The e ects of road pricing on driver behavior and air pollution
Matthew Gibson

a Department

a,∗

, Maria Carnovale

b

of Economics, Williams College, Schapiro Hall, 24 Hopkins Hall Dr., Williamstown MA

b CERTeT,

01267
Università Bocconi and Duke University, Duke Box 90239, Durham, NC 27708

Abstract
Exploiting the natural experiment created by an unanticipated court injunction, we evaluate
driver responses to road pricing.

We  nd evidence of intertemporal substitution toward

unpriced times and spatial substitution toward unpriced roads. The e ect on tra c volume
varies with public transit availability. Net of these responses, Milan's pricing policy reduces
air pollution substantially, generating large welfare gains. In addition, we use long-run policy
changes to estimate price elasticities.

Keywords:

road pricing, tra c policy, air pollution

1. Introduction
Growing air pollution, congestion, and accident externalities from vehicle tra c have
produced increasing interest in policy remedies. Beijing and Mexico City bar vehicles from
their roads on some days based on their license plate numbers (Davis, 2008; Viard and Fu,
2014; Wang et al., 2014). Many German cities have created Low Emissions Zones (Wol ,
2014), which prohibit dirtier vehicles within their borders. Stockholm, London, and Milan
charge fees to enter congested downtown areas. In the US, the Department of Transportation
is currently sponsoring a large number of road pricing experiments, including San Francisco's
Golden Gate Bridge, Interstate 95 near Miami, SR520 near Seattle, and Interstate 35W near
Minneapolis (DeCorla-Souza, 2004; Xie, 2013). Economists have raised concerns over nonprice policies because behavioral responses can be so large that net policy bene ts may be
zero, or even negative (Davis, 2008; Gallego et al., 2013). Theory suggests that road pricing
might be more e cient (Vickrey, 1963; Arnott et al., 1993), but this prediction depends on
driver responses. On which margins do drivers respond to road pricing, and how large are
such responses?

∗

Corresponding author
Email addresses:

mg17@williams.edu

(Matthew Gibson),

maria.carnovale@duke.edu

(Maria

Carnovale)

Preprint submitted to Elsevier

May 15, 2015

 1 INTRODUCTION
Confounding factors typically make tra c policies di cult to evaluate. Drivers know the
policy start date well in advance and may begin to adjust their behavior beforehand, which
attenuates estimated e ects. Municipalities typically increase public transit service at the
same time they implement road pricing or a driving restriction. This makes it impossible
to estimate the e ect of the policy in isolation. For example, Eliasson et al. (2009) point
out that Stockholm expanded bus service at the same time it implemented a congestion
charge. Because the buses used for the expansion were older and dirtier, the reduction in
emissions within the charge area was muted. Milan  rst implemented a congestion charge
concurrent with,  tra c calming measures, new bus lanes, increased bus frequency, increases
in parking restrictions and fees, and medium-term policies such as park-and-ride facilities
and underground network extensions  (Rotaris et al., 2010).
To address these identi cation challenges, we exploit a natural experiment: in late July
2012, an Italian court unexpectedly suspended Milan's road pricing policy, called  Area
C.  The city reinstated pricing eight weeks later.

Using unique tra c data at 15-minute

resolution, our study examines behavioral responses to Milan's policy, which requires drivers
entering the city center to pay

¿5 on weekdays 7:30AM-7:30PM. Drivers respond to pricing

in two ways: 1) shifting trips to the unpriced period, just before 7:30AM or after 7:30PM;
and 2) driving around the boundary of the priced area.
Net of these behavioral responses, we  nd the Area C policy reduces vehicle entries into
the priced area by 14.5 percent and air pollution by 6 to 17 percent.

The latter e ect is

large, particularly given that the priced region is just  ve percent of Milan's land area and the
city has an unusually clean vehicle  eet. Using a well-identi ed US estimate of willingness
to pay from Bayer et al. (2009) and scaling for income in Milan, we calculate that this
pollution reduction increases welfare by approximately $3 billion annually. Routes without
public transit experience large tra c changes from pricing, while those with public transit
experience much smaller changes. We provide evidence that this surprising result may arise
from residential sorting: residents who live near public transit may strongly prefer public
transit. In addition, we use changes in Milan's pricing policy across the 2008-2011 and 2012
periods to estimate elasticities: city-center entries by charged vehicles decrease .3 percent in
response to a one percent price increase.
This study contributes to the empirical literature on second-best road pricing policies
(Small et al., 2005; Small and Verhoef, 2007; Xie, 2013).

Closely related to our analysis

are Olszewski and Xie (2005), which analyzes the cordon charge and expressway pricing in
Singapore, Santos and Fraser (2006) and Santos (2008) on the London cordon charge, and
Eliasson et al. (2009) on the Stockholm cordon charge. These studies  nd cordon charges
do reduce tra c within the priced area.

Also related are Foreman (2013) and Small and

2

 2 BACKGROUND
Gomez-Ibanez (1998), which  nd evidence of intertemporal substitution in response to timevarying tolls. Our work complements the theoretical literature on second-best road pricing
(Lévy-Lambert, 1968; Marchand, 1968; Verhoef et al., 1996), particularly the literature on
cordon charges (Mun et al., 2003; Verhoef, 2005). Finally, we contribute to the literature
on environmental e ects of tra c policies. Many such studies have found no evidence of air
quality improvements (Transport for London, 2005, 2008; Invernizzi et al., 2011). Authors
commonly attribute this to driver substitution behaviors or exploitation of policy loopholes
(Davis, 2008; Gallego et al., 2013). In important work, Wol  (2014)  nds that German Low
Emissions Zones reduce the concentration of particles with a diameter of 10 microns or less
(PM10) by approximately 9 percent; this study is particularly signi cant given e orts by
European cities to meet stringent air quality standards.
Our study is unique in obtaining unconfounded causal estimates of behavioral responses
to road pricing and net road pricing e ectiveness.

This is the  rst analysis to examine

removal, rather than imposition, of a tra c policy. Other studies have used indirect measures
of tra c (such as gasoline sales or vehicle registrations) or hourly vehicle counts, but to the
best of our knowledge ours is the  rst to combine direct, high-resolution measures of tra c
volume with air pollution data. Finally, our  nding that the net e ect of pricing varies with
public transit availability is novel. It contributes to the literature on public transit and air
quality (Friedman et al., 2001) and adds a new dimension to the literature on tra c policies.
The remainder of the paper proceeds as follows. Section 2 provides policy background and
describes the natural experiment. Section 3 covers data, Section 4 describes our estimating
equations, and Section 5 discusses results. Section 6 concludes.

2. Background
Located in the center of Milan, Area C includes approximately 8.2 square kilometers (5
percent of city land area) and 77,000 residents (6 percent of population).
follows the

The boundary

Cerchia dei Bastioni, the route of the walls built under Spanish control in 1549.

Many of the portals still stand today, though the walls are largely gone. Figure A2 illustrates
the area.
Milan provides high levels of public transit, including four subway lines, 19 tram lines,
120 bus lines, and 4 trolley lines.

Together these lines transport 700 million passengers

across 155 million kilometers per year. The 80-kilometer subway network is larger than all
other Italian subways combined (Azienda Transporti Milanesi, 2013). Public transit has a
41 percent mode share in the city, followed by cars at 30 percent, walking at 17 percent,
bicycles at 6 percent, and motorbikes at 6 percent (Martino, 2012). The average round-trip

3

 2 BACKGROUND
commute in Milan takes 53 minutes, comparable to US cities like Dallas (52 minutes), Seattle
(55 minutes), and Los Angeles (56 minutes; Toronto Board of Trade, 2011).
Milan is one of the most polluted large cities in Europe. From 2002 through 2010 the
city exceeded the EU standard for PM10 on an average of 133 days per year (Danielis et al.,
2011). Since the mid 1990s the city has experimented with tra c policies intended to curb
its air pollution problem. Milan's  rst major road pricing program, called Ecopass, ran from
January 1, 2008 to December 31, 2011. Drivers paid a fee to enter Area C that varied with the
emissions from their vehicles. Vehicles meeting the Euro 3 standard paid nothing, while the
dirtiest diesel vehicles paid

¿10.

1

The charge applied weekdays 7:30AM-7:30PM. Drivers

could pay by internet, phone, or at the bank.

The city enforced the charge using license

plate-reading cameras located at the 43 entrances to Area C (Danielis et al., 2011). Drivers
who entered without paying faced  nes of

¿70-¿275 (la Repubblica, 2008). Approximately

2 percent of entering vehicles each day incurred  nes (Martino, 2012).
In June 2011 the voters of Milan overwhelmingly approved continued road pricing, with
79 percent in favor (Danielis et al., 2011).

2

As of January 16, 2012, the city implemented a

¿5

congestion charge for most vehicles entering Area C weekdays 7:30AM-7:30PM. This policy
was named Area C.

3

Motorcycles and public vehicles (e.g.

ambulances) were exempted.

4

Administrative details were largely the same as those for Ecopass. Drivers gained the option
to pay by direct debit, using a radio re ector placed in the vehicle (similar to FasTrak or
E-ZPass in the US). Violators were  ned

¿87 (Carra, 2012).

On July 25, 2012, a court unexpectedly suspended the Area C congestion charge in
response to a lawsuit by Mediolanum Parking (Povoledo, 2012).

More than ten previous

lawsuits against Ecopass and Area C had failed, so the suspension provoked surprise from the
press (Carra and Gallione, 2012). Charge enforcement halted the next day, July 26. There
was no press coverage prior to the court injunction, suggesting the decision was completely
unanticipated. The duration of the suspension was unknown and some observers believed
it would be permanent (Carra, 2012). Political forces marshaled on both sides. The mayor
declared,  We will save Area C.  Meanwhile the opposition called suspension the  death 
of Area C,  the defeat of ideological fervor and the victory of Milan's productivity and

1 Vehicles built prior to imposition of EU emissions standards were prohibited from October 15 through
April 15. Drivers received a 50% discount on the  rst 50 entries and a 40% discount on the next 50 entries.
Residents of Area C were also eligible for discounts (Rotaris et al., 2010).

2 49 percent of voters participated. The referendum did not specify the exact form the continued program

would take.

3 Vehicles classi ed diesel Euro 3 or below, or gasoline Euro 0 or below, were prohibited. Private vehicles

over 7m long were also prohibited. Scooters, motorcycles, and alternative-fuel vehicles, including hybrids,
were exempted. Residents paid

¿2 per entry (City of Milan, 2012; Milan Tourism, 2012).

