See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/320657797

Public Transit and Air Pollution

Article  in  SSRN Electronic Journal · January 2017

DOI: 10.2139/ssrn.3049945

CITATIONS READS
9 10,542

3 authors, including:

Nicholas Rivers Soodeh Saberian

University of Ottawa University of Ottawa
86 PUBLICATIONS   1,756 CITATIONS    8 PUBLICATIONS   241 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Soodeh Saberian on 09 January 2018.

The user has requested enhancement of the downloaded file.

 Public Transit and Air Pollution*
Nicholas Rivers„ Soodeh Saberian… Brandon Schaufele§

October 9, 2017

Abstract

Advocates of public transit frequently tout improved air quality as a primary bene-
fit. Yet little is known about the causal impacts of public transit on local air pollution.
Exploiting variation in transit availability resulting from work stoppages in 18 Cana-
dian cities between 1974–2011, this study identifies the effect of public transit on air
pollution. Our findings indicate that transit leads to a 3.5 part-per-billion increase in
nitrogen oxides while having no statistically significant effect on carbon monoxide or
PM2.5 . Estimates are robust to a series of specification and placebo tests and magni-
tudes are consistent with a calibrated simulation model. Overall, the results suggest
that expanding the current configuration of public transit in North American cities is
unlikely to yield improvements in local air quality.

Keywords: Air pollution, urban transportation, public transit.

JEL codes: R40, Q53, C54.

* We are grateful to Dana Anderson, Werner Anteweiler, Justin Beaudoin, Ross McKitrick and Jay
Shimshack for providing helpful comments and suggests. Earlier drafts of this paper have been presented
at the Canadian Economics Association Annual Meetings and the Canadian Resource and Environmental
Economics Annual Meetings. All errors are ours.
„ Graduate School of Public and International Affairs and Institute of the Environment, University of
Ottawa, [email protected]
… Department of Economics, University of Ottawa, [email protected]
§ Ivey Business School, University of Western Ontario, [email protected].
1 Introduction
Traffic-related air pollution leads to adverse health outcomes,1 deteriorated cognitive per-
formance2 and reduced labor productivity.3 Curbing vehicular emissions has, as a result,
become a central goal of environmental policy with investment in public transit emerging as
a broadly endorsed strategy for improving traffic-related air quality (Beaudoin, Farzin and
Lawell, 2015).4 Yet, it is unclear whether expanded public transit actually decreases local
pollution. Like other light-duty vehicles, buses burn fossil fuels and trains require electricity
whose generation may originate from high-emission sources. Advocates implicitly assume
that enhanced public transit will motivate a sufficiently large number of commuters to sub-
stitute from cars to buses and trains, ultimately improving air quality. The fundamental law
of road congestion, however, contends that expanding public transit will have little effect
on the number of vehicle kilometres traveled by cars (Downs, 1962; Duranton and Turner,
2011): improved transit may induce some commuters to substitute from cars to buses or
trains, but a large latent group of drivers will quickly occupy the freed capacity.
This paper estimates the causal relationship between the provision of publicly provided
transit and air pollution. Using daily and hourly pollution data, we exploit as-good-as-
randomly assigned transit union strikes in 18 Canadian cities. Across the entire sample, we
show that transit work stoppages lead to a large and statistically significant decrease in at-
mospheric nitrogen oxide (NOX ) concentrations equal to 3.5 parts per billion (ppb). Transit
strikes also result in an imprecisely estimated, statistically insignificant increase in carbon
monoxide (CO) and a small and statistically insignificant decrease in PM2.5 concentrations.
Investment in public transit therefore entails more NOX pollution while having little mean-
ingful effect on other pollutants. Thus, the prospect for an environmentally advantageous
pollution swap, where pollution generated by private vehicles is swapped for emissions from
1
A range of literature explores the link between traffic-related air pollution and health – see, for example,
(e.g., Currie and Neidell, 2005; Chay and Greenstone, 2003; Künzli et al., 2000; Krämer et al., 2000; Currie
et al., 2009; Neidell, 2009; Xia et al., 2015; Jerrett et al., 2014).
2
Lavy, Ebenstein and Roth (2014), for instance, shows that exposure to carbon monoxide and PM2.5
lead to lower test scores for college entrance exams in Israel, while Bharadwaj et al. (2017) find that fetal
pollution exposure leads to lower fourth grade test scores in Santiago, Chile.
3
Chang et al. (2016) and Zivin and Neidell (2012), for example, illustrate that exposure to ozone reduces
the productivity of agricultural workers in California.
4
For example, environmental groups in the city of Toronto have endorsed transit’s air-quality benefits,
stating that without further public transit investment, “Toronto’s air would be significantly dirtier and we
would be emitting hundreds of thousands of tonnes of additional greenhouse gas (ghg) emissions into the
atmosphere” (Toronto Environmental Alliance, 2010). Likewise, in December 2016, the city of Paris made
public transportation free to riders in an effort to combat severe air pollution (Sharman, 2016).

1
public transit leading to an overall improvement in air quality, is limited. Rather, increasing
the number of buses on the road deteriorates a city’s overall air quality. A stylized analyti-
cal model, calibrated to match Canadian commuting behavior (the setting for the empirical
component of the paper), supports the econometric results: it is likely that public transit in
North America results in increased NOX emissions. On average, the current configuration
of public transit in North American cities does not improve local air quality.
A burgeoning literature links public transit to pollution.5 Friedman et al. (2001) investi-
gate the impacts of a city-wide change in transport infrastructure, including increased public
transit provision and severe downtown driving restrictions that were put in place during the
1996 Summer Olympic Games in Atlanta, and find that pollution monitors in downtown
Atlanta recorded a 27.9% decrease in peak daily ozone concentrations. Chen and Whalley
(2012) find that the the opening of Metropolitan Rapid System in Taipei reduced carbon
monoxide (CO) concentrations by 9 to 14% (while having no impact on other pollutants),
while Lalive, Luechinger and Schmutzler (2013) show that increasing the frequency of rail
services in Germany is associated with lower concentrations of CO and NOX . Bel, Holst et al.
(2015) find that Mexico City’s “Bus Rapid Transit” system is effective at reducing NOX , CO
and PM2.5 . At the global scale, Gendron-Carrier et al. (2017) use satellite imaging and a
panel of international subways expansions to demonstrate a 4% reduction in airborne partic-
ulates.6 In contrast, in an international panel of 75 cities, Hilber and Palmer (2014) highlight
that, as the number of cars increases, annual NOX concentrations decrease. Likewise, using
data from a panel of 96 US cities, Beaudoin and Lin Lawell (2016) find that increases in
the provision of public transit result in a small deterioration in urban air quality. Much
of this research however focuses on isolated geographical locations, potentially limiting the
generalizability, or uses annual data, making identification of the impact of public transport
difficult.
This paper’s conclusions run counter to much of the existing literature on public transport
and air quality. Rather than improving air quality, we demonstrate that public transit
increases local pollution levels (in particular, NOX concentrations). We posit two likely
reasons for these differences. First, we examine a panel of North American cities. Compared
5
A broader literature connects public transit to traffic congestion – for example, Anderson (2014); Winston
and Maheshri (2007); Aftabuzzaman, Currie and Sarvi (2010). Likewise, several recent studies link traffic
congestion to different health outcomes (Levy, Buonocore and Von Stackelberg, 2010; Knittel, Miller and
Sanders, 2011; Currie and Walker, 2011).
6
However, it is worth noting that Gendron-Carrier et al. (2017) regard the 4% estimate as too large to
be plausible, given current technologies.

2
to Asia or Europe, North America has unique traffic patterns, geography and economic
structure. These structural factors are likely the major basis for the discrepancies. Indeed,
Beaudoin and Lin Lawell (2016) who also study public transit in North American cities,
finds conclusions that are aligned with ours. Second, while much of the literature focuses
on the impact of new subway lines on air pollution (Gendron-Carrier et al., 2017; Chen and
Whalley, 2012; Lalive, Luechinger and Schmutzler, 2013), we focus instead on existing public
transport systems in North America, which are dominated by buses. Buses typically burn
diesel, while subways use electricity. Hence, different patterns of results are expected.
Methodologically, the paper most similar to ours is Bauernschuster, Hener and Rainer
(2017).7 Bauernschuster, Hener and Rainer (2017) show that a public transit strike is ac-
companied by a 14% increase in particulate pollution (PM10 ) in the five largest German
cities. Like us, they use work actions for econometric identification. Yet, despite the out-
ward similarity of the two papers, the contexts and implications are quite different. First,
Bauernschuster, Hener and Rainer (2017) examine extremely short work stoppages. All of
the strikes in their data span less than 24 hours, with many as brief as two hours or less.
Strikes in our analysis are longer, lasting several weeks on average. This means that the
two papers are estimating fundamentally different parameters: while it is reasonable to de-
lay commuting in response to a very brief strikes such as those in Bauernschuster, Hener
and Rainer (2017), such avoidance behaviour is much more challenging for strikes that last
several weeks. Further, as mentioned, caution should be exercised when generalizing from
large German metropolises to the North American context as the configuration of cities,
baseline levels of pollution, economic structure and travel mode shares are different. For
example, while diesel passenger vehicles are common in European countries, they represent
a tiny niche in North America. Gasoline and diesel vehicles emit distinct pollutant mixes,
so this implies that transit displaces different pollutants on the two continents. Moreover,
given that our conclusions are very different than those of Bauernschuster, Hener and Rainer
(2017) (and several of the other papers), this suggests that a consensus on the relationship
between public transit and air quality is likely context-specific and that additional research
on the topic is warranted.
Identifying the causal effect of public transit on air pollution is challenging for three rea-
sons. First, the supply of public transit is not randomly assigned and is generally confounded
7
Our paper is also similar to Anderson (2014), who identifies the effect of public transit using public
transit strikes. Anderson (2014) uses the LA transit strike of 2003 to measure the impact of transit on
congestion, whereas we focus on air quality.

