Price Comparison and Pricing Strategies

A Case Study:
the Italian Motorway Refuelling Market
Luca Aguzzoni

February 2009
Abstract
The availability of price information is crucial in shaping mar-
ket competition. Policy interventions aimed at informing con-
sumers about prices and making possible price comparison have
the potential to lower search cost and empower consumers. This
might in turn trigger competition among retailers that compete
to attract informed (and active) consumers. This paper studies
the implementation of a consumer policy aimed at facilitating
fuel price comparison on Italian toll motorway. We empirically
test the eect of introducing price comparison on the price nal
consumer face.
Very preliminary version
Key words: Consumer Policy, Treatment Evaluation, Price Dis-
persion, Gasoline Market, Price Comparison

[email protected], European University Institute, Florence

1 Introduction
Availability of price information plays a crucial role in shaping market com-
petition. There are several reasons why rms might not want to provide price
information to the market and why consumers might not search all the avail-
able prices. Firms might nd it protable not to disclose price information
as to retain market power and charge higher prices to captive consumers. On
the other hand, consumers might not search for prices because they entail
positive search cost and might settle to buy a product at an higher price
even though the same product was on sale at a lower price from another
retailer. For instance Diamond (1971) describes a situation in which con-
sumers are uninformed about prices and have positive search cost. Firms
sell an homogeneous product and have identical marginal cost. If consumers
visit randomly the rst retailers then it is optimal for the rms to charge the
monopoly price and for the consumer to search only one rm. This result is
dened as the Diamond Paradox as we just need the presence of a positive
(possibly very low) search cost as to make optimal for the rms to charge
the monopoly price.
Such a situation describes extremely well the market under study in this
paper, the (Italian) motorway refueling market. A market that each year
satisfy more than 10% of all the automotive refueling needs in Italy (Unione
Petrolifera 2006, 2007). In such a market retailers (rms) do not have any
incentive to supply price information to consumers and consumers face search
cost (mostly in terms of cost of time) that make it optimal for them to just
shop randomly at the rst retailer.
In April 2007 a law approved by Italian parliament (Law 2 April 2007
n.40) provided some guidelines to motorway concessionaire about some reme-
dies to put in place as to increase price information and make possible price
comparison, as to foster price competition among retailers.
An highway concessionaire Autostrade per lItalia, the main Italian toll-
motorway concessionaire, in July 2007 implemented such measures on its
own network as to oer a service to its customer, before setting o for sum-
mer holidays. The consumer policy intervention consisted in: 1) providing a
price comparison website were drivers could consult oered fuel prices before
setting o; and most importantly 2) the installation of roadway price com-
parison panel as to compare the available prices at the next four consecutive
refuelling stations. The implementation of such a policy was welcomed by
1
consumer association that since long time had demanded the deployment of
such comparison devices, and expected savings of up to 8 euro cents per liter
were foreseen
1
.
In this paper we use econometric methods, typically used in the pro-
gram/treatment evaluation literature, to empirically estimate the impact of
the price comparison policy, on the service stations that were required to
comply to it. We take as outcome variable of interest the nal price paid by
consumers and we estimate the eect of price comparison on price levels.
To perform the empirical analysis we collect the price information ex-
ploiting the availability of the price comparison website. We then exploit
some feature of the strategy used for the deployment of the roadway price
comparison panels, which happened in stages, to identify the eect of the
intervention (treatment).
Our identication strategies are based on a matching on observables esti-
mator for the rst stage of panel deployment; and on a dierence in dierence
estimator for the second stage of panel deployment.
In the empirical analysis we do not nd evidence that price comparison
panels had the power to lower consumer prices. On the contrary we nd
evidence that price comparison panel might be associated with slightly higher
prices. We also nd evidence that service stations concerned by the policy
intervention change their prices more frequently than station unaected by
the policy change, but still with no eect on nal consumer price.
2 Related Literature
This paper mainly relates to the growing literature that looks at the persis-
tence of price dispersion, even for homogeneous products,.in today consumer
markets. Such a literature revamped during the last decades and looks with
great attention at the role that available information technologies can play
in solving long term market information problems. Indeed, after the mass
adoption of internet by consumers in the late nineties and the success of
online shopping economists looked with increased attention at the dynam-
ics of online prices. In particular internet was thought of having reduced if
not brought to zero the search cost entailed by consumers when comparing
prices (The Economist, November 20 1999, cited in.Baye, Morgan, Scholten
2004). This should have in turn been reected in reduced price dispersion, at
1
Carlo Rienzi for CODACONS (Coordination of the Assotiations for the Protection of
Consumers), 17 July 2007, ANSA news
2
least in online markets. Since then several papers attempted to test whether
online markets were indeed close to perfect competition as it was originally
thought (among these papers Brynjolfsson and Smith 1999, Ellison and El-
lison 2001, and Baye Morgan Scholten 2004). For instance, Baye Morgan
Scholten in their paper analyzes prices of over 1000 products over a period
of eight months, recording more than 4million price information. In their
paper they empirically study both the hypothesis of price dispersion and of
price convergence. However, they dont nd evidence of price convergence.
During the eight months under study the gap between the two lowest prices
did not signicantly get smaller.
Thus not even when online consumers have access to a clearinghouse, a
platform on which rms post their prices and consumers at once can get
information on all the prices, we observe price convergence and the law of
one price does not hold.
In our paper we study a similar technological platform but this time
intended for road refueling and not for online shopping. Thus we test what
is the potential for information technologies when they are applied to real
market (vs the online virtual marketplace).
As the other studies nds other consideration than prices have to play
a role in consumer considerations, even for highly standardized goods, and
persistent dierences between informed and uninformed consumers are still
the key to understand price dispersion.
3 Case Study Framework
In this section we give an overview of the case study. The scope of our
analysis is to evaluate the eect of a consumer policy, implemented in a
specic market (motorway
2
gasoline refuelling) taking as given the market
structure in place. This section rst describes the policy and its practical
implementation. Then, it briey describes the Italian motorway refuelling
market with the purpose to explain the environment under study, motivate
the relevance of our case study and collocate it in the wider scenario of
consumer policy evaluation and retailer gasoline market competition.
2
In this paper we use the words "motorway" and "highway" interchangeably to refer
to pay toll high capacity roads designed to carry fast motor trac.
3
3.1 The Policy Prescription
In early 2007 the Italian government, through a decree-law
3
, committed to
foster competition and increase consumer protection in some consumer sen-
sitive markets, among which the gasoline market
4
. The parliament then,
some months later, approved the government decree and turned it into law
(Law 2 April 2007 n.40
5
). From Art.2, of the law, (the one regulating gaso-
line market) we can read the objective of the legislator as follow: (1) foster
competition, and (2) price transparency; (3) guarantee an adequate level of
knowledge about cost of service, and (4) facilitate the comparison of alter-
native oers. Although, the policy objective are stated clearly the law is not
so clear about the measures to be taken as to reach the proposed objectives.
Indeed, the law limits only to recommend the dissemination of information
about prices (even in comparative form) using the already available channels
or by predisposing new ones. The law then delegates to an Interministerial
Committee of Economic Planning (CIPE) the denition of the specic guide-
lines. Accordingly, the CIPE, in July 2007, published its guidelines were it
prescribed concessionaires of main national roads (pay-toll or not) to pre-
dispose a price comparison information system. Nonetheless, also the CIPE
guidelines delegated to a forthcoming act from the Ministry of Transport the
denition of the exact procedures to follow.
In the mean time, in anticipation to both the CIPE guidelines and the
exact ministerial specications, Autostrade per lItalia (henceforth ASPI), the
largest Italian pay-toll motorway concessionaire, decided to implement the
proposed price information measures before the 2007 summer holiday (taking
place usually in August, when million of drivers use toll-motorways to reach
their holiday destination). As it appears clear from ASPI press releases of the
time the decision to act in advance of further regulation or specication was to
oer an information service to their customer. This decision has indeed been
presented, and marketed, directly by ASPI, within the category of Customer
Information Services together with the decision to install information panel
about trac conditions and the decision to oers several other customer
oriented products to facilitate (and incentive) motorway driving
6
.
