Who Benefits from Information Disclosure? The Case of Retail Gasoline∗ Fernando Luco Department of Economics Texas A&M University April 6, 2017

Abstract This paper uses the sequential implementation of regulation that required gas stations in Chile to post prices on a government website, to study whether disclosure of information increases competition or facilitates coordination among firms. Using a difference-in-difference approach, I show that margins increased on average by 10 percent following the introduction of the website. However, while margins increased across the whole country, they increased the most in areas with low or non-existent consumer search (low-income areas) and increased the least and even decreased in areas with significant consumer search (high-income areas). JEL Codes: D22, D43, D83, L12, L41. Keywords: Information disclosure, competition, coordination, search costs.

In the last decades, several countries have implemented mandatory information disclosure policies meant to allow consumers to compare prices and product attributes across sellers and to give firms the incentives to improve the quality of their products and intensify ∗

This paper is a revised version of the third chapter of my Ph.D. dissertation at Northwestern University

and was previously circulated as “Mandatory Price Disclosure and Competition.” I am especially grateful to Igal Hendel, Rob Porter, and Aviv Nevo for their guidance and generosity. I also thank Laura Doval, ´ Ren´e Leal, Jorge Lemus, Guillermo Marshall, Alvaro Parra, Steve Puller, and attendants at the IIOC 2015, Hal White 2015 Antitrust Conference, and the NBER Winter IO meeting 2016 for helpful comments and suggestions; and Sergio Campam´ a for sharing his data. All errors are my own. E-mail: [email protected].

1

competition. Examples of these interventions include markets such as ready-mixed concrete, retail gasoline, supermarkets, alcoholic beverages, food snacks, and restaurants, in countries such as Australia, Chile, Denmark, Israel, Italy, South Korea, and the United States. What governments often do not have in mind, however, is that information disclosure allows firms to monitor their rivals’ actions and, potentially, it may facilitate coordination. In the end, whether disclosure intensifies competition or facilitates coordination crucially depends on who uses the newly disclosed information. In this paper, I study how competition changed in the Chilean retail-gasoline industry after the Chilean government passed regulation in February of 2012, requiring gas stations to post their prices on a government website and to keep prices updated as they changed at the pump.1 The website was introduced on March 1st, 2012, and during its first month, it only published information for the region where the capital (Santiago) is located. The rest of the country was added sequentially to the website in the following months according to a schedule set by the government. By July of 2012, the website contained information from the whole country. The Chilean government introduced the website for two reasons.2 First, the website would allow the Comisi´ on Nacional de Energ´ıa (CNE, National Energy Commission) to have price information in real time that would be used to evaluate the performance of the market and to forecast prices. Second, it would allow consumers to access geocoded price information for all gas stations in the country, as well as information on stations’ characteristics. In this setting, information disclosure may have both pro- and anti-competitive effects. On the one hand, disclosure may intensify competition if consumers benefit from lower search costs. Furthermore, competition may increase as well if stations use the website to differentiate from each other by publishing information about the services they offer. On the other hand, if stations have easy access to price information, disclosure may facilitate coordination. Furthermore, if there is heterogeneity in the magnitude of the effects across 1

The website is www.bencinaenlinea.cl. Gas stations have to update their prices with a maximum lag of

15 minutes between the time the price was changed at the station and the time it is reported on the website. 2 See “Informaci´ on sobre el Sistema de Informaci´ on en l´ınea de Precios de Combustibles” at http://www.bencinaenlinea.cl/web2/normativa.php.

2

locations, disclosure policies may have important distributional consequences. For these reasons, I study how information disclosure affected competition in the Chilean retail-gasoline industry and who benefited from it, asking two questions. First, what is the impact of disclosure on a firm’s margin and on margin dispersion? Second, how does the impact of disclosure on local market outcomes vary with the intensity of consumer search? To answer these questions, I combine several datasets that allow me to study how margins evolved between January 2010 and December 2013 across six of the largest cities in Chile.3 I identify the impact of disclosure on the intensity of competition exploiting the sequential implementation of the disclosure mechanism. The results provide strong evidence in favor of disclosure softening competition, as margins increased on average by 10 percent across all cities, with no changes in margin dispersion. Furthermore, I show that this increase in margins is not explained by alternative hypotheses such as an increase in differentiation, or changes in brand- or city-specific pricing behavior. A possible explanation for the increase in margins that followed the implementation of disclosure is the little success of the government in nudging consumers to search for lower prices. Indeed, the median number of visits per day that the website received over the sample period was 947. In this setting, both the lack of consumer search and the possibility of observing rivals’ prices online, allows firms to sustain higher payoffs by using the website as a coordination and monitoring mechanism (see Section 2 for an example). To study whether the impact of disclosure on margins is heterogeneous across locations and how it is related to consumer search, I use data on the locations of consumers when they search for prices through a smartphone app, to create measures of search intensity in the neighborhood of gas stations. Using these measures of search intensity, I find that while margins increased across the whole country, they increased the most in areas with low or non-existent consumer search (low-income areas), while they increased the least, and even decreased, in areas with high search intensity (high-income areas). This suggests that when consumer search intensity is low, the supply-side response to disclosure (coordination) 3

Santiago (the capital), Valpara´ıso, Rancagua, Talca, Concepci´ on, and Punta Arenas. These cities are the

only ones for which data is available for the period before the website was introduced and they represented 59.3 percent of the population of the country in 2012. The area of Santiago (or Gran Santiago) represented 41 percent of the total population of the country in 2012.

3

dominates, but when consumer search intensity increases, it has the potential to overcome price coordination and increase competition. Finally, the results also show that the relationship between margin dispersion and consumer search follows an inverted U, with dispersion increasing with search for low levels of search intensity and decreasing for higher levels of search intensity as more consumers become informed. In this context, this paper makes two main contributions. First, I show that whether the demand- or supply-side response to disclosure dominates, crucially depends on the intensity of local consumer search. Second, I show that disclosure policies may have important distributional consequences depending on who has access to the newly disclosed information. That is, while most of the literature has focused on the average impact of disclosure on market outcomes, having data on local consumer-search intensity allows me to study how the impact of disclosure on market outcomes varies with the intensity of consumer search. Though the findings reported in this paper suggest that disclosure decreased the intensity of competition in the Chilean retail-gasoline industry, the literature has reported mixed findings about the effect of disclosure on market outcomes.4 Albæk et al. (1997) use price data of ready-mixed concrete from Denmark to show that, in the year following the implementation of a disclosure policy that informed the prices of some types of concrete, average prices of the informed categories increased by 15–20 percent, while prices of uninformed categories only increased by 1–2 percent. Dranove et al. (2003) study how health-care report cards affected patient outcomes and matching between patients and providers. They show that following the introduction of report cards, healthier patients received a larger fraction of surgeries, while sicker patients showed worse health outcomes. Borenstein (1998) describes the Airline Tariff Publishing Case of the 1990s and discusses how rapid communication may lead to price coordination. Finally, in similar environments to the one described in this paper, Jang (2014) and Hong (2014) study how the smartphone penetration rate is related to price dispersion and markups in the South Korean gasoline industry; and Rossi and Chintagunta (2015) study how the introduction of signs containing price information from nearby stations affected competition in the Italian motorway. Jang (2014) 4

For a comprehensive review focused on health care, education, and finance, see Dranove and Jin (2010)

and the references therein.

4

uses data on daily prices and volume to show that on average both price dispersion and markups slightly increased with smartphone penetration. Hong (2014) focuses on a slightly different time period and finds that on average both price dispersion and margins decreased with smartphone penetration. Rossi and Chintagunta (2015) show that on average margins decreased by 20 percent following the implementation of the signs. A second branch of the literature has focused on how consumers react to disclosure, either mandatory or voluntary. Mathios (2000) studies the differential impact of voluntary versus mandatory information disclosure in the salad-dressing market. Jin and Leslie (2003) study how displaying hygiene quality cards in restaurants windows, affected consumers and firms. Bollinger et al. (2011) study how calorie-posting by Starbucks affected both consumer behavior and profits. All of them find responses consistent with consumer changing their behavior to choosing higher quality products. This paper is also related to the literature that studies pricing in the retail-gasoline industry. Borenstein and Shepard (1996) study how expectations about future profits and costs affect current margins. Wang (2009) studies how regulation that determines when and how frequently gas stations are allowed to change their prices affects their pricing strategies. Lewis (2008) studies how price dispersion depends on both differentiation and local competition. Lewis (2011) studies how the path of prices affects search intensity and future price responses to changes in costs. Clark and Houde (2014) study pricing behavior when collusion collapsed in Qu´ebec’s retail-gasoline market. Byrne and de Roos (2016) show that after gas stations in Perth, Australia, had to start informing authorities about the prices they would charge in the following day, it took them twelve years to learn how to use the mechanism to coordinate as an effective cartel. Finally, Lemus and Luco (2017) study how, after the implementation of disclosure in the Chilean retail-gasoline industry, stations in some markets became price leaders and this price leadership led to higher margins in those markets relative to markets with no leaders. Finally, this paper is also related to the empirical literature on search. In particular, this paper explains the observed changes in the intensity of competition as a consequence of (the lack of) consumer search. In this setting, this paper is related to Brown and Goolsbee (2002) that show that while prices of insurance policies that were available online decreased

5

significantly with the introduction of the Internet, prices of policies that were not available online were not affected by the introduction of the Internet. They also show that the relationship between price dispersion and consumer search followed an inverted U. Baye et al. (2004) show that price dispersion is an equilibrium phenomenon that does not disappear over time. Finally, Sorensen (2000) shows that both prices and dispersion decrease with the frequency of purchase. The paper continuous as follows. Section 1 presents the data sources used in the paper and describes the industry. Section 2 presents an example that illustrates how information disclosure may affect competition. Section 3 studies how the level and dispersion of margins changed following the implementation of information disclosure. Section 4 studies how consumer-search behavior affected competition during the disclosure period. Finally, Section 5 presents the conclusions of the paper.

1

Industry and Data

1.1

The Chilean Gasoline Industry

Chile is a net importer of oil and most of the oil it imports is imported by a state-owned company called ENAP (85% in 2012). The rest is directly imported by distributors. Because ENAP competes with international producers, its pricing policy is determined by international prices. Since 2009, ENAP has offered three different prices that vary according to the price of oil in the Mexican Gulf and the type of purchase, but not the volume purchased. The first price corresponds to purchases made more than 45 days prior to expected delivery. A second price is offered to those willing to sign a long-term contract and it consists on a discount over the first price, regardless of the volume purchased. Finally, for delivery within 45 days, ENAP charges a spot price that varies with international prices and inventory.5 The Chilean gasoline industry has three levels. The first level is the refinery stage, with oil either refined into gasoline by ENAP or distributors importing directly from international markets. The second level corresponds to distributors. As of 2012, there where four main 5

See ENAP (2010), ENAP’s CEO’s presentation to the Chilean Competition Court in September 2010,

pages 2 and 3.

6

distributor companies in Chile: Copec, Petrobras, Shell, and Terpel. Distributors can sell in the industrial market or in the retail market through gas stations. The retail market corresponds to the third level of the industry and it is the object of study in this paper. Gas stations can be branded or unbranded (independent). Branded stations can either be owned by the distributor or independently owned. In either case, a branded station has an exclusivity contract with the distributor to sell gasoline of that brand. Finally, because most fuel products are provided by ENAP, gasoline is a homogeneous product, so stations and brands compete on service and location. In Chile, most stations sell gasoline of 93, 95, and 97 octanes, and diesel. In this paper I focus on gasoline of 93 octanes because it accounted for 53 percent of all gasoline sold in the country 2013.