4 This category includes mopeds and powered scooters.

4

 4 ESTIMATION
good sense  (Carra, 2012). The city altered neither public transit service nor parking fees
in response to the injunction.

On September 6, the city announced the charge would be
5

reinstated as of September 17, 2012.

For a timeline of these events, see Figure 1.

3. Data
Our tra c data come from AMAT and the Settore Piani cazione e Programmazione
Mobilità e Trasporto Pubblico Comune di Milano. For Area C, we have entries by vehicle type
and entry portal at 15-minute resolution, 2008-2012. There are 43 entry portals. These data
are recorded by the license plate cameras used to enforce the Area C charge. In addition, we
have counts of passing vehicles at 15-minute resolution, 2008-2012. These data are measured
6

by 748 buried sensors, mostly outside Area C.

Table 1 reports descriptive statistics for both

data sets at the daily level (aggregating over sensors/cameras and 15-minute intervals).
Our pollution and weather data come from ARPA Lombardia, the provincial air quality
agency. We have pollution and weather data at the monitor level, from 2003 through February 2013. Measured pollutants include carbon monoxide (CO), particles 10 microns or less
in diameter (PM10), and particles 2.5 microns or less in diameter (PM2.5). CO is measured
hourly, while particulates are measured daily. There are eight pollution measurement stations in the city of Milan proper (see Figure 2), of which two are inside Area C. The number
of monitors varies by pollutant and over time, as not all stations monitor all pollutants.
Table 1 provides descriptive statistics at the monitoring station-day level. The rightmost
column includes EU pollution standards for comparison. The European Commission (EC)
has the power to levy large  nes against non-attainment cities. For example, the EC  ned
Leipzig

¿700,000 per non-attainment day for failing to meet the PM10 standard (Wol  and

Perry, 2010).

4. Estimation
To explore the e ect of policy suspension on tra c volume we estimate a series of equations within the following framework:

5 The reinstated charge now ends at 6PM on Thursdays, rather than at 7:30PM as before (Corriere della
Sera, 2012a). Other features are unchanged.

6 According to AMAT, the buried sensors are less accurate than the cameras. Neither buried sensor data

nor camera data are available prior to 2008 (the cameras had not yet been installed and activated).

5

 4 ESTIMATION

traf f ict = β ∗ suspensiont + λ ∗ suspensiont ∗ wkendt
+ γ̄ ∗ time FEt + θ̄ ∗ date trendt + η̄ ∗ weathert + εt
The

traf f ic

(1)

variable measures either Area C entries or passing cars, over a day or a 15-

minute period, with

t

indexing days.

The vector

time FE

includes  xed e ects (FE) for

year, month, week, weekend, day of week, and holidays, plus interactions of weekend with
year. In addition, it includes dummies for the two-week interim period between Ecopass and
Area C and the interaction of the interim period with weekend.
non-random, we brie y analyze it in Section 5.5.1. The vector

7

While the interim period is

date trend

is a polynomial

trend in date, which controls for long-run trends not captured by our dummies for year,
month, etc.

8

In our primary results below we report estimates using a 7th-degree trend,

following Davis (2008). Weather controls comprise ten-piece linear splines in temperature
and positive precipitation. We control for weather because it plausibly in uences the choice
of public versus private transportation, or car versus motorcycle. The

suspension

variable

is a dummy equal to one for the period July 26, 2012 through September 16, 2012, when
the charge was suspended. The error term

ε

includes shocks to tra c not captured by our

controls, for example, an unusually bad auto accident or the Pope's visit on June 2, 2012.
In this and all subsequent equations, the coe cient of interest is

β,

the weekday e ect of

charge suspension. Weekends were unpriced both under the Area C policy and during charge
suspension. Intuition suggests the weekend e ect (β

+ λ)

will be negative if there is some

scope for weekday-weekend substitution, zero if there is little such scope.
The key identifying assumption underlying both equation 1 and subsequent models is
the exogeneity of the

suspension

variable.

That is, we assume that conditional on our

rich seasonal and weather controls, the timing of charge suspension is unrelated to other
determinants of tra c volume and pollution.

This is reasonable because the charge was

suspended unexpectedly by a court, as discussed in Section 2.
For the analysis of spatial substitution, we estimate two panel models at the sensor-day
level, with sensor  xed e ects. The  rst speci cation is as follows (s indexes sensor):

7 We do not explicitly control for Ecopass because of the year dummies 2008-2011.
8 The date trend is constructed as follows. We de ne a variable t equal to one for the  rst date in our
data, two for the second date, and so on. We then control for polynomials in

6

t.

 4 ESTIMATION

traf f icst = β̄ ∗ suspensiont ∗ distances + λ̄ ∗ suspensiont ∗ wkendt ∗ distances
+ αs + γ̄ ∗ time FEt + θ̄ ∗ date trendt + η̄ ∗ weathert + εst
In equation 2

distances

(2)

is a vector of dummies for sensors in several distance bins, where

distance is measured from the outside of the Area C boundary.

For example, this vector

includes a dummy that equals one if a sensor is within one kilometer outside the Area C
boundary, zero otherwise. Interacting these distance dummies with the charge suspension
dummy allows us to estimate the e ect of charge suspension both inside Area C and in
di erent donut-shaped regions outside Area C. The second speci cation is similar, but instead
of grouping sensors by distance, we group them into ring roads (unpriced, larger roads
encircling the priced area) and other roads (described in more detail in Section 5).
To analyze heterogeneity by public transport availability, we estimate a panel version of
equation 1 (p indexes portal), including portal  xed e ects

αp :

traf f icpt = β ∗ suspensiont ∗ pubtransp + λ ∗ suspensiont ∗ pubtransp ∗ wkendt
+ αp + γ̄ ∗ time FEt + θ̄ ∗ date trendt + η̄ ∗ weathert + εpt
In the equation above,

pubtransp

(3)

is a vector containing a dummy for the presence of public

transit, and another for the absence of public transit.

We also estimate versions of the
9

model comparing portals with and without bus, tram, and metro service.

By interacting

these measures of public transit availability with the suspension dummy, we can identify
cross-sectional di erences in responses to charge suspension that arise from portal-speci c
di erences in public transit access.
To investigate the e ect of suspension on daily average pollution we estimate the following
equation:

9 For example, a given portal will have

bus equal to 1 if a bus line crosses the boundary of Area C through

that portal. This is a simpli cation that ignores the e ect of being near (but not on ) a bus line. If the two
e ects have the same sign, as is plausible, this speci cation will bias us against  nding a di erence between
portals with and without a bus line. Similarly, the

tram

and

metro

question passes directly through or beneath the portal in question.

7

variables equal 1 only if the mode in

 5 RESULTS

ln(avg _pollution)t = β ∗ suspensiont + λ ∗ suspensiont ∗ wkendt
+ γ̄ ∗ time FEt + θ̄ ∗ date trendt + η ∗ ln (avg _pollutiont−1 ) + δ̄ ∗ atmospheret + εt
(4)
The dependent variable is the log average level of a pollutant measured over a day, with

t

indexing days. We conduct the analysis in logs to make the estimates for di erent pollutants more easily comparable. To avoid the endogeneity problems that arise in a dynamic
panel speci cation, we average over monitors and estimate the model separately for each
10

pollutant and area of Milan.

In order to control for the persistence of pollutants emitted

on the previous day, we include one lag of the dependent variable.

The lagged pollution

variable also controls for the previous day's atmospheric conditions, avoiding the need for
functional form assumptions on lagged atmospheric variables. ARPA normalizes the pollution measurements for temperature and pressure. The vector

atmosphere

includes 4-knot

cubic splines in humidity, wind speed, solar radiation, and precipitation, plus a dummy for
positive precipitation. As in equation 1, our speci cation also includes a 7th-degree trend in
date and time  xed e ects.

5. Results
5.1. Tra c
We  rst provide some semi-parametric evidence on the e ect of charge suspension for
vehicle types subject to the charge (buses and motorcycles are excluded).
the residuals from equation 1, omitting the

suspension

variable.

Figure 3 plots

We  t separate degree-

zero local polynomials for the period June-July 2012 (Area C pricing), August-September
(suspension of pricing), and October-November (pricing reinstated).

The graph shows a

sharp increase in weekday entries into Area C upon charge suspension, consistent with a
surprise announcement.

All three  tted lines are  at; there is no evidence of a seasonal

trend in the residuals before, during, or after charge suspension.

This indicates that our

time  xed e ects, together with a polynomial trend in date, are e ectively controlling for
seasonal patterns that might otherwise bias our estimates. There are several large positive
residuals between the Sept.

17 reinstatement of pricing and Oct.

1.

This may re ect

commuters delaying a mode switch before purchasing an October public transit pass.

10 Estimation results from a dynamic panel speci cation are available upon request. They are extremely
similar, as the asymptotic bias is of order 1/T (Nickell, 1981) and our data contain thousands of days.

8

 5.1 Tra c

5 RESULTS

Table 2 records results from our linear model for all vehicles, charged vehicles (buses and
motorcycles excluded), motorcycles (including mopeds and scooters), and other vehicles (primarily police cars and ambulances, which are exempt from the charge). Charge suspension
results in approximately 27,000 additional entries per day (14.5 percent) and the estimate
is statistically signi cant at the one percent level. The composition of entries also changes.
Entries by charged vehicles increase by roughly 29,000 (19 percent) while entries by motorcycles, which are exempt from the Area C charge, fall by roughly 2,000 (92 percent). The
latter result is not statistically signi cant. The estimate for other vehicles, predominantly
public vehicles like police cars, is small in magnitude and not statistically distinguishable
from zero at any conventional signi cance level. This provides a placebo test, as drivers of
public vehicles are exempt and so do not face a price change from charge suspension. The
weekend e ect of charge suspension is the sum of the estimates in the  rst and second rows.
Estimated weekend e ects are positive but not statistically signi cant under a zero null
hypothesis. This implies there is little scope for weekday-weekend substitution. Weekend
estimates for our subsequent models are similarly indistinguishable from zero and we omit
them for concision. We employ Newey-West standard errors to account for autocorrelation in

εt

out to seven lags. Because of occasional missing data, these standard errors will be biased

slightly downward. We have also estimated our models with standard errors clustered at the
week level and the results (available upon request) are not meaningfully di erent. Note that
the assumption of independence across clusters fails for days near the boundary of a week,
so this is not our preferred method of estimating standard errors.
Table A2 compares our estimated change in vehicle entries to results for 10 other pricing
policies. Eight  nd e ects broadly comparable in magnitude, ranging from

−22 percent.