3
with other factors that influence air quality (Chen and Whalley, 2012; Beaudoin, Farzin and
Lin, 2014). The supply of buses, for instance, tends to be greater in cities where the con-
gestion is also high. This prototypical endogeneity problem explains the focus on short-run
changes in air quality: parameter identification at these time scales is more clearcut. Second,
the impacts of public transit on air quality depend on the ability and willingness of com-
muters to substitute between modes of transportation. Obtaining reliable, high frequency
data on private and public transit usage has proved challenging. Without these data, it is
difficult to eliminate plausible alternative explanations for regression results. Limited data
also prevents credible estimation of important structural elasticities. Third, variation in the
demand for public transit is confounded with other covariates which influence air quality.
Anderson (2014), for example, demonstrates that mode demand is time-varying. Demand
for public transit is higher during rush hours when pollutant concentrations are also ele-
vated. City-specific economic conditions likewise lead to greater demand for buses and more
emissions from industrial production. Ignoring the potential endogeneity of public transit
investment can lead to large variation in policy-relevant parameters. Beaudoin, Farzin and
Lawell (2015), for example, show up to a 40% understatement of the benefits from transit in
a regression of air pollution on transit. We avoid many of these problems by combining high
frequency data with good-as-random variation in transit availability. While our results focus
on the short-run response to changes in transit availability, we believe that our compelling
identification strategy along with high resolution data make this paper a valuable contribu-
tion to the existing body of research in this area. Moreover, we view our paper as providing
evidence on the impact of transit on pollution in North America, a region with few existing
studies.
The remainder of the paper contains four sections. Section 2 presents a stylized model
that outlines the scope for an environmentally beneficial pollution swap as envisioned by
public transit advocates. Section 3 describes our primary econometric model including our
identification strategy and data. Section 4 presents our results. This section includes several
robustness checks and an investigation of intra-day heterogeneity. Section 5 concludes.

2 Scope for a Pollution Swap

Research demonstrating how public transit improves air quality implicitly makes assumptions
about commuters’ willingness to substitute between traveling and other goods and, more
importantly, their willingness to substitute between taking transit and using private vehicles.

4
Improved air quality from expanded transit systems depends on the prospect of a pollution
swap whereby pollution from cars is swapped for (hopefully less) pollution from public transit
vehicles. The potential for this substitution, of course, depends on commuters’ responsiveness
to changes in key attributes of transit such as its availability, quality and price. Section 3 uses
reduced form econometric models to estimate the effect of public transit on air quality, but
we begin by presenting a stylized model that outlines the mechanism through which public
transit affects air quality. Parameterizing this model using plausible values for Canada, we
illustrate that investments in public transit can indeed improve air quality, yet the scope for
an environmentally beneficial pollution swap is limited. Our later empirical results further
demonstrate that the necessary substitutions between public transit and private vehicles do
not appear to be observed in our data.

2.1 Analytical Model

Aggregate vehicular emissions depend on both technological and behavioral factors. Charac-
teristics such as vehicle age, vehicle size, trip frequency and average speed influence overall
emission rates. Fuel-type, in particular, has a central role in the level of different types emis-
sions. Panel A of Table 1 shows the per mile emission rates for NOX , CO, and PM2.5 , from
a model year 2010 diesel fueled bus and gasoline fueled car, based on estimates from Cai,
Burnham and Wang (2013).8,9 Per mile traveled, buses emit 1.3 grams of NOX , 1.1 grams of
CO, and 22 milligrams of PM2.5 . Cars, in contrast, emit substantially less NOX at 0.1 grams
per mile, three times as much CO at 2.9 grams per mile, and three times less PM2.5 at 7
milligrams per mile.10,11
8
In North America, buses typically use diesel, while private vehicles burn gasoline. Emissions per mile
are estimates for mileage-weighted lifetime emissions for each vehicle.
9
We concentrate on three main “criteria” pollutants that are released by vehicles: NOX , CO, and
PM2.5 (note that PM2.5 is both released directly by vehicles as well as formed indirectly in secondary
reactions). The choice of pollutants is motivated by data availability; however, cars and buses have differ-
ential emission rates for other pollutants as well. For example, cars emit more volatile organic compounds
per mile, while per mile large particulate matter emissions are greater with buses.
10
On an equal horsepower basis, gasoline engines emit roughly 28 times as much CO as diesel engines,
because the latter burn their fuel in excess air to ensure full combustion (Krivoshto et al., 2008). Particulate
matter can be both released directly from engines as well as formed through chemical reactions from primary
pollutants. The estimates in Cai, Burnham and Wang (2013) focus only on the former, so likely underestimate
total vehicle particulate matter emissions.
11
In Canada, about 40% of all public transit riders make at least a portion of their journey by rail transit
as opposed to bus transit. Subways and light rail typically uses electricity as an energy source. While
there is heterogeneity across regions, the majority of Canada’s electricity is generated by sources with no air
pollution emissions. We therefore disregard emissions from urban rail transit.

5
 Technological change dramatically improved both diesel and gasoline engine emission
rates over the past decades. Figure 1 illustrates how lifetime mileage-weighted emissions
changed between 1990 and 2010. In general, emissions are between 80-98% lower on a per-
mile basis. Panel A shows how CO emissions from diesel buses fell from 7.5 grams to 1 gram
per vehicle mile. Gasoline fueled cars experienced an even larger improvement, with CO
emissions decreasing from roughly 15 grams to 3 grams per vehicle mile. Panels B and C
display similar trends for NOX and PM2.5 . The NOX emissions from buses decreased from
nearly 25 grams per mile to 2 grams per mile, while the PM2.5 emission rate declined from
1 gram per mile to virtually zero by 2007.
Still, while there are obvious improvements in the emission rates of both passenger vehi-
cles and diesel buses, variation remains in their relative improvements over the period. Any
improvements in air quality due to a pollution swap must arise from the relative reduction in
bus emissions compared with cars. The dashed black lines in Figure 1 shows the time series
of the relative emission rates of diesel buses and gasoline-fueled cars. Unlike the secular
trend in absolute emission rates, there is no obvious pattern to relative emissions. For NOX ,
both gasoline cars and diesel bus emissions have improved, yet the rate of improvement was
more rapid for cars. As a result, between 2000 and 2010, NOX per bus-mile relative to the
NOX per car-mile increased substantially. The opposite pattern is true for PM2.5 : since 2007,
the figure shows a dramatic improvement in the emissions rate for diesel buses relative to
gasoline cars.
The technological parameters provide clear predictions about relative emission rates per
mile traveled, but neglect important behavioral adjustments made by commuters in re-
sponse to observed congestion, cost of travel and other factors. Regardless of technological
improvement, the overall effect of public transit availability on air quality is ambiguous since
it depends on these difficult to observe variables. Consider, for instance, several prospective
implications of an increase in gasoline taxes. If buses are under capacity, a gasoline tax
may motivate commuters to substitute from cars to buses. No change in NOX would be
expected (as the excess bus capacity is filled), but we might observe a reduction in CO as
fewer car trips are taken. This is the heart of the pollution-swap hypothesis: there is a
combination of technological and behavioral parameters that imply investments in public
transit can improve air quality. It is reasonable to imagine an alternative scenario, how-
ever: an identical car-to-bus substitution with an at-capacity transit system may induce
more bus trips as transit authorities add buses to meet the increased demand. This would,
ultimately, lead to an increase in NOX . Moreover, incorporating the fundamental law of

6
road congestion hypothesis (Downs, 1962; Duranton and Turner, 2011) with this at-capacity
scenario, any freed road space resulting from the initial car-to-bus substitution is occupied
with latent demand from passenger vehicles. The net effect then is an increase in NOX and
no change in CO. The point is that the relationship between public transit and air quality
– and whether there is scope for an environmentally beneficial pollution swap – therefore
depends on both technological parameters and behavioral responses. Historically it has been
challenging to obtain reliable, time-varying ridership and capacity data on public transit
usage. Aggregates are available, but as Anderson (2014) demonstrates, there is substantial
time-dependent variation in transit demand. Use of aggregate capacity and ridership values
therefore may introduce large biases in estimating the impact of transit on pollution.
A stylized model formalizes the interactions between commuter choice and technology
and how these determine the relationship between changes in the characteristics of public
transit and changes in emissions. A representative consumer chooses transport services (T )
and other goods (X) to maximize utility (subject to a standard budget constraint):

U = U (T, X).