3
Decree Law 31st January 2007 n.7
4
Other markets considered were: Fixed Line and Mobile Phone, Internet Services, Car
Insurance, Mortgages, Airline Taris, and "best before date" in food products.
5
Law 40/7 henceforth
6
ASPI, both during the summer period and other period of the year launched several
customer oriented initiatives (for instance free coe between 00.00-05.00, trac infor-
mation in English, dedicated area in stopping area for babies and pets, help in travel
planning). Source:Press conference presentations: "Via Libera allestate" 2007; "La via
4
There is no evidence that ASPI received particular pressures from public
bodies (i.e. Government or parliament) to implement the price information
policy, as at January 2009 it is still the only highway concessionaire to have
acknowledged Law 40/7 and complied to it
7
. We take these facts as evidence
that the ASPI acting in the design and implementation stage was mostly
independent and driven by internal company considerations.
In practice what ASPI did during the months between April and July 2007
was: 1) First, to create a software platform that stations
8
managers could
use to communicate, in real time, the fuel prices oered at their premises;
2) Second, post these prices (Figure 2, Appendix) , catalogued by motor-
way code, kilometer and direction, in an apposite section within the ASPI
website (www.autostrade.it); 3) Finally, they started to install physical price
comparison panels
9
( like the one in gure 3 and 4 Appendix).along their
motorway network
Thus, the two price comparison measures adopted are:1) Price comparison
website; and 2) Physical price comparison panels. Although both measures
oer the same type of informative content they dier under some dimensions.
The comparison website is accessible only through an internet connection
(thus there is no simultaneity between information and purchase
10
), lists all
the stations on the ASPI network, but it entails some positive search cost (be
aware of the service, time to open the browser, locate the desired highway
and pool of stations). On the other hand, the road comparison panel only
lists 4 consecutive stations, it is available for free to everyone (driving by), it
entails almost no search cost
11
, and the price information and the purchase
decision can be potentially simultaneous
12
. Given these characteristics we
per lestate" 2008; Website www.autostrade.it
7
Eventually it happened that the specications adopted by ASPI do not match the
one eventually approved in the CIPE guidelines of July 2007. The CIPE specications set
a maximum of 3 consecutive stations for each comparison panel (while ASPI panel has
4 consecutive stations). Moreover ASPI panel mark with an highly visible green dot the
cheapest station and this is not required by CIPE guidelines.
8
We use the words "station", "refueling station", "service station", "retailer" , "lling
station" interchangeably to refer to a facility that oer the refueling service.
9
We use the words "physical" or "road" comparison panel to refer to a tangible price
comparison device installed next to the roadway.
10
Although recent development in mobile technologies make it possible to browse the
web also on the move
11
There could be an attention cost. To process the information on the panel the driver
has to divert some of his cognitive ability from driving to the acquisition of the infor-
mations. Still we assume this cost is a fraction of the cost required to access the online
version.
12
The stations listed on the panel are usually within a distance of 2 to 100 km.
5
assume that the latter measure (the road panel) has the highest potential
to disseminate price information and provide price comparison among close
substitutes.
The two measures also dier in the way they have been introduced. In-
deed, shortly after the creation of the software platform the website com-
parison was already online covering all the refuelling stations on the ASPI
network. On the other hand, the deployment of the road panels could not
be as instantaneous. It was indeed a long process that started in summer
2007, with only few panels installed, and came to an end in late 2008 when
eventually all stations on the ASPI network were covered by the road panels.
This timing dierence implied that some stations were required to post their
prices both on the online comparison website and on the road comparison
panel since July 2007 while, at the same time, some other stations only had
to post their prices on the comparison website, and were covered by road
comparison panel only at a later stage.
In our analysis we argue that we can exploit this timing dierence to esti-
mate the impact of the introduction of the physical price comparison panel.
To do that we exploit the price comparison website, for the purpose of data
collection
13
. Then we use information about the timing of deployment of
new road panel as to identify the eect of the policy. The following section
describes ASPI deployment decision (of road comparison panels) and it ex-
plains how we can exploit it to estimate the impact of this price information
measure.on market prices.
3.2 ASPI Policy implementation
Dierently from the price comparison website the deployment of the road
price comparison panel was not simultaneous for all the stations on the ASPI
network. Institutional and physical constraints together with corporate de-
cisions commanded that the deployment took place in several stages. As we
discussed above, ASPI had since the beginning all the intention to deploy
the price comparison panels as soon as possible on its network. However,
it was not possible to install all the panels in such a short time because of
limitation imposed by procurement law. Indeed, concessionaire of pay-toll
highways are required to issue a call for tender, with European wide public-
ity, for works with starting value higher than Euro 221,000. Therefore ASPI
13
The introduction of the price comparison website was simultaneous for all the stations
on the ASPI network. We assume the eect of the website is the same for all the station
and the estimate we nd take as baseline the case in which prices are posted on the web.
6
decided to split the deployment in two stages: First stage, deploying only
10 panels as to overcome the limitation imposed by the public procurement
law; Second stage,for the remainder of the project, issue a call for tender,
with European wide publicity, assign it and completing the deployment to
cover all the ASPI network. For the purpose of our analysis we were able to
identify two clear waves (or phases) of deployment. The rst wave took place
in July 2007 in which all the rst 10 panels were installed; the second wave
took place in July 2008 when other 24 panels were installed (Table 5 shows
the panel position and time of deployment). We then exploit these two waves
to empirically test the impact of the introduction of this price information
remedy.
3.2.1 Phase 1- A Quasi Experiment?
During the rst phase only 10 price comparison panels were installed, cover-
ing 38 service stations As we said above this resulted as an outcome of the
trade-o between the limitations imposed by the institutional constraints and
ASPI commitment to oer an information service to its customers before the
2007 summer holiday. Given the low numbers of panel installed ASPI had
to make a location decision for these rst panels. Such location decision
was not random. Indeed, the objective was to expose the highest number of
customer to this new service. Hence, ASPI targeted the highway sections
14
with the highest road trac levels. These turned out to be the outbound
highway sections close to the largest cities along the ASPI network. The
cities targeted in the rst wave of deployment are Naples, Rome, Florence,
Bologna and Milan. Although not random, the location decision, in this rst
wave of deployments, was thus based on an observables (the trac level, and
possibly the outbound direction). However as long as we nd other highway
sections with similar trac levels and market conditions we can argue that
conditioning on the observable trac level the decision to treat
15
some ser-
vice station, instead of others, was indeed random. In the empirical analysis
section we exploit this feature and use a matching on observable estimator
to estimate the impact of the policy during the rst phase.
14
We use the words "section" and "segments" interchangeably to refer to a part of
motorway between an entry point and the rst exit point on the same direction. Usually
there is never more than one service station for a single segment. Usually for each stations
on one segment there is one on the opposite segment, on the opposite side of the motorway.
15
We adopt the treatment evaluation terminology where we refer to treated units when-
ever we mean those units that are directly aected by the policy. In our case treated
stations are those stations that are required to post their prices on the road comparison
panel.
7
3.2.2 Phase 2
After July 2007 no other panel was installed until July 2008 when the second
wave of deployment started. As mentioned above the second phase started
only after the assignment of the full completion procurement, for which ASPI
had to issue a call for tender with European publicity. The aim of the
second phase was to nish the project started in 2007 and cover all the
ASPI network with road price comparison panels (a total of 53 panels were
planned) within the end of 2008. Of this second phase we analyze a specic
wave of deployment that took place in July 2008 just before the summer
break. In this wave a total of 24 road panels were installed (concerning other
65 stations) bringing the total number of active price comparison panels to
34 (concerning a total of 103 stations). Again our objective is to estimate the
impact of the introduction of these panels. Dierently from the rst wave
at this stage we do not know the process behind the selection of the new
panel location. Thus we cannot use the matching on observable estimator
that we use in the rst phase estimation. However for the second wave we
have available a richer set of information. We indeed recorded the daily
prices, for all stations on the ASPI network, since July 2007. Thus once the
second wave took place we are able to compare the prices before and after the
installation of the new panels (something we could not do for the rst phase
due to lack of before price information) and therefore estimating the impact
of the road panels on the new treated stations. In the empirical analysis
section, given the data available we dene and use a dierence in dierence
(DID) estimator to estimate the eect of the policy change.