1.2

The Policy Intervention

On February 1st, 2012, CNE passed “Resoluci´on No. 60” (Decree N. 60) creating the “Sistema de Informaci´on en L´ınea de Precios de los Combustibles en Estaciones de Servicios” (the Online Price Information System for Fuel Products sold in Gas Stations), announcing the creation of the website www.bencinaenlinea.cl and establishing the way in which the intervention would be rolled out.6 The decree established that the system would be rolled out sequentially across the country during a five-month period, with different groups of regions entering the system each month.7 Figure 1 shows the order in which the different regions entered the system. The first region to enter was the Metropolitan Region, that includes the capital, Santiago. It was followed by four regions entering the system in April: Coquimbo, Valpara´ıso, Libertador Bernardo O’Higgins, and Maule. In May, the two regions located at the extreme south were added: Regi´ on de Ays´en del General Carlos Ib´an ˜ez del Campo and Regi´ on de Magallanes y de la Ant´artica Chilena. In June, four 6 7

See http://www.bencinaenlinea.cl/web2/archivos/RE CNE N60 Sistema de Precios en Linea DO.pdf. Chile is administratively divided in fifteen regions, with each region further divided in provinces and

provinces divided into municipalities. In most cases, a city and a municipality overlap perfectly. The only exception is Santiago that is composed by 37 municipalities, 26 of which are urban and the rest rural. For this reason, when referring to cities other than Santiago, distinguishing between cities and municipalities is irrelevant.

7

regions of the central-south part of the country were added: B´ıo B´ıo, Araucan´ıa, Los Lagos, and Los R´ıos. In July, the northern part of the country was added to the system. The order of the intervention was decided by the CNE and, to the best of my knowledge, only technical considerations influenced it. Finally, the government decided to enforce the policy through another agency (Superintendencia de Electricidad y Combustibles) that, among other things, has to enforce the fifteen-minute window-limit that stations have to update prices in the website after they change prices at the pump.

1.3

Data

This paper combines six sources of data. The first dataset corresponds to data published on the website described in the previous section. This dataset was provided by the CNE and it contains information about stations’ characteristics and locations. The second dataset is a survey that contains price information for both periods, before and after the introduction of price disclosure. This survey was done by the Servicio Nacional del Consumidor (SERNAC, Consumer National Service). The dataset contains weekly price information for 43 stations in Santiago (almost 10 percent of the gas stations in Santiago and 19 percent of the gas stations in those municipalities in Santiago that are covered by the SERNAC dataset) and monthly price information for stations in other cities of the country (covering between 50 and 81 percent of all stations in those cities). In its original version, the dataset covers the period 2005–2013. I focus on the period between January 2010 and December 2013 for two reasons. First, to minimize the possibility that changes in competition may be caused by changes in market structure, as I cannot control for changes in market structure before the introduction of the website, and only imperfectly since then. Second, because the stations included in the SERNAC sample before 2010 varied significantly. The third dataset is generated by the CNE—but using the stations in the SERNAC survey (for both weekly and monthly samples). In this dataset, the CNE publishes the average margin for the stations in the SERNAC survey, for stations in Santiago and five other cities. In this dataset, margins are measured as the difference between the retail price

8

(at the station level) and the sum of the wholesale price (refinery price) and taxes.8 Because the wholesale price is common to all stations in a city, it is possible to recover the sum of distribution costs, the margin of the distributor, and the margin of the station.9 For this reason, in estimation it will be important to control for changes in distribution costs that may lead to changes in our measures of margins. To do this, I include, in all regressions, the interaction between oil prices and the distance from every station to the main pipeline in the country.10 Hence, changes in our measure of margins due to changes in transportation costs will be captured by this interaction. The fourth dataset consists on the location where search requests were executed through a smartphone app. I use this data to define measures of search intensity in the neighborhood of each gas station. Finally, I use two datasets that provide additional information at the municipality level. One dataset is published by the “Subsecretar´ıa de Telecomunicaciones”, the government agency in charge of all telecommunications across the country, and contains information on the number of fixed Internet connections for all municipalities during the period of the analysis.11 The other dataset is the SINIM dataset (“Sistema Nacional de Informaci´on Municipal” or National Municipal Information System), generated by the “Subsecretar´ıa de Desarrollo Regional y Administrativo”, the government agency in charge of overseeing 8

This information is published at http://www.cne.cl/estadisticas/energia/hidrocarburos under the name

“Margen bruto semanal para combustibles en Santiago” (weekly gross margin in Santiago) and “Margen bruto nominal regional” (regional gross margin). 9 Even though large branded companies may have access to different prices as long as they can guarantee demand for gasoline, their prices are indexed in the same way as prices for every other company (and are the same among those who can guarantee their demand). Hence, even though margins could depend on the specific brand, the way margins change as a function of the wholesale price is common to all brands. Furthermore, in this paper I focus on stations that belong to the four biggest distributors, all of which should have the same wholesale price. 10 Every Wednesday, ENAP announces the wholesale price that it will charge starting Thursday at midnight. This price is defined as the price at the receiving end of the main pipeline, located in the municipality of Maipu in the city of Santiago. For this reason, I compute distances from every station to this location. 11 See http://www.subtel.gob.cl/estudios-y-estadisticas/internet/. The website also contains aggregate information on mobile Internet access, though these data are not available at the level of the municipality of origin but for the whole country. For this reason, I cannot use these data in the analysis.

9

municipalities. This dataset provides information on poverty rates, population, and the fraction of the population that lives in rural areas, among other administrative data.12 It is important to note that I do not have access to volume data, so the analysis focuses on how margins changed following disclosure and how these changes are related to station and market characteristics and to consumer-search behavior. Table 1 reports the summary statistics of the station-level data. The table is divided into three panels that differ on the variables of interest. The first panel reports statistics on margins and shows that these were of the order of 70.35 Chilean pesos per liter, or 8.99 percent of the retail price, with significant dispersion. The second panel summarizes station characteristics and shows that 43 percent of stations had a convenience store, almost 5 percent had a pharmacy, 38 percent had public restrooms, 34 percent had a repair shop, 18 percent offered self-service pumps (in addition to full-service), and 91 percent were opened the whole day. I later use this information to study whether changes in margins may be associated with increasing differentiation through the website, as consumers may become aware of these characteristics when searching for prices.13 Finally, the third panel summarizes the number of monthly search requests executed in the neighborhood of each gas station, for different distance thresholds. The table shows that, on average, there is little consumer search, though the standard deviation and the range are large, evidence of significant heterogeneity across locations. Table 2 report statistics at the market level. Because there is not a unique way to define markets, in estimation I follow two approaches. One approach is to define markets using a fix radius around a gas station. This has the benefit of simplicity, but as markets are defined around every gas station, a significant number of stations are considered multiple times. Furthermore, for any radius r, stations that are separated by almost 2r would still be considered as being within the same market. For this reason, I also follow an alternative approach, inspired in graph theory, that uses the concept of cliques. In this case, a market is 12

See http://datos.sinim.gov.cl/. Note, however, that there is little time-series variation in the data. For

this reason, I only include these data in specifications that check the robustness of the results. 13 The website also contains information on two additional variables: whether a station accepts cash and whether it accepts credit or debit cards as means of payment. I do not include these in the analysis as essentially all stations accept them, leaving no meaningful variation left.

10

defined as the largest set of stations that are all within a specific distance of each other. The implementation is more complex than using a fix radius but it has the benefit of minimizing the number of stations that is considered multiple times. To fix ideas, consider the example presented in Figure 2. In this example, all stations are within a two-kilometer radius of the station represented by a black circle. However, the stations represented by squares are further than two kilometers away from the stations represented by triangles. If markets are defined by a radius, all these stations would be in the same market. Furthermore, if a market is defined for any of the stations represented by triangles, the other triangle and the circle would be considered too, in addition to any station within two kilometers of the station of reference (not presented in the figure). Hence, different markets may differ only in whether a single station is added or dropped. However, when using cliques, only two markets are defined, with only one station considered twice. Specifically, one market considers the stations represented by the triangles and the circle, while the other considers the stations represented by the squares and the circle. This is, only the circle is considered twice, making the cliques approach more attractive. For this reason, and because in this application the results do not change significantly, I present the results of the cliques approach in the main text and refer the reader to the Online Appendix for the results that follow from using a fixed radius.14 Table 2 reports statistics for the market-level data, using cliques to define these markets. The first panel reports statistics of market structure and shows that, on average, markets have seven stations, though there is significant dispersion across markets. Also, on average, markets have more Copec stations than stations of any other brand. The second panel focuses on margin dispersion and shows that the range of margins varies between 0 and 63 14

One could follow two other approaches. First, one could define markets exploiting administrative bound-

aries such as municipalities (the smallest administrative unit in Chile). However, as most cities other than Santiago correspond to a single municipality, this approach would result in these cities having a single market with a relatively large number of stations. For this reason, I do not consider this approach. Second, one could follow an approach such as the one in Rossi and Chintagunta (2015), where markets are defined by authorities as the signs that provided price information, did so for four stations at a time. This, however, is not possible in the Chilean case as the disclosure website provides information at the region and municipality level, in addition to providing an interactive map.

11

Chilean pesos per liter, with the average being 5.53. The mean interquartile range is 4.11 and the mean standard deviation of margins is 3.76 Chilean pesos per liter.

1.4

The Cities in the Data

The cities considered in this paper are determined by the data published by the CNE, that provides information for cities in six regions of the country. These cities are regional capitals, meaning that the regional administrative offices are located within them and they concentrate most of the population and services in the region. Indeed, these cities are among the largest of the country, concentrating 59.3 percent of the estimated population of the country for 2012. Table A.1, in the Online Appendix, summarizes demographic information at the municipality level obtained from the SINIM (2016) dataset. The table shows that, on average, these municipalities have slightly less than 200,000 people, the mean poverty rate is 12 percent, and very few people live in rural areas (and none in the City of Santiago). Hence, the cities studied in this paper are large relative to those in the rest of the country.

2

How can Information Disclosure Affect Competition?

In this section, I present a simple model that illustrates how the introduction of a website that publishes price information may affect market outcomes. I do this using a simple framework that builds on Campbell et al. (2005) and Schultz (2005). Campbell et al. (2005) follow both Stahl (1989) and Green and Porter (1984) to show that though in a static context increased transparency may intensify competition, in a dynamic context transparency may facilitate collusion. Schultz (2005) also studies whether increasing transparency makes it easier to sustain collusion, and finds that, while increasing transparency increases the benefits from undercutting a rival during a collusive stage, it also reduces punishment payoffs. In the model, firms that are coordinating their pricing strategies must decide whether to continue in such an agreement or deviate and enjoy higher payoffs until their deviation is observed. In this context, the main implication of the model is that shortening the length of time during which firms can enjoy deviation profits makes deviations less attractive,

12

as while deviation payoffs decrease, punishment payoffs remain unchanged. In turn, this results in firms being able to coordinate at higher prices, which were not sustainable when the time required to observe a deviation was longer. Hence, the set of prices sustainable under coordination is endogenous and depends on both the fraction of informed consumers and the time it takes to observe a deviation from the collusive agreement. Though in practice stations may differentiate from each other in terms of location, brand, and service, among others, in the model that I present here, consumers perceive differentiation based only on location. To keep things simple, I model the problem as one of spatial competition and focus on price competition given location. For simplicity, assume a linear city of length one with two firms, A and B, that compete by choosing prices pA and pB , and have no production costs. I assume that firm A is located at 0 and firm B at 1.15 In presenting the model, I follow the notation in Schultz (2005). To visit a station, consumers must pay a linear transportation cost. In this section, I assume that conditional on distance to a station, consumers value both stations equally. That is, a consumer located at k, who purchases from firm A, receives indirect utility equal to v − pA − tk. Instead, if she purchases from B, her utility is v − pB − t(1 − k). In the Appendix, I present an extension of the model that allows for further differentiation, for example on quality, by assuming that vA > vB . Because the main results are the same in the two models, here I present the simpler version that assumes vA = vB = v and refer the interested reader to the Appendix. I also assume that there are two types of consumers, informed and uninformed. Informed consumers are aware of the prices that firms charge, while uninformed consumers are not. I assume that a consumer is informed with probability φ ∈ (0, 1). In this setting, an informed consumer will be indifferent between purchasing from either firm if v − pA − txI = v −pB −t(1−xI ). Solving for xI results in the familiar expression xI (pA , pB ; t) =

pB −pA 2t

+ 21 ,

where t corresponds to the cost incurred in transportation. Uninformed consumers have expectations about firms’ prices. In this case, the expected indirect utility from purchasing from firm A is v − peA − txU , while if purchasing from B, it is v − peB − t(1 − xU ). Hence, the uninformed consumer who is indifferent between purchasing 15

Nothing changes when assuming firms are located at 0 < a < b < 1.