Transport services are provided by through public transit (B) or private cars (D), such that:

T = T (B, D).

Assuming constant elasticity of substitution functions allows us to express the demand for
driving and public transit (relative to benchmark demand) as:
 
(σ −σ )
B = B(pB , pD , pX ) = cσUU cT T U p−σ
B
T

 
(σ −σ )
D = D(pB , pD , pX ) = cσUU cT T U p−σ
D
T
,

where cU is the price index (general cost function for utility), cT is the price of transport
services,12 and pD and pB are the price of driving and public transport, respectively.13 The

1
12
Given the constant elasticity of substitution function adopted, cU = θX p1−σ X
U
+ θT cT1−σU 1−σU and
1
cT = θD p1−σ + θB p1−σ

1−σ
T T T . The θ parameters refer to benchmark cost shares.
D B
13
The price of driving and public transport can include both pecuniary as well as non-pecuniary compo-
nents, including time costs, access costs, and other costs (Anderson, 2014; Parry and Small, 2009). In our
simple model, we do not connect the demand for private transport to congestion, such that the price of public
and private transport in our model is exogenous. It is possible to endogenize the price of private transport
to reflect congestion costs in a simple manner, by allowing the price of private transport to be a function of

7
parameters σU and σT reflect the elasticities of substitution between transport and other
goods and between driving and public transit, respectively. These elasticities of substitution
are critical for connecting commuter behavior to public transit and air quality.
Total emissions are the sum of pollution from buses and driving:

E = EB + ED
= BφB /ζB + DφD /ζD ,

where φ is the per-mile emissions of pollutants from public and private vehicles, and ζ is the
occupancy rate of each vehicle.14
This set-up enables us to derive an expression for the change in emissions that results
from a change in the price, availability or other characteristic of public transit. (Although
the empirical results later in the paper are based on a non-marginal supply shift, a price
change is used in this analytical model to tractably capture changing supply conditions.)
Changes in the price (availability) of public transit affects emissions as follows:

dE φB ∂B(pB , pD , pX ) φD ∂D(pB , pD , pX )
= +
dpB ζB ∂pB ζD ∂pB

This expression contains both technological and behavioral parameters. The first right-hand
side term is the product of change in the demand for bus services due to the change in the
price of public transport, ∂B(pB∂p,pBD ,pX ) , multiplied by the emissions rate for public transit, φζBB .
Increasing the price of transit causes transit demand to decrease, so this term is negative.
The second right-hand side term is the change in car travel due to the change in public
transit price, ∂D(pB∂p,pBD ,pX ) , multiplied by the emissions rate for passenger car travel, φζDD . For
an increase in transit price, car demand will increase as long as cars and buses are substitutes.
Hence, this term is positive. The sign of the overall change in emissions – i.e., the effect of
public transit availability on air pollution – depends on the relative rate of emissions from
cars compared with public transport and the magnitude of the relevant elasticities.
Given the assumed constant elasticity of substitution functional forms, it is possible to
write a closed-form solution for the elasticity of transport sector emissions with respect to
traffic volume, pD = Dψ (where we assume public transit does not cause congestion, for simplicity). A value
of ψ > 1 implies that additional driving imposes costs on others, which increase the price of driving. We do
not focus on this model variant, but do include a simulation results in which we allow ψ to take on positive
values below.
14
It is possible that vehicle occupancy rates are endogenous, but for simplicity, we do not consider that in
the simple model.

8
public transport price, ηZp . This expression is:

dE pB EB
≡ ηZp = θT θB σU + θB (σT − σU ) − σT . (1)
dpB E E

The response in emissions following a change in the price of public transport, ηZp , depends on
the following parameters: the elasticity of substitution between transport services and other
goods (σU ), the elasticity of substitution between transport modes (σT ), the cost shares of
all transport (θT ) and public transit (θT ) and the share of public transit in total transport
emissions ( EEB ).
The prospect of a pollution swap and an overall improvement in air quality from an
expansion of the transit system depends on the values of these parameters. To give a sense
of the magnitude and direction of the change in emissions due to a change in public transit
availability, we conduct a numerical simulation. Calibrated values for these parameters are
shown in Table 1. These coefficients are from several sources. Cost shares of transportation
and of driving in total transportation, shown in Table 1 Panel B, are from the 2014 Canadian
Survey of Household Spending.15 Panel C displays the elasticities of substitution. The
elasticity of substitution between transit and driving is set at 0.4 to match an elasticity of
transport demand typically found in the literature (see, for example, Litman, 2004). The
elasticity of substitution between travel and other goods is chosen to generate an own-price
elasticity of driving of -0.18, a value also consistent with the literature (Gillingham, 2014).
The pollution rates for cars and buses are from Cai, Burnham and Wang (2013), who produce
lifetime mileage-weighted emission rates for diesel buses and gasoline cars.16 The selected
emission rates correspond to the 2010 model year. We assume that 40% of total transit
passenger-miles are provided by (zero emission) passenger rail, and the remaining 60% are
provided by diesel bus, which reflects average Canadian configuration. The mode share for
public transit reflects the mode share in Canada’s larger cities.17 Finally, occupancy rates
of passenger cars and urban transit buses are from Office of Energy Efficiency (2014) and
determined by dividing passenger kilometres of travel by vehicle kilometres of travel.
Based on these parameters, we start by simulating the change in emissions that result
from doubling the price of public transit, pB . Our econometric models are identified using
15
Results are given here: http://www.statcan.gc.ca/daily-quotidien/160212/dq160212a-eng.htm.
16
As noted previously, we assume that urban rail produces no emissions.
17
Public transit mode share is at or above 20% in Canada’s largest cities. See https://www12.statcan.
gc.ca/nhs-enm/2011/as-sa/99-012-x/2011003/c-g/c-g01-eng.cfm. Note that the cost share of transit
is 0.1 and the mode share is 0.2, so by implication transit costs 2 times as much per kilometre as driving.

9
public transit strikes which completely remove public transit, so this does not perfectly
parallel the subsequent reduced form results. Still, the simulation yields useful intuition on
the scope of the effect of a large change in public transit availability.
Given the model’s parameterization, the results suggest that doubling the price of pub-
lic transport reduces NOX emissions from the transport sector by 2.2% and increases CO
emissions by a smaller amount equal to 1.9%. There is also a small, 0.7% reduction in
PM2.5 emissions. In other words, in the base case with realistic parameter values, one type
of pollution is swapped for another: an increase in CO emissions from cars is exchanged for
a decrease in NOX emissions from buses.
As is apparent in equation (1), the scope of a pollution swap clearly depends on the
elasticities of substitution. Figure 2 plots sensitivity analysis for these two parameters. The
left-hand panel illustrates how much the elasticity of emissions with respect to the price
of public transport changes with different values of the elasticity of substitution between
public and private transport, σT . High values of σT indicate that commuters are more
willing to switch between cars and buses. Buses produce more NOX emissions than cars,
so this switch entails a reduction in NOX that is more pronounced at higher elasticities.
In contrast, car travel produces more CO than bus travel, so the opposite is true for this
pollutant. When consumers are unwilling to switch between buses and cars – i.e., σT is small
– doubling the price of bus travel, unambiguously leads to improved air quality as both CO
and NOX emissions decline, hence a pollution swap can lead to improved overall air quality.
Figure 2 also shows what we would expect at the benchmark values where σT = 0.4. The
dashed line illustrates this base case. In this scenario, doubling the price of buses, reduces
the demand for bus trips and hence reduces NOX emissions. At the same time it increases the
demand for car trips and associated CO emissions. Given this elasticity of substitution and
other parameters values, the model predicts virtually no change in PM2.5 . The right-hand
panel of Figure 2 provides similar sensitivity analysis for σU , the degree to which consumers
reduce overall travel demand in response to an increase in bus prices. Values of σU greater
than approximately 0.5 unambiguously reduce all types of pollution. Again however, at
the benchmark parameterization, doubling the price of buses predicts less NOX pollution, a
slight increase in CO pollution, but little change in PM2.5 .
Table 2 reports additional sensitivity analyses where we vary assumptions about bus and
car occupancy rates, elasticities of substitution between travel and other goods and between
driving and transit, congestion feedback, as well as the share of rail in benchmark public
transport. These results suggest that, in general, increases in the price of public transit

10
reduce NOX emissions, but increase CO emissions. For instance, if we double the occupancy
rate of buses, ζB , from 12.5 to 25, and then double the price of bus travel, NOX falls by 0.3%
while CO increases by 2.0%. This NOX -for-CO pattern is maintained across most parameter
combinations. Effects on PM2.5 emissions are ambiguous and depend on model parameters.
In the final row of the table, we allow the price of driving to respond to the quantity of
driving, to roughly reflect congestion. This model crudely captures the “fundamental law of
road congestion,” which suggests that driving volumes are unchanged due to transit provision
(Duranton and Turner, 2011).18 In this scenario, changes in the price of public transit have
little impact on the demand for driving, and the results suggest that all three pollutants
experience declines due to increases in public transport price. In other words, under most
plausible scenarios, the best that investment in public transit can do is to swap one type of
pollutant for another, but unambiguous improvements in air quality due to investments in
public transit are unlikely. The next section then aims to to measure what actually happens
to air quality what transit is temporarily eliminated due to transit strikes.

