each health treatment in the health data, we use the list constructed in (1) to identify all of the schools in that patient’s zip code.
. We assign health treatments to districts based on the following rules: a. If all schools in the patient’s zip code belong to the same dis-
trict, we assign this health record fully to that district. b. If the patient’s zip code contains schools from multiple dis-
tricts, we assign shares of the outcome to each district in proportion to the share of schools (per district) in the patient’s zip code. For example, if a zip code contains 2 schools from dis- trict A and 3 schools from district B, we would assign 2/5ths of all health outcomes in that zip code to district A and 3/5ths to district B.
c. If there are no schools physically located in a patient’s zip code, we assign the health treatment for this patient to the nearest school district, as measured by distance from the centroid of the zip code to the edge of a school district.
.2. Weather data
All regression specifications contain mean monthly tempera- ure and total monthly precipitation derived from the US Historical limatology Network. To map from individual weather stations o school districts, we simply assign school districts to the clos- st weather station, where “closest” is determined by the Vincenty ormulae for geodesic distances. In the final analysis, data from 10 istinct weather stations in the Puget Sound region are matched to he 53 school districts for each month.
eferences
dar, S., et al., 2008. Predicting airborne particle levels aboard Washington State school buses. Atmospheric Environment 42 (33), 7590–7599.
merican Lung Association, 2008. Diesel Exhaust and Air Pollution, http://www.lungusa.org/site/pp.asp?c=dvLUK9O0E&b=36089 (accessed August 08).
ehrentz, E., et al., 2004. Measuring self-pollution in school buses using a tracer gas technique. Atmospheric Environment 83 (23), 3735–3746.
oyer, M., Lyons, K., 2004. Washington State Clean School Bus Program: Report to 2005 Legislature. WA State Dept. of Ecology Publication Number 04-02-029.
runekreef, B., Holgate, S.T., 2002. Air pollution & health. Lancet 360, 1233–1242.
A Air Resources Board, 2002. Public Hearing to Consider Amendments to the Ambient Air Quality Standards for Particulate Matter and Sulfates. CA Air Resources Board and Office of Environmental Health Hazard Assessment Staff Report, Available at http://www.arb.ca.gov/research/aaqs/std-rs/pm-final/pm- final.htm.
han, C.C., et al., 1991. Driver exposure to volatile organic compounds, CO, ozone, and NO2 under different driving conditions. Environmental Science and Tech- nology 25, 964–972.
hay, K.Y., et al., 2003. The clean air act of 1970 and adult mortality. Journal of Risk and Uncertainty 27 (3), 279–300.
W
W
alth Economics 30 (2011) 987– 999 999
hay, K.Y., Greenstone, M., 2003. The impact of air pollution on infant mortality: evidence from geographic variation in pollution shocks induced by a recession. The Quarterly Journal of Economics 118 (3), 1121–1167.
ohen, A.J., Nikula, K.J., 1999. The health effects of diesel exhaust: laboratory and epidemiology studies. In: Holgate, S.T., et al. (Eds.), Air Pollution and Health. Academic Press, San Diego, pp. 707–745.
ommittee on Environmental Health, 2004. Ambient air pollution: health hazards to children. Pediatrics 114, 1699–1707.
urrie, J., Neidell, M., 2005. Air pollution and infant health: what can we learn from California’s recent experience? Quarterly Journal of Economics 120 (3), 1003–1030.
urrie, J., Walker, W.R., 2011. Traffic congestion and infant health: evidence from E-ZPass. American Economic Journal: Applied Economics 3 (1), 65–90.
urrie, et al., 2009a. Air pollution and infant health: lessons from New Jersey. Journal of Health Economics 28 (3), 688–703.
urrie, et al., 2009b. Does pollution increase school absences? Review of Economics and Statistics 91 (4), 682–694.
ockery, D., 2009. Health effects of particulate air pollution. Annals of Epidemiology 19 (4), 257–263.
auderman, W.J., et al., 2000. Association between air pollution and lung function growth in southern California children. American Journal of Respiratory and Critical Care Medicine 162 (4), 1383–1390.