13

from A or B is given by xU (peA , peB ; t) =

peB −peA 2t

+ 12 . Following Schultz (2005), I assume

expectations to be rational. In this setting, the equilibrium of the static game depends on the level of differentiation. Schultz (2005) shows that, for low levels of differentiation, the game may not have an equilibrium in pure strategies but rather in mixed strategies. On the other hand, for higher levels of differentiation, the static game does have an equilibrium in pure strategies. For the purposes of this paper, studying the equilibrium in pure strategies is enough to show how information disclosure affects payoffs. For this reason, in what follows I focus on the equilibrium in pure strategies. In the context of the dynamic game proposed here, the equilibrium of the static game can be used to characterize the punishment stage. In the collusive stage, firms will charge the same prices, so price dispersion is not observed as long as firms are symmetric in terms of locations. As the model in the Appendix shows, this changes if vA > vB , and price dispersion arises in the collusive equilibrium. The other results remain the same. When differentiation arises only because of the distance from a consumer to the station, if there is no coordination and serving all consumers is profitable, one-period best-responses   are given by p∗i (p−i ; t, φ) = 12 p−i + φt , for i ∈ {A, B}. Letting N denote Nash outcomes, one-period Nash-equilibrium prices are given by pN i =

πiN =

t φ

and profits are then given by

t 16 2φ .

Consider now the case in which firms coordinate to charge prices above the Nashequilibrium prices. Firms would, if possible, coordinate on monopoly prices. This, however may not always be possible. Indeed, costly verification of deviations may result in firms charging prices below those of a monopoly but above those of the Nash equilibrium. Without loss of generality, assume the coordination price is pC , such that pC > pN . In this context, profits under coordination are given by πiC =

pC 2 .

It remains to be shown whether coordination is at all possible in this context. To do this, we need to determine the optimal deviation and, given this deviation, we need to check 16

To see this, note that each firm will receive half of the uninformed consumers regardless of the prices it

charges. Regarding informed consumers, firm A will receive all consumers located to the left of the indifferent consumer while firm B receives those located to the right.

14

if coordination is sustainable. For this reason, assume that firm B charges pC . What is the D C optimal deviation for firm A? Firm A chooses pD A such that pA = arg maxp p·D(p, p ; t, φ). 2 (φpC +t) . The solution to this problem implies that deviation profits are given by πiD = 8tφ

Finally, if a firm decides to deviate, it will do so until the deviation is observed by its rival, because, by assumption, deviations are always observed—but with a lag. In this context, coordination is sustainable if Z

|

0

∞

pC −rz e dz 2 {z }

Payoffs under coordination

≥

Z

|0

z∗


2 Z ∞ φpC + t) −rz t −rz e dz + e dz , 8tφ z ∗ 2φ {z } | {z }

Payoffs under deviation

Punishment payoffs

where z ∗ represents the instant at which a deviation, that started at z = 0, is observed.17 It is explicit in this last expression (an incentive-compatibility constraint, IC) that the punishment strategy is to play the Nash-equilibrium outcome. It is well known that other strategies, such a stick-and-carrot strategy, may also be employed. However, I focus on this simpler strategy to keep the presentation clear. We can rewrite the IC constraint as φ2 (pC )2 (1 − e−rz ) − 2φpC t(1 + e−rz ) + t2 (1 − e−rz ) + 4te−rz ≤ 0, ∗

∗

∗

∗

which defines all prices that can be sustained under coordination, as a function of the interest rate r, transportation cost t, and, more importantly, both the fraction of informed consumers, φ, and the length of time that it takes to observe a deviation, z ∗ . Figure 3 shows how the set of sustainable prices changes as time to detection changes, for different values of the fraction of informed consumers and transportation costs. The figure shows that, regardless of the parameter values, as time to detection increases, the set of prices that can be sustained under coordination is a subset of the set of prices that was sustainable under shorter time to detection. More importantly, the maximum sustainable price under coordination is a decreasing function of time to detection. This is important because there are many prices that could be sustained under coordination for any given time to detection, so it is natural to focus on the highest feasible one. That this price decreases 17

The same model could be used in discrete time, assuming that the website reduces the number of periods

it takes to observe a deviation. However, it is easier to use a continuous-time framework to show that the same results hold for any arbitrary length of time.

15

with time to detection means that the introduction of the website, conditional on the fraction of informed consumers not changing, facilitates coordination among firms. In addition, the figure also shows that the set of prices that can be sustained under coordination gets smaller as the fraction of informed consumers increases. In this sense, the model suggests that there is a smooth transition going from (almost) everybody being uninformed to (almost) everybody being informed. If everybody is informed, though, only the transportation cost and the time to detection determine the set of feasible prices under coordination (under the assumption that the market is covered). Finally, as firms become less differentiated (i.e., as t decreases), the maximum feasible price under coordination decreases and the new set of sustainable prices turns to be a subset of the original set. Overall, both the example presented here and the one in the Appendix provide the foundation for a strong intuition. That is, payoffs under coordination depend on both the fraction of consumers who are informed about prices (i.e., consumers who actively search) and the time it takes to detect a deviation by one of the players. For a fixed time to detection, a larger fraction of informed consumers intensifies price competition, a consequence of a demand-side response to information disclosure. At the same time, for a fixed fraction of informed consumers, a shorter time to detecting a deviation allows firms to coordinate on higher prices. Hence, whether information disclosure intensifies competition or facilitates coordination depends on whether the effect on competition induced by the demand-side response is larger or smaller than the effect induced by the supply-side response.

3

Information Disclosure and the Intensity of Competition

I now turn to quantifying how mandatory price disclosure impacted competition in the Chilean retail-gasoline industry. In this section, I empirically examine whether the overall effect of disclosure was to increase or decrease the intensity of competition. Section 4 studies whether there is heterogeneity in responses across markets and how this heterogeneity is related to consumer search. To estimate the overall effect of disclosure on competition, I estimate regressions of the

16

form Yit = β0 + β1 1{Website operative}it + Xit′ γ + ξi + ηt + ǫit ,

(1)

where Yit corresponds to the outcome of interest (the natural logarithm of inflation-adjusted margins and measures of margin dispersion), the indicator function takes the value of one if station i is located in an area in which the website is operative at time t and zero otherwise, Xit corresponds to time-varying station characteristics (such as the interaction between distance to the main pipeline and oil prices), and ξi corresponds to unobserved station characteristics, which I assume constant over time and I control for by means of station fixed effects. Finally, ηt corresponds to time fixed effects that allow to take country-wide shocks, that affect the whole industry, into account. Regarding identification, it is clear that simple pre- and post-analysis at the city level could confound a number of factors, which would prevent me from interpreting the effects as representing changes in competition due to the interaction between consumer search and the supply-side pricing. For this reason, I exploit the roll-out of the policy and use a difference-in-difference approach that allows me to control for potentially confounding effects. Furthermore, in most specifications I include a rich set of station or market fixed effects, though in some specifications I drop these fixed effects and introduce either station or market characteristics to study whether product differentiation through the release of information in the website may explain the observed changes in margins. Finally, all specifications control for changes in transportation costs by including the interaction between oil prices and the distance from a station to the main pipeline. Because the policy was implemented across the whole country during a five-month window, it is important to clarify two aspects of the empirical approach that I follow. First, because of the roll-out period, the number of cities in the control group decreases from month to month as new cities are added to the disclosure mechanism. This feature is not related to the sample used in the analysis but to the nature of the policy design, as the whole country entered the disclosure mechanism during this five-month period. Second, because the policy was implemented by geographic areas that group different regions of the country, a large number of stations is simultaneously affected by the policy. In this setting, considering these stations as independent from each other, even if they are in different regions and 17

separated by hundreds of kilometers, could result in estimating incorrect standard errors. For this reason, all standard errors are clustered at the “area of intervention” level, where an area is the group of regions that entered the system at the same time (see Figure 1). This approach is a conservative one in that stations that are physically separated by hundreds of kilometers are included in the same cluster, resulting in the largest estimated standard errors.18 The drawback of following this conservative approach is that the number of clusters is small. For this reason, I follow Cameron et al. (2008) and also use the Wild bootstrap to correct for the small number of clusters.19 Hence, all output tables report standard errors clustered at the area of intervention level in parenthesis, while a subset of regressions also report the p-value associated with the Wild bootstrap approach in square brackets.

3.1

Impact of Disclosure on Station’s Margins

I start studying how disclosure affected margins by estimating different specifications of Equation 1. In all specifications, the dependent variable is the natural logarithm of margins but the covariates differ across specifications. To include Santiago with the same data frequency as the other cities, I only use the first observation of every month for every station in Santiago.20 The results are reported in Table 3. The difference across specifications is either the inclusion of different time trends or the sample considered. In all cases, standard errors are clustered at the “area of intervention” level. Columns 3 and 4 also report the p-value associated with the Wild Bootstrap. Table 3, column 1, includes the disclosure dummy and the controls described above. It does not, however, include any time trend. In this case, the results show that disclosure increased margins by 7 percent, which is significant both in statistical and economic terms. Table 3, columns 2 and 3, replicate column 1 but include different time trends. Column 2 uses a common trend across all regions, while column 3 allows the trend to be different in 18 19

An alternative would be, for example, to cluster at the region level. All implementations of the Wild bootstrap use the cgmwildboot command in Stata using 1,000 repli-

cations. 20 Lemus and Luco (2017) show that in Chile 88 percent of gas stations change prices once per week. Hence, though the frequency of the data may not be ideal, the loss is not as important as it would be if the object of study was the retail-gasoline market in, for instance, the United States, where stations may change prices several times in a day.

18

each region of the country. The results show that including either trend makes the estimated coefficient on disclosure to increase to 10 percent. Finally, column 3 also reports the p-value associated with the Wild bootstrap approach and shows that the significance of the results is robust to this specification. The results reported above show that the implementation of disclosure is associated with a significant increase in margins that is consistent with a decrease in the intensity of competition. At this point, however, it is not possible to argue whether these results are explained by demand- or supply-side responses to disclosure. However, before discussing the potential mechanisms behind the observed changes in margins, I explore a number of robustness checks and discuss potential threats to identification and how these are considered in estimation. I then return to studying the mechanism behind the observed changes in margins. First, it is crucial to discuss where identification is coming from in this application. Indeed, because the policy was rolled out over a five-month period, by August 2012 the whole country was already under the disclosure policy. Hence, even though there may be a lag between the moment the website was operative and when both consumers and firms started using it, most of the variation in the data should come from the months during which the system was rolled out. In other words, the variation in treatment would not allow to identify changes in competition caused by disclosure from, for example, an hypothetical decrease in the elasticity of demand that could take place in 2013. For this reason, in Table 3, column 4 replicates column 3 but it drops all observations following August 2012. In this case, we find that disclosure increased margins by 9 percent. This is of particular importance as it clearly shows where identification is coming from and provides support for the empirical approach followed. In other words, though the additional data allows to fully take into account any lag in learning about the policy and how to use the website, column 4 shows that most of the effect was already present by a month after the whole country was under the new policy, allowing for clean identification of the impact of information disclosure on competition.21 21

Dropping observations following July 2012 rather than August 2012 results in an estimated coefficient

of the disclosure dummy equal to 8.2 percent. I included August in specification 4 to consider those stations

19

Second, having established that the results are robust to eliminating a significant fraction of the data, I now turn to consider threats to identification and how these are addressed in estimation. I start noting that the measure of margins used in this paper is the difference between the retail and wholesale cost at the city level. This means that margins are equal to the sum of the margins of the distributor, the station, and distribution costs. Hence, variation in any of these would result in observed variation in our measure of margins. Though there is no reason to believe that margins of the distributor may have changed during this time period, distribution costs vary with fuel prices. Hence, even if distribution margins are constant over the implementation of the policy, changes in distribution costs may lead to overestimating the impact of disclosure on competition if these changes are not considered in estimation. For this reason, all specifications reported in this paper, included those reported above, control for distribution costs by including the interaction between oil prices and the distance from a station to the main refinery of the country as a covariate. In this way, under the assumption that distribution margins did not change over the implementation of the policy, any observed change in our measure of margins is associated with changes in margins at the station level. Another threat to identification is that both fuel prices are affected by seasonality and that, at the same time, stations differ in observable and unobservable attributes. In estimation this is taken into account with a rich set of time and station fixed effects that take into account shocks that may affect the whole country as well as time-invariant station characteristics. To study how robust the main specification is to the inclusion of these regressors, Table 4 reports the estimated coefficients of a number of regressions that drop some of these controls. The table replicates, in column 5, the preferred specification of Table 3 (column 3), to facilitate comparison. The results show that while the inclusion of time fixed effects plays a significant role in reducing the standard errors of the estimated coefficients, it does not affect their magnitude. Indeed, across all specifications that do not include trends, disclosure is associated with a 7 percent increase in margins. As shown above, including the trends increases the estimated coefficient to 10 percent, showing that that did report prices in July but not necessarily in the first week of the month in which they entered the system.