3 Empirical Approach
The analytical model illustrates the mechanism linking public transit to air quality. This
channel relies on several hard-to-identify parameters. We therefore investigate the reduced
form relationship between transit and air quality using good-as-random variation in transit
availability. We first present our econometric strategy and then discuss the data.

3.1 Econometric Strategy

Transit strikes, where public transit systems cease operations due to labor negotiations,
produce a quasi-random variation in transit availability. This allows the effect of transit
availability on air pollution to be cleanly identified. Transit strikes typically reduce public
transit service to near zero in a local region19 and, as transit strikes are temporary, they
18
Duranton and Turner (2011) research focuses on the long run, while our empirical analysis is a short-run
analysis. We include this scenario for completeness.
19
Often public transit authorities will continue to provide service to elderly and special needs commuters.
In one case that we know of included in our sample (the Vancouver public transit strike in 2001), transit
strikes shut down bus public transit but subway public transit was unaffected. Further, it is possible for
private firms to provide temporary services in the absence of publicly-provided transit, but private operators
are rare in Canada. Given that public transit may not be fully eliminated, our estimates should be interpreted
as conservative (lower bound), intent-to-treat effects that result from imperfect compliance.

11
are unlikely to induce residential sorting. Moreover, while transit work actions potentially
affect traffic and air pollution in a region, they are unlikely to be affected by air pollution, so
causality is uni-directional.20 Transit strikes are therefore an exogenous source of variation
in transit availability that allows the short-run effect of public transit on air pollution to be
recovered.
There is an important caveat to this source of identifying variation however. This identi-
fication strategy yields the short-run response to a change in public transit availability. This
short-run response does not perfectly map onto long-run outcomes and, in general, it is the
long-run impact of public transit investment that is important from a public policy perspec-
tive. Yet, as mentioned, it is empirically challenging to estimate the long-run response of air
pollution to changes in public transit and the bulk of research on this topic estimates short-
run effects (e.g., Bauernschuster, Hener and Rainer, 2017; Chen and Whalley, 2012). While
our estimates are not the ideal policy-relevant parameter of interest, we do believe that the
plausibility of identification nevertheless allows us to provide useful results. In addition, the
setting of our results enables us to build on prior research that identifies short-run outcomes.
For example, Bauernschuster, Hener and Rainer (2017) use extremely short transit strikes
(less than a day; typically one to two hours) to reach conclusions about the relationship
between air pollution and transit availability. These very short-run responses may differ
quite significantly from our longer-run responses to changes in public transit. Likewise,
Chen and Whalley (2012) use a regression discontinuity design to measure the impact of a
subway opening on air pollution in China. Given this design, identification comes from the
immediate change in air pollution following a subway opening. In our study, many of the
transit strikes last for several weeks and, as a result, induce behavioral responses that can be
contrasted with those of Bauernschuster, Hener and Rainer (2017) and Chen and Whalley
(2012) as estimates that more closely parallel the long-run responses of interest.
Our primary empirical specification employs fixed effects to control for selection on un-
observables, by leveraging a large dataset comprised of daily (and hourly) air pollution data
in multiple cities across multiple years. We estimate the effect of public transit strikes on
the ambient concentration of pollution emissions by comparing measured pollution concen-
trations during a strike with pollution concentrations in the same city in the same year when
no strike is in place and on the same date in other cities with operating transit. We then
interpret the coefficient on our strike variable as the effect of an exogenous change in local
20
Although transit strikes are likely strategically timed, we can condition on observable determinants of
timing, as we discuss below, such that we argue that transit strikes are conditionally exogenous.

12
transit availability on air quality.
Our main specification is:

Ecymd = δ · strikecymd + Wcymd β + φcy + θymd + cymd . (2)

where Ecymd , the dependent variable, is average daily NOX , CO, or PM2.5 pollution concen-
trations in city c in a particular year, y, month, m, and day, d.21 It is important to highlight
a key difference between (2) and the discussion in Section 2. The model in the prior sec-
tion focused on transport sector emissions, whereas this econometric model uses ambient
pollution concentrations as a dependent variable (we use the same notation for each). The
two are related, but cities have other sources of emissions in addition to those from the
transportation sector and other variables such as weather affect pollution concentrations. So
there is an imperfect correlation between emissions and concentrations.
strikecymd is a dummy variable that takes a value of one if there is a transit strike on a
specific date, ymd, in municipality c, and zero otherwise. Our coefficient of interest is δ. δ
measures the change in pollution as a result of a transit strike. This can be interpreted as
the negative of the short-run change from an expanded public transit system. δ is a reduced
form coefficient that incorporates the behavioral and technological factors that govern the
relationship between buses, private vehicles, and air pollution. We estimate the model
separately for NOX , CO and PM2.5 , obtaining separate δ coefficients for each. Further,
while equation (2) is our preferred specification, we vary the resolution of our fixed effects
in robustness checks as a test of the sensitivity of our results to alternative identifying
assumptions.
There are a series of other variables in (2). Wcymd is a vector of weather covariates includ-
ing daily maximum temperature, mean temperature, mean temperature squared, minimum
temperature, daily precipitation and precipitation squared. Including weather is important
because air pollutants, once released, can be catalyzed by sun and high temperature. More-
over, precipitation can remove air pollution from the atmosphere. Conditioning on weather
further helps to address potential bias in coefficients that could be introduced if transit au-
thorities plan transit strikes to take advantage of particular weather events (for instance,
choosing to start a strike on a day when poor weather is forecast).22 φcy is a city-by-year
variable that captures time-varying city-specific trends. General economic and labor market
21
While the main results use daily data, we also include results that use hourly pollution data.
22
In an unreported robustness check, we test the sensitivity of our results to omitting all weather covariates.
The coefficient on strikes is not significantly affected by this change.

13
conditions influence pollution emissions (and hence concentrations), so this effect controls for
these unobservables. θymd is a day-of-sample fixed effect that captures other time-varying
confounders that are common across our panel. Holidays, for example, tend to have less
traffic than normal workdays. There is also seasonality in pollution variation and θymd helps
to capture these time-varying unobservables. cymd is an error term. All standard errors are
clustered at the city level allowing for arbitrary temporal correlation in error term.
The identifying assumption in (2) is that, conditional on included covariates, the specific
timing of strikes in each municipality is as good-as-randomly assigned. In other words, we
assume that the timing of strikes is not chosen in response to (or correlated with) observed
or predicted levels of air pollution. On the whole, this assumption is mild, but potential
violations could occur if strikes were timed to create maximum disruption for a city, for
example by holding a strike on the first day of September (when students return to school).
The data suggest that this is not the case: there is no statistical pattern to the start date of
strikes in the data (results are presented below). More importantly, our empirical strategy,
which includes date fixed effects, controls for the variation in pollution concentrations by date
(there is a separate fixed effect for every day of the sample in our preferred specification). In
addition, to measure the sensitivity of our results to this identifying assumption, we conduct
falsification tests where we implement “fake” strikes, which are shifted in time relative to
the “real” strikes we observe in our data. We expect and find no effect due on pollution
emissions attributable to a “fake” strike in these placebo models.