−3

percent to

These include the London (−18 percent) and Stockholm (−22 percent) cordon

charges studied by Santos (2008) and Eliasson et al. (2009), respectively. Singapore's central
Restricted Zone yields two appreciably larger estimates,−44 percent and

−52 percent.

Small

and Verhoef (2007) suggest this policy produced such a dramatic response because the charge
was initially set extremely high.
To examine intertemporal substitution, Figure 4 plots the coe cients from a series of
96 regressions, with each 15-minute interval of the day modeled separately.

11

The estimates

show intertemporal substitution in both the morning and the evening. Charge suspension
results in approximately 500 fewer entries (23 percent) in the 15 minutes just before the

11 We use Newey-West standard errors to account for serial correlation. For most 15-minute intervals, serial
correlation falls to near zero after 7 lags. For the period 11:30PM-5:15AM, however, there are spikes in serial
correlation at 14, 21, and 35 days. We hypothesize that this results from the preponderance of public and
commercial vehicles during this window.

9

 5.1 Tra c

5 RESULTS

charge begins at 7:30AM and just after it ends at 7:30PM. This indicates that under the
charge, drivers were shifting trips into these unpriced periods. Indeed in the morning the
negative estimates are statistically distinguishable from zero (at the 5 percent level) for the
entire hour 6:30-7:30AM. Charge suspension increases entries during the 7:30AM-7:30PM
period that is priced under the Area C policy, consistent with the daily average estimates
reported in Table 2.

The increases achieve local maxima just after 7:30AM and before

7:30PM, suggesting some intertemporal substitution by commuters. The hours 9AM-3PM,
however, see roughly uniform increases in tra c under charge suspension.

This indicates

that non-commuters comprise a large share of marginal drivers.
Such a pattern of responses is the inverse of what is often called  peak spreading. 

In

theory peak spreading a ects driver welfare through two channels: 1) by reducing trip duration; and 2) by rescheduling trips (Arnott et al., 1993; Lindsey and Verhoef, 2000). The
former welfare e ect is positive, but the sign of the latter is theoretically ambiguous. Spreading the peak tra c load increases aggregate schedule delays

12

, but pricing better aligns trip

times with drivers' values of schedule delay (Arnott and Kraus, 1998; Lindsey and Verhoef,
2000). Because most air pollution emitted by vehicles is persistent within a day (Seinfeld
and Pandis, 2012), peak spreading may not change welfare along this dimension.
Finally we investigate spatial substitution toward roads outside Area C. Table 3 presents
results from a panel model at the sensor-day level, estimated from the buried sensor data.
Note these data measure passing cars per unit time and the resulting estimates are not
directly comparable to those from camera data. Tra c at the average sensor increases by
469 vehicles per day (8 percent) and the estimate is statistically signi cant at the one percent
level. This overall result conceals an interesting spatial pattern. Consistent with the models
based on camera data, suspension of the charge increases tra c inside Area C. Tra c on
the roads within one kilometer outside the Area C boundary, however, decreases by 1,061
vehicles per day (approximately 18 percent). This estimate is signi cant at the ten percent
level. Both point estimates for roads more than two kilometers outside the boundary are
positive, with one statistically signi cant at  ve percent and the other not signi cant. This
is consistent with an increase in radial trips (e.g. commutes from a residential neighborhood
into the center) from charge suspension. Overall this pattern of results suggests that some
drivers respond to the charge by driving around Area C. For drivers seeking to avoid the
priced area, the natural route typically involves the Circonvallazione Esterna, a ring of larger
roads located .6km-2km outside the Area C boundary. Table 3 shows the estimated e ect of

12 In keeping with the theoretical literature, by  delay  we mean a deviation from the desired arrival time,
either earlier or later.

10

 5.1 Tra c

5 RESULTS

charge suspension on these ring roads is large, negative, and signi cant at the  ve percent
level. Some of this decrease may re ect reduced circumferential commuting to public transit
stations. Evaluating this type of spatial substitution has proved di cult in other settings
due to confounding factors. In London, for example, the city substantially improved ringroad infrastructure because Transport for London anticipated spatial substitution (Santos,
2004).

To the best of our knowledge, ours is the  rst study to recover an unconfounded

driver response on this dimension.
In interpreting these results, it is reasonable to ask whether they capture the shortrun response to a pricing holiday or a long-run response.

The initial six-month trial of

the Stockholm cordon charge provides some evidence on this point. Eliasson et al. (2009)
observe,   . . . there was some doubt as to whether any tra c reduction would actually take
place during a brief and transient trial. Could it be that people would decide to `sit out'
the trial period without changing their travel habits? We now know that the trial indeed
had an immediate e ect.  E ects from Stockholm's initial trial proved very similar to longrun e ects (Börjesson et al., 2012).

In Milan the suspension of the Area C charge was

widely publicized, so the vast majority of residents knew about the change.

Evidence on

residents' expectations is qualitative and limited. Press accounts suggested Area C pricing
might not return (Corriere della Sera, 2012b), but there was likely a range of beliefs about
this.

The key question is not, however, whether residents expected the suspension to be

permanent, but whether they behaved as though it were. We have some suggestive evidence
on this point from Figure 3. If residents exhibited habit persistence or slowly updated their
beliefs about the suspension, we would expect an upward trend in the residuals during the
suspension period. No such trend is apparent; instead the full magnitude of driver responses
emerges immediately. This pattern also implies that switching costs across modes, routes,
and travel times are not  rst-order considerations for marginal drivers. Therefore we believe
our estimates largely re ect long-run behavior.
There is one respect in which the observed behavior during suspension likely does not
correspond to long-run behavior: vehicle portfolios. Intuition suggests they adjust slowly.
Given the possibility of renewed pricing, a risk-averse resident of Milan might well have been
reluctant to purchase a new vehicle in response to charge suspension. If vehicle portfolios
constrained a reasonably large fraction of the population, our estimates represent lower
bounds on the magnitude of long-run e ects. While our use of a natural experiment does
incur this cost on the external validity dimension, it brings o setting bene ts on the internal
validity dimension (e.g. avoidance of policy endogeneity concerns).
It is impossible to conduct a full welfare analysis of the Area C policy using our data, but
previous work casts light on some of its e ciency properties. Verhoef (2005) studies a cordon

11

 5.2 Interaction with public transit

5 RESULTS

charge in a general-equilibrium model of a monocentric city with endogenous population
density. He  nds the optimal cordon location is at 22 percent of the distance from the city
center to the city limits. We can compare the location of Milan's cordon to this benchmark.
Like many older European cities, Milan
q is monocentric. From the land area of Area C, we
can calculate an idealized radius of

8.2km2
π

= 1.62

for the city limits, we obtain an idealized radius of

1.62km
yields a ratio of
7.59km
located.

= .21.

kilometers. Proceeding in like fashion

q

181km2
π

= 7.59

kilometers.

Dividing

This suggests that at minimum Milan's cordon is not badly

We can also evaluate the level of Milan's charge.

Mun et al. (2003) simulate a

cordon charge using data from Osaka and  nd the optimal charge is equivalent to roughly 30

¿9. While Milan di ers from
Osaka on many dimensions, it is possible that a charge modestly above the current ¿5 level
minutes' worth of labor income. In Milan this would be about

would increase welfare. Verhoef (2005)  nds that an optimal cordon charge achieves 88-90
percent of the gains from a  rst-best pricing policy. De Borger and Proost (2001)  nd an
optimal combination of a cordon charge and parking fees can achieve 70 percent of  rst-best
gains. Taken together, this body of research implies that the Area C pricing policy may be
reasonably e cient.

5.2. Interaction with public transit
In addition, we investigate the interaction of charge suspension and public transit availability. To that end we estimate a panel model with a portal-day as the unit of observation.
The results in Table 4 indicate commuters on routes with public transit available respond
much less to the suspension of the charge.

Portals on a metro line, for example, show a
13

response that is not statistically distinguishable from zero.

There are at least two plausible explanations for these results. The  rst relies on cost
di erences. Assume an identical distribution of preferences for driving on two routes, one
with public transit ("Route A") and one without ("Route B"). If a sorting equilibrium holds,
commuters on the two routes must achieve equal utility. This implies that if Route A has
cheap public transit, it must have expensive car travel. This could be a direct result of public
transit, as when road lanes are devoted to tram lines, or a product of transit planning, as
when metro lines are placed beneath more congested roads. If a city applies the same charge
to cars on both routes, the percentage price change for Route A is much smaller and theory
predicts a smaller tra c response.
Alternatively, the results in Table 4 could spring from residential sorting (preference
heterogeneity). Suppose people with strong preferences for public transit live near Route A.

13 Portals with public transit may still be generating welfare changes if the composition of tra c is changing,
but we cannot evaluate this with our data.

12

 5.3 Price response

5 RESULTS

They might not own cars. They might, for example, dislike the claustrophobic conditions that
prevail on buses and trains at rush hour. Such individuals might be relatively unresponsive
to changes in the price of driving. Assume the initial cost of driving is the same for both
routes.

Then for a given road price change, there will be more infra-marginal drivers on

Route B than Route A.
It is di cult to choose between these explanations using the available data, but Table
A1 provides suggestive evidence. When we interact charge suspension with a time-invariant
measure of rush-hour congestion,

14

the response is larger for congested portals without public

transit than for uncongested portals without public transit.

(Congestion does not matter

for portals with public transit.) This result is inconsistent with an explanation based

solely

on cost di erences, which would predict smaller responses on more congested routes.

It

provides some evidence of preference heterogeneity, but does not exclude the possibility that
cost di erences drive some of the responses to charge suspension. For further discussion, see
Appendix Section A.1.

5.3. Price response
Milan's sequence of tra c policies, including both Ecopass and Area C, presents an
opportunity to recover another important feature of driver behavior: price responsiveness.
Under Ecopass (2008-2011), the weighted average weekday price for passenger vehicles was

¿0.72 (author's calculation, based on Rotaris et al. 2010). In January 2012 the
Area C policy increased the weekday price to ¿5. This provides potentially exogenous price
approximately

variation, although the usual concerns about policy endogeneity obtain.