20

the impact of disclosure on competition is robust across specifications. Finally, to study whether the results are robust to the cities considered in the analysis, Table 5 replicates the specification in column 3 of Table 3, but dropping one city at a time. The results show that, though the point estimates do change when the cities under consideration change, the main finding remains robust. This is, regardless of which city is dropped from the analysis, margins increased between 8.8 and 12.8 percent. After identifying the average treatment effect of the website, I now explore possible demand- and supply-side mechanisms. Here I consider three possible mechanisms in addition to gas stations using the website to soften competition. First, gas stations may use the website to reveal information about the services they offer. This would allow gas stations to further differentiate from each other by providing consumers with information about station characteristics. Second, though I have already shown that the effect of disclosure on competition is not determined by what may have happened in a single region of the country, I study whether the findings are explained by changes in the pricing behavior of a single brand of gas stations, across the whole country. Finally, I consider the case that a merger that took place in 2013 may explain the observed changes in margins (though column 4 in Table 3 suggests this is not the case). For this reason, I now turn to study if any of this alternatives provides a better explanation of the data than gas stations using the website to soften competition. I start studying whether gas stations used the website to further differentiating from each other. This would happen if consumers become aware, through the website, of services that are offered by some stations and not by others. To take this into account, Table 6 replicates column 3 of Table 3 but it drops the station fixed effects and introduces station characteristics as regressors. Column 1 shows that none of the coefficients on characteristics is significant, neither before nor after disclosure, which is evidence against increasing differentiation through the disclosure mechanism. The results do not change when brand fixed effects are added (column 2).22 I now turn to studying how disclosure is associated with changes in margins across 22

An implicit assumption here is that gas stations did not start to offer some services because they could

be advertised through the website, but that station characteristics are fixed over time.

21

brands. The idea is that if the observed changes can be explained by a change in pricing behavior by a single firm, it would be evidence against increasing coordination through the disclosure mechanism. On the other hand, if changes are common across brands, then increasing coordination seems more likely. The results are reported in the first three columns of Table 7. Across these specifications, the omitted category corresponds to Copec stations. In the first two specifications, I replace station fixed effects with brand dummies. In addition, column 2 adds the interaction between brand dummies and the disclosure dummy as regressors. Column 1 shows that dropping the station fixed effects and introducing the brand dummies has no impact on the coefficient of disclosure. It also shows that margins are similar across brands. The estimates in column 2, however, are more interesting. Indeed, the estimated coefficients reported in column 2 show that margins were similar across brands not only before disclosure but also afterwards. This is, margins change similarly across all brands after disclosure was implemented and the significance of this finding remains unchanged when using the Wild bootstrap. Finally, column 3 brings back the station fixed effects and drops the brand dummies, while keeping the interaction between the brand dummies and the disclosure dummy. The results are similar to the ones in the previous columns: margins increased significantly and similarly across all brands. This suggests that the decrease in the intensity of competition that followed the introduction of the website was homogeneous across brands, consistent with the website facilitating tacit coordination. I finish this section addressing another potential mechanism that could explain the observed changes in margins. This is that in 2013, the Chilean Supreme Court ruled against a previous decision of the Chilean Tribunal de Defensa de la Libre Competencia (the Chilean Competition Court) and authorized, with a number of remedies, a merger between Shell and Terpel. The ruling by the Supreme Court took place in January of 2013 while the merger took place in June of the same year. I take this into account in a number of ways. First, in columns 4 and 5 of Table 7, I replicate the specification in column 3 of Table 3 but I either add a dummy variable that takes the value of one for stations involved in the merger starting in June of 2013 (column 4) or simply drop the data for periods after the merger took place (column 5). Second, I replicate columns 4 and 5 but I either redefine the dummy to take the value of one starting in January, when the merger was authorized by the Supreme

22

Court (column 6), or simply drop all data following the merger authorization (column 7). The results, reported in the last four columns of Table 7, show that the estimated impact of disclosure on competition remains unchanged across the different specifications. This is not surprising taking in consideration the results reported in column 4 of Table 3. Finally, I take a dynamic difference-in-difference approach to determine how long it took for margins to increase following the introduction of disclosure. To do this, I define t = 0 to be the moment at which disclosure was implemented in each city. Hence, t = 0 happens at different points in time across cities. Then, I created indicator variables that take the value of one for each month around the implementation of disclosure in each city and add these indicators as regressors. The indicators cover the period between five months before the implementation of disclosure to 10 and more months after disclosure was implemented. These are labeled t = −5, −4, . . . , ≥ 10. The omitted category is all months before the first indicator turns on in each city. The results, presented in Figure 4, are both interesting and important. Indeed, the figure shows that margins increased significantly starting in the second month since the implementation and remain stable for the whole period that followed the policy implementation. This is consistent with the estimates reported in column 4 of Table 3. At the same time, because most of the variation that identifies the policy implementation corresponds to the roll-out period, it is not surprising that standard errors increase afterwards. However, the main message of the figure is that, following the implementation of disclosure, margins increased significantly and remained at a higher (stable) level during the post-rollout period. Summarizing, the results presented in this section show that margins increased by 10 percent on average following the introduction of disclosure and that the increase i) is not explained by further differentiation through the website, ii) it was common across brands, iii) not determined by a merger that took place in 2013, iv) it was not limited to a single city, and v) it took place starting in the second month since the policy was implemented. This suggests that the observed increase in margins may have been caused by increasing coordination through the disclosure mechanism.

23

3.2

Impact of Disclosure on Margin Dispersion

Having established that information disclosure increased margins in the Chilean retailgasoline industry, I now turn my attention to studying how it may have affected margin dispersion. Because gasoline is a relatively homogeneous product and gas stations differentiate from each other on attributes such as location and service, studying if dispersion was affected by disclosure, is informative about the overall competitiveness of the industry and the evolution of the entire distribution of margins. For this reason, this section is related to a large body of literature that has studied whether price dispersion is a persistent phenomenon. Among others, Sorensen (2000), Lewis (2008), Lewis (2011), and Chandra and Tappata (2011) have studied how price dispersion in related to consumer search, local competition, and consumer heterogeneity, without actually using search data but testing the implications of different search models. Building on this literature, this section focuses on whether margin dispersion changed with disclosure. Later, Section 4 turns to study how both the level of margins and their dispersion are related to consumer search. In this section, I study how disclosure affected margin dispersion using a similar approach to the one followed above, but using three measures of margin dispersion (the range, the interquartile range, and the standard deviation), because different measures of dispersion may lead to different conclusions (for instance, the range of margins might decrease while the standard deviation might increase). As mentioned above, to study how disclosure affected margin dispersion, it is necessary to define markets. I do this in two ways: defining markets using cliques and using a two-kilometer radius around each gas. Because cliques minimize the number of stations considered more than once and the results do not change when using the alternative definition, I discuss in detail the results that use the cliques definition and refer the reader to the Online Appendix for the results using the two-kilometer radius. Table 8 reports the estimated parameters of different specifications of Equation 1 that use a different measure of dispersion as the dependent variable. Panel A focuses on the range of margins. Column 1 regresses the range of margins on the disclosure dummy, the interaction between oil prices and the average distance from the market to the main pipeline, and market and time fixed effects. The estimates show that that dispersion did not change, on average, with disclosure. Column 2 adds region-specific trends and shows 24

similar results. Column 3 replaces the market fixed effects with region fixed effects and adds the number of stations in the market as regressors. The results show that controlling for the number of stations in the market does not change the estimated coefficient of the disclosure dummy. Column 4 replaces the number of stations with the number of stations of the different brands and shows that the estimated coefficient on the disclosure dummy remains unchanged. Finally, columns 3 and 4 also report the p-values associated with the Wild bootstrap, confirming the previous results. Similar results are reported in panels B and C, that study how the interquartile range of margins and the standard deviation of margins changed, on average, following disclosure. The tables show that none of estimated disclosure dummies is significantly different from zero, suggesting that, on average, margin dispersion did not change following the implementation of information disclosure. This result, together with those reported in the previous section, suggest that the whole distribution of margins shifted to the right following the implementation of disclosure. Finally, Table A.2 in the Online Appendix reports the estimated coefficients associated with the same set of regressions but using a two-kilometer radius around each gas station to define markets. The table shows that regardless of the specification, none of the disclosure dummies is significant, suggesting that margin dispersion did not change, on average. For this reason, the next section turns to study how both the level of margins and their dispersion is related to the intensity of local consumer-search behavior.

4

Consumer Search and Margins

Having established that, on average, margins increased by 10 percent following the introduction of disclosure and that margin dispersion did not change, I now turn to study whether there was heterogeneity in responses to disclosure across locations and, in particular, how this heterogeneity is related to consumer search. This is important because as disclosure was implemented through a website, it is likely that consumers in high-income areas of the country may have easier access to the newly disclosed information than consumers in low-income areas. This suggests that disclosure policies, such as the one studied here, may

25

have important distributional consequences. In this section, I introduce two additional sources of data that allow me to study the heterogeneity of the impact of disclosure on market outcomes across locations. The first one corresponds to the location at which consumers executed search requests using an app. The second one corresponds to demographic information at the municipality level. The analysis is divided in four steps. I start studying how margins at the station level vary with the intensity of consumer search. I show that, though margins increased across the whole country, they increased the least and even decreased in areas with high consumer search. On the other hand, margins increased the most in areas with low consumer search. What makes this finding more interesting is that low-search areas are low-income areas as well, which suggest that there are significant distributional effects associated with disclosure policies. Then I turn my attention to how dispersion of margins was affected by local search behavior and show that, though the overall relationship between margins and search is that of an inverted U, dispersion starts to decrease at levels of search that are not often observed in my dataset. Hence, in most locations considered in this paper, dispersion increases as search increases, only decreasing in areas with relatively high consumer search. Finally, I turn to quantify the impact of disclosure on both consumers and stations and show that, though consumers would benefit from searching more, the magnitudes are relatively small conditional on the observed prices. On the other hand, stations benefit from disclosure significantly. I start studying how the intensity of local-search behavior affected competition in the Chilean retail-gasoline industry. To do this, I start plotting maps of search intensity together with municipality boundaries and their mean household income as reported by the CASEN 2011 survey. Figure 5 presents two of these plots for the City of Santiago in different days, and show that consumer search was mostly concentrated in high-income areas. This suggests that though gas stations may always use the website to monitor rivals’ prices, significant consumer search may only take place if the newly disclosed information is easily accessible. In terms of where search took place, this finding should not be surprising as smartphone penetration is likely to be higher in these areas as well. To further study the relationship between search and margins, I now estimate different

26

specifications that relate margins to income and search behavior.23 First, column 1 in Table 9 reports the estimated coefficients of regressing the natural logarithm of margins on disclosure, the interaction of disclosure and mean household income at the municipality level, and a series of fixed effects (because mean household income does not change in my sample, it cannot be included together with municipality fixed effects but only interacted with the disclosure dummy). The results show that, on average, margins increased by 10 percent, but that they increased the least and even decreased in areas with higher income. Indeed, the results show that a one-standard-deviation increase in income is associated with a 4 percent margin increase, rather than the average 10 percent, while in areas with higher income margins even decreased. This means that though margins increased across the whole country, they increased the most in low-income areas and even decreased in some high-income areas. Column 2 also includes other regressors at the municipality level (the number of fixed Internet connections per 1,000 people, the poverty rate, and the fraction of the population in rural areas) and shows that the main result remains unchanged: margins increased, on average, by 10 percent, but the increase was larger in low-income areas. I now turn to consumer search behavior. I do this regressing the natural logarithm of a station margin on the number of search requests that were executed within a certain radius of that station in the previous month, the square of the number of search requests, income, the square of income, and the same regressors included in the previous regressions. Importantly, I also include as regressors the number of stations within a given distance threshold, as consumers may be more likely to search in areas with more stations (this also implies that station fixed effects cannot be included). Also, because current search behavior is likely to be endogenous (i.e., consumers may search more when prices are higher), I use the lagged number of search requests as regressors. This has the benefit that while lagged search is not affected by current prices, current margins are likely to be affected by past search if consumers visit stations based on past search behavior. The results are reported in Table 10 and in Figure 6, that show that margins decreased almost monotonically with search intensity, consistent with what has been reported in both the theoretical and empirical work on search (see Stahl, 1989; Brown and Goolsbee, 2002, among others). 23

In all these regressions, continuous regressors are standardized to facilitate interpretation.