3.2 Data
Our research design requires data on air pollution, weather and work actions. These are
assembled from several sources.
Air pollution data are from Canada’s National Air Pollution Surveillance Program (NAPS).23
NAPS was established to provide long-term air quality data at a uniform standard from
across Canada. Launched in 1969 with 36 air quality monitoring stations, it has expanded
significantly over time. Our data spans 1974 through 2011 for 18 municipalities for CO. For
NOX , it runs from 1980 through 2011 for 13 cities. Measurement of PM2.5 runs from 1997
through 2011 for 9 municipalities.24 NAPS monitoring stations measure ambient pollution
23
A description of NAPS, along with supporting documentation regarding monitoring protocols and sta-
tion locations, is available at Environment Canada’s website: http://www.ec.gc.ca/rnspa-naps/Default.
asp?lang=En&n=5C0D33CF-1.
24
These include Toronto, Montreal, Vancouver, Ottawa, Calgary, Edmonton, Hamilton, Quebec City,
Kitchener-Waterloo, London, Victoria, Oshawa, St. Catherines, Halifax, Windsor, Saskatoon, Regina, Win-

14
concentrations for a large number of substances. Most stations continuously monitor CO,
NOX , ground level ozone, sulfur dioxide and particulate matter (PM10 and PM2.5 ). Air sam-
ples are also periodically analyzed at a central lab facility to identify cumulative levels of
over 100 additional substances. We focus on CO, NOX and PM2.5 as these are the primary
pollutants commonly emitted by vehicles.25
Assembling the dataset required assigning pollution monitoring stations to cities. Figure
3 illustrates the procedure for mapping monitoring stations to city centers. Air pollution
monitoring stations are given as black crosses. Fixed radii circles were drawn around each
city centre (as measured by the coordinates of the City Hall). The top panel shows both
the dispersion of monitoring stations and cities in our data. The bottom panel provides an
example of the fixed radii mapping assignment of stations to cities. For several small cities,
only a single monitoring station is captured within the circle. Larger cities such as Toronto
and Vancouver have several monitoring stations recording air quality. In these cases, the
closest monitoring station to city centre is used. As NOX and PM2.5 disperse easily while CO
is a local pollutant (Gaur et al. (2014) and Rattigan et al. (2010)), we further investigate the
sensitivity of our analysis to the choice of pollution monitoring stations, by estimating results
on a sub-sample of monitoring stations that are within 2 kilometres of major thoroughfares.
In robustness checks, we also test the sensitivity of our results to different ways of assigning
pollution monitors to cities. In particular, we use the inverse-distance weighted average of
all pollution monitoring stations within 5 and 10 miles of the city centre, rather than simply
assigning the closest monitor to the city centre. We show that our results are not affected
by the choice of assignment mechanism.
Information on transit strikes are from Employment and Social Development Canada’s
Workplace Information Division. This division tracks collective agreements and maintains
a database of all work stoppages categorized according to North American Industry Classi-
fication System (NAICS) codes. We use NAICS code 485110: urban transit systems. Our
data on work stoppages capture, among other variables, the dates that the action began and
nipeg and St. John’s. Cities range in size from approximately 5 million inhabitants (Toronto) to 0.2 million
(St. John’s). Montreal, Quebec City and Victoria are excluded from the NOX analysis as data are missing
for these cities. The differences in the number of cities across samples is attributable to the absence of strikes
in some municipalities for periods in which we have air pollution data.
25
Ozone is a key vehicle-related pollutant as well. However, ozone is not released directly by vehicles but
instead forms from a reaction between NOX and volatile organic compounds in the presence of sunlight and
heat (i.e., it is not a primary pollutant). The chemical transformation is highly non-linear, and so predicting
and modeling the effect of vehicles on ozone emissions is challenging. For this reason, we focus on primary
pollutants in this study.

15
ended, the number of workers affected, the location and the names of the transit corporation
as well as the union. There were 105 municipal transit strikes between 1974 and 2011. There
is little systematic tendency for strikes to occur in any particular season and, given gaps in
pollution data, not all strikes are common across pollutants. Figure 4b plots the number of
observed and initiated strikes (i.e., the first day of work stoppage) for each month. The figure
shows roughly 8.75 strikes have been initiated in each month. Visually there appears to be a
slight tendency for strikes to start in Fall, but this tendency is not statistically significant.26
Our regressions use a variety of fixed effects to control for most systematic seasonal trends.
On average each strike lasts for 19.19 days, with a minimum of 1 day and a maximum of 87
days. Figure 4a illustrate a histogram of strike lengths over all cities while Figure 4c shows
the cross-sectional variation in observed strikes by city. As shown, over the period of 1974
to 2011 there were 57 observed strikes in Montreal while only 1 strike occurred in Regina.
Measurement of pollution is correlated with weather. Data on meteorological conditions
are obtained from Environment Canada.27 To control for the effect of weather conditions
on the ambient air pollution, polynomials of maximum, minimum and mean temperature as
well as precipitation are included. As with the NAPS monitoring stations, meteorological
conditions are assigned to city centers using information from the closest weather stations.
Table 3 presents some key information for the NOX , CO, and PM2.5 in our sample. There
are three panels. Panel A presents an overview of the sample. There are 11,216 unique days
for 13 cities in the NOX sample. For CO, there are 18 cities and 13,038 days. Pollution
monitors require maintenance and occasionally fail to record the level of ambient pollution.
The NAPS system records these as missing values which we drop from our analysis. We are
unaware of any systematic bias associated with these missing values. There are 944 strike
work-days in our NOX and 411 stoppage days in our PM2.5 sample, a considerable reduction
from the 2,155 in the CO sample. A shorter period of analysis and the lack of NOX data for
cities in Quebec account for the reduced number of strike days.
Panel B presents summary statistics for our dependent and weather variables. The mean
NOX concentration is 35.97 parts per billion (ppb) with a standard deviation of 31.52. For
CO, the mean value is 0.96 parts per million (ppm) with a standard deviation of 0.98. For
PM2.5 the mean value is 7.80 µ/m3 with a standard deviation of 6.38.
Finally, Panel C shows the residual variation in the dependent variable after it is re-
26
A Pearson Chi-squared test fails to reject the null hypothesis of equal proportions of strikes by month:
2
χ = 14.09, df = 11, p = 0.22.
27
Climate data is available at Environment Canada’s website: http://climate.weather.gc.ca/.

16
gressed on the suite of fixed effects and weather controls from (2). These include NAPS-year
and day-of-sample fixed effects as well as maximum, minimum and mean city-specific daily
temperatures, a quadratic in mean temperature, precipitation and a quadratic in precipita-
tion. Even with this rich set of fixed effects, substantial residual variation is apparent. Fixed
effects and weather explain 47% of the variation in NOX , 65% of the variation in CO and
41% of the variation in PM2.5 .

4 Results
Four sets of results are presented. We first provide our main results plus several robustness
checks. We then examine intra-day heterogeneity and the implications of public transit on
maximum daily pollution readings. Next, technological change has had a large role in the
relationship between transit and air quality, so we separately examine pre- and post-2000
effects. Finally, we show selected falsification tests. Throughout, the message is consistent:
public transit increases ambient NOX concentrations while having little to no effect on CO
or PM2.5 .

4.1 Main results

Table 4 shows the main results using the daily mean pollution level as the dependent variable.
Nine columns are presented, three for each pollutant. The different columns alter the source
of identifying variation by using different combinations of time and location fixed effects. All
standard errors are clustered at the city level to control for arbitrary temporal correlation
in error term.28
Columns (1) to (3) of Table 4 show results for NOX emissions, with all three columns
showing stable estimates. In (1), using date and NAPS-year fixed effects, the removal of
transit improves ambient NOX concentrations by 3.8ppb. The corresponding models for CO
and PM2.5 are in columns (4) and (7). A transit strike leads to a 0.02ppm increase in CO,
but the standard error is large. Likewise, removing buses reduces PM2.5 by 0.8 µg/m3 , again
with a large standard error.
Changing the source of identifying variation has little influence on any of the point
28
Models are also run that allow for spatial correlation by clustering on year-month and temporal and
spatial correlation by clustering at city-year and city and year. The standard errors are smaller than those
obtained by clustering by NAPS monitoring station suggesting that the confidence intervals presented here
may be conservative.

17
estimates or standard errors. Columns (2) and (3) show that a strike is associated with
statistically significant decrease of 3.77 to 3.98 ppb in NOX concentrations. In (5) and (6),
CO concentrations still increase with transit strikes, but again the parameters are imprecisely
estimated. Finally, removal of public transit leads to a decrease in PM2.5 of 0.5 to 0.8 µg/m3 ,
but this reduction is not statistically distinguishable from zero.
It is worth noting that the estimates from the econometric models corroborate those
generated from the simulation model. Table 3 shows that the mean NOX concentration in
our data is 36.0 ppb, so a 3.8 ppb reduction suggests that removal of transit reduces NOX by
approximately 10%. This is of similar magnitude, albeit slightly larger, than the estimates
presented in Table 2 which reflect a non-complete removal of publicly provided transit.
Further, while the point estimates on CO concentrations are not distinguishable from zero,
the econometric model implies a roughly 2% reduction in CO concentrations due to buses.
Again, this is approximately within the range of the values in Table 2. Similar to the CO
estimates, the PM2.5 coefficients are imprecise. Moreover, Table 2 does not provide a clear
prediction on the magnitude or sign of a change in PM2.5 . The imprecision of the econometric
point estimates warrants care, yet, the general pattern of results suggests that a pollution
swap – NOX for CO – is possible. Unfortunately, the strongest conclusion emerging from our
results is that NOX is reduced when public transit service is cut; we cannot be unequivocal
in our assertions about CO given the impreciseness of the estimates.

Sensitivity to Pollution Monitor Assignment

As described in Section 3, we impute a city’s pollution level by using the value from the
closest monitor to the city centre. Table 5 uses three alternative methods of assigning
pollution to cities. Column (1), (3) and (7) use the inverse-distance weighted average values
of all monitors within 5 mile radius of city centroids. Columns (2), (5) and (8) use the
inverse-distance weighted average within a larger 10 mile radius for each pollutant. Finally,
column (3), (6) and (9) restrict the analysis to monitors near major thoroughfares. NOX and
PM2.5 disperse more easily while CO is a local pollutant, so the location of the pollution
monitor has the potential to notably influence the results.
The NOX estimates are consistent with Table 4. The coefficients show a reduction of
between 3.6 and 3.9ppb in NOX concentrations associated with a public transit strike. Pre-
cision remains a problem for CO and PM2.5 . While the sign and magnitude of the point
estimates are similar to the models in Table 4, the confidence intervals are wide and include
zero. Monitor assignment is immaterial to the results.