3.3 The motorway refuelling market
This sections provides some information about the Italian highway gasoline
market. It is beyond the scope of this paper to illustrate in details and model
the market structure in place and their implication for our results. We take
the market structure as given and we simply assess the eectiveness of the
introduction of what can be seen as a consumer market remedy. That in our
case takes the form of a road price comparison panel. We remain agnostic
about the role of each players (retailers, oil company, highway concessionaire)
in the reaction to the treatment as we are only interested on the eect on
the nal price paid by consumers. Still this section oer an overview of this
specic market as to guide us in interpreting and evaluate our ndings.
In this paper we study the Italian motorway refueling market and more
precisely the price competition between retailers located on the pay-toll high-
8
ways. In Italy there are more than 6500 Km of pay-toll highways and al-
though this network only accounts for the 2% of the national road surfaces,
its roads sustains about 25% of the national transportations needs
16
. ASPI
is the main pay-toll highway concessionaire and has concession for roughly
3000 Km (almost half of the entire national network). Its network covers
almost all the country with exception of very few regions (Sardinia, Sicily,
Calabria, Trentino Alto Adige) On Italian highways there are more than 450
service stations (of which 210
17
on the ASPI network) that sell a range of fuel
products (typically: unleaded, premium unleaded, diesel, premium diesel).
The stations operating on the highways represent only the 2% of the total
service stations operating nationwide, however they supply more than 10% of
total fuel consumption (respectively 6% for unleaded and 15% for diesel fuel)
(Unione Petrolifera 2006 2007). By volume the most sold fuel on the high-
way is diesel that accounts for more than 75% of total fuel supply (Unione
Petrolifera 2007).
The range of fuels sold by each lling stations is considered homogeneous
(for instance oil companies even share reneries in some cases). That is,
within each category of fuel, products oered by dierent brands are quali-
tative the same, the only dierences that might arise come from brand dif-
ferentiation not related to the quality of the fuel (advertisement, corporate
social responsibility, loyalty programme).
There are eight major brands that operates on the Italian highways:
AGIP, ESSO, ERG, SHELL, Q8, TOTAL, API/IP, TAMOIL (dierently
from the ordinary roads on the highways there are very few "independent"
retailers
18
). All these competitors are vertically integrated rms that are ac-
tive at every stage from the production to the distribution process. For what
concerns the end market they all can rely on an extensive network of service
stations distributed all over the country. Such stations can be directly owned
by the oil companies or given in concession to third parties that owns and
manage them.
The price setting happens in usually in two stages. In the rst stage the
oil company indicates to the station manager a "suggested price". At the
second stage the station manager can discretionally change that price, within
a range imposed by the oil company. This range is implicitly determined by
16
Source AISCAT Association of Italian Highway Concessionaire, www.aiscat.it
17
year 2007
18
Usually known as "pompe bianche" (white pumps). These independent retailers buy
the fuel at the wholesale market, directly from reneries, and then sell it. They are
usually characterized by very low expenditure in marketing or branding and are popular
for oering lower prices or discounts
9
two contractual conditions: the lower bound is given by the price at which
the station manager buys the fuel (assuming they do not sell at a loss); the
upper bound is usually a ceiling on the price the station manager can practice
(usually determined by the oil company in relation to the "suggested price").
Thus the station managers freedom in setting prices appears to be some-
how limited, although oil companies seems to sustain that at the station level
managers can still pursue an independent pricing strategy (Autorit Garante
del Mercato b). Nonetheless oil companies retain powerful instruments to in-
uence these possible independent strategies (i.e. the suggested price and the
contractual relation linked to it). Indeed by looking at past publications of
"suggested prices"
19
(not anymore available after a recent Antitrust ruling)
it appears clear that although oil companies publish a reference "national
suggested price", for all the category of fuels and dierent types of service
(self-service vs full service), they also set a variety of price dierentials (to be
applied to the reference price) targeting smaller groups of stations (usually
dened by location) It is not rare that oil companies set a suggested price
almost for each single station (this is exactly the case for some stations lo-
cated on the highways). However, for the sake of our analysis we are not too
concerned about the actual shares of power in the pricing decision. As we
have access to the price posted by the station manager (that in turn reects
the prices asked at the pump) we only focus on the nal price as it has all
the information we need to perform our study.
4 Case Study Analysis
In the following sections we rst describe our sample, we then present some
descriptive statistics on the refueling market, and then we employ economet-
ric techniques to estimate the impact of the price comparison policy.
4.1 The Service Station Sample
To conduct our analysis we collected information on 178 service stations
(about 40% of total Italian motorway service stations and 85% of ASPI
service station) operating on the ASPI network. These stations were selected
for two reasons: 1) availability of price information; 2) location. As we have
already mentioned, we collected the daily price information directly from the
ASPI website
20
. Since ASPI was the only highway concessionaire to post
19
Available at on the Staetta Quotidiana an energy sector magazine
20
We programmed a "spider" that every day downloaded price information from ASPI
website (www.autostrade.it)
10
daily prices online we had to restrict our analysis only to this subsample of
Italian service stations. For what concern the location, among the sample of
all the service stations along the ASPI network, we selected only those located
in places relevant for our analysis (i.e. either they were treated stations or
had the potential of being included in the control sample).
However in the empirical analysis we will not use all the 178 stations
indeed depending on the estimation technique we use and depending on the
assumption we make we construct every time an appropriate control group
that we then compare to the treatment group.
For what concern the time dimension of our analysis we have collected
price information, for unleaded and diesel fuel (both in the self-service typol-
ogy
21
), for a total of 255 days. The days are not consecutive (there are some
gaps
22
) and span from 26/10/2007 until 11/11/2008. Again in the econo-
metric analysis we will not make use of all the time information as we will
have to discard some time periods or days
23
, but this will be made clear in
the next sections.
Together with the price and time information we collected other data
useful for our analysis. First of all we were able to get information about the
average daily transit for each sections of the motorway on which the service
station are located. Although this data do not give us a day of the week
variation or a time variation we consider them sucient because they provide
us with the observables that the decision maker (ASPI) declared to be using
in the selection of the treated group during the rst wave of deployment.
We also collected demographic data at the regional and provincial level that
we present in the descriptive analysis but that do not directly enter in our
econometric estimation.
4.2 Descriptive Statistics
In this section we present some descriptive statistics that guide us dening
the methods we use in the econometric analysis and interpreting the relative
ndings.
21
Usually stations oers both self and full service. However we restirct our analysis only
to self service for two reasons. Price comparison website and road panel only advertise
self service price and full service price are not advertised.
22
For instance during summer 2008 when the second wave of panel deployment took
place no data were collected. Since we could not recover exactly on which day each new
panel was turned on we decided to restart collecting data once the second wave was over
and we knew all the new panel to be active.
23
For instance some Sundays there were some nation wide promotions on refueling. We
discard the Sundays from our sample because we might pick up price eects that are not
related to the introduction of the price comparison panels.
11
Service stations located along the toll motorways, on average, sell almost
three times more unleaded fuel and eight times more diesel fuel compared to
the average station located in a non toll street (Table 1). This could be taken
as evidence that stations along the motorway can experience higher variations
in total prots due to small variations in nal price.(holding constant the
demand). However, stations along the motorway are usually of bigger size
and have also to remunerate the motorway concessionaire, in addition to the
oil company and both this factor might depress the size of nal prots.