We estimate the

elasticity of vehicle entries with respect to price using a variant of equation 1, replacing the
dependent variable with log entries and the policy variables with log price. This speci cation
does not rely on the natural experiment that forms the basis of our other results. Table 5
reports results.

A one percent price increase decreases entries by charged vehicles by .3

percent and the estimate is signi cant at the one percent level.

This estimated response

captures the net e ect of two changes: the increase in pecuniary cost and the decrease in
time cost (from reduced congestion) under the Area C policy. Given that the Area C charge
is a relatively small part of total trip cost, which includes time, fuel, and depreciation, this
demand response is large.

For passenger and commercial vehicles elasticities are -.17 and

14 We standardize entries at the portal-15 minute-lane level, then average the resulting values during rush
hour periods (over days). Portals with lower values (low rush hour throughput relative to the portal average)
are likely more congested. We de ne a congestion dummy equal to 1 for portals in the bottom 10 percent of
the distribution.

13

 5.4 Pollution

5 RESULTS

-.47, respectively, with both estimates signi cant at the  ve percent level.

15

The estimate

for other vehicles is a placebo test of the e ect of passenger prices on entries by exempt
municipal vehicles; we  nd no e ect. Our elasticity estimates are necessarily local and may
not obtain outside the range of prices observed in our data. They suggest, however, that a
modest price increase, e.g. from

¿5 to ¿6, might produce substantial additional reductions

in Area C entries.
If the change in price from Ecopass to Area C is conditionally exogenous, our estimate
captures an internally valid causal e ect. The question of external validity remains, however.
Theory predicts that demand elasticity will vary with income, preferences, the availability
of substitutes, and other factors. To provide a qualitative sense of such factors, Table A2
puts our estimated elasticity in the context of estimates developed in other studies, using
data from other locations. At -.3 our overall estimate is modestly larger than most previous
 ndings for cordon charges, which generally range from -.2 to -.1. Estimates for expressway
and bridge tolls exhibit more variation, ranging from -.56 to -.06. While some are similar to
our result (e.g. Small et al. (2006) for California State Route 91), others are substantially
larger or smaller. For example, Odeck and Brathen (2008)  nd an average price elasticity of
-.56 on Norwegian toll roads. Such larger estimates may re ect the often-greater availability
of close substitutes for single-facility tolls than for cordon charges.
Some cities manipulate parking prices, rather than road prices, in order to optimize travel
demand. Studies of such policies typically estimate parking demand, rather than demand for
travel on a given road or demand for cordon crossing. While these estimates are not directly
comparable to ours, they may provide an instructive benchmark.

In a survey of research

 ndings, the US Transit Cooperative Research Program found a range of parking demand
elasticities from -.6 to -.1, with -.3 the mode (Vaca and Kuzmyak, 2005). Using data from
exogenous changes in San Francisco parking prices, Pierce and Shoup (2013) estimate an
average demand elasticity of -.4.

5.4. Pollution
Table 6 reports the pollution e ect of charge suspension on weekdays, estimated using
equation 4. We focus on CO, PM10, and PM2.5 because these pollutants have direct, negative health e ects (Seaton et al., 1995) and all are closely associated with vehicle emissions
(Gallego et al., 2013). Estimates show statistically signi cant increases in CO and PM10,
both inside and outside Area C, in the 6 to 17 percent range. These magnitudes are similar
to those from our tra c models. The point estimate for PM2.5 outside Area C is greater

15 We do not separate passenger and commercial vehicles in our analysis of tra c volume because under
the Area C policy they both face the same

¿5 price.

14

 5.4 Pollution

5 RESULTS

at 21 percent, but the standard error is large and the estimate is signi cant only at the ten
percent level. This imprecision may stem from the much shorter period over which PM2.5
data have been collected. For CO we can also estimate the e ect for monitors located on
the ring roads (Circonvallazione Esterna). This estimate is near zero, which roughly accords
with our tra c results in Table 3. As the half-lives of commonly regulated air pollutants
are measured in hours or days (Seinfeld and Pandis, 2012), the observed pollution increases
likely derive from additional trips and mode shifting, rather than from trip rescheduling.
These pollution e ects are large, particularly given that the priced area is small (5 percent
of land area) and Milan has an unusually clean vehicle  eet. Milan's earlier Ecopass policy,
which applied from 2008 through 2011, created incentives for drivers to purchase cleaner
vehicles and many did so (Rotaris et al., 2010).

This means that for a given number of

foregone trips, the e ect on pollution would have been smaller in 2012 than in 2007. Like
our tra c estimates, our pollution estimates are lower bounds on long-run e ects because of
the potential for vehicle portfolios to change over the long run.
In order to evaluate the welfare e ects of these air pollution changes, we require an estimate of willingness to pay for reductions in PM10 pollution. We adopt estimated annual
willingness to pay of $148.70 per person per µg/m3 (in 1982-1984 dollars) from Bayer et al.
(2009), who use data from US metropolitan statistical areas. By accounting for migration
costs and instrumenting for ambient pollution, this study overcomes several important identi cation challenges.

For comparison, note that the meta-analysis by Smith and Huang

(1995)  nds a mean marginal willingness to pay of $110 (in 1982-1984 dollars) per µg/m3
TSP reduction in US cities. While this estimate is meaningfully smaller than the one from
Bayer et al. (2009), this is unsurprising for two reasons: 1) the downward biases in the OLS
hedonic speci cations analyzed by Smith and Huang (1995); and 2) the higher real income
of the US population in the data used by Bayer et al. (2009). The rough similarity between
the Bayer et al. (2009) and Smith and Huang (1995) estimates provides some reassurance
that our choice is reasonable.
Naturally the use of a willingness to pay estimate from the United States raises bene t
transfer concerns. Kaul et al. (2013)  nd that transfer errors are typically smaller for function
transfers than for value transfers. While a full function transfer is beyond the scope of the
present exercise, we can scale the Bayer et al. (2009) estimate to account for local income
in Milan. As suggested by Ready and Navrud (2006), we employ a PPP-adjusted exchange
rate and  nd that average income in Milan was roughly 85 percent of US income in 2007
(Hammitt and Robinson, 2011; OECD, 2011). Both Smith and Huang (1995) and Hammitt
and Robinson (2011)  nd that the income elasticity of willingness to pay for air pollution
reduction is small in wealthier nations, with the latter noting that US agencies often use an

15

 5.5 Robustness checks

5 RESULTS

income elasticity of approximately .5. Multiplying

$149 ∗ (1 − (.15 ∗ .5)) yields approximately

$138. Converted to 2014 dollars, this becomes $327.
With this  gure in hand, we can compute the aggregate welfare e ects of the PM10
changes in Milan from the Area C policy. Mean PM10 concentration in our data is

44µg/m3 outside. The concentration
7.5µg/m3 , respectively. Approximately

48µg/m3

inside Area C,

changes implied by our estimates are

1.9µg/m3

77,000 people live in Area C and 1.2

and

million outside.

The implied welfare gain from the Area C policy is approximately $48

million inside Area C and $2.94 billion outside, for a total of $3 billion.

This estimate is

very large relative to the annual PM10 bene t  gures used by transportation researchers in
cost-bene t analysis of the Ecopass policy, which have typically been in the range of

¿0.4-1.3

million (Rotaris et al., 2010; Danielis et al., 2011). We note that the median transfer error
identi ed by Kaul et al. (2013) is 39 percent, and this does suggest some caution.

Even

allowing for the possibility of large transfer error, however, our welfare estimate is an order
of magnitude larger than those in Rotaris et al. (2010) and Danielis et al. (2011).
The  nding that pricing reduces air pollution both inside and outside Area C speaks to
an important distributional question. Opponents of the Area C policy have argued that it
improves air quality in an a uent area while doing nothing to address the remainder of the
city (Danielis et al., 2011). The estimates in Table 6 provide evidence against this claim.
While spatial substitution may reduce air quality near ring roads, the policy improves air
quality in other locations, both inside and outside Area C. Voting behavior is consistent
with such a widespread improvement. In a 2011 referendum on road pricing, more than 79
percent voted in favor (Danielis et al., 2011), an outcome that would be unlikely if only Area
C residents bene ted from the policy. The successful referendum is somewhat surprising in
view of the generally unfavorable public attitudes toward road pricing.

As in Stockholm,

residents experienced the bene ts of road pricing during a trial period before voting and this
may have been in uential (Eliasson, 2008; Harsman and Quigley, 2010).

5.5. Robustness checks
5.5.1. Tra c
We estimated all models with the following trends in date: 1) no trend; 2) linear trend;
3) 4th-degree trend; and 4) 7th-degree trend.

In nearly all cases the choice of trend had

negligible in uence on the sign, magnitude and signi cance of the estimates.
In addition, we compare our primary estimates to those from the interim period between
the end of the Ecopass policy and the start of the Area C policy (January 1-15, 2012).
During this time drivers could enter the city center without paying, but this period raises
identi cation concerns.

First, it was not randomly timed and therefore the potential for

16

 5.5 Robustness checks

5 RESULTS

unobserved confounders (like changes in bus service) is greater than for the period of our
natural experiment. Second, because the return of pricing was assured, questions of habit
formation and switching costs are more problematic.

Third, because this period lasted

only two weeks, these models have less statistical power.