27

The estimated coefficients imply that a one-standard-deviation increase in the number of requests executed withing 3 kilometers (5 kilometers) of a gas station is associated with 8 percent (11 percent) lower margins. This means that even little consumer search has the potential to intensify price competition, though, as seen above, most search requests took place in high-income areas. We can therefore conclude that price disclosure not only decreased the intensity of competition, but that it did so affecting low-income areas the most. This is probably the case because consumers in low-income areas may not have easy access to the disclosed information, while gas stations do. On the other hand, because consumers in high-income areas do have easy access to price information, search costs decrease and the demand-side response to disclosure is able to counteract the supply-side response to it. Finally, adding other demographics to the regressions does not change the results (see Table A.3 and Figure A.1 in the Online Appendix). I now turn to studying how margin dispersion is related to consumer search. As reported in the previous section, margin dispersion seemed not to have changed, on average, following the introduction of disclosure. However, we have already shown that there was heterogeneity in how margins changed across locations depending on the intensity of consumer-search behavior. It is therefore important to know if this was also the case with margin dispersion. Table 11 reports the estimated coefficients of regressions that mimic column 1 in Table 9, with one difference: the dependent variable is a different measure of margin dispersion in each column. As before, the results show that dispersion did not change, on average, with disclosure. However, dispersion did increase with income. In addition, the findings described above show that that search was more significant in high-income areas as well and resulted in margins increasing the least and even decreasing in these area, which suggests that margin dispersion should have increased in areas with more search. To determine to what extent this is the case, I now replicate Table 10 to study how margin dispersion and consumer search are related. The results, reported in Table 12 and Figure 7, show that margin dispersion increased with search, though with two exceptions, the point estimates are not significant. At the same time, the coefficients on the square of the number of requests are always negative, but not significant. Figure 7 reports the marginal effect of search on margin dispersion for the different distance thresholds. The figure shows that margin

28

dispersion first increases with the number of search requests and later decreases, resulting in an inverted U (see Stahl, 1989; Brown and Goolsbee, 2002; Chandra and Tappata, 2011, among others). Overall, though the confidence intervals on the marginal effects are large, the estimated relationship suggests that at low levels of consumer search, margin dispersion increased as search increased. However, when the number of search requests was large, margin dispersion started to decrease. Finally, essentially identical results are obtained when other demographics are added as regressors (see Table A.5 and Figure A.2 in the Online Appendix). So far, we have shown that margins increased with the implementation of disclosure, but increased the least and even decreased in areas in which consumers engaged in significant search behavior. For this reason, I finish this section quantifying to what extent consumers would benefit from engaging in search. To do this, I compute the expected gain from buying from the cheapest station in a market, where markets are defined using the cliques approach. Here, gains from search are defined as the difference between what a consumer would pay if purchasing at random versus visiting the cheapest gas station in the market. Because the cliques approach results in narrowly defined markets and prices are kept fixed as they are observed in the data,24 the computed gains from search are a lower bound relative to what a consumer would obtain if she were to search across the different markets she would go through in a normal commute (something that the website allows) or if stations were to react to the increase in search behavior. In this setting, the results show that gains from search are, on average $2.9 Chilean pesos per liter. This is equivalent to 15 dollars in savings per year, for a car with a 50-liter tank that is filled once a week. Though this does not seem impressive, it is important to remember that it is a lower bound as consumers are likely to go through multiple markets during their normal commute. Indeed, if one were to consider gains from search at the city level (something that consumers in cities other than Santiago may consider), gains from search increase to $7 pesos per liter, or 35 dollars in savings per year.25 Hence, gains from search exist and, though they are relatively modest, 24 25

This is, I do not allow gas stations to react to the change in consumer behavior. This number excludes the City of Santiago as though consumers are likely to commute across the city

during a normal work day, they are unlikely to cross the city to find the cheapest gas station. When considering the City of Santiago in the analysis, the mean gain from search increases to $8.8 Chilean pesos

29

they are consistent with those reported in other studies in similar settings (i.e., Jang, 2014). On the other hand, a back-of-the-envelope calculation using total volume of gasoline sold in 2013, shows that the average gas station increased its profits by US$27,000 in that year alone, suggesting that gas stations benefit the most from the introduction of the website, as they can use it to monitor the prices charged by their rivals and increase overall payoffs.26 Overall, the findings presented in this paper are consistent with both a supply- and a demand-side response to the implementation of information disclosure, with each of these dominating in different cases. Indeed, in those areas in which there is little or no consumer search—low-income areas—margins increased significantly while dispersion did not change, consistent with firms using the website to keep track of rivals’ prices. However, in areas with significant consumer search—that are high-income areas—margins increased the least and even decreased, while dispersion increased, suggesting that consumer search is strong enough to overcome coordination among firms.

5

Concluding Remarks

This paper studies how the sequential implementation of an information-disclosure mechanism affected consumer and firm behavior in the Chilean retail-gasoline industry, as well as the distributional consequences of information disclosure. In February 2012, the Chilean government mandated all gas stations in the country to post their prices on a government website and to update prices every time they changed at the pump. The system was first implemented in the capital, Santiago, in March 2012, and later, sequentially, throughout the rest of the country. By July 2012, the system was working in the whole country. Because disclosure may be pro- or anti-competitive depending on whether consumer search—the demand-side response to disclosure—or price coordination—the supply-side per liter or 44 dollars in savings per year. 26 This number is computed using data from different sources. Total volume sold in Chile, in cubic meters, is published at http://www.cne.cl/wp-content/uploads/2015/05/Venta mensual combustibles7.xls. This shows that 53 percent of gasoline sold in Chile corresponded to gasoline of 93 octanes. I use this number and the number of gas stations that operated in Chile in 2013 to calculate the average number of liters sold by gas stations that year. Then, I calculated the additional revenues associated with the increase in margins estimated above. Finally, I use the exchange rate of December 30th, 2013 to compute gains in dollars.

30

response—dominates, I focus on two questions that allow me to quantify the effect of disclosure on competition. First, what is the impact of mandatory disclosure on a firm’s margin and on margin dispersion? Second, how do margins and margin dispersion vary with the intensity of consumer search? The results show that disclosure decreased the intensity of competition, though these results vary depending on whether the supply- or the demand-side response dominated in different areas of the country. Indeed, margins increased by 10 percent on average relative to the period before disclosure was implemented, while margin dispersion did not change. However, margins decreased in some high-income areas and increased in low-income areas. At the same time, consumer search was significantly more intensive in high-income areas relative to low-income areas. Hence, when consumer search is significant, it can overcome the decrease in competition that results from firms optimally responding to disclosure. However, if consumer search is not significant, disclosure may result in decreasing the intensity of competition. Furthermore, disclosure may be associated with important distributional consequences, because while customers in high-income areas may have benefited from the implementation of disclosure, customers in low-income areas were hurt by it. Finally, the results also show that the relationship between margin dispersion and consumer search follows an inverted U, with dispersion first increasing in search and later decreasing as more consumers become informed. A critical question is whether the observed change in behavior is a consequence of optimal responses to disclosure or to other confounding factors. Based on a number of robustness checks, such as testing for increasing differentiation, changes in market structure, and changes in firm- or city-specific pricing behavior, this paper concludes that it is unlikely that factors other than disclosure may have generated the results presented here. Regarding disclosure policies in general, this paper shows that mechanisms that increase market transparency may increase competition and benefit consumers only if consumers can easily access and use the disclosed information. If this is not the case, the supply-side response to disclosure (coordination) is likely to dominate and the intensity of competition will decrease. Hence, this paper provides evidence that suggests that policy makers should consider access to the newly disclosed information to be of major importance.

31

References Albæk, Svend, Peter Møllgaard, and Per B. Overgaard, “Government-Assisted Oligopoly Coordination? A Concrete Case,” The Journal of Industrial Economics, 1997, 45 (4), 429–443. Baye, Michael R., John Morgan, and Patrick Scholten, “Price Dispersion in the Small and in the Large: Evidence from an Internet Price Comparison Site,” The Journal of Industrial Economics, 2004, 52 (4), 463–496. Bollinger, Bryan, Phillip Leslie, and Alan Sorensen, “Calorie Posting in Chain Restaurants,” American Economic Journal: Economic Policy, September 2011, 3 (1), 91–128. Borenstein, Severin, “Rapid Communication and Price Fixing: The Airline Tariff Publishing Company Case,” in John E. Kwoka and Lawrence J. White, eds., The Antitrust Revolution: The Role of Economics, 3rd ed., Oxford University Press, 1998. and Andrea Shepard, “Dynamic Pricing in Retail Gasoline Markets,” The RAND Journal of Economics, 1996, 27 (3), pp. 429–451. Brown, Jeffrey R. and Austan Goolsbee, “Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry,” Journal of Political Economy, 2002, 110 (3), 481–507. Byrne, David P and Nicolas de Roos, “Learning to coordinate: A study in retail gasoline,” 2016. Cameron, A Colin, Jonah B Gelbach, and Douglas L Miller, “Bootstrap-based improvements for inference with clustered errors,” The Review of Economics and Statistics, 2008, 90 (3), 414–427. Campbell, Colin, Gautam Ray, and Waleed A. Muhanna, “Search and Collusion in Electronic Markets,” Management Science, 2005, 51 (3), pp. 497–507.

32

Chandra, Ambarish and Mariano Tappata, “Consumer search and dynamic price dispersion: an application to gasoline markets,” The RAND Journal of Economics, 2011, 42 (4), pp. 681–704. Clark, Robert and Jean-Fran¸ cois Houde, “The Effect of Explicit Communication on pricing: Evidence from the Collapse of a Gasoline Cartel,” The Journal of Industrial Economics, 2014, 62 (2), 191–228. Dranove, David and GZ Jin, “Quality disclosure and certification: Theory and practice,” Journal of Economic Literature, December 2010, 48 (4), 935–963. , Daniel Kessler, Mark McClellan, and Mark Satterthwaite, “Is More Information Better? The Effects of “Report Cards” on Health Care Providers,” Journal of Political Economy, 2003, 111 (3), 555–588. ENAP, “Presentaci´ on ante el Tribunal de Defensa de la Libre Competencia,” September 2010. Testimony by Rodrigo Az´ocar Hidalgo. Green, Edward J. and Robert H. Porter, “Noncooperative Collusion under Imperfect Price Information,” Econometrica, 1984, 52 (1), pp. 87–100. Hong, Woo-Hyung, “Do Smartphones Spur Competition? Evidence from the Korean Retail Gasoline Market,” October 2014. Jang, Youngjun, “The Effect of the Internet ad Mobile Search Technologies on Retail Markets: Evidence from the Korean Gasoline Market,” November 2014. Jin, Ginger Zhe and Phillip Leslie, “The Effect of Information on Product Quality: Evidence from Restaurant Hygiene Grade Cards,” The Quarterly Journal of Economic, 2003, 118 (2), 409–451. Lemus, Jorge and Fernando Luco, “The Dynamics of Pricing with Frequent Cost Changes,” March 2017. Mimeo. Lewis, Matthew S, “Price Dispersion and Competition with Differentiated Sellers,” The Journal of Industrial Economics, 2008, LVI (3), 654–679. 33

, “Asymmetric price adjustment and consumer search: An examination of the retail gasoline market,” Journal of Economics & Management Strategy, 2011, 20 (2), 409–449. Mathios, Alan D., “The Impact of Mandatory Disclosure Laws on Product Choices: An Analysis of the Salad Dressing Market,” Journal of Law & Economics, October 2000, XLIII, 651–678. Rossi, Federico and Pradeep K. Chintagunta, “Price Transparency and Retail Prices: Evidence from Fuel Price Signs in the Italian Motorway,” Journal of Marketing Research, 2015, (in press). Schultz, Christian, “Transparency on the consumer side and tacit collusion,” European Economic Review, 2005, 49 (2), 279–297. SINIM, “Sistema Nacional de Informaci´on Municipal,” Technical Report, SUBDERE, Ministerio del Interior 2016. Sorensen, Alan T., “Equilibrium Price Dispersion in Retail Markets for Prescription Drugs,” Journal of Political Economy, August 2000, 108 (4), 833–850. Stahl, Dale O., “Oligopolistic Pricing with Sequential Consumer Search,” The American Economic Review, 1989, 79 (4), pp. 700–712. Wang, Zhongmin, “(Mixed) Strategy in Oligopoly Pricing: Evidence from Gasoline Price Cycles Before and Under a Timing Regulation,” Journal of Political Economy, 2009, 117 (6), pp. 987–1030.