Table 1: Average Fuel Sales per stations 2006 (ml litres)
location toll motorway all other
service stations #. 461 21989
Unleaded 2.22 0.73
Gasoline 6.19 0.74
Total 8.41 1.47
Source: elaboration from Unione Petrolifera 2006, 2007
If we look at price levels we can observe a peculiar characteristics of the
retail refuelling market: prices seldom move at the station level. For instance
if we plot the retail prices against the oil company "suggested price" ( the
price that should reect variation in costs) we observe that prices at stations
level follows almost one to one the variations in the suggested price without
having signicant independent movements. Figure 1 shows this relation,
over a period of two months, for two representative AGIP and two Tamoil
retailers, where the two solid lines are the prices set by the oil companies and
the other two lines are the retail prices . From the gure it seems that once
the oil company suggested price change the retailer x a margin (positive or
negative) and keep the price xed until a new movement in the suggested
price take place.
Thus, dierently from other papers that have looked at the refuelling
market (ystein and Frode, 2008), we do not observe signicant (non-cost
related) price variation neither within the same day nor within the same week.
For instance we have also checked (Table 6 Appendix) if price changes happen
systematically on a specic day of the week. We found that on average, over
the period we consider, 84% of the times stations do not change prices, 9%
of times there was a price increase and 7% of the times there was a price
decrease. The other evidence we found is that almost no station increases
the price on a Sunday. Whereas it does not seems there is any preferred day
for a price increase or a price reduction.
Table 2 (below) presents some descriptive statistics about the sample of
12
Figure 1: Suggested and Retailer Price Variation
Agip Retailers
1.32
1.34
1.36
1.38
1.4
suggested
retailer 1
retailer 2
Tamoil Retailers
1.32
1.34
1.36
1.38
1.4
1
/
1
4
/
1
7
/
1
1
0
/
1
1
3
/
1
1
6
/
1
1
9
/
1
2
2
/
1
2
5
/
1
2
8
/
1
3
1
/
1
3
/
2
6
/
2
9
/
2
1
2
/
2
1
5
/
2
1
8
/
2
2
1
/
2
2
4
/
2
2
7
/
2
1
/
3
suggested
retailer1
retailer2
service stations. We can see that during the rst wave of panel deployment
about 21% of service stations are treated whereas after the second wave a
total of 62% of station are treated.
Agip appears to be the predominant brand with almost 30% of service
stations with that brand. Tamoil is the second brand with a share of 20%
and Esso follows closely at 16%. The majority of the stations in our sample
are located along the two main ASPI motorways the A1(33%) and the A14
(30%). About the geographical distribution we can see that the majority of
service stations are located in the Centre (43%) another 37% are located in
the North and the remaining 21% is located in the South of Italy.
On average stations in our sample have more than twenty thousand vehi-
cles transiting on their segment of the motorway every day (although only a
fraction actually stops at the station premises). About 76% of these transits
are made by light weight vehicles (cars or vans) and the rest is made up
by heavy weight vehicles (trucks). However stations dier widely in their
average transits and the transit distribution looks very right skewed.
13
Table 2: The Service Stations Sample
outcome var. Mean
1
Std. Dev. outcome var. Mean
1
Std. Dev.
treat. wave 1 0.21 0.41 treat. wave 2 0.62 0.49
Agip 0.27 0.45 A1 0.33 0.47
Api 0.04 0.19 A8 0.03 0.18
Erg 0.07 0.25 A11 0.03 0.18
Esso 0.16 0.37 A12 0.02 0.15
Ip 0.00 0.00 A13 0.04 0.21
Kuwait 0.08 0.28 A14 0.30 0.46
Shell 0.06 0.23 A16 0.07 0.25
Som 0.01 0.07 A26 0.06 0.24
Tamoil 0.20 0.40 A27 0.02 0.15
Total 0.12 0.33 A30 0.02 0.15
light transit 17737 11162 South 0.21 0.41
heavy transit 5541 3127 Centre 0.43 0.50
total transit 23274 13768 North 0.37 0.48
1
Averages over 178 service stations on ASPI network
4.2.1 Observed Price Dispersion
In this section we provide some information about the average price disper-
sion for refueling services. In the econometric analysis we estimate the impact
that price comparison panels have on prices for nal consumers. Thus in or-
der to say something about the magnitude of this eect we need to present
some statistics about the average price dispersion that a customer might ex-
pect to nd when lling her car. Table 7 (Appendix) presents some estimate
of the observed price dispersion. In this table we present an estimate derived
from our sample and we compare it against estimates of price dispersion we
found in other studies. In our measure of price dispersion we constructed
a sample of service stations and we computed the range (Max price - Min
price) for the service stations competing in the same market
24
. We found
that on average a customer on the highway can expect prices to vary by less
24
The sample consisted of the service stations of our ASPI sample that were not aected
by the rst wave of price comparison panel deployment. Prices refer to a single day
6/2/2008 (between Phase 1 and Phase 2) and 84 service stations were included in the
sample.
We computed the range instead of other measures of dispersion to compare our results
with those of other studies that also computed the range. However for our sample we can
compute the standard deviation that we use later.
The stations are dened to be competing in the same market when they are within a
distance of less than 100 km.
14
than 2 euro cents among the dierent retailers on a 100 Km interval. This
imply a maximum potential savings of less than 1 euro for a typical car rell
of 40 liters.
We can then compare our results with those of other studies. If we look
at the average range of price dispersion on another toll motorway
25
we found
estimates very similar to ours. On the other hand if we look at price range
estimates for service stations operating in normal road we found a much
higher price dispersion. This dispersion goes from a lower bound of 4 euro
cents per liter (a savings of almost 2 euros for a 40 liter rell) if we look at
data from ISTAT (Italian National Institute of Statistics)
26
; until an upper
bound of more than 10 euro cents a liter (savings of about 5 euros for a 40
liter rell) if we look at a report from the FIGISC(Italian Federation of the
Relling Service Station Managers).
Thus we can conclude that on average customers on a toll-motorways can
expect to nd a lower price dispersion (lower scope for savings) than the one
they can otherwise nd on normal road.
Another price comparison we might be interest in is to compare the level
of fuel prices on the highway and on the urban area. This to possibly say
something about the possible degree of competition we can expect between
service stations on the urban area and those on the highway. Table 8 (Ap-
pendix) presents the mean price for unleaded and diesel fuel in ve provincial
area (Milan, Bologna, Florence, Rome, Naples). The mean price level is re-
ported for both the service stations operating on the toll motorway and for
those operating on normal streets. By comparing the means we can observe
that the average price level is higher among the motorway service stations.
However there is a signicant overlap between the two price distributions (for
both unleaded and diesel fuel). Thus there seems to be scope for competi-
tion between motorway and normal street relling
27
. For instance retailers
operating on motorway could use the price comparison panel as to attract fre-
quent travellers by oering a price comparable or lower to the prices available
on normal roads.
25
A22, managed by Autostrada del Brennero SpA. On the website
http://www.autobrennero.it/ they publish a weekly report on fuel prices.
26
ISTAT collects this data to produce its price ination indexes.
27
Moreover the mean for normal street retailers might be underestimated. Indeed we
could only get price information about the twenty cheapest service stations operating at
the province level.
15
4.2.2 Motorway use and Customer attitude to fuel rell
In this section we provide some evidence about the average use
28
of the
motorway and customers attitude to self service relling. During the year
2007 the average travel on the ASPI network was 80 Km long. Respectively
75 Km for light weight vehicles (that represents the 80% of total transits) and
99.7 Km for heavy weight vehicles. During the year the trac level seems
to be quite constant (except for a peak of light weight travels in August) A
substantial share of trips, 1/3 for light and 1/4 for heavy weight vehicles, is
less than 25 Km long and these trips are mainly concentrated around the
metropolitan areas on both inbound and outbound directions.