auderman, W.J., et al., 2004. The effect of air pollution on lung development from 10 to 18 years of age. New England Journal of Medicine 351 (11), 1057–1067.
anke, K., Propper, C., Henderson, J., 2009. Do current levels of air pollution kill? The impact of air pollution on pollution mortality in England. Health Economics 18 (9), 1031–1055.
awryk, N.J., et al., 1996. Concentrations of volatile organic compounds in the pas- senger compartments of automobiles. Environmental Science and Technology 30 (2), 810–816.
ave, J.R., et al., 1999. The cost of treating patients with community-acquired pneu- monia. Seminars in Respiratory and Critical Care Medicine 20 (3), 189–197.
cConnell, R., et al., 2002. Asthma in exercising children exposed to ozone: a cohort study. Lancet 359, 386–391.
cCreanor, J., et al., 2007. Respiratory effects of exposure to diesel traffic in persons with asthma. New England Journal of Medicine 357 (23), 2348–2358.
onahan, P., 2006. School bus pollution report card. In: Union of Concerned Scien- tists , May 2006.
oretti, E., Neidell, M., 2011. Pollution, health, and avoidance behavior: evidence from the Ports of Los Angeles. Journal of Human Resources 46 (1), 154–175.
eidell, M., 2004. Air pollution, health, and socio-economic status: the effect of outdoor air quality on childhood asthma. Journal of Health Economics 23 (6), 1209–1236.
earson, R.L., et al., 2000. Distance-weighted traffic density in proximity to a home is a risk factor for leukemia and other childhood cancers. Journal of the Air Waste Management Association 50 (2), 175–180.
eden, D.B., 2002. Pollutants and asthma: role of air toxics. Environmental Health Perspectives 110 (S4), 565–568.
eters, J.M., et al., 2004. Epidemiological investigation to identify chronic effects of ambient air pollutants in Southern California. California Air Resources Board and the California Environmental Protection Agency, Contract No. 94-331.
ortney, P.R., 2000. Air pollution policy. In: Portney, P.R., Stavins, R.N. (Eds.), Public Policies for Environmental Protection. RFF Press, Washington.
abin, L.D., et al., 2005. Characterizing the range of children’s air pollutant exposure during school bus commutes. Journal of Exposure Analysis and Environmental Epidemiology 15, 377–387.
chlenker, W., Walker, W.R., 2010. The effect of airports on air quality and respira- tory problems. Columbia University Working Paper.
hikiya, D.C., et al, 1989. In-vehicle air toxics characterization study in the South Coast Air Basin. Final Report, South Coast Air Quality Management District.
pix, C., et al., 1998. Short-term effects of air pollution on hospital admissions of respiratory diseases in Europe: a quantitative summary of APHEA study results. Archives of Environmental Health 53 (1), 54–64.
tanford, R., et al., 1999. The cost of asthma in the emergency department and hospital. American Journal of Respiratory and Critical Care Medicine 160 (1), 211–215.
.S. Environmental Protection Agency, 1999. The Benefits and Costs of the Clean Air Act 1990–2010. Government Printing Office, Washington, EPA-410-R-99-001.
argo, J., et al., 2002. Children’s Exposure to Diesel Exhaust on School Buses. Envi- ronment and Human Health, Inc., North Haven.
ilhelm, M., Ritz, B., 2003. Residential proximity to traffic and adverse birth out- comes in Los Angeles County, California, 1994–1996. Environmental Health Perspectives 111 (2), 207–216.
- School buses, diesel emissions, and respiratory health
- 1 Introduction
- 2 Background
- 2.1 School buses and diesel emissions
- 2.2 Diesel emissions and respiratory health
- 2.3 Clean school bus initiatives
- 2.4 Retrofit timing and scope
- 2.5 Linking retrofits and health
- 3 Data
- 4 Methods
- 4.1 Two-period, two-group difference-in-differences
- 4.2 Multiple period approaches
- 5 Results
- 6 Robustness
- 6.1 Robustness: identifying assumptions
- 6.2 Robustness: specification
- 7 Discussion and interpretation
- Appendix A Data appendix
- A.1 Health data
- A.2 Weather data
- References