Nonetheless the interim period

provides a rough benchmark against which to evaluate our main results. Table A3 shows
that the estimated e ects on Area C entries for all vehicles and charged vehicles are similar
in magnitude to our primary results and statistically signi cant at the one percent level.
The estimated e ect on motorcycle entries is positive and signi cant for the interim period,
which di ers from our primary result and does not accord with theory. The positive sign
could re ect the fact that riding a motorcycle in a lower-tra c environment is both safer
and more enjoyable.
Table A4 reports spatial substitution results for the interim period. The pattern of results
accords with those from our natural experiment, but the decrease in tra c on ring roads is
no longer statistically signi cant. Similarly, Table A5 shows how e ects on Area C entries
during the interim period vary with public transit availability. (These estimates come from
the same model as those in Table 4.) Again the pattern of results is strongly similar to those
from our natural experiment, with portals lacking public transit seeing greater increases in
tra c.
Taken together, Tables A3, A4, and A5 demonstrate that estimates from the interim
period (in January) are quite similar to those from our natural experiment (July-September).
This pattern suggests that seasonality in the elasticity of demand for Area C trips is not
a  rst-order concern. Thus our primary estimates provide evidence on driver responses to
pricing that generalizes beyond the time of year at which the natural experiment happened
to occur. There remains the possibility of bias from seasonal trends in the level of demand.
As discussed in Section 5.1 and illustrated in Figure 3, our time  xed e ects and polynomial
trend in date appear to e ectively control for such trends. Nonetheless we describe seasonal
trends in more detail here. Figure A3 shows the seasonal pattern of entries into Area C. The
period of the natural experiment includes Italy's traditional vacation season, which sees far
fewer Area C entries in three August weeks. The  ve remaining weeks of the experiment,
however, include some of the busiest weeks of the year (in September). Given the pattern in
Figure A3, any failure of our seasonality controls will bias the magnitude of our estimates
downward.
Figure 5 displays the  e ect  of a placebo charge suspension each year 2008-2011 on Area
C vehicle entries. There is no evidence of an increase in Area C entries during the placebo
periods; if anything they show slight decreases. Similarly, Table A6 reports estimated e ects
of a placebo suspension for the same dates in 2011, rather than 2012. Estimated magnitudes

17

 6 CONCLUSION
are much smaller than those in our main results and not one is signi cant.

5.5.2. Pollution
Table 7 reports the estimated e ects of placebo suspensions for the same dates 2008-2011
(rather than 2012). Half the estimates are negative and most are statistically insigni cant,
which aligns with the placebo tests from our tra c models and suggests that our main results
are not driven by misspeci cation. The estimated placebos for CO inside Area C (2011) and
PM10 outside Area C (2009) are positive and statistically signi cant, which recommends
some caution in interpreting our corresponding primary estimates.

6. Conclusion
Our analysis uses a natural experiment to examine behavioral responses and recover
causal e ects of Milan's Area C road pricing policy. We  nd the policy reduces tra c and
pollution considerably.

Drivers respond with intertemporal substitution toward unpriced

times and spatial substitution toward unpriced roads outside the charge area. In addition, we
show that the e ect of pricing on tra c depends on the availability of public transportation.
Routes without public transit experience large tra c changes from the Area C charge, while
those with public transit experience much smaller changes. We also use long-run changes in
Milan's pricing policies to estimate elasticities of tra c with respect to the charge: entries
by charged vehicles decrease .3 percent in response to a one percent price increase.

This

estimate captures the net e ect of an increase in the charge and the resulting decrease in
time cost from reduced congestion.
Our  ndings are relevant for policy design.

Theory predicts that the substitution be-

haviors we observe would occur under both optimal and second-best policies, but cities can
tailor policy to manage their magnitudes. Cities like Milan, with  xed cordon charges, might
reduce intertemporal substitution and move closer to the theoretical optimum by charging
a lower but non-zero price for  shoulder  periods adjacent to peak periods.

Some drivers

might still choose the shoulder period, or switch back to the peak period, but others might
switch to public transit or carpool. A city might reduce spatial substitution by expanding
the geographic area subject to pricing, such that driving around the priced area would be
impractical. Alternatively it might improve roads likely to see policy-driven tra c increases,
as London did prior to introducing its cordon charge (Santos, 2004).
Our public transit results also have policy implications.

Because responses to pricing

vary with transit availability, welfare impacts from pricing will be spatially heterogeneous.
Policymakers may wish to consider these distributional impacts when designing a road pricing
policy.

More generally, our results suggest that road pricing and public transit may be

18

 6 CONCLUSION
substitutes, at least within cities. In areas that already have high levels of public transit,
there may be limited scope for reducing tra c via road pricing.
We  nd suspension of the charge increased weekday concentrations of CO by 6 percent
and PM10 by 17 percent. This is a remarkable change in air quality, given: 1) the charge
area represents only 5 percent of Milan's land area, and a smaller fraction of the broader
metropolitan area; and 2) it is a lower bound on the potential long-run increase. Our estimate
is still more surprising in light of the city's relatively clean vehicle  eet. Previous welfare
analyses of Milan's Ecopass policy have found net bene ts of approximately

¿7-12 million

per year, even placing extremely low values on air pollution reductions (Danielis et al., 2011).
Using an adjusted willingness to pay for PM10 reduction from Bayer et al. (2009), we  nd
that the Area C policy produces a $3 billion welfare gain from air pollution reductions alone.
Pollution e ects from similar policies in cities with dirtier  eets could well be larger. More
congested cities would also tend to see larger welfare gains. Among the most congested large
world cities are Istanbul, Mexico City, and Rio de Janeiro (TomTom, 2014). Among large
US cities, New York, Los Angeles, and Chicago see the highest welfare losses from congestion
(Lomax et al., 2012). Our results suggest there is scope for road pricing, even in second-best
form, to produce very large welfare gains in such environments.

Acknowledgements
The authors thank Prashant Bharadwaj, Jennifer Burney, Richard Carson, Andrew
Chamberlain, Julie Cullen, Gordon Dahl, Jamie Mullins, Kevin Roth, Lanfranco Senn, and
two anonymous referees for valuable advice. We also thank AMAT and the Settore Piani cazione e Programmazione Mobilità e Trasporto Pubblico Comune di Milano for data and
assistance.

References
Anderson, M. L., 2014. Subways, strikes, and slowdowns: The impacts of public transit on
tra c congestion. The American Economic Review 104 (9), 2763 2796.
Arnott, R., Kraus, M., 1998. When are anonymous congestion charges consistent with
marginal cost pricing? Journal of Public Economics 67, 45 64.
Arnott, R., Palma, A. D., Lindsey, R., 1993. A Structural Model of Peak-Period Congestion:
A Tra c Bottleneck with Elastic Demand. The American Economic Review 83 (1), 161 
179.
Azienda Transporti Milanesi, 2013. Company Pro le. Tech. rep.

19

 6 CONCLUSION
Bayer, P., Keohane, N., Timmins, C., Jul. 2009. Migration and hedonic valuation: The case
of air quality. Journal of Environmental Economics and Management 58 (1), 1 14.
Börjesson, M., Eliasson, J., Hugosson, M. B., Brundell-Freij, K., Mar. 2012. The Stockholm
congestion charges 5 years on. E ects, acceptability and lessons learnt. Transport Policy
20, 1 12.
Carra, I., June 2012. Il sindaco: Salveremo Area C. la Repubblica.
Carra, I., Gallione, A., June 2012. Milano scon tta dal parcheggiatore sospesa l' area antitra co. la Repubblica.
City of Milan, 2012. Area C. Accessed December 1, 2012.
URL

http://www.comune.milano.it

Corriere della Sera, Sep. 2012a. Area C, torna il ticket anti tra co da 5 euro Al giovedì
telecamere spente alle 18.
Corriere della Sera, Jul. 2012b. Il Consiglio di Stato sospende Area C Da questa mattina
telecamere spente.
Danielis, R., Rotaris, L., Marcucci, E., Massiani, J., 2011. An economic, environmental and
transport evaluation of the Ecopass scheme in Milan: three years later.
Davis, L., Feb. 2008. The E ect of Driving Restrictions on Air Quality in Mexico City.
Journal of Political Economy 116 (1), 38 81.
De Borger, B., Proost, S., 2001. Reforming transport pricing in the European Union: A
modelling approach. Edward Elgar Publishing.
DeCorla-Souza, P., 2004. Recent U.S. Experience: Pilot Projects. Research in Transportation
Economics 9 (04), 283 308.
Eliasson, J., Nov. 2008. Lessons from the Stockholm congestion charging trial. Transport
Policy 15 (6), 395 404.
Eliasson, J., Hultkrantz, L., Nerhagen, L., Rosqvist, L. S., Mar. 2009. The Stockholm
congestion-charging trial 2006: Overview of e ects. Transportation Research Part A: Policy and Practice 43 (3), 240 250.
Finkelstein, A., 2009. E-ZTax: Tax Salience and Tax Rates. The Quarterly Journal of Economics 124 (3), 969 1010.

20

 6 CONCLUSION
Foreman, K., 2013. Crossing the Bridge:

The E ects of Time-Varying Tolls on Curbing

Congestion.
Friedman, M. S., Powell, K. E., Hutwagner, L., Graham, L. M., Teague, W. G., 2001. Impact
of changes in transportation and commuting behaviors during the 1996 Summer Olympic
Games in Atlanta on air quality and childhood asthma. Journal of the American Medical
Association 285 (7), 897 905.
Gallego, F., Montero, J.-P., Salas, C., Nov. 2013. The e ect of transport policies on car use:
Evidence from Latin American cities. Journal of Public Economics 107, 47 62.
Goh, M., 2002. Congestion management and electronic road pricing in Singapore. Journal
of Transport Geography 10, 29 38.
Hammitt, J. K., Robinson, L. A., 2011. The Income Elasticity of the Value per Statistical Life: Transferring Estimates between High and Low Income Populations. Journal of
Bene t-Cost Analysis 2 (1), 1 29.
Harsman, B., Quigley, J. M., 2010. Political and Public Acceptibility of Congestion Pricing:
Ideology and Self-Interest. Journal of Policy Analysis and Management 29 (4), 854 874.
Invernizzi, G., Ruprecht, A., Mazza, R., De Marco, C., Mo£nik, G., Sioutas, C., Westerdahl,
D., Jul. 2011. Measurement of black carbon concentration as an indicator of air quality
bene ts of tra c restriction policies within the ecopass zone in Milan, Italy. Atmospheric
Environment 45 (21), 3522 3527.
Jones, P., Hervik, A., 1992. Restraining car tra c in European cities: an emerging role for
road pricing. Transportation Research Part A: Policy and Practice 26 (2), 133 145.
Kaul, S., Boyle, K. J., Kumino , N. V., Parmeter, C. F., Pope, J. C., 2013. What can
we learn from bene t transfer errors? Evidence from 20 years of research on convergent
validity. Journal of Environmental Economics and Management 66 (1), 90 104.
la Repubblica, Jan. 2008. Milano scatta l 'ora dell'Ecopass multe salate per chi sgarra.
Lévy-Lambert, H., 1968. Tari cation des Services à Qualité Variable Application aux Péages
de Circulation. Econometrica: Journal of the Econometric Society 36 (3), 564 574.
Lindsey, C. R., Verhoef, E. T., 2000. Tra c Congestion and Congestion Pricing.
Lomax, T., Schrank, D., Eisele, B., 2012. 2012 Urban Mobility Report. Tech. rep., Institute,
Texas Transportation.