For what concerns the drivers relling habits we can get some information
from two surveys published by ACI (Italian Automotive Club). These studies
report that about 40% of drivers favour self service relling always or often
while another 33% opt for the self service only occasionally. The lower price
of the self service seems to be determinant in the choice of the service for
about 45% of the respondent whereas the rest favoured it for its exibility.
The same studies also provide some information about the customer loyalty
and the choice of the service stations. They report that proximity and lower
price are the two determinants for the choice of the relling station. In the
urban area most of the drivers always rell from the same station whereas
when outside the urban area there is no delity to the single service station
but there is some brand loyalty..
4.3 Econometric Analysis
In this section we use econometric techniques to estimate the eect of the
introduction of fuel price comparison panels. In particular the interest is in
estimating the average eect of the policy on the retailers that are required
to post their prices (i.e. average treatment on the treated ATT ). Following
the potential outcome framework (widely used in treatment evaluation lit-
erature), we are interested in estimating the impact of a binary treatment
(1 = 1 post price; 1 = 0 do not) on a continuous scalar outcome of interest
1 (i.e. fuel price per liter). For each individual retailer i. i = 1. .... ` , we
can dene the two following potential outcomes: 1
i
(0) if the retailer i does
not post her prices; 1
i
(1) if the retailer i post her prices. If both outcomes
were observable, and we were interested in computing the eect of treatment
28
Data come from two surveys: Autostrade per lItalia 2007, Conference Presentation
"Estate 2007. "Via libera in sicurezza "; and Autostrade per LItalia 2008, Conference
Presentation "La via per lestate. Le vacanze iniziano in autostrada".
16
on treated units, we could just compare 1
i
(1) and 1
i
(0) for each treated re-
tailer i. However, at any given point in time, only one outcome realizes for
each retailer i and such a comparison is impossible. For instance for all the
treated retailer only the outcome 1
i
(1) is observed and for the non-treated
retailers only the outcome 1
i
(0). To overcome this impossibility we use a
matching and a dierence in dierence estimator and below we explain how
we implement them to estimate the ATT.
4.3.1 Phase 1-Matching on Observables
During the rst phase the treated group consists of only 38 stations and the
non treated group of 140 stations. Regarding the data available for this stage
we have: daily, station level, price information for 206 days; average daily
transit data, for each segment of the motorway on which stations are located;
and some local geographical and demographic characteristics. The daily
prices available do not cover the period before the start of the policy thus
we cannot implement any sort of before after estimation that are typically
performed in treatment evaluation literature. Still the time dimension of the
(price) dataset let us monitor the retailers reaction to the policy over time.
As we mentioned above, the location choice for the rst ten price com-
parison panels was not completely random. Therefore if we simply compare
the price level of the treated and non-treated group (possibly controlling for
some observables), the resulting estimates would be biased. For the estima-
tion of the treatment eect we adopt a dierent identication strategy using
a matching estimator. The use of matching was directly suggested by the
rule that ASPI followed in the deployment of the rst ten price comparison
panels. Such a rule says that they targeted the outbound city motorway
segments with the highest levels of trac as to maximize the number of
customer that could see the panel. Thus the selection choice was based on
some observable characteristics and our decision to implement a matching
estimator followed naturally.
Matching estimator requires the unconfoundedness assumption (1), that
state that assignment to treatment is determined by observable variables and
conditioning on these variables assignment to treatment is random. This
implies that conditioning on the covariate A the selection into treatment is
not correlated with the outcome 1 .
1 (0). 1 (1) ? 1jA (1)
This assumption is untestable (Imbens 2005) and in the next sections we
argument when we think this assumption holds in our setting depending on
17
the covariates we include in the vector A. The other condition we need for
matching is the overlap or matching assumption (2).
0 < Pr[1 = 1jA] < 1 (2)
Assumption 2 states that conditioning on the observables A the proba-
bility of receiving the treatment (1 = 1) is greater than 0 but less than 1.
This assumption states that for a given value of A there are at least some
observations from both the treatment and control group. This is crucial for
matching and ensures that for every treated observation there is a matched
non-treated observation with the same observables.
We use the simple matching estimator (3) proposed by Abadie and Imbens
(2001) and compute it on STATA using the command
29
provided by Abadie,
Drukker, Herr, and Imbens (2003).
11 =
1
`
1
N
1
X
i:W
i
=1
n
1
i

[
1
i
(0)
o
(3)
This estimators computes the average, across all treatment units i (i =
1. .... `
1
), of the dierence between the observed outcome 1
i
and the esti-
mated counterfactual
[
1
i
(0), where the counterfactual
[
1
i
(0), for each treated
unit i. is given by:
[
1
i
(0) =
1
`(i)
M(i)
X
m=1
1
m
(4)
that is
[
1
i
(0) is the average outcome of the `(i) (untreated) units matched
with the treated unit i (where `(i) 0).
In our analysis we dene three distinct matching strategies depending on
the observables we, in turn, include in the conditioning vector A. The main
problem we face in our empirical analysis is the small size of our sample,
both for the treated group and the potential control group. This limits the
number of covariates we can include in the vector A. Indeed as we increase
the number of covariates in A the overlapping condition fails to hold as it
becomes increasingly dicult to nd a match for each treated unit. Thus
taking into account the size of our sample we can only match on few observ-
ables. Still we argue that in our setting very few observables are correlated
with the treatment decision (essentially average transit level). Nonetheless,
there might be several omitted observables that might be correlated with
29
STATA command nnmatch see Abadie, Drukker, Herr, and Imbens (2003)
18
the outcome variable but, we argue, if we perform a careful selection of our
control group we can control for it.
Table 3: Matching Results
Matching Outcome
1
ATT (euro cents) T stat
Specication 1 unleaded -0.38 -2.10

diesel 0.04 0.34
Treated=38 change unl 3.13 2.02

Control=92 change dies 4.52 2.83

Specication 2 unleaded -0.05 -0.39
diesel 0.16 0.91
Treated=20 change unl 1.82 0.90
Control=20 change dies 3.32 1.65

Specication 3 unleaded -0.30 -1.27
diesel 0.08 0.50
Treated=38 change unl 2.17 1.00
Control=25 change dies 4.28 1.81

1
Outcome variables: unleaded and diesel refer to average price level, average across time day1 day 206
change unl and change dies percentage of days in which a price change happened

signicant at 5% level

signicant at 10% level
Matching Specication 1 In the rst matching specication we only
condition on one variable: the average daily transits on a specic motorways
segment. In chart 6 (Appendix) we can see how the overlapping assumption
is satised when we restrict our analysis to those segments with more than
twenty thousands average daily transits. The assumption is that for service
stations on motorway segments with similar transit levels the selection into
treatment, made by policy maker, was random. In this strategy we assume
that other observables do not aect neither the selection into treatment nor
the outcome variable. For instance here we do not control for the fact that
all treated stations are located on segments of the motorway on the city
outbound direction, whereas within the control group, some stations are
located on city inbound sections, some on city outbound sections and some
other far from urban areas. Here the assumption is that motorway relling
stations make price settings decision that are not inuenced by their location
relative to main urban areas, their price setting decision only depends on the
average vehicles that transit by their premises. Also we are not considering
19
other observable like brand or geographical position
30
, as we assume these do
not play an important role once we control for average transit level. Table
3 presents the results of the matching estimator. We present the eect of
treatment on the treated for four outcome variables of interest. The rst two
unleaded and diesel refer to the average price level for unleaded and diesel
fuel. The other two variables change unl and change dies are a measure of
the frequency of price change at the retailer level.