34

Tables and Figures Table 1: Summary statistics: Station-level data Mean

Median

Std. Deviation

Min

Max

Real margins (CLP $/liter)

70.35

73.51

22.75

5.55

139.67

Real margin as % of retail price

8.99

9.35

2.80

0.70

16.53

Margins

Number of observations

5795

Station characteristics (% of stations) Convenience store

42.97

0

49.70

0

1

Pharmacy

4.69

0

21.22

0

1

Public restrooms

37.50

0

48.60

0

1

Repair shop

33.59

0

47.41

0

1

Self-service pumps

17.97

0

38.54

0

1

Open 24 hours

91.41

1

28.14

0

1

1 kilometer

100.93

44

211.53

0

4034

3 kilometers

730.73

282

1402.93

2

10991

5 kilometers

1645.86

486

3175.29

2

22865

Search requests close to a station (within a month)

Number of observations

2279

Note: CLP$ stands for Chilean pesos.

35

Table 2: Summary statistics: Market-level data Mean

Median

Std. Deviation

Min

Max

Number of stations

6.99

7

2.98

2

13

Number of Copec stations

1.14

1

1.17

0

5

Number of Petrobras stations

0.75

1

0.82

0

4

Number of Shell stations

0.94

1

0.93

0

4

Number of Terpel stations

0.46

0

0.70

0

2

Market structure

Number of markets

101

Margins (CLP$ per liter) Range

5.53

2.17

8.47

0

62.70

Interquartile range

4.11

1.12

6.77

0

62.70

Standard deviation

3.76

2.19

4.58

0

33.99

Number of observations

4683

Note: In this table, markets are defined using the cliques approach. CLP$ stands for Chilean pesos.

36

Table 3: Effect of disclosure on log(marginit ) (1)

(2)

(3)

(4)

log(marginit ) Disclosure

0.0717

0.104

0.100

0.091

(0.0227)∗

(0.0144)∗∗∗

(0.0147)∗∗∗

(0.0279)∗∗

[0.000]∗∗∗

[0.000]∗∗∗

Station FE

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Common trend

No

Yes

No

No

Region-specific trend

No

No

Yes

Yes

Observations dropped

Post August 2012

N

5795

5795

5795

3777

Mean margins (CLP$ per liter)

70.35

70.35

70.35

69.71

R2

0.725

0.739

0.754

0.857

Note: The dependent variable in all regressions is log(marginit ), with margins measured in Chilean pesos per liter. Standard errors (in parentheses) are clustered at the area of intervention level. Specifications 3 and 4 also report, in square brackets, the p-value associated with the Wild Bootstrap estimation with 1, 000 replications and clustering at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The Wild bootstrap is implemented

using the cgmwildboot command in Stata. CLP$ stands for Chilean pesos. Specification 4 drops all observations following August 2012.

37

Table 4: Main specification with and without controls (1)

(2)

(3)

(4)

(5)

log(marginit ) 0.0689

0.0649∗

0.0685

0.0717∗∗

0.100∗∗∗

(0.0619)

(0.0233)

(0.0624)

(0.0225)

(0.0147)

Station FE

Yes

Yes

Yes

Yes

Yes

Year and Month FE

No

Yes

No

Yes

Yes

Cost Controls

No

No

Yes

Yes

Yes

Region-specific trends

No

No

No

No

Yes

N

5795

5795

5795

5795

5795

Mean margins (CLP$ per liter)

70.35

70.35

70.35

70.35

70.35

R2

0.692

0.724

0.692

0.725

0.754

Disclosure

The dependent variable in all regressions is log(marginit ), with margins measured in Chilean pesos per liter. Standard errors (in parentheses) are clustered at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

38

p < 0.01.

Table 5: Effect of disclosure on log(marginit ) excluding one city at a time (1)

(2)

(3)

(4)

(5)

(6)

log(marginit ) Excluding:

Valpara´ıso

Rancagua

Talca

Concepci´on

Punta Arenas

Santiago

Disclosure

0.105

0.0905

0.102

0.0883

0.128

0.104

(0.0214)∗∗

(0.0184)∗∗

(0.0175)∗∗

(0.0172)∗∗

(0.0237)∗∗

(0.0135)∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

Station FE

Yes

Yes

Yes

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Region-specific trends

No

No

No

No

Yes

Yes

N

4906

4825

5224

4789

5420

3811

R2

0.781

0.726

0.757

0.745

0.773

0.483

Note: The dependent variable in all regressions is log(marginit ), with margins measured in Chilean pesos per liter. Standard errors (in parentheses) are clustered at the area of intervention level. All specifications report p-values, in square brackets, associated with the Wild bootstrap estimation with 1, 000 replications and clustering at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The Wild bootstrap

is implemented using the cgmwildboot command in Stata.Each specification excludes a city as specified at the top of the corresponding column. CLP$ stands for Chilean pesos.

39

Table 6: Effect of disclosure and station characteristics on log(marginit ) (1)

(2)

log(marginit ) Disclosure

Convenience store

Convenience store× Disclosure

Pharmacy

Pharmacy× Disclosure

Public restrooms

Public restrooms× Disclosure

Repair service

Repair service× Disclosure

Has self-service pumps

Has self-service pumps× Disclosure

Open 24 hours

Open 24 hours × Disclosure

0.188∗∗

0.175∗∗

(0.0515)

(0.0483)

0.0677

0.0709

(0.0490)

(0.0466)

-0.0183

-0.0232

(0.0215)

(0.0212)

-0.167

-0.174

(0.133)

(0.128)

0.152

0.163

(0.0938)

(0.0971)

0.0193

0.0309

(0.0246)

(0.0269)

-0.0363∗

-0.0500

(0.0151)

(0.0213)

0.0350

0.0338

(0.0283)

(0.0259)

-0.0347∗

-0.0322∗∗

(0.0116)

(0.00633)

-0.0242∗∗

-0.0208∗∗∗

(0.00436)

(0.00151)

0.0207

0.0128

(0.0241)

(0.0159)

0.0733

0.0656

(0.0568)

(0.0558)

-0.0670

-0.0588

(0.0645)

(0.0611)

Cost controls

Yes

Yes

Brand FE

No

Yes

Brand FE× Disclosure

No

Yes

Region FE

Yes

Yes

Year and Month FE

Yes

Yes

Region-specific trends

Yes

Yes

N

5676

5676

Mean margins (CLP $ liter)

70.35

70.35

R2

0.678

0.678

Note:

The dependent variable in both regressions is

log(marginit ), with margins measured in Chilean pesos per liter. Standard errors, in parentheses, are clustered at the area of intervention level.

∗

p < 0.10,

40

∗∗

p < 0.05,

∗∗∗

p < 0.01.

Table 7: Disclosure, Brands, and Mergers: Robustness Checks on the Effect of Disclosure on log(marginit ) (1)

(2)

(3)

(4)

(5)

(6)

(7)

log(marginit ) Disclosure

Petrobras

Shell

Terpel

0.103

0.0951

0.0879

0.102

0.103

0.100

0.104

(0.0140)∗∗∗

(0.0195)∗∗

(0.0244)∗∗

(0.0137)∗∗∗

(0.0133)∗∗∗

(0.0147)∗∗∗

(0.0109)∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]

[0.000]∗∗∗

-0.000594

-0.0100

(0.000957)

(0.0139)

[0.370]

[0.516]

-0.0202

-0.0271

(0.0288)

(0.0320)

[0.516]

[0.516]

0.00154

0.00899

(0.0225)

(0.0318)

[0.758]

[0.514]

Petrobras×Disclosure

Shell×Disclosure

Terpel×Disclosure

0.0198

0.0257

(0.0302)

(0.0335)

[0.514]

[0.514]

0.0154

0.0260

(0.00930)

(0.0159)

[0.248]

[0.248]

-0.0145

-0.00922

(0.0209)

(0.0202)

[0.598]

[0.714]

Merged in June 2013

0.0244 (0.0264) [0.514]

Merger approved in January 2013

0.026 (0.0114) [0.256]

Station FE

No

No

Yes

Yes

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Region-specific trends

Yes

Yes

Yes

Yes

Yes

Yes

Observations dropped

Post June 2013

Yes Post January 2013

N

5795

5795

5795

5795

5486

5795

5260

Mean margins (CLP$ per liter)

70.35

70.35

70.35

70.35

70.41

70.35

70.33

R2

0.668

0.668

0.754

0.754

0.766

0.754

0.770

Note: The dependent variable in all regressions is log(marginit ), with margins measured in Chilean pesos per liter. Standard errors, in parentheses, are clustered at the area of intervention level. Specifications 2 to 8 also report, in square brackets, the p-value associated with the Wild bootstrap estimation with 1, 000 replications and clustering at the area of intervention level. Wild bootstrap is implemented using the cgmwildboot command in Stata.

41

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The

Table 8: Effect of Disclosure on Margin Dispersion Panel A: Range

(1)

(2)

(3)

(4)

0.955

0.266

0.281

0.269

(1.037)

(0.846)

(0.798)

(0.835)

[0.594]

[0.736]

Disclosure

Mean dependent variable

CLP$5.525 per liter

N

4683

4683

4683

4683

R2

0.368

0.424

0.307

0.336

0.658

0.0686

0.0506

0.0490

(0.964)

(0.884)

(0.833)

(0.880)

[0.716]

[0.716]

Panel B: Interquartile range Disclosure

Mean dependent variable

CLP$4.110 per liter

N

4683

4683

4683

4683

R2

0.309

0.356

0.200

0.210

0.305

-0.0472

-0.0938

-0.0881

(0.703)

(0.607)

(0.554)

(0.596)

[0.736]

[0.736]

3555

3555

Panel C: Standard deviation Disclosure

N

3555

Mean dependent variable

3555

CLP$0.759 per liter

R2

0.293

0.326

0.230

0.235

Market FE

Yes

Yes

No

No

Region FE

No

No

Yes

Yes

Region-specific trend

No

Yes

Yes

Yes

Number of stations in the market

No

No

Yes

No

Number of stations of each brand in the market

No

No

No

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Note: The dependent variable is a different measure of margin dispersion in each panel (range, interquartile range, and standard deviation). Standard errors (in parentheses) are clustered at the area of intervention level. Specifications 3 and 4 also report, in square brackets, the p-value associated with the Wild bootstrap estimation with 1, 000 replications and clustering at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01.

The Wild Bootstrap is implemented using the cgmwildboot command in Stata. CLP$ stands for Chilean pesos.