18
4.2 Hourly Pollution Concentrations
Strikes vary at the daily level in our data, so our preferred models concentrate on daily
average pollution levels. Intra-day heterogeneity, especially with respect to heavy traffic
periods, may however have distinct patterns that are not captured by the daily average.
To investigate this intra-day heterogeneity, we estimate (2) for each of the 24 hours in a
day, where, for instance, the 7:00am interval represents the hour from 7:00 to 7:59am. Rather
than using daily average pollution concentration as the dependent variable, the pollution
concentration for that specific hour is used. All control variables remain the same. In
particular, we use our preferred specification, which includes weather controls, date fixed
effects and pollution monitor-year fixed effects. Separate regressions are conducted for each
of NOX , CO, and PM2.5 for each hour.
Figure 5 plots the point estimates from these regressions. A transit strike causes a
statistically significant reduction in atmospheric NOX concentrations throughout the day.
These reductions are only statistically significantly different than zero during the standard
workday however. Outside of the 7:00am to 5:00pm stretch, the parameters are too imprecise
to be statistically distinguishable from zero. The patterns for CO and PM2.5 illustrate the
difficulty in making claims about these pollutants. In both cases, the point estimates, which
are signified by the dark circles, fluctuate around zero, while the error bars are wide. It
does appear, however, that CO increases during the morning and afternoon travel peak in
response to a strike, while PM2.5 decreases (although, as shown by the figure, coefficients
are not statistically significant). In general, these hourly regressions reinforce the results
from the earlier tables: public transit increases NOX pollution but has little effect on CO or
PM2.5 .
The results in Figure 5 suggest that removing public transit has larger effects on atmo-
spheric pollution concentrations during rush hours. Table 6 looks at this from a slightly
different perspective. It presents results using the daily maximum pollution concentration
as the dependent variable. Daily maximum pollution levels typically occurs during rush hour
periods, but this is not strictly the case. Columns (1) to (3) show point estimates for the
effect of a transit strike on daily maximum pollution readings. Several elements of these
models are worth highlighting. First, compared with the preferred specifications in Table 4,
the point estimates are notably larger. The strike coefficient in the NOX model has nearly
doubled. For CO, the point estimate is seven times larger when using daily maximum pol-
lution than in Table 4, while the parameter on PM2.5 increased nearly threefold. Statistical
significance remains elusive of CO and PM2.5 . Further, given the disproportionate increase

19
in the standard error in the NOX model, we are able to claim less about the effect of public
transit on daily maximum pollution concentrations. It is only possible to reject a null hy-
pothesis of no effect at a 10% level. Nonetheless, the results do tend to support the general
inference that public transit does little to improve local air quality.

4.3 Transit Strikes and Technological Change

In the discussion of the analytical model, we highlighted the large changes in transporta-
tion technology over the past 20 years. During this period, both cars and buses became
substantially cleaner. In this section, we examine whether these changes in technology are
observable in the data. In particular, we sub-divide our sample into two periods – pre-2000
and post-2000.29 We then conduct separate regressions on each of these sub-samples.
Figure 1 provides suggestive predictions for the results. While, say, the NOX emission
rate for both cars and buses improved, the relative performance for buses compared to cars
fluctuated over time. In the pre-2000 period, buses emitted between 10 to 25 times more
NOX compared to cars. After 2000, this ratio spikes to more 40 times on a per mile basis
before declining.30 Thus we might expect the impact on a strike on NOX to be larger
after 2000 even though post-2000 buses are cleaner than pre-2000 buses. This prediction
is confirmed in Table 7. Columns (1) and (4) illustrate the effect of a transit strike on
NOX pollution before and after 2000, respectively. A work stoppage leads to a small 0.3 ppb
reductin in NOX in the early period. The latter period shows a much larger effect of 10.1
ppb.
The opposite is pattern holds for PM2.5 . Figure 1 suggests a continuous performance
improvement in the emissions rate of buses relative to cars post-2000. We therefore expect a
transit strike to have a beneficial impact on PM2.5 before 2000 and less so later in the sample.
Again, this prediction is confirmed in Table 7, which indicates a statistically significant
reduction in PM2.5 of 3.7 µg/m3 before 2000. After 2000, the coefficient is not meaningfully
different from zero.
Results for CO are less intuitive. Figure 1 suggests a large improvement in CO emissions
for buses relative to cars during the latter period of our sample, so we expect that a transit
strike will have more detrimental impacts on CO later in the sample. This is not what
29
We choose this date as it roughly represents the mid-point of our observations over all three pollutants.
30
It is important to note that the emissions rates presented in Figure 1 are a mileage-weighted average
over a vehicle’s entire lifetime, not average emissions rates for the on-road vehicle stock (which we do not
observe). As a result, the temporal correspondence between emissions from on-road transit and predicted
lifetime-weighted emissions rates is unlikely to be perfect.

20
we find. Instead, Table 7 shows that after 2000, a transit strike yields a reduction in CO
emissions, suggesting that transit increases emissions. These coefficients are at odds with
the technological predictions.31 Yet, it is important to emphasize that the coefficient on
the pre-2000 CO impact is not precisely estimated and the sign on the post-2000 period
reinforces the general conclusion that public transit does not improve local air quality.

4.4 Placebo Models

Table 3 shows substantial residual variation in pollution concentrations even after controlling
for a wide array of fixed effects. One concern with our research design may be that the
transit strike indicator is actually capturing other systematic unobservable variables. We
claimed that transit strikes caused a change in NOX concentrations but had no statistically
identifiable causal effect on CO or PM2.5 . Table 8 probes this claim via a falsification test,
whose motivation is that we should not observe an effect where we do not expect one.
Table 8 presents three columns of results for placebo transit strikes, where we time shift
strikes in the dataset. We replace our real strike variable with a fake strike variable that
occurs exactly one calendar year later. Column (1) shows the effect of the placebo transit
strike on NOX . The coefficient remains negative but is less than one third the magnitude
of our preferred specification and is not statistically distinguishable from zero. For CO,
the point estimate actually increases in magnitude, but remains statistically insignificantly
different from zero. Finally, for PM2.5 the point estimate is of a different sign from the
estimates in Table 4 and is likewise imprecise. Notwithstanding the residual variation in the
dependent variables, we do not appear to detect an effect where one is not expected. We
view this as corroborating our conclusions and as support the causal interpretation of the
main models.

5 Conclusion
This paper measures the effect of public transit on air quality in 18 major Canadian cities.
Transit strikes are used to identify the causal short-run effect of buses on local air pollution.
Our main result is that removing public transit causes a 3.5 part per billion, or about
10%, short-run reduction in ambient NOX concentrations. These reduced form results are
31
The scenario in Table 2 in which we crudely simulate congestion yields predictions that are closer in line
with these findings.

21
robust to different methods of assigning pollution to cities, to several distinct sources of
identifying variation and also holds up to placebo tests. A calibrated simulation model
further suggests that it possible for transit strikes to improve air quality via a pollution
swap, but demonstrates that environmentally advantageous swaps are unlikely to observed
at benchmark values in North America. Overall, this study suggests that, in contrast to
existing studies that take place in European cities or focus on the expansion of subways,
increasing public transit capacity in it’s current configuration in North America is unlikely
to improve local air quality.

22
Bibliography
Aftabuzzaman, Md, Graham Currie, and Majid Sarvi. 2010. “Evaluating the con-
gestion relief impacts of public transport in monetary terms.” Journal of Public Trans-
portation, 13(1): 1.

Anderson, Michael L. 2014. “Subways, strikes, and slowdowns: The impacts of public
transit on traffic congestion.” American Economic Review, 104(9): 2763–96.

Bauernschuster, Stefan, Timo Hener, and Helmut Rainer. 2017. “When labor dis-
putes bring cities to a standstill: The impact of public transit strikes on traffic, accidents,
air Pollution, and health.” American Economic Journal: Economic Policy, 9(1): 1–37.

Beaudoin, Justin, and Cynthia Lin Lawell. 2016. “Is public transit’s “green” reputa-
tion deserved? Evaluating the effects of transit supply on air quality.” Working paper,
University of California at Davis.

Beaudoin, Justin, Y Hossein Farzin, and C-Y Cynthia Lin. 2014. “Evaluating Public
Transit Investment in Congested Cities.” Working paper, University of California at Davis.

Beaudoin, Justin, Y Hossein Farzin, and C-Y Cynthia Lin Lawell. 2015. “Public
transit investment and sustainable transportation: A review of studies of transit’s impact
on traffic congestion and air quality.” Research in Transportation Economics, 52: 15–22.