From this rst specication we can see that the introduction of the price
comparison panel seems to have no statistically signicant eect on diesel
price whereas it seems to be associated with a (statistically signicant) de-
crease in unleaded fuel price of 0.4 euro cents. If we compare this gure with
the average price dispersion we expect to nd on a motorway we can see that
this eect account for roughly 20% of the observed price dispersion. However
for the nal consumer such a results would be reected only in a potential
savings of less than 20 euro cents for a typical rell of 40 liters, a savings that
does not seem to have a great potential in attracting customers, not even the
most price sensitive. Also we nd a positive, and statistically signicant,
ATT for the two price change variables. Thus it seems that retailers that
post their prices on the price comparison panels change their prices more
frequently (respectively 3% more for unleaded and 4.5% more for diesel).
The above results are based on comparisons of individual retailers Phase
1 average prices. However, comparing the mean might not be the best mea-
sure as in the aggregation over time we might lose the price dierentials that
characterize the day to day competition. Thus we estimated the ATT also
for all the 206 days in the Phase 1 sample. Figure 7 (Appendix) shows these
estimates and the respective interval of condence (at the 95% condence
level). From the chart we can see that there is not even a single day in
which we can reject the null hypothesis that the ATT is equal to zero. This
evidence suggest us to be careful when interpreting the estimates obtained
when looking only at the average price.
Matching Specication 2 In the second matching specication we adopt
a dierent matching strategy. Here the aim is to pick, for each treated sta-
tion, an untreated station that can be as close to a clone as possible. This
reduces the dimensionality of our control group and treatment group however
we hope to gain in the precision of our estimates, if we select a good clone.
Our strategy is to match each treated station with the respective untreated
30
We do control only for ve regions since we know they have a special tax regulation
for unleaded gasoline, but not for diesel.
20
station located exactly on the opposite side of the motorway
31
. By doing so
when we dierence the fuel prices we hope to dierence out all possible local
market condition that might enter the price setting decision. Stations lo-
cated on opposite sides indeed face the same labour market condition, have
similar geographical characteristics and have very comparable transit lev-
els. Essentially this matching strategy leaves out only two potential sources
of heterogeneity: brand dierentiation and type of trac (inbound vs out-
bound). We are less concerned by not controlling for brand dierentiation
as we do not nd evidence that retailers brand is associated with price dif-
ferences once we control for stations and locations eects. Moreover as we
mentioned above only a small fraction of customers seems to base their pur-
chasing decision on brands when driving outside an the urban area and this
might hold even strongly when the purchase is on a motorway. It is the latter
source of unobserved heterogeneity (inbound vs outbound) that concerns us
the most. Indeed, we could argue that the ow of trac going out of a large
urban area has dierent characteristics from the ow of trac entering. Still,
from the evidence we presented we can infer that most of the vehicles going
into (out) the urban area, at a certain point in time, are also the same vehi-
cles that will leave (enter) the urban area at a later point in time. This holds
both for long distance travels and also for local travels. Thus we can say
that, although at dierent point in time, most of the potential customers of
a refueling station are also potential customer of the refueling station on the
opposite side of the motorway. We can also rule out the possibility that local
motorway trac does not rell on motorway stations. Indeed we provided
evidence that there is an overlap between the motorway price distribution
and the urban area price distribution. Hence the price comparison panel
could even be used by retailers as to attract an higher share of these frequent
drivers than can promptly, and at a zero cost, compare the prices available
on the motorway with those available on the urban area.
Thus if we are condent that refuelling stations located on opposite side
of the motorway face almost exactly the same market conditions we can be
condent we have found almost a clone for each treated stations.
From table 3 we can see that, given this matching specication, we do not
nd any statistically signicant eect of the price comparison panel for three
of the four outcome of interest. Indeed we only nd a signicant eect (at
10% signicance level) for the frequency of change of diesel fuel price, where
treated stations change diesel price 3% more times than untreated stations.
Also, we do not nd any statistically signicant eect if we look at the
31
Some treated stations are excluded because we nd occurences were opposite stations
are both treated
21
daily estimates (Figure 8 Appendix) over the 206 days period under analysis.
4.3.2 Matching Specication 3
In the third and last matching specication we match treated and control
units under two dimensions: the average transit level and the outbound
direction (exactly the two observable on which policy maker, ASPI, based
her selection decision). In practice here we make the assumption that, even
though two opposite refueling stations face some similar market conditions,
they face a completely dierent demand given by the direction of the trac
ow. Thus potential customers driving toward (outward) the urban area
have dierent characteristics. In addition, we also control for transit levels
as this is the observable that mostly driven the selection into treatment. The
use of two conditioning variables has the eect of drastically reducing the
control group below the size of the treatment group.
If we look at the estimates of the ATT for the average price level we do
not observe any statistically signicant eect neither when we compare the
average prices (tables 3) nor when we look at the daily estimates (Figure 9
Appendix).
Again we only found a statistically signicant ATT (at 10% level of signif-
icance) for the frequency of price changes. Where again we nd that treated
stations seem to change diesel prices 4% more than untreated stations but
with no eect on price level.
4.3.3 Phase 2-Dierence in Dierence
During the second phase of panel deployment another 24 panels were de-
ployed along the ASPI motorway network (as of summer 2008). This time
(apparently) no clear rule was followed in the location decision as at this
stage the objective was to nish the work started in Phase 1 (summer 2007)
and treat all the stations of the network. Since we started collecting price
information from October 2007, once the new price comparison panels were
turned on we found ourselves with the data to perform a before after esti-
mation. Indeed in the second phase we can fully exploit the time dimension
of our dataset and estimate the impact of the price comparison panel on the
new wave of treated stations. Dierently from the rst phase in the second
phase the selection decision was not based on any observables that we can
exploit, however this does not concerns us since the time dimension let us
compare the same units, although at dierent point in time. We thus use a
dierence in dierence estimator to estimate the ATT.
If we denote with 1
G
it
the outcome of a unit i. that belongs to the group
22
G = ft:cct. co:t:o|g . at time t = f0. 1g then we can dene the following
dierences:
1
treat
i0
1
treat
i1
(5)
1
control
i0
1
control
i1
(6)
where the dierence 5 is the change in outcome, over time, for a treated
unit and dierence 6 is the change in outcome for a control unit. If we take
another dierence as in 7 we can dene the dierence in dierence estimator
as:
o
DID
= 1
treat
i0
1
treat
i1

1
control
i0
1
control
i1

(7)
Thus by taking the rst two dierences we control for the group specic
eects and by further dierencing we control for the time specic eects that
are common for both groups leaving out the eect of treatment on the treated
group.
We estimate this eect using the following OLS regression:
1
G
it
= c + ,G
i
+ t + o(t G
i
) + c
i
(8)
where the parameter o is our parameter of interest as it identies the
average eect of treatment on the treated. A crucial assumption for the
unbiasedness of this estimator is the parallel-trend assumption that is the
time trend is the same across the treatment and control group and it is equal
to .
To estimate the ATT we thus have to dene a treatment and control
group. We include in the treatment group all the refueling stations that
received the treatment at the beginning of Phase 2, and had not received
the treatment in Phase 1, this group consists of 65 refueling stations. As a
control group we select the 38 stations that received the treatment in Phase
1. Although these stations are actually receiving a treatment in Phase 2 they
were already receiving it in Phase 1. Thus for these stations we expect no
change to occur once the treatment is extended to the next 64 stations
32
.
Since control stations were already treated we assume that any change we
nd between period 1 and 2 are attributable to the time trend. As a control
group we could not use the other 75 ASPI stations we had in our sample.
32
Actually there could be some general equilibrium eects however we have to assume
away any implication of general equilibrium and assume that units are independent.
23
Indeed these latter stations were supposed to receive the treatment anytime
in the months between August and December 2008
33
.
Table 4 presents the results of the dierence in dierence estimation. The
rst two columns report the results from two regressions where we regressed
the dependent variable (daily unleaded and diesel price in euro cents) on the
period, group and treatment dummy, controlling for clustering. Instead, in
the third (and fourth) column the dependent variables is average detrended
unleaded (diesel) price (in euro cents).