42

Table 9: Margins, Income, and Demographics. Dependent variable is log(marginit ) (1)

(2)

log(marginit ) Disclosure

Disclosure×Income

0.101

0.0982

(0.0145)∗∗∗

(0.0108)∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

-0.0604

-0.0758

(0.0074)∗∗∗

(0.0190)∗∗

[0.054]∗

[0.082]∗

Disclosure×Number of Fixed Internet connections

0.0165 (0.0085) [0.246]

Disclosure×Poverty rate

-0.0056 (0.0359) [0.712]

Disclosure×Rural population

-0.0665 (0.0208)∗∗ [0.224]

Region-specific trends

Yes

Yes

Cost controls

Yes

Yes

Comuna FE

Yes

Yes

Year and Month FE

Yes

Yes

N

5795

Mean real margin

CLP$70.35

R2

0.740

0.742

Note: Standard errors, in parentheses, are clustered at the area of intervention level. Both specifications also report the p-value associated with the Wild Bootstrap estimation in squared brackets.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The

Wild bootstrap is implemented using the cgmwildboot command in Stata with 1, 000 replications. All demographic variables are standardized. CLP$ stands for Chilean pesos.

43

Table 10: Margins and Consumer-Search Behavior. Dependent variable is log(marginit ) (1)

(2)

(3)

(4)

(5)

(6)

-0.0277

-0.112

(0.0083)∗∗

(0.0104)∗∗∗

[0.002]∗∗∗

[0.002]∗∗∗

log(marginit ) Number of requests within 1 km

-0.0099

-0.0039

(0.0036)∗

(0.0031)

[0.002]∗∗∗

[0.120]

Number of requests within 1 km (square)

-0.0004 (0.0005) [0.094]∗

Number of requests within 3 km

-0.0175

-0.0819

(0.0116)

(0.0132)∗∗∗

[0.162]

[0.002]∗∗∗

Number of requests within 3 km (square)

0.0118 (0.0015)∗∗∗ [0.000]∗∗∗

Number of requests within 5 km

Number of requests within 5 km (square)

0.0174 (0.0017)∗∗∗ [0.000]∗∗∗

Income (standardized)

Yes

Yes

Yes

Yes

Yes

Yes

Income squared

No

Yes

No

Yes

No

Yes

Number of stations within the distance threshold

Yes

Yes

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Region FE

Yes

Yes

Yes

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Yes

Yes

Region-specific trend

Yes

Yes

Yes

Yes

Yes

Yes

0.576

0.580

N

2278

Mean dependent variable

CLP$72.17 pesos per liter.

R2

0.572

0.573

0.573

0.577

Note: The dependent variable in all specifications is log(marginit ). The sample consists only on post-intervention observations, starting in June 2012 (when the app data started to be recorded). The number of requests in all specifications is standardized and corresponds to the number of requests executed in the previous month, as described in the text. Specifications 1, 3, and 5 do not include squared values while specifications 2, 4, and 6 do. All specifications report standard errors in parenthesis and p-values, associated with the Wild bootstrap, in squared brackets. In all cases, clustering is at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The Wild Bootstrap was implemented using the cgmwildboot

command in Stata, with 1, 000 replications. CLP$ stands for Chilean pesos.

44

Table 11: Margin Dispersion and Household Income (1)

(2)

(3)

Range

Interquartile Range

Std. Deviation

0.175

0.0108

-0.0936

(0.7291)

(0.7424)

(0.4938)

[0.742]

[0.862]

[0.766]

0.742

0.806

0.531

(0.1613)∗∗

(0.1511)∗∗

(0.1020)∗∗

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

Market FE

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Region-specific trends

Yes

Yes

Yes

N

4683

4683

3555

Mean dependent variable (CLP$ per liter)

5.525

4.110

3.759

R2

0.428

0.360

0.331

Disclosure

Disclosure×Income

Note: The dependent variable in all specifications is measured in Chilean pesos per liter. Standard errors, in parenthesis, are clustered at the area of intervention level. p-values, reported in squared brackets, are associated with the Wild Bootstrap and clustering at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The Wild bootstrap was implemented in Stata

using the cgmwildboot command with 1, 000 replications. In all specifications income is standardized. CLP$ stands for Chilean pesos.

45

Table 12: Margin Dispersion and Consumer-Search Behavior Range (1) Requests within 1 km

Requests within 1 km, squared

(2)

Interquartile range (3)

(4)

(5)

(6)

Std. Deviation (7)

0.681

0.891∗

1.145

(0.516)

(0.290)

(0.527)

0.0224

-0.0334

-0.0851

(0.0229)

(0.0196)

(0.0515)

Requests within 3 kms

Requests within 3 kms, squared

(8)

46

1.714

1.982

1.413

(1.453)

(1.081)

(1.070)

-0.173

-0.254

-0.134

(0.204)

(0.164)

(0.179)

Requests within 5 kms

Requests within 5 kms, squared

(9)

3.068

3.329∗

2.295

(1.804)

(1.360)

(1.145)

-0.477

-0.603

-0.355

(0.309)

(0.260)

(0.234)

Income

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Income squared

No

Yes

No

No

Yes

No

No

Yes

No

Number of stations within the distance threshold

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Region FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Region-specific trends

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

N

1820

1820

1820

1820

1820

1820

1427

1427

1427

Mean dependent variable

8.017

8.017

8.017

5.887

5.887

5.887

5.140

5.140

5.140

R2

0.323

0.328

0.330

0.267

0.272

0.277

0.317

0.319

0.324

Note: The dependent variable in all specifications is measured in Chilean pesos per liter. The sample consists only on post-intervention observations, starting in June 2012 (when the app data started to be recorded). All specifications report standard errors, clustered at the area of intervention level, in parentheses.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. In all specifications the number of requests is standardized. CLP$

Figure 1: Administrative division of Chile and policy roll-out

The figure shows the main administrative divisions of Chile (15 regions) and the roll-out of the policy under study.

Figure 2: Comparing market definitions: cliques and radius

The figure compares the consequences of using different market definitions. In the figure, all stations are within a circle with a two kilometer radius centered at the station represented by the black circle. However, the stations represented by black triangles and black squares are more than two kilometers away from each other. Defining markets using a two kilometer radius from the black circle puts all stations within the same market. On the other hand, using cliques results in two markets, one including the circle and the squares and another one including the circle and the triangles.

47

Figure 3: Prices sustainable under coordination as a function of parameter values and the time it takes to detect a deviation t = 0.1 P

P

P

10000

10000

10000

9000

9000

9000

8000

8000

8000

7000

7000

7000

6000

6000

6000

5000

5000

5000

4000

4000

4000

3000

3000

3000

2000

2000

1000

1000

0

2000 1000

0 0

10

20

30

40

50

60

70

80

90

100

0 0

10

20

30

40

50

60

70

80

90

Time

100

0

10

20

30

40

50

60

70

80

90

Time

(a) φ = 0.1

100

Time

(b) φ = 0.5

(c) φ = 0.9

t = 0.5 P

P

P

10000

10000

10000

9000

9000

9000

8000

8000

8000

7000

7000

7000

6000

6000

6000

5000

5000

5000

4000

4000

4000

3000

3000

2000

2000

1000

1000

0

3000 2000 1000

0 0

10

20

30

40

50

60

70

80

90

100

0 0

10

20

30

40

50

60

70

80

90

Time

100

0

10

20

30

40

50

60

70

80

90

Time

(d) φ = 0.1

100

Time

(e) φ = 0.5

(f) φ = 0.9

t=1 P

P

P

10000

10000

10000

9000

9000

9000

8000

8000

8000

7000

7000

7000

6000

6000

6000

5000

5000

5000

4000

4000

4000

3000

3000

2000

2000

1000

3000 2000

1000

0

1000

0 0

10

20

30

40

50

60

70

80

90

100

0 0

10

20

30

40

50

60

70

80

Time

(g) φ = 0.1

90

100

Time

(h) φ = 0.5

0

10

20

30

40

50

60

70

80

90

100

Time

(i) φ = 0.9

The figure shows all sustainable prices under coordination, as a function of the time it takes firm B to observe a deviation by firm A, for different combinations of parameters. The colored area represents the set of prices that are sustainable under coordination. All figures assume r = 0.0001. t represents transportation costs and φ is the fraction of consumers that are informed.

48

Coefficient Change in margins (percentage)

Figure 4: Dynamic difference-in-difference

.4

.2

0

−.2

−.4

−.6 −5 −4] −3 −2 −1 0 1 2 3 4 5 6 7 8 Months before and after disclosure was introduced in each city Baseline category is all months before t−5

9

>=10

The figure shows the estimated coefficients and confidence intervals for different time dummies that take the value of one for the specified period and zero otherwise. All dummies are defined at the city level, meaning that, for example, t = 0 is equal to one for each city at a different point in time, according to the roll-out of the website. The regression replicates those in Table 3 but defining the different pre- and post-disclosure dummies for each month around the policy implementation. The baseline category is all observations before the first indicator turns on.

49

Figure 5: Consumer Search Behavior in the City of Santiago

(a) July 26th, 2012

(b) November 27th, 2012

The figure shows heat-intensity maps representing consumer-search intensity in the City of Santiago, together with municipality boundaries and mean household income at the municipality level. More intense and filled colors represent higher intensity. More transparent shades represent lower search intensity. Darker colors within municipality boundaries represents higher mean household income.

4.3

4.2

4.25

4.1

4

3.9

4.3

Linear predictiction: log(margin)

4.3

Linear predictiction: log(margin)

Linear predictiction: log(margin)

Figure 6: Margins and Consumer Search Behavior (marginal effects)

4.2

4.15

4.25

4.2

4.15

4.1

3.8 4.05 0

1000 2000 3000 Number of search requests within one kilometer

(a) 1 km

4000

4.1 0

1000 2000 3000 Number of search requests within one kilometer

4000

(b) 3 km

0

1000 2000 3000 Number of search requests within one kilometer

4000

(c) 5 kms

The figure shows the marginal effect of consumer search on stations’ margins and its 95% confidence interval. The underlying regressions correspond to those reported in Table 10.

50

Figure 7: Number of requests and Margin Dispersion (marginal effects) Range 25

20 16

Linear predictiction: Range

Linear predictiction: Range

15

10

Linear predictiction: Range

14

20

12

10

15

10

8

6

5

5 0

500

1000 1500 Number of search requests within 1 kilometers

2000

0

(a) 1 km

5000 Number of search requests within 3 kilometers

10000

0

5000

(b) 3 km

10000 15000 Number of search requests within 5 kilometers

20000

25000

(c) 5 kms

Interquartile range 14

15

10

15

12

Linear predictiction: Interquartile range

Linear predictiction: Interquartile range

Linear predictiction: Interquartile range

20

10

8

6

10

5

4

5

0 0

500

1000 1500 Number of search requests within 1 kilometers

2000

0

(d) 1 km

5000 Number of search requests within 3 kilometers

10000

0

5000

(e) 3 km

10000 15000 Number of search requests within 5 kilometers

20000

25000

(f) 5 kms

Standard Deviation 12

10

5

12

10

10

Linear predictiction: Std. Deviation

Linear predictiction: Std. Deviation

Linear predictiction: Std. Deviation

15

8

6

8

6

4 4 0

2 0

500

1000 1500 Number of search requests within 1 kilometers

(g) 1 km

2000

0

5000 Number of search requests within 3 kilometers

10000

(h) 3 km

0

5000

10000 15000 Number of search requests within 5 kilometers

20000

25000

(i) 5 kms

The figure shows the marginal effect of consumer search on margin dispersion and its 95% confidence interval. The underlying regressions correspond to those reported in Table 12.

51

Transparency and Price Dispersion This appendix presents an extension of the model presented in section 2 that allows for price dispersion. The model presented here differs from that in the main text because it allows stations to differ in, for example, quality. To take this into account, consider the case where vA > vB . Then, assuming that firm A is located at 0 and firm B at 1, a consumer located at k, who purchases from firm A, receives indirect utility equal to vA − pA − tk, while if she purchases from B, her utility is vB − pB − t(1 − k). Assuming, as in the main text, that there are two types of consumers, informed and uninformed, it is possible to show that an informed consumer will be indifferent between purchasing from either firm if vA − pA − txI = vB − pB − t(1− xI ). Solving for the location of the indifferent consumer gives xI (pA , pB , ∆; t) =

∆+pB −pA 2t

+ 12 , where ∆ = vA − vB > 0 and

t corresponds to the cost incurred in transportation. That is, ∆ corresponds to additional utility from purchasing from A because of, for example, its perceived higher quality. To keep things simple, I assume that uninformed consumers behave exactly the same as in the main text and half of them visit each station regardless of the price difference. Note, however, that for half of the uninformed consumers to visit each station in each period, it is necessary that these consumers are uninformed not only about prices but also about quality differences. Allowing for uninformed consumers to perceived quality differences changes the fraction of these consumers that visits each station, but it does not affect the results.27 For this reason, in what follows I assume that uninformed consumers do not perceive quality differences. In this setting, it is possible to show that the one-period profits associated with static N = Nash competition are given by πA

(3t+∆φ)2 18φt

N = and πA

(3t−∆φ)2 18φt .