21

 6 CONCLUSION
Marchand, M., 1968. A Note on Optimal Tolls in an Imperfect Environment. Econometrica
36 (3), 575 581.
Martino, A., 2012. Milano: From Pollution Charge to Congestion Charge. Tech. rep., TRT
Trasporti e Territorio.
Meland, S., 1995. Generalised and advanced urban debiting innovations: the GAUDI Project.
III: The Trondheim toll ring. Tra c Engineering & Control 36 (3), 150 155.
Milan Tourism, 2012. Area C. Accessed December 1, 2012.
URL

http://www.turismo.milano.it

Mun, S.-i., Konishi, K.-j., Yoshikawa, K., 2003. Optimal cordon pricing. Journal of Urban
Economics 54, 21 38.
Nickell, S., 1981. Biases in Dynamic Models with Fixed E ects. Econometrica 49 (6), 1417 
1426.
Odeck, J., Brathen, S., 2008. Travel demand elasticities and users attitudes: A case study of
Norwegian toll projects. Transportation Research Part A: Policy and Practice 42, 77 94.
OECD, 2011. OECD Regions at a Glance. Accessed April 9, 2015.
URL

http://www.oecd-ilibrary.org/sites/reg_glance-2011-en/03/06/index.html

Olszewski, P., Xie, L., 2002. Tra c Demand Elasticity with Respect to Road Pricing   Some
Evidence from Singapore. In: Proceedings of the International Conference on Seamless
and Sustainable Transport. No. November. pp. 217 228.
Olszewski, P., Xie, L., Aug. 2005. Modelling the e ects of road pricing on tra c in Singapore.
Transportation Research Part A: Policy and Practice 39 (7-9), 755 772.
Pierce, G., Shoup, D., 2013. Getting The Prices Right. Journal of the American Planning
Association 79 (1), 67 81.
Polak, J., Meland, S., 1994. An assessment of the e ects of the Trondheim Toll Ring on travel
behaviour and the environment. In: Towards an intelligent transport system. Proceedings
of the  rst world congress on applications of transport telematics and intelligent vehiclehighway systems, November 30 - 3 December 1994, Paris, volume 2.
Povoledo, E., Aug. 2012. E ort to Cut Milan Tra c Halted as Court Favors Garage.
Ramjerdi, F., Minken, H., Ostmoe, K., 2004. Norwegian Urban Tolls. Research in Transportation Economics 9 (04), 237 249.

22

 6 CONCLUSION
Ready, R., Navrud, S. l., 2006. International bene t transfer: Methods and validity tests.
Ecological Economics 60 (2), 429 434.
Rotaris, L., Danielis, R., Marcucci, E., Massiani, J., Jun. 2010. The urban road pricing
scheme to curb pollution in Milan, Italy. Transportation Research Part A: Policy and
Practice 44 (5), 359 375.
Santos, G., 2004. Urban Road Pricing in the U.K. Research in Transportation Economics
9 (04), 251 282.
Santos, G., 2008. London Congestion Charging. Brookings-Wharton Papers on Urban Affairs, 177 234.
Santos, G., Fraser, G., 2006. Road pricing: lessons from London. Economic Policy (April),
263 310.
Seaton, A., Godden, D., MacNee, W., Donaldson, K., 1995. Particulate air pollution and
acute health e ects. The Lancet 345 (8943), 176 178.
Seinfeld, J. H., Pandis, S. N., 2012. Atmospheric chemistry and physics: from air pollution
to climate change. John Wiley & Sons.
Small, K. A., Gomez-Ibanez, J. A., 1998. Road Pricing for Congestion Management: The
Transition from Theory to Policy. In: Road Pricing, Tra c Congestion and the Environment: Issues of E ciency and Social Feasibility. Vol. 2. pp. 107 123.
Small, K. A., Verhoef, E. T., 2007. The Economics of Urban Transportation.
Small, K. A., Winston, C., Yan, J., 2005. Uncovering the Distribution of Motorists' Preferences for Travel Time and Reliability. Econometrica 73 (4), 1367 1382.
Small, K. A., Winston, C., Yan, J., 2006. Di erentiated Road Pricing, Express Lanes, and
Carpools:

Exploiting Heterogeneous Preferences in Policy Design. Brookings-Wharton

Papers on Urban A airs, 53 96.
Smith, V. K., Huang, J.-c., 1995. Can Markets Value Air Quality?

A Meta-Analysis of

Hedonic Property Value Models. Journal of Political Economy 103 (1), 209 227.
TomTom, 2014. Tomtom tra c index. Accessed April 11, 2015.
URL

http://www.tomtom.com/en_gb/trafficindex/#/list

Toronto Board of Trade, 2011. Toronto as a Global City: Scorecard on Prosperity. Tech. rep.

23

 6 CONCLUSION
Transport for London, 2005. Environment Report 2005. Tech. rep.
Transport for London, 2008. Environment Report 2008. Tech. rep.
Tretvik, T., 2003. Urban road pricing in norway: Public acceptibility and travel behavior.
in: Acceptability of transport pricing strategies. In: MC-ICAM Conference, Acceptability
of Transport Pricing Strategies.
Vaca, E., Kuzmyak, J. R., 2005. TCRP Report 95. Tech. rep.
Verhoef, E. T., 2005. Second-best congestion pricing schemes in the monocentric city. Journal
of Urban Economics 58, 367 388.
Verhoef, E. T., Nijkamp, P., Rietveld, P., 1996. Second-best congestion pricing: the case of
an untolled alternative. Journal of Urban Economics 40, 279 302.
Viard, V. B., Fu, S., 2014. The E ect of Beijing's Driving Restrictions on Pollution and
Economic Activity.
Vickrey, W., 1963. Pricing in Urban and Suburban Transport. The American Economic
Review 53 (2), 452 465.
Wang, L., Xu, J., Qin, P., 2014. Will a Driving Restriction Policy Reduce Car Trips?
The Case Study of Beijing, China. Transportation Research Part A: Policy and Practice
67 (September), 279 290.
Wol , H., 2014. Keep your clunker in the suburb: Low-emission zones and adoption of green
vehicles. The Economic Journal 124 (578), F481 F512.
Wol , H., Perry, L., Aug. 2010. Trends in clean air legislation in europe: Particulate matter
and low emission zones. Review of Environmental Economics and Policy 4 (2), 293 308.
Xie, C., 2013. Dynamic Decisions to Enter a Toll Lane on the Road.

24

 6 CONCLUSION

Figures & tables
Figures

6/12/2011

12/31/2011

7/26/2012

1/16/2012

Ci
w ty
ill an
re no
tu u
rn nc
9/ es
17 p
ol
Ar
ic
ea
y
C
po
lic
y
re
tu
rn
s

Co
Ar ur
ea t s
C usp
po e
lic nd
y s

Ec
In op
te as
rim s
e
pe nd
rio s
d
be
gi
In
ns
t
Ar er
ea im
C pe
po rio
lic d
y en
be d
gi s
ns

//
1/1/2008

Ro
re ad
fe p
re ric
nd in
um g

Ec

op
a

ss

be

gi
n

s

Figure 1: Time line of road pricing in Milan

9/6/2012

9/17/2012

Figure 2: Air pollution monitoring stations in Milan

Green circles represent the locations of pollution monitoring stations operated by ARPA Lombardia. Not all stations monitor
all pollutants. The via Senato and Verziere stations are inside Area C. The Piazza Zavattari, viale Marche, and viale Liguria
stations are on the ring roads (Circonvallazione Esterna). The Piazza Abbiategrasso, Pascal Citta Studi, and Parco Lambro
stations are outside the ring roads.

25

 6 CONCLUSION

Pricing

Suspension

Pricing reinstated

2

2

01
de
01

oc
01

au
01

c2

t2

01
g2

n2
ju
01

01

2

−40000
01
2

Residual − Area C entries
−20000
0
20000

40000

Figure 3: E ect of Area C charge suspension on vehicle entries

Residuals are from equation (1), with the charge suspension dummy variable excluded. The dependent variable is daily entries
into Area C by vehicles subject to the charge.

Each point represents one observation (one day).

Fitted lines are based on

separate degree-zero local polynomial regressions for pre-suspension, suspension, and post-suspension periods. Time controls
include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year, a holiday dummy, a
7th-degree time trend in date, and a dummy for the unpriced January 2012 interim period. Weather controls comprise ten-piece
linear splines in temperature and positive precipitation.

26

 6 CONCLUSION

1000

Figure 4: E ect of Area C charge suspension on vehicle entries, by 15-minute interval

Priced

Unpriced

Change in # of Area C vehicle entries
−500
0
500

Unpriced

0

4

8
12
16
Hour of the day (0 = midnight)

20

24

Dependent variable is Area C entries by vehicles subject to charge. Estimates are from equation (1), estimated separately for
each 15-minute interval. Whiskers represent Newey-West standard errors multiplied by 1.96. The lag length is 35 for hours
23.5-5.25, 7 otherwise.  Unpriced  and  priced  labels refer to status of a given interval under the Area C policy, which prices
entries 7:30AM-7:30PM.

27

 6 CONCLUSION

Figure 5: E ect of placebo suspensions on Area C vehicle entries

01

01

de

oc

t2

c2

00

00

9

9
00
g2
01

01

au

ju

de
01

9

0

50000

2009

n2 −50000
00
9

8
00
c2

00
t2
oc
01

1
20
1

20

ec
01
d

01
o

01

01

au

ct

n2

g2
0

11

11

01
1

−50000

0

50000

2011

ju

ec

01
o

01
d

ct

20

20
1

0

10

10
g2
0
au
01

ju
01

8

8
00
g2
au
01

2010

Residual − # of vehicles
n2 −50000
0
50000
01
0

01

ju

Residual − # of vehicles
n2 −50000
0
50000
00
8

2008

Residuals are from equation (1). Dependent variable is daily entries into Area C by vehicles subject to charge. Each point
represents one observation (one day). Fitted lines based on separate degree-zero local polynomial regressions for pre-placebo,
placebo, and post-placebo periods (which correspond to the dates of the 2012 natural experiment). Time controls include 0-4
year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year, a holiday dummy, a 7th-degree time
trend in date, a dummy for the unpriced January 2012 interim period, and a dummy for the June-Sept. 2012 charge suspension.
Weather controls comprise ten-piece linear splines in temperature and positive precipitation.