From the results we can see that from period 0 (pre August 2008) to
period 1 (after August 2008) on average unleaded became cheaper and diesel
became more expensive. We do nd statistically signicant groups eect
with the treated group on average charging more for unleaded (roughly 1 euro
cents more) whereas for diesel there is not a statistically signicant dierence
between the two groups. For what concern our parameter of interest the ATT,
we nd that it is positive and statistically signicant (at 5% signicance
level) Thus it seems that the introduction of the price comparison panel is
associated with an increase in both unleaded and diesel price at the treated
refueling stations. Respectively, on average, unleaded costs about 0.75 euro
cents more and diesel costs around 0.45 euro cents more. Thus it seems that
the introduction of the policy instead of lowering price for nal consumers
had the eect of raising it (although of a small magnitude, 30 euro cents
more for a 40 liter unleaded lling and 20 euro cents more for a 40 liters
diesel lling).
Table 4: Dierence in Dierence Results
1) 2) 3) 4) 5) 6)
N 22864 22864 216 216 216 216
Dep. Var unleaded diesel avg unl avg dies change unl change dies
Phase -0.72 0.87 -0.50 0.60 0.11 0.10
(-7.84) (9.31) (-1.55) (3.28) 6.13 (4.66)
Group 0.99 -0.09 0.88 -0.26 -0.07 -0.09
(4.37) (-0.56) (3.12) (-1.62) (-4.31) (-4.75)
Treatment 0.68 0.37 0.78 0.55 0.04 0.05
(3.38) (1.89) (1.95) (2.41) (1.81) (1.84)
cons 138.03 131.96 -1.49 -1.89 0.22 0.27
(927.17) (1725.57) (-6.53) (-14.56) (16.91) (17.98)
t-stat in parenthesis
33
We could have used them as control only if we had known exactly in which day the
price comparison panels for these stations were turned on.
24
Table 4 in column 5 and 6 also present the dierence in dierence results
for two regressions in which the outcome variable of interest is the average
frequency of price change (respectively for unleaded and diesel fuel). Before
we found that during Phase 1 the treated units changed diesel price more
frequently (around 4% more than control units). Also in Phase 2 we nd
similar results. Both diesel and unleaded price seems to vary more often after
the treatment occurs (respectively 4% and 5% more, statistically signicant
at 10% level).
5 Conclusion
This paper studies the introduction and the eect of a consumer market in-
tervention aimed at helping consumer obtaining price information and make
price comparisons. This policy intervention targeted fuel prices and was
implemented on the Italian motorway network managed by Autostrade per
lItalia SpA (ASPI), the main Italian motorway concessionaire, with a net-
work of roughly 3000 Km. At the end of the intervention around 200 re-
fueling stations, operating on ASPI network, were concerned by the policy
that required them to post their fuel prices (self service unleaded and diesel
fuel) both on an online price comparison website and on some 53 road price
comparison panels.
We exploit some feature of the deployment strategy adopted to install
the road price comparison panels as to estimate the eect of this latter in-
tervention. While we exploit the availability of the price comparison website
as to collect the daily price information that we later use in the empirical
analysis. Of the deployment process of the 53 road comparison panels we
can identify two clear waves of deployment: the rst in July 2007 were 10
panels are installed on motorway sections with high level of trac (aecting
38 service stations); and the second in July 2008 were other 24 panels are
installed (aecting other 65 stations).
We use two dierent estimation methods, one for each wave of deploy-
ment, to identify the average eect of the intervention (treatment) on the
concerned stations (treated stations). For the rst wave of deployment we
argue that selection on observable is in place. We thus use a matching es-
timator to identify the (sample) average eect of treatment on the treated
(ATT). We adopt three dierent matching strategies depending on which
observables (and unobservables) we think aects the selection into treatment
and/or the outcome of interest as to control for observed and unobserved het-
erogeneity between our treatment and control group. For the second wave
25
of deployment we do not nd evidence of selection on observable. Rather,
treated stations seems to be picked without a specic rationale. However,
since our dataset let us perform a before after comparison (not possible for
the rst phase) we use a dierence in dierence estimator to compute the
ATT. We dene a treatment group and a control group and assuming that
the parallel trends condition is satised we estimate the eect of the policy
for these new treated stations.
From the empirical analysis we found that contrary to the policy maker
expectations the introduced measure had very little power in reducing nal
consumer prices. Indeed if before the introduction of the policy some con-
sumer association were forecasting possible savings of about 8 euro cents per
liter, we found a contrasting evidence. Our evidence suggests that on average
the rst ten panels seem not to have had any statistically signicant eect in
lowering the nal price to consumers. If we turn our attention to the second
wave of 24 road comparison panels we even nd a statistically signicant
increase in price of almost 1 euro cents per liter.
The only eects that the road comparison panels seems to have is that
treated retailers change more often the posted price (around 4% more than
an untreated retailers).
Motorway refueling market has all the characteristics to be a market in
which fuel retailers enjoy some degree of market power as consumers typically
make uninformed choices (prices information usually takes place only at the
station premises) and it is costly to search more retailers (more costly than in
urban areas). Thus, this intervention could have had the potential to activate
consumers and in turn activate price competition between retailers. However
this study does not nd any evidence that consumers have beneted from
such a policy.
Several reasons could explain why such a policy failed to achieve its objec-
tive. For instance, although there are many independent retailers the major
oil companies, through contractual relationship, might limit the scope of price
competition at the retailers level and actually impose their own prices. On
the other hand it could also be the case that even in a competitive settings
the maximum savings that a consumer can achieve are too limited to trigger
an active market search for a low price (for instance even a savings of 8 euro
cents per liter, the one hoped for by consumer association, would mean a
saving of only 3,2 euro for a 40 liter rell). Or it could also be the case that
it takes time before consumers learn how to use these market remedies and
therefore it takes even more time before the retailers observe the implication
of such intervention on their business.
26
Hence, further research is needed to understand why this policy did not
work in these early stages and how it can be tweaked to benet consumers
the most.
27
References
[1] AISCAT (Associazione Italiana Societ Concessionarie Autostrade e
Trafori) www.aiscat.it
[2] Ansa a, "BENZINA:CODACONS,CON CARTELLONI RISPARMI
FINO A 8 CENT/LITRO", ANSA Notiziario Generale in Italiano
July 17, 2007 Tuesday 6:13 PM CET
[3] Ansa b, "BENZINA: AUTOSTRADE, VIA A PRIMI CARTELLONI
SEGNALAPREZZI", ANSA Notiziario Generale in Italiano July
17, 2007 Tuesday 5:22 PM CET
[4] Autorit Garante Concorrenza e Mercato a, I681-PREZZI DEI CAR-
BURANTI IN RETE, Provvedimento n. 16370
[5] Autorit Garante Concorrenza e Mercato b, Presentazione degli impegni
ai sensi dellart. 14ter della L. n. 287/1990 Provvedimento n. 16370
(I681-Carburanti in Rete)
[6] Autostrade per lItalia 2007, Conference Presentation "Estate 2007. Via
libera in sicurezza "
[7] Autostrade per LItalia 2008, Conference Presentation "La via per
lestate. Le vacanze iniziano in autostrada".
[8] Michael R. Baye, John Morgan, Patrick Scholten, Price Dispersion In
The Small And In The Large: Evidence From An Internet Price
Comparison Site, Journal of Industrial Economics, vol. 52(4), De-
cember 2004, pp. 463-496.
[9] Legge 40/07, conversione del dlg n. 77 del 2 Aprile 2007
[10] Staetta Quotidiana, I "Prezzi Consigliati" dei carburanti, N.25
6/2/2007 , N.20 30/1/2007, N.15 23/1/2007, N.10 16/1/2007
[11] ystein Foros and Frode Steen, Gasoline prices jump on mondays: an
outcome of aggressive competition?, CEPR 2008 Discussion Paper
series No.6783
[12] Andrea Shepard. Price Discrimination and Retail Conguration, The
Journal of Political Economy, Vol. 99(1), February 1991, pp. 30-53.