Bel, Germà, Maximilian Holst, et al. 2015. “Evaluation of the Impact of Bus Rapid
Transit on Air Pollution.” University of Barcelona, Research Institute of Applied Eco-
nomics.

Bharadwaj, Prashant, Matthew Gibson, Joshua Graff Zivin, and Christopher

Neilson. 2017. “Gray matters: fetal pollution exposure and human capital formation.”
Journal of the Association of Environmental and Resource Economists, 4(2): 505–542.

Cai, H, A Burnham, and M Wang. 2013. “Updated emission factors of air pollutants
from vehicle operations in GREETT M using MOVES.” Argonne National Laboratory.

Chang, Tom, Joshua Graff Zivin, Tal Gross, and Matthew Neidell. 2016. “Par-
ticulate Pollution and the Productivity of Pear Packers.” American Economic Journal:
Economic Policy, 8(3): 141–69.

23
Chay, Kenneth Y, and Michael Greenstone. 2003. “The impact of air pollution on
infant mortality: Evidence from geographic variation in pollution shocks induced by a
recession.” Quarterly Journal of Economics, 118(3): 1121–1167.

Chen, Yihsu, and Alexander Whalley. 2012. “Green Infrastructure: The effects of Ur-
ban mass transit on air quality.” American Economic Journal: Economic Policy, 3(4): 58–
97.

Currie, Janet, and Matthew Neidell. 2005. “Air pollution and infant health: What
can we learn from California’s recent experience.” Quarterly Journal of Economics,
3(120): 1003–1030.

Currie, Janet, and W Reed Walker. 2011. “Traffic congestion and infant health: Evi-
dence from E-ZPass.” American Economic Journal: Applied Economics, 1(3): 65–90.

Currie, Janet, Eric A Hanushek, E Megan Kahn, Matthew Neidell, and Steven G
Rivkin. 2009. “Does pollution increase school absences?” The Review of Economics and
Statistics, 91(4): 682–694.

Downs, Anthony. 1962. “The law of peak-hour expressway congestion.” Traffic Quarterly,
16(3).

Duranton, Gilles, and Matthew A Turner. 2011. “The fundamental law of road con-
gestion: Evidence from US cities.” American Economic Review, 101(6): 2616–2652.

Friedman, Michael S, Kenneth E Powell, Lori Hutwagner, LeRoy M Graham, and

W Gerald Teague. 2001. “Impact of changes in transportation and commuting behaviors
during the 1996 Summer Olympic Games in Atlanta on air quality and childhood asthma.”
Jama, 285(7): 897–905.

Gaur, Abhishek, SN Tripathi, VP Kanawade, Vinod Tare, and SP Shukla. 2014.

“Four-year measurements of trace gases (SO2, NOx, CO, and O3) at an urban location,
Kanpur, in Northern India.” Journal of Atmospheric Chemistry, 71(4): 283–301.

Gendron-Carrier, Nicolas, Marco Gonzalez-Navarro, Stefano Polloni, and

Matthew A Turner. 2017. “Subways and urban air pollution.”

Gillingham, Kenneth. 2014. “Identifying the elasticity of driving: evidence from a gasoline
price shock in California.” Regional Science and Urban Economics, 47: 13–24.

24
Hilber, Christian AL, and Charles Palmer. 2014. “Urban development and air pol-
lution: Evidence from a global panel of cities.” Grantham Research Institute on Climate
Change and the Environment Working Paper series.

Jerrett, Michael, Rob McConnell, Jennifer Wolch, Roger Chang, Claudia Lam,
Genevieve Dunton, Frank Gilliland, Fred Lurmann, Talat Islam, and Kiros
Berhane. 2014. “Traffic-related air pollution and obesity formation in children: A longi-
tudinal, multilevel analysis.” Environ Health, 13(1): 49.

Knittel, Christopher R, Douglas L Miller, and Nicholas J Sanders. 2011. “Cau-

tion, drivers! Children present: Traffic, pollution, and infant health.” National Bureau of
Economic Research.

Krämer, Ursula, Thilo Koch, Ulrich Ranft, Johannes Ring, and Heidrun
Behrendt. 2000. “Traffic-related air pollution is associated with atopy in children liv-
ing in urban areas.” Epidemiology, 11(1): 64–70.

Krivoshto, Irina N, John R Richards, Timothy E Albertson, and Robert W

Derlet. 2008. “The toxicity of diesel exhaust: implications for primary care.” The Journal
of the American Board of Family Medicine, 21(1): 55–62.

Künzli, Nino, Reinhard Kaiser, Sylvia Medina, M Studnicka, Olivier Chanel,

Paul Filliger, Max Herry, F Horak, V Puybonnieux-Texier, Philippe Quénel,
et al. 2000. “Public-health impact of outdoor and traffic-related air pollution: A European
assessment.” The Lancet, 356(9232): 795–801.

Lalive, Rafael, Simon Luechinger, and Armin Schmutzler. 2013. “Does Supporting
Passenger Railways Reduce Road Traffic Externalities?”

Lavy, Victor, Avraham Ebenstein, and Sefi Roth. 2014. “The Impact of Short Term
Exposure to Ambient Air Pollution on Cognitive Performance and Human Capital For-
mation.” National Bureau of Economic Research.

Levy, Jonathan I, Jonathan J Buonocore, and Katherine Von Stackelberg. 2010.

“Evaluation of the public health impacts of traffic congestion: a health risk assessment.”
Environmental health, 9(65): 1–12.

Litman, Todd. 2004. “Transit price elasticities and cross-elasticities.” Journal of Public
Transportation, 7(2): 3.

25
Neidell, Matthew. 2009. “Information, Avoidance behavior, and health the effect of ozone
on asthma hospitalizations.” Journal of Human Resources, 44(2): 450–478.

Office of Energy Efficiency. 2014. “Comprehensive Energy Use Database.”

Parry, Ian WH, and Kenneth A Small. 2009. “Should urban transit subsidies be re-
duced?” American Economic Review, 99(3): 700–724.

Rattigan, Oliver V, H Dirk Felton, Min-Suk Bae, James J Schwab, and Ken-
neth L Demerjian. 2010. “Multi-year hourly PM 2.5 carbon measurements in New York:
diurnal, day of week and seasonal patterns.” Atmospheric Environment, 44(16): 2043–2053.

Sharman, Jon. 2016. “Paris makes all public transport free in

battle against ’worst air pollution for 10 years’.” Independent,
“http://www.independent.co.uk/news/world/europe/paris-public-transport-free-air-
pollution-spike-a7460191.html”.

Toronto Environmental Alliance, TEA. 2010. “Clearing the Air on the TTC.”
“http://www.torontoenvironment.org/campaigns/transit/clearingtheair”.

Winston, Clifford, and Vikram Maheshri. 2007. “On the social desirability of urban
rail transit systems.” Journal of Urban Economics, 62(2): 362–382.

Xia, Ting, Monika Nitschke, Ying Zhang, Pushan Shah, Shona Crabb, and Alana
Hansen. 2015. “Traffic-related air pollution and health co-benefits of alternative transport
in Adelaide, South Australia.” Environment international, 74(1): 281–290.

Zivin, Joshua S Graff, and Matthew J Neidell. 2012. “The impact of pollution on
worker productivity.” American Eeconomic Review, 102(7): 3652–3673.

26
Figures

Figure 1: Per-Mile Lifetime Mileage-Weighted Emission Rates

Emission rate (g CO/vehicle mile)

15

Bus/car emissions
10
0.8

5
0.6

0 0.4
1990 1995 2000 2005 2010
Vehicle model year
Emission rate (g NOX /vehicle mile)

25
40
20
Bus/car emissions
15 30
10
20
5

0
10
1990 1995 2000 2005 2010
Vehicle model year
Gasoline passenger vehicle
Diesel transit bus
Relative bus/car
Continuted on following page. . .

27
 Figure 1: Per-Mile Lifetime Mileage-Weighted Emission Rates.
Continued from previous page.

Emission rate (g PM2.5 /vehicle mile)

1 80

Bus/car emissions
0,8
60
0,6

0,4 40

0,2
20
0
1990 1995 2000 2005 2010
Vehicle model year
Gasoline passenger vehicle
Diesel transit bus
Relative bus/car

Notes: Source: Cai, Burnham and Wang (2013).

The red and blue solid lines indicate the emission rate of diesel transit buses and gasoline passenger vehicles,
respectively, and correspond to the left axis. The dashed black line measures the relative per mile bus to car
emissions and corresponds to the right axis.

28
 Figure 2: Elasticity of Pollution Emissions With Respect to Price of Public Transport

0.10 0.10
CO

PM2.5
0.00 0.00
CO
ηZp

NOX PM2.5
−0.10 −0.10
NOX

−0.20 −0.20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
σT σU
Notes: The vertical dashed lines indicate assumed benchmark elasticities of substitution.

29
 Figure 3: Mapping Pollution Stations to Cities

Notes: Each city in our analysis is represented by a circle while air pollution monitoring stations are black
crosses. Pollution monitoring stations were mapped to cities according to several weighting procedure ac-
cording to their distance from each city’s center.