Tables
Table 1: Descriptive statistics, daily level

Units

Mean

Std. dev.

Min

Max

N

EU standard

-

169,744

47,628

3,905

261,172

1,737

-

Passing vehicles

-

2,585,316

1,348,275

37,412

5,918,492

1,754

-

CO

mg/m3

1.26

.67

0

7.6

17,625

10

PM10

µg/m3

47.66

31.07

0

276

9,091

40

PM2.5

µg/m3

33.74

26.56

0

177

2,550

25

Precipitation

mm

2.07

6.91

0

121.2

30,929

-

Area C entries

All statistics calculated over daily means. The EU standard for CO is based on a rolling 8-hour mean, while those for PM10
and PM2.5 are based on annual means.

28

 6 CONCLUSION

Table 2: E ect of Area C charge suspension on vehicle entries

Charge suspension

Charge suspension*weekend

All vehicles
∗∗∗
26725.2

Charged vehicles
∗∗∗
29266.1

Motorcycles
-1920.9

-62.69

(5059.5)

(3275.8)

(2447.3)

(54.50)

3295.2

∗
171.4

∗∗
-19590.1

-23094.3

∗∗∗

Other vehicles

(9090.5)

(5865.1)

(3566.0)

(98.93)

Year, month, week, DoW FEs

Yes

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Yes

Observations
2

R

1737

1737

1720

1737

0.808

0.805

0.785

0.901

Dependent variable is daily Area C entries.  Other  vehicles are primarily public vehicles like police cars and ambulances, which
are exempt from the charge. Each column is a single model corresponding to equation (1).  Charge suspension  corresponds
to the weekday e ect, while the sum of  Charge suspension  and  Charge suspension*weekend  corresponds to the weekend
e ect. All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year,
a holiday dummy, a 7th-degree time trend in date, and a dummy for the unpriced January 2012 interim period.

Weather

controls comprise ten-piece linear splines in temperature and positive precipitation. Newey-West standard errors with 7 lags in
parentheses. Signi cance denoted by:

* p < 0.10, ** p < 0.05, *** p < 0.01.

14.5 percent of the mean.

29

The e ect on entries by all vehicles is approximately

 6 CONCLUSION

Table 3: Weekday e ect of Area C charge suspension on sensor-level tra c volume, by distance outside Area
C boundary

Vehicle count
∗∗∗
469.8

All roads

Vehicle count

Vehicle count

(131.3)

∗∗∗
1063.2

Area C

(337.1)
0-1km outside boundary

-1061.1

∗

(587.0)
1-2km outside boundary

-161.5
(361.0)

∗∗
606.8

2-4.2km outside boundary

(258.3)
>4.2km outside boundary

515.3
(391.2)

∗∗∗
469.2

Non-ring roads

(158.4)

∗∗
-2433.7

Ring roads

(1020.0)
Year, month, week, DoW FEs

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Weather controls

Yes

Yes

No

803086

801442

801442

0.085

0.093

0.093

Observations
2

R

Dependent variable is daily count of vehicles passing over sensor. Each column is a single model corresponding to equation
(2). All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year, a
holiday dummy, a 7th-degree time trend in date, and a dummy for the unpriced January 2012 interim period. Weather controls
comprise ten-piece linear splines in temperature and positive precipitation. Distance measured from the outside of the Area C
boundary. Distance dummies set at the 25th, 50th, and 75th percentiles.  Ring roads  denotes the Circonvallazione Esterna,
a ring of unpriced, larger roads located .6km-2km outside the Area C boundary. Standard errors clustered at sensor level in
parentheses.

Signi cance denoted by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

approximately 8 percent of the mean.

30

The overall e ect in the leftmost column is

 6 CONCLUSION

Table 4:

Weekday e ect of Area C charge suspension on portal-level vehicle entries, by public transit

availability

Vehicle count
∗∗∗
883.9

No metro

Vehicle count

Vehicle count

Vehicle count

(160.4)
Metro

259.9
(266.5)

No bus

905.6

∗∗∗

(148.6)
Bus

374.3

∗

(211.7)

∗∗∗
788.6

No tram

(152.3)

∗∗
586.1

Tram

(247.1)

∗∗∗
1063.2

No public trans.

(208.9)
Public trans.

518.3

∗∗∗

(133.4)
Year, month, week, DoW FEs

Yes

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Yes

71862

71862

71862

71862

0.407

0.407

0.406

0.407

Observations
2

R

Dependent variable is daily Area C entries through a given portal. Each column is a single model corresponding to equation
(3).

All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year,

a holiday dummy, a 7th-degree time trend in date, and a dummy for the unpriced January 2012 interim period.

Weather

controls comprise ten-piece linear splines in temperature and positive precipitation. Standard errors clustered at portal level in
parentheses. Signi cance denoted by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

31

 6 CONCLUSION

Table 5: Price elasticity of Area C vehicle entries

All charged
∗∗∗
-0.304

Passenger
∗∗∗
-0.171

Commercial
∗∗
-0.467

-0.0222

(0.0944)

(0.0330)

(0.221)

(0.0341)

Month, week, DoW FEs

Yes

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Yes

log(Price)

Observations
2

R

Other

1147

1147

1147

1147

0.458

0.418

0.662

0.557

Dependent variable is log daily Area C entries. Each column is a single model corresponding to equation (1), but with the
policy variables replaced by the log of weekday average price:
under Ecopass,

¿5

¿0.72

for passenger vehicles and

¿3.52

under Area C. Estimated elasticities re ect tra c response to two changes:

for commercial vehicles

1) a 1 percent increase in

the Area C cordon charge; 2) the resulting cost decrease from reduced congestion and travel time. Weekends were unpriced
under both policies and are excluded from the sample. The January 2012 interim period and the July-September 2012 natural
experiment period were also unpriced. All speci cations include 11 month, 51 week, and 4 day of week FEs, a holiday dummy,
and a 7th-degree time trend in date. Co-linearity with price variation prevents the inclusion of year dummies. Weather controls
comprise ten-piece linear splines in temperature and positive precipitation.  Other  vehicles are primarily public vehicles like
police cars and ambulances, which are exempt from charge.
denoted by:

*

Standard errors clustered at the year-week level.

Signi cance

p < 0.10, ** p < 0.05, *** p < 0.01.

Table 6: Weekday pollution e ect of Area C charge suspension, by location

Area C

Ring roads

ln(CO)
∗∗
0.0606

ln(PM10)

(0.0248)

(0.0407)

ln(PM2.5)

0.0404

0.0182
0.0205

Outside

0.1696

∗∗

0.2139

∗

(0.0676)

(0.1210)

Lagged pollution

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Year, month, week, DoW FEs

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Dependent variable is daily log average pollution in a given area of Milan. Pollution normalized for temperature and pressure by
ARPA. Each estimate comes from a di erent regression corresponding to equation (4). Speci cations include 10 year, 11 month,
51 week, and 5 day of week FEs, a weekend dummy interacted with year, a holiday dummy, a 7th-degree trend, a dummy for
the unpriced January 2012 interim period, and 1 lag of log average pollution. Weather controls include 4-knot cubic splines in
humidity, wind speed, solar radiation, and precipitation, plus a dummy for positive precipitation. Newey-West standard errors
with 1 lag in in parentheses. Signi cance denoted by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

32

 A

Table 7: Weekday pollution e ects of placebo suspensions

CO

PM10

0.0463

-0.0642

(0.0247)
∗
-0.0358

(0.0472)

PM25

2008
Area C
Ring roads

(0.0206)
Outside

-0.0495

0.0474

(0.0681)

(0.1399)

2009
Area C
Ring roads

-0.0095

0.0464

(0.0198)
∗∗
0.0433

(0.0428)

(0.0182)
Outside

0.1231

∗∗

(0.0591)

0.1492
(0.0917)

2010
Area C

-0.0182

-0.0718

(0.0254)

(0.0493)

Ring roads

0.0063
(0.0184)

Outside

-0.0258

-0.0498

(0.0736)

(0.1054)

2011
Area C

∗∗∗
0.1007

-0.0516
(0.0492)

Ring roads

(0.0177)
∗∗∗
-0.0506
(0.0190)

Outside

0.0142

0.0697

(0.0849)

(0.1138)

Dependent variable is daily log average pollution pollution in a given area of Milan.

Pollution normalized for temperature

and pressure by ARPA. Each estimate comes from a di erent regression corresponding to equation (4). Placebo suspensions
run July 27-September 16.

Speci cations include 10 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy

interacted with year, 7th-degree trend, a holiday dummy, a dummy for the unpriced January 2012 interim period, a dummy
for the June-Sept. 2012 charge suspension, and 1 lag of log average pollution. Weather controls include 4-knot cubic splines in
humidity, wind speed, solar radiation, and precipitation, plus a dummy for positive precipitation. Newey-West standard errors
with 1 lag in parentheses. Signi cance denoted by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

Appendix A
A.1 Public transit results
The results in Table A1 provide evidence of some preference heterogeneity (sorting).
They are inconsistent with an explanation solely based on cost di erences, which would

33

 A.1 Public transit results

A

predict smaller responses on more congested routes, but do not exclude the possibility that
cost di erences drive some of the responses to charge suspension.
To see this, consider the framework of Anderson (2014), who derives a condition for
choosing rail over driving (simpli ed here):
time costs of driving and the subway.

crail −cdrive ≤ P0 , where crail and cdrive denote the

P0 is the  scal cost di erence between modes converted

to units of time. Suppose three driver types, A, B, and C as illustrated in Figure A1, all of
whom initially take the subway.

16

Drivers B and C face higher subway time cost because they

must commute circumferentially to the subway line. Assume B and C are close enough that
this cost is the same (crail,C
Provided

cdrive,A

is not too

= crail,B ). Then we have crail,C − cdrive,C < crail,B − cdrive,B ≤ P0 .
small, crail,A − cdrive,A < crail,C − cdrive,C < crail,B − cdrive,B ≤ P0 .

This framework partially reproduces the public transit results of Table 4; it predicts the
marginal drivers will be of types B and C, not A. But it implies more infra-marginal drivers
on less congested roads (more type B than C), which does not match the pattern of Table
A1.

Figure A1: Marginal drivers

16 While this example is obviously stylized, the radial layout mimics the actual pattern of roads in Milan.

34

 A.2 Supplemental  gures

Table A1:

A

Weekday e ect of Area C charge suspension on portal-level vehicle entries, by public transit

availability and congestion

Vehicle count
∗∗∗
1012.6

No public trans.