[13] Imbens, G.W. (2005), Semiparametric Estimation of Average Treatment
Eects under Exogeneity: a Review, The Review of Economics and
Statistics
[14] Brynjolfsson, Erik and Michael D. Smith, Frictionless Commerce? A
Comparison of Internet and Conventional Retailers, MIT Sloan
School of Managment Working Paper, July 2000.
28
[15] Ellison, Glenn and Sara Fisher Ellison, Search, Obfuscation, and Price
Elasticities on the Internet, mimeo, MIT, January 2001.
[16] Diamond P., A Model of Price Adjustment, Journal of Economic Theory,
3, 156-168.
[17] Unione Petrolifera, Notizie 4/2006, www.unionepetrolifera.it
[18] Unione Petrolifera, Notizie 5/2007, www.unionepetrolifera.it
29
A Appendix
Table 5: Panel Location
Phase id Highway id Km direction
disp01 A1 Milano-Napoli 9.30 south
disp02 A1 Milano-Napoli 183.35 south
disp03 A1 Milano-Napoli 307.55 south
disp04 A1 Diramazione Roma sud - GRA 17.14 south
Phase 1 disp05 A1 Milano-Napoli 748.75 north
disp06 A1 Diramazione Roma nord - GRA 9.90 north
disp07 A1 Milano-Napoli 259.70 north
disp08 A1 Milano-Napoli 177.00 north
disp09 A8 Milano-Varese 2.00 north
disp10 A14 Bologna-Taranto 38.90 south
disp11 A1 Milano-Napoli 79.80 north
disp12 A8 Milano-Varese 13.15 south
disp13 A4 Torino-Trieste 19.20 est
disp14 A13 Bologna-Padova 104.85 south
disp15 A13 Bologna-Padova 0.82 north
disp16 A14 Bologna-Taranto 95.60 north
disp17 A1 Milano-Napoli 340.30 north
disp18 A11 Firenze- Pisa Nord 0.15 west
disp19 A11 Firenze- Pisa Nord 81.15 est
disp20 A1 Milano-Napoli 432.07 north
disp21 A1 Milano-Napoli 515.80 south
Phase 2 disp22 A1 Milano-Napoli 516.00 north
disp23 A16 Napoli- Canosa 2.58 est
disp24 A16 Napoli- Canosa 53.45 west
disp25 A14 Bologna-Taranto 498.15 north
disp26 A14 Bologna-Taranto 468.20 south
disp27 A14 Bologna-Taranto 411.77 north
disp28 A14 Bologna-Taranto 359.75 south
disp29 A14 Bologna-Taranto 293.57 north
disp30 A16 Napoli- Canosa 131.60 est
disp31 A16 Napoli- Canosa 158.46 west
disp32 A14 Bologna-Taranto 664.00 south
disp33 A14 Bologna-Taranto 703.56 north
disp34 A14 Bologna-Taranto 555.43 south
Source: Autostrade per lItalia
30
Table 6: Relation between day of the week and price change
Day Obs. Price Increase Price Reduction No Price Change
N. mean N. mean N.
Monday 1320 113 0.014 46 -0.008 1161
(9%) (3%) (88%)
Tuesday 1320 149 0.010 110 -0.006 1061
(11%) (8%) (80%)
Wednesday 1320 185 0.008 112 -0.007 1023
(14%) (8%) (78%)
Thursday 1320 152 0.011 49 -0.006 1119
(12%) (4%) (85%)
Friday 1320 81 0.013 68 -0.008 1171
(6%) (5%) (89%)
Saturday 1320 124 0.011 141 -0.007 1055
(9%) (11%) (80%)
Sunday 1320 19 0.007 98 -0.014 1203
(1%) (7%) (91%)
total 9240 823 624 7793
(9%) (7%) (84%)
31
Table 7: Price Dierentials
Source Fuel Service Road Range (e cents) Avg. Save (e) Period
(Max-Min) 40 liters rell
ISTAT
1
Unleaded self non-toll 4.2 1.69 10/07-10/08
ISTAT Diesel self non-toll 4.7 1.88 10/07-10/08
ISTAT Unleaded full non-toll 3.5 1.40 10/07-10/08
ISTAT Diesel full non-toll 3.9 1.55 10/07-10/08
FIGISC
2
Unleaded full non-toll 11.3 4.53 19/02/2008
FIGISC Diesel full non-toll 13.1 5.22 19/02/2008
PrezziBenzina
3
Unleaded self non-toll 5.4 2.16 20/11/2008
PrezziBenzina Diesel self non-toll 6.1 2.44 20/11/2008
A22
4
Unleaded self toll 1.6 0.64 28/04/2008
A22 Diesel self toll 1.6 0.64 28/04/2008
A22 Unleaded full toll 1.7 0.68 28/04/2008
A22 Diesel full toll 1.6 0.64 28/04/2008
ASPI subsample
5
Unleaded self toll 1.9 0.77 06/02/2008
ASPI subsample Diesel self toll 1.7 0.66 06/02/2008
1
ISTAT (Italian National Institute of Statistics) monthly price collection;
2
FIGISC ( Italian Federation of the Relling Service Station Managers) La distribuzione carburanti: libro bianco sulla concorrenza, 3/4/2008
3
Website were registered users can self report fuel prices (www.prezzibenzina.it)
4
Pay toll motorway (313 Km long) managed by Autostrada del Brennero SpA. Publish weekly price report on www.autobrennero.it
5
Sub sample of ASPI service stations
Table 8: Price Comparison Motorway Price Vs Normal Street
Unleaded Diesel
Normal road
1
Motorway
2
Normal road Motorway
Province mean s.e. mean s.e. mean s.e. mean s.e.
Bologna 1.177 0.005 1.194 0.004 1.171 0.005 1.188 0.004
Firenze 1.174 0.005 1.190 0.007 1.166 0.005 1.183 0.008
Milano 1.166 0.003 1.193 0.004 1.159 0.003 1.186 0.004
Napoli 1.235 0.009 1.240 0.005 1.209 0.014 1.208 0.006
Roma 1.173 0.002 1.188 0.004 1.165 0.002 1.178 0.003
1
Mean computed over the twenty cheapest service stations (at the province level).
Data are self reported by users on the website www.prezzibenzina.it
2
Mean computed at the province level for the service stations in our sample operating on the ASPI motorway
32
F
i
g
u
r
e
2
:
S
c
r
e
e
n
s
h
o
t
A
S
P
I
p
r
i
c
e
c
o
m
p
a
r
i
s
o
n
w
e
b
s
i
t
e
(
s
o
u
r
c
e
:
w
w
w
.
a
u
t
o
s
t
r
a
d
e
.
i
t
)
33
Figure 3: Price Comparison Panel (source: www.autostrade.it)
Figure 4: Example of Panel Station Location
Figure 3 shows the design of price comparison panels installed along the
ASPI motorway network. As we can see from the gure, the panel post the
brand and prices (self service unleaded and diesel fuel) of four consecutive
stations. The stations are ranked by distance to the panel (with the closest
being rst) and the cheapest station is highlighted by a green dot next to its
price. ASPI ocials, when enquired about the panel design, reported that
the decision to post only four prices is an outcome of a trade-o between
posting many prices (as to oer more information, but with come physical
constraints) and assuring a minimum level of comparison among brands.
Since on the ASPI network there is a maximum of three consecutive stations
all from the same brand they then adopted the 4 stations design. Both in the
design stage and in the location decision the managers of the gasoline stations
were not involved. They are only responsible for the communication of the
34
prices through the software platform (that serves both the price comparison
website and for the price comparison panel).
Figure 5: Italian toll-motorway network (source: www.autostrade.it)
Figure 6: Selection into treatment and transit variable
35
Figure 7: Matching on Average Transit Variable
36
Figure 8: Matching service stations on opposite side
37
Figure 9: Matching on Average Transit Variable (only outbound Stations)
38