30
 (a) Duration of Strikes in Days

(b) Frequency and Total Strike Days by Month

31

(c) Cross-Sectional Distribution of Strikes

Figure 4: Duration, Frequency and Distribution of Strikes Across Cities and Months
 Figure 5: Change in Hourly Pollutant Concentration due to Transit Strike
NOx CO

PM2.5

This figure plots point estimates from hourly regressions for each of the 24 hours in the day, using the level of
pollution concentration in that hour as the dependent variable. Gray whiskers are 95% confidence intervals.

32
Tables

Table 1: Parameters used to Calibrate Simulation Model

Symbol Value Description

A. Pollution rates
φN
B
OX
1.31 Per mile diesel bus NOX emissions (g/mile)
N OX
φD 0.12 Per mile gasoline passenger car NOX emissions (g/mile)
CO
φB 1.10 Per mile diesel bus CO emissions (g/mile)
CO
φD 2.87 Per mile gasoline passenger car CO emissions (g/mile)
P M 25
φB 21.8 Per mile diesel bus PM2.5 emissions (mg/mile)
P M 25
φD 7.1 Per mile gasoline passenger car PM2.5 emissions (mg/mile)
- 0.4 Share of public transit miles provided by passenger rail
B. Cost shares and operational parameters
θX 0.8 Cost share for other goods
θT 0.2 Budget share for transport
θD 0.9 Cost share of driving in transport
θB 0.1 Cost share of transit in transport
ζB 12.5 Bus occupancy
ζD 1.58 Car occupancy
γB 0.2 Bus mode share
C. Behavioural parameters
σU 0.2 Elasticity of substitution between transport and other goods
σT 0.4 Elasticity of substitution between transit and driving

33
 Table 2: Sensitivity of the Simulation Model to Calibrated Parameters

Scenario ∆ NOX ∆ CO ∆ PM2.5

Baseline -2.2% 1.9% -0.7%
ζB = 25 -0.3% 2.0% -1.3%
ζD = 1 -0.8% 1.9% 1.2%
σU = 0.4 -3.5% 0.5% -0.7%
σT = 0.8 -2.8% 4.5% 2.4%
No rail -4.3% 1.7% 0.1%
ψ = 100 (congestion) -3.8% -0.1% -1.2%
The results show the simulated change in transportation sector NOX , CO andPM2.5 emissions that result
from a doubling of public transit fares. “No rail” is a simulation where all public transport is provided by
bus, with no rail public transport. ψ = 100 is a simulation in which the price of driving is endogenous with
the elasticity of price with respect to driving given by ψ.

34
Table 3: Key Information on Samples for NOX , CO and PM2.5
Concentrations

Panel A – Overview of Sample

NOX CO PM2.5
Days in sample 11,216 13,038 4,964
Number of strike-days 944 2,155 411
Number of strikes 83 105 25
Number of cities 13 18 9

Panel B – Summary Statistics

Mean Std Dev
NOX (ppb) 35.97 31.52
CO (ppm) 0.96 0.98
PM2.5 (µ/m3 ) 7.80 6.38

Precipitation (mm) 2.37 5.96

Maximum temperature (◦ C) 11.51 11.73
Minimum temperature (◦ C) 1.86 10.80
Mean temperature (◦ C) 6.70 11.09

Panel C – Variation Explained by Fixed Effects and Weather

Dependent variable R-squared
Daily average NOX concentration 0.47
Daily average CO concentration 0.65
Hourly PM2.5 concentration 0.41
Sources: Air pollution data are obtained from Canada’s National Air Pollution
Surveillance Program (NAPS). We obtain data on transit strikes from Employment
and Social Development Canada’s Workplace Information Division. Weather data
came from the Environment Canada.

Panel C regresses the dependent variable on NAPS-year, month-day and hour

fixed effects as well as the full suite of weather controls as described in equation (2).
Substantial residual variation remains for NOX , PM2.5 and CO.

35
 Table 4: Change in Daily Pollutant Concentrations due to Transit Strikes

NOX CO PM2.5
(1) (2) (3) (4) (5) (6) (7) (8) (9)

Transit Strike -3.771∗∗ -3.906∗∗ -3.988∗ 0.0180 0.0129 0.0371 -0.841 -0.832 -0.586
[1.695] [1.685] [2.108] [0.0673] [0.0668] [0.109] [0.715] [0.729] [0.506]

36
Observations 109,613 109,613 109,613 160,352 160,352 160,352 46,734 46,734 46,734
Weather controls 3 3 3 3 3 3 3 3 3
Date fixed effects 3 - - 3 - - 3 - -
NAPS-id fixed effects - 3 3 - 3 3 - 3 3
NAPS-year fixed effects 3 3 - 3 3 - 3 3 -
Year-month fixed effects - - 3 - - 3 - - 3
Month-day fixed effects - 3 - - 3 - - 3 -
Notes: Dependent variable is daily pollutant concentrations. Weather covariates include temperature, quadratic temperature, min-
imum temperature, precipitation and quadratic precipitation. All weather covariates are daily average. Values in parentheses are
standard errors clustered by naps-id. ∗ significant at 10% ∗∗ significant at 5% ∗∗∗ significant at 1%.
 Table 5: Change in Daily Average Pollution Concentrations due Transit Strikes: Alternative Pollution Monitoring
Stations

NOX CO PM2.5
5 miles 10 miles Close to 5 miles 10 miles Close to 5 miles 10 miles Close to
w. avg. roads w. avg. roads w. avg. roads
(1) (2) (3) (4) (5) (6) (7) (8) (9)

37
Transit Strike -3.655∗∗ -3.910∗∗ -3.574∗∗ 0.0308 0.0366 0.0711 -0.810 -0.763 -0.229
[1.717] [1.667] [1.715] [0.0554] [0.0544] [0.102] [0.635] [0.637] [0.655]

Observations 109,613 111,515 99,420 160,352 162,512 149,987 46,734 47,189 39,179
Weather controls 3 3 3 3 3 3 3 3 3
NAPS-year fixed effects 3 3 3 3 3 3 3 3 3
Date fixed effects 3 3 3 3 3 3 3 3 3
Notes: Dependent variable is daily pollutant concentrations. Weather covariates include temperature, quadratic temperature, minimum
temperature, precipitation and quadratic precipitation. All weather covariates are daily average. Values in parentheses are standard errors
clustered by naps-id. ∗ significant at 10% ∗∗ significant at 5% ∗∗∗ significant at 1%.
Table 6: Change in Maximum Daily Pollutant Concen-
trations due to Transit Strikes

(1) (2) (3)

NOX CO PM2.5
Transit strike -7.152∗ 0.126 -2.265
[3.969] [0.128] [2.058]
Observations 109,613 160,352 47,920
Weather controls 3 3 3
NAPS-year fixed effects 3 3 3
Date fixed effects 3 3 3
The dependent variable is maximum daily concentration of
NOX , CO and PM2.5 . Weather covariates include temperature,
quadratic temperature, minimum temperature, precipitation and
quadratic precipitation. All weather covariates are daily average.
Values in parentheses are standard errors clustered by NAPS
monitors. ∗ significant at 10% ∗∗ significant at 5% ∗∗∗ significant
at 1%.

38
Table 7: Change in Average Daily Pollutant Concentrations due to Transit Strikes
Before and After 2000

Before After
(1) (2) (3) (4) (5) (6)
NOX CO PM2.5 NOX CO PM2.5
Transit strike -0.310 0.041 -3.699∗ -10.140∗ -0.056∗ 0.163
[1.548] [0.045] [0.414] [2.021] [0.023] [0.461]

Observations 63,305 113,730 4,177 46,308 46,622 38,389

Weather controls 3 3 3 3 3 3
Naps-year fixed effects 3 3 3 3 3 3
Date fixed effects 3 3 3 3 3 3
Notes: Dependent variable is daily pollutant concentrations. Weather covariates include temper-
ature, quadratic temperature, minimum temperature, precipitation and quadratic precipitation.
All weather covariates are daily average. Regressions also control for public holiday and day of
week fixed effects. Values in parentheses are standard errors clustered by naps-id. ∗ significant
at 10% ∗∗ significant at 5% ∗∗∗ significant at 1%.

39
 Table 8: Change in Daily Average Pollution Concentra-
tions Using Time Shifted Placebo Strikes

(1) (2) (3)

NOX CO PM2.5
Transit strike -1.121 0.0347 0.294
[1.055] [0.0304] [0.261]
Observations 116,053 167,077 76,056

Weather controls 3 3 3
NAPS fixed effects 3 3 3
Year-month fixed effects 3 3 3
The dependent variable is daily average pollutant concentration
of NOX , CO and PM2.5 . Weather covariates include temperature,
quadratic temperature, minimum temperature, precipitation and
quadratic precipitation. All weather covariates are daily average.
Values in parentheses are standard errors clustered by NAPS mon-
itors. ∗ significant at 10% ∗∗ significant at 5% ∗∗∗ significant at
1%.

40

View publication stats