(211.4)

∗∗∗
541.5

Public trans.

(149.3)
No public trans. * congested

∗∗∗
843.9
(159.7)

Public trans. * congested

-189.1
(534.9)

Year, month, week, DoW FEs

Yes

7th-deg. trend in date

Yes

Weather controls

Yes

Observations
2

71862

R

0.407

Dependent variable is daily Area C entries through a given portal. Each observation is a portal-day. Each column is a single
model corresponding to equation (3). Congested dummy equals 1 for portals where avg standardized peak (8-9:30AM, 5:458PM) volume is below 10th percentile. All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend
dummy interacted with year, a holiday dummy, a 7th-degree time trend in date, and a dummy for the unpriced January 2012
interim period. Weather controls comprise ten-piece linear splines in temperature and positive precipitation. Standard errors
clustered at portal level in parentheses. Signi cance denoted by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

A.2 Supplemental  gures
Figure A2: Milan's Area C

Area C is the area of the Milan city center that has been priced under the Ecopass and Area C policies. Numbered circles
represent entry portals where the charge is applied using license plate-reading cameras.

35

 A.2 Supplemental  gures

A

c

v
no

de
01

t
01

oc
01

se
p

g
au

01

l
01

ju
01

n
ju
01

ay
m

r
01

ap
01

ar
m

fe

b
01

01

01

ja

n

50000

100000

Area C entries
150000
200000

250000

Figure A3: Area C vehicle entries by day of year, 2008-2011

Day of year
Figure shows average number of Area C entries by day of year 2008-2011, as measured by license plate readers. Vertical lines
delimit the period of the 2012 charge suspension.

36

 Table A2: Comparison of tra c volume e ects and elasticities to other empirical road pricing studies
Paper

Location

Policy

Volume change

Price elasticity

Gibson and Carnovale 2015

Milan

Cordon charge

+14.5%

-.30

Jones and Hervik (1992)

Alesund

Toll

-

-.45

Jones and Hervik (1992); Ramjerdi et al. (2004)

Oslo

Cordon charge

-10% to 0%

-.22, -.03

Polak and Meland (1994); Meland (1995)

Trondheim

Cordon charge

-10%

-.10

Small and Gomez-Ibanez (1998)

Autoroute A1 Paris-Lille

Variable toll

-4%

-.16

Small and Gomez-Ibanez (1998)

Singapore Restricted Zone

Cordon charge (1975)

-44%

-

Goh (2002); Olszewski and Xie (2002)

Singapore Expressways

Toll (1995)

-16%

-.22 to -.15

Olszewski and Xie (2002)

Singapore Restricted Zone

Cordon charge (1976, 1989)

-52% to -10%

-.22 to -.32

Tretvik (2003); Ramjerdi et al. (2004)

Bergen

Cordon charge

>-3%

-

37

Small et al. (2006)

Orange County SR91

Variable toll

-

-.36

Odeck and Brathen (2008)

Norway, various

Toll

-

-.56 to -.82

Santos (2008)

London

Cordon charge

-18%

-

Eliasson et al. (2009)

Stockholm

Cordon charge

-22%

-

Finkelstein (2009)

Various US

Toll

-

-.06

Foreman (2013)

SF Bay Bridge

Variable toll

-9%, -4%

-.08

Xie (2013)

Minneapolis I-394

Variable toll

-

-.14

A.3 Supplemental tables

A.3 Supplemental tables

With the exception of our result, papers are listed in order of publication. Volume changes are from the introduction of pricing or a price change. The volume estimate from
our study is positive because we examine the removal of a pricing policy, rather than the imposition of one. Note that our volume change estimate is from a natural experiment,
while our price elasticity relies on long-run policy variation. Small and Verhoef (2007) suggest the very large e ect of Singapore's initial pricing policy was due to the extremely
high charge, which they characterize as far above the second-best optimum level. We separate estimates based on the initial 1977 introduction of pricing in Singapore (Small and
Gomez-Ibanez, 1998) from those based on later price changes (Goh, 2002; Olszewski and Xie, 2002). Foreman (2013) provides two estimates: -9% from a regression discontinuity
identi cation, and -4% from a di erence-in-di erences identi cation. The rightmost column contains elasticities of tra c volume (cordon crossings or volume on a segment)
with respect to price. Unless otherwise noted, they re ect both the direct e ect of a toll increase (negative) and the rebound e ect from reduced congestion (positive). They
generally do not correspond to structural parameters. Elasticities are not available for all of the policies. In some cases authors do not provide su cient information to calculate
them. In others the relevant toll rises from zero to a positive number, leaving the percentage change unde ned. In particular, the elasticity estimate of Olszewski and Xie (2002)
comes from later price variation, not the initial introduction of the Singapore cordon charge. The Finkelstein (2009) estimate is an average over 33 US facilities.

A

 A.3 Supplemental tables

A

Table A3: Comparison of Area C charge suspension to unpriced January 2012 interim period

Charge suspension

Interim period

All vehicles
∗∗∗
26725.2

Charged vehicles
∗∗∗
29266.1

Motorcycles
-1920.9

-62.69

(5059.5)

(3275.8)

(2447.3)

(54.50)

38692.4

∗∗∗

31237.0

∗∗∗

7924.1

∗∗

Other vehicles

13.11

(8308.6)

(6166.6)

(3978.2)

(68.76)

Year, month, week, DoW FEs

Yes

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Yes

Observations
2

R

1737

1737

1720

1737

0.808

0.805

0.785

0.901

Charge suspension estimates are identical to those in main results (Table 2). Dependent variable is daily Area C entries. Each
column is a single model corresponding to equation (1). All speci cations include 4 year, 11 month, 51 week, and 5 day of
week FEs, a weekend dummy interacted with year, a holiday dummy, and a 7th-degree time trend in date. Weather controls
comprise ten-piece linear splines in temperature and positive precipitation.

Other vehicles are primarily public vehicles like

police and ambulances, which are exempt from charge. The interim period January 1-15, 2012, between the Ecopass and Area
C policies, was unpriced. Newey-West standard errors with 7 lags in parentheses. Signi cance denoted by:
0.05, *** p < 0.01.

38

*

p < 0.10, ** p <

 A.3 Supplemental tables

A

Table A4: Weekday e ect of unpriced January 2012 interim period on sensor-level tra c volume, by distance
outside Area C boundary

All roads

Vehicle count
∗∗∗
1067.6

Vehicle count

Vehicle count

(127.4)

∗∗∗
1597.4

Area C

(437.7)
0-1km outside boundary

583.8
(494.3)

1-2km outside boundary

892.9

∗∗∗

(344.6)

∗∗∗
978.0

2-4.2km outside boundary

(308.1)
>4.2km outside boundary

1402.0

∗∗∗

(310.5)

∗∗∗
1330.2

Non-ring roads

(150.8)
Ring roads

-231.0
(697.1)

Year, month, week, DoW FEs

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Weather controls

Yes

Yes

No

803086

801442

801442

0.085

0.093

0.093

Observations
2

R

Dependent variable is daily count of vehicles passing over sensor. Each column is a single model corresponding to equation (2).
All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year, a holiday
dummy, a 7th-degree time trend in date, and a dummy for the July-September 2012 charge suspension.

Weather controls

comprise ten-piece linear splines in temperature and positive precipitation. Distance measured from the outside of the Area C
boundary. Distance dummies set at the 25th, 50th, and 75th percentiles.  Ring roads  denotes the Circonvallazione Esterna, a
ring of unpriced, larger roads located .6km-2km outside the Area C boundary. The interim period January 1-15, 2012, between
the Ecopass and Area C policies, was unpriced. Standard errors clustered at sensor level in parentheses. Signi cance denoted
by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

39

 A.3 Supplemental tables

A

Table A5: Weekday e ect of unpriced January 2012 interim period on portal-level vehicle entries, by public
transit availability

No metro

Vehicle count
∗∗∗
800.9

Vehicle count

Vehicle count

Vehicle count

(118.9)
Metro

577.1

∗∗∗

(153.9)
No bus

822.4

∗∗∗

(131.2)
Bus

571.6

∗∗∗

(130.3)

∗∗∗
791.5

No tram

(128.1)

∗∗∗
598.6

Tram

(133.8)
No public trans.

913.3

∗∗∗

(148.2)

∗∗∗
633.5

Public trans.

(109.2)
Year, month, week, DoW FEs

Yes

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Yes

71862

71862

71862

71862

0.407

0.407

0.406

0.407

Observations
2

R

Dependent variable is daily Area C entries through a given portal. Each column is a single model corresponding to equation
(3).

All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with year,

a holiday dummy, a 7th-degree time trend in date, and a dummy for the July-September 2012 charge suspension. Weather
controls comprise ten-piece linear splines in temperature and positive precipitation. The interim period January 1-15, 2012,
between the Ecopass and Area C policies, was unpriced. Standard errors clustered at portal level in parentheses. Signi cance
denoted by:

*

p < 0.10, ** p < 0.05, *** p < 0.01.

40

 A.3 Supplemental tables

A

Table A6: E ect of 2011 placebo suspension on Area C vehicle entries

All vehicles
Placebo suspension (2011)

Charged vehicles

Motorcycles

Other vehicles

2296.6

419.8

1815.7

-77.80

(5826.7)

(3992.3)

(2297.9)

(48.72)

Year, month, week, DoW FEs

Yes

Yes

Yes

Yes

7th-deg. trend in date

Yes

Yes

Yes

Yes

Weather controls

Yes

Yes

Yes

Yes

Observations
2

R

1737

1737

1720

1737

0.808

0.805

0.785

0.901

Dependent variable is daily Area C entries.  Other  vehicles primarily public vehicles like police cars and ambulances, which are
exempt from charge. Each column is a single model corresponding to equation (1). Placebo suspension runs July 27-September
16, 2011.

All speci cations include 4 year, 11 month, 51 week, and 5 day of week FEs, a weekend dummy interacted with

year, a holiday dummy, a 7th-degree time trend in date, a dummy for the unpriced January 2012 interim period, and a dummy
for the June-Sept.

2012 charge suspension.

Weather controls comprise ten-piece linear splines in temperature and positive

precipitation. Newey-West standard errors with 7 lags in parentheses. Signi cance denoted by:
< 0.01.

41

*

p < 0.10, ** p < 0.05, *** p