Consider now the case in which firms coordinate to charge prices above the Nashequilibrium prices. Without loss of generality, assume that coordination prices are pC  A C C C and pB . In this context, profits of firm A under coordination are given by πA = pA 12 + 27

Specifically, if uninformed consumers observe quality differences, then the fraction of uninformed con!

sumers that visits station A increases from

1−φ 2

to

1−φ 2

1+

∆ t

. This does not add any interesting feature

to the analysis but it complicates the algebra. For this reason, I assume that uninformed consumers do not perceive quality differences.

52

C ∆+pC B −pA φ 2t

!

.

As before, for coordination to be possible it is necessary that if firm B is charging pC B, firm A does not find it profitable to deviate. To determine when this is the case, it is necessary to find the most profitable deviation and determine whether firm A would like to deviate if firm B follows the collusive agreement.28 What ! is the optimal deviation for firm A if B charges its collusive price? pD A = D = under deviation are equal to πA

1 2

t + φ(∆ + pC B ) , which implies that A’s profits

2 (t+φ(∆+pC B )) 29 . 8φt

In this context, coordination is sustainable if ! ! Z z∗ Z ∞ Z ∞  C ∆ + pC (3t + ∆φ)2 −rz 1 C 1 −rz C B − pA pA e dz ≥ t + φ(∆ + pB ) e−rz dz + +φ e dz , 2 2t 2 18φt 0 z∗ 0 {z } {z } | {z } | | Payoffs under coordination

Payoffs under deviation

Punishment payoffs

where z ∗ represents the instant at which a deviation, that started at z = 0, is observed. As

before, it is explicit in this last expression (an incentive-compatibility constraint, IC) that the punishment strategy is to play the Nash-equilibrium outcome. Now the incentive-compatibility constraint can be written as !2

! pC C C A − 1 + φ(∆ − pA + pB ) + 2

(∆ + pC B )φ + t 8φt

(1 − e−rz ) +

(∆φ + 3t)2 −rz e ≤ 0, 18φt

which defines all pairs (PAC , pC B ) that can be sustained under coordination, as a function of parameter values and the time it takes to observe a deviation. Figure 8 shows these prices as a function of the time it takes to observe a deviation, for a fixed combination of (t, r, ∆), for three values of φ (0.1, 0.5, and 0.9). In the figure, the blue area corresponds to the maximum prices that can be charged under collusion when φ = 0.1. That is, the set of all possible collusive prices is given by the area below the different surfaces. In this way, the figure shows that as the fraction of informed consumers increases, the area of collusive prices shifts inwards, regardless of the length of time it takes to observe a deviation. Indeed, 28

This is a consequence of all deviations being observed. Hence, if a firm deviates, it will do so to its most

profitable deviation to maximize the profits it will receive before the game moves to the punishment stage. 29 C C Note that when ∆ = 0 and pC profits in each stage are equal to those presented in the main A = pB = p text for the corresponding stage.

53

Figure 8b shows a different view of the same figure and shows that when φ is 0.5, the area of collusive prices shrinks to that in red, and when φ is 0.9, so most consumers are informed, the area of collusive prices shrinks to the yellow one, all in the interior of the area covered by cases with a smaller fraction of informed consumers. The figure also shows that price dispersion arises in equilibrium. This is easier to observe in Figure 8b. Indeed, in this figure, the axis on the left (that measures the deepness of the graph) corresponds to pC B , the horizontal axis is time to detect a deviation (measured from right to left), and the vertical axis is pC A . The figure clearly shows that all possible pairs of collusive prices are below the “45-degree” plane, meaning that firm B always charges prices below those charged by A. The figure also shows that as the length of time it takes to detect a deviation decreases (moving left to right in Figure 8b), it is apparent that price dispersion decreases as well as the slope of the surface that represents potential collusive prices increases, bringing the area closer to the 45-degree plane. Furthermore, the figure shows that when the length of time it takes to observe a deviation increases, prices that were sustainable under shorter detection times are no longer sustainable, making collusion unfeasible if it takes (relatively) too long to observe a deviation. Finally, Figure 8c shows what happens when differentiation increases. In this case, the blue surface represents the maximum sustainable prices for φ = 0.5 and ∆ = 50 (recall that ∆ = vA − vB ), while the red surface, that is in the interior of the blue one, represents prices for the same fraction of informed consumers and ∆ = 200. The figure shows that as differentiation increases, the set of sustainable prices shifts inwards, making collusion less likely to happen. Overall, this example provides similar information to that presented in the text regarding what happens with collusive prices when the time it takes to detect a deviation decreases. In addition, this example suggests that price dispersion decreases as more consumers become informed and that, conditional on the fraction of informed consumers, increasing differentiation makes collusion more difficult.

54

Figure 8: Prices sustainable under coordination with vertical differentiation Parameters: {t = 0.1, r = 0.0001, ∆ = 50}

C PA C PA

C PB

C PB

Time

←− Time

(a) φ = 0.1, φ = 0.5,

(b) φ = 0.1, φ = 0.5,

φ = 0.9

φ = 0.9

C PA

C PB

←− Time

(c) ∆ = 50, ∆ = 200

C The figure shows all pairs (pC A , pB ) that can be sustained under coordination, as a function of the length of

time it takes to firm B to observe a deviation from firm A, for different fractions of informed consumers.

55

Online Appendix: Not For Publication

Who Benefits from Information Disclosure? The Case of Retail Gasoline Fernando Luco

A

Figures and Tables

Figure A.1: Margins and Consumer-Search Behavior (marginal effects).

Robustness

checks 4.3

4.3

4.2

4

3.8

Linear predictiction: log(margin)

Linear predictiction: log(margin)

Linear predictiction: log(margin)

4.25

4.2

4.1

4.2

4.15

3.6 4.1

3.4

4 0

1000

2000 3000 Number of search requests within one kilometer

4000

5000

(a) 1 km

4.05 0

1000

2000 3000 Number of search requests within three kilometers

4000

5000

(b) 3 km

0

2000 4000 6000 Number of search requests within five kilometers

8000

(c) 5 kms

The figure shows the marginal effect of consumer search on stations’ margins and the 95% confidence interval. The underlying regressions correspond to those reported in Table A.3.

ii

Figure A.2: Margin Dispersion and Consumer-Search Behavior (marginal effects). Robustness checks Range 25

20 16

Linear predictiction: Range

Linear predictiction: Range

15

10

Linear predictiction: Range

14

20

12

10

15

10

8

6

5

5 0

500

1000 1500 Number of search requests within 1 kilometers

2000

0

(a) 1 km

5000 Number of search requests within 3 kilometers

10000

0

5000

10000 15000 Number of search requests within 5 kilometers

(b) 3 km

20000

25000

(c) 5 kms

Interquartile range 14

15

10

5

15

12

Linear predictiction: Interquartile range

Linear predictiction: Interquartile range

Linear predictiction: Interquartile range

20

10

8

6

10

5

4 0 0

500

1000 1500 Number of search requests within 1 kilometers

2000

0

(d) 1 km

5000 Number of search requests within 3 kilometers

10000

0

5000

10000 15000 Number of search requests within 5 kilometers

(e) 3 km

20000

25000

(f) 5 kms

Standard Deviation 12

12

10

8

6

10

Linear predictiction: Std. Deviation

12

Linear predictiction: Std. Deviation

Linear predictiction: Std. Deviation

14

8

6

10

8

6

4 4

4 2 0

500

1000 1500 Number of search requests within 1 kilometers

2000

(g) 1 km

0

5000 Number of search requests within 3 kilometers

10000

0

5000

10000 15000 Number of search requests within 5 kilometers

(h) 3 km

20000

25000

(i) 5 kms

The figure shows the marginal effect of consumer search on margin dispersion and the 95% confidence interval. The underlying regressions correspond to those reported in Table A.5.

iii

Table A.1: Summary statistics. Demographic information by year across municipalities Poverty rate

Population

(%)

Rural population (% of total)

Municipality/Year

2013

2010

2013

2010

2013

2010

Valpara´ıso

16.13

21.00

267,853

272,543

1.56

1.96

Rancagua

8.99

11.70

253,189

245,476

0.04

0.04

Talca

17.51

19.30

253,742

242,473

0.05

0.05

Concepci´ on

21.48

14.10

230,255

228,651

0.03

0.03

Punta Arenas

5.42

8.60

125,712

124,949

0.02

0.02

Santiago

5.71

7.80

156,049

167,867

0

0

Cerro Navia

14.64

18.20

129,630

136,044

0

0

Conchal´ı

10.83

11.60

101,796

109,891

0

0

Estacion Central

17.61

9.60

107,335

113,839

0

0

Independencia

8.23

8.60

48,565

52,616

0

0

La Cisterna

7.46

12.30

68,370

72,95

0

0

La Florida

9.21

9.70

396,684

399,177

0

0

La Granja

15.94

23.20

120,144

124,985

0

0

La Reina

7.12

2.50

94,037

96,232

0

0

Las Condes

1.38

1.30

291,971

286,204

0

0

Maipu

9.20

6.40

931,211

805,503

0

0

˜ noa Nu˜

5.16

2.80

140,531

147,380

0

0

Quinta Normal

11.44

7.50

83,187

88,801

0

0

Recoleta

11.53

10.90

119,303

127,347

0

0

San Joaqu´ın

26.87

13.10

73,197

79,272

0

0

San Miguel

12.97

5.10

68,855

72,062

0

0

Note: The table summarizes demographic information obtained from the SINIM (2016) dataset for each of the municipalities in the analysis. The first five municipalities correspond to cities other than Santiago. All municipalities that follow, starting from Santiago, are within the urban area of the City of Santiago (which explains why the rural population is zero for all of them). iv

Table A.2: Effect of disclosure on margin dispersion, with markets defined using a twokilometer radius around gas stations (1)

(2) Range

Disclosure

Number of stations

(3)

(4)

(5)

(6)

Interquartile range

Standard deviation

-0.690

-0.652

-0.359

-0.296

-0.359

-0.335

[0.524]

[0.524]

[0.802]

[0.802]

[0.524]

[0.660]

0.925

0.0322

0.0956

[0.000]∗∗∗

[0.770]

[0.204]

Number of Copec stations

Number of Petrobras stations

Number of Shell stations

Number of Terpel stations

0.378

-0.106

0.0308

[0.522]

[0.250]

[0.282]

0.315

0.0928

0.0197

[0.000]∗∗∗

[0.638]

[0.572]

0.758

0.252

0.156

[0.000]∗∗∗

[0.000]∗∗∗

[0.000]∗∗∗

-0.104

0.0384

-0.105

[0.488]

[0.776]

[0.646]

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Region

Yes

Yes

Yes

Yes

Yes

Yes

Year and Month

Yes

Yes

Yes

Yes

Yes

Yes

Region-specific trends

Yes

Yes

Yes

Yes

Yes

Yes

N

5725

5606

5725

5606

5507

5388

Mean dependent variable

CLP$ 10.46

R2

0.452

0.463

CLP$ 5.876 0.352

0.362

CLP$ 4.434 0.408

0.421

Note: In this table, markets are defined using a two-kilometer radius around gas stations, instead of cliques. In columns 1 and 2, the dependent variable is the range of margins at the market level. In columns 3 and 4, the dependent variable is the interquartile range of margins. In columns 5 and 6, the dependent variable is the standard deviation of margins. In all regressions, the dependent variable is measured in Chilean pesos per liter. Inv all columns, p-values are reported in square brackets. p-values are associated to the Wild bootstrap estimation with 1, 000 replications and clustering at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The Wild bootstrap is imple-

mented using the cgmwildboot command in Stata. CLP$ stands for Chilean pesos.

Table A.3: Margins and Consumer-Search Behavior. Dependent variable is log(marginit ) (1)

(2)

(3)

log(marginit ) Requests within 1 km

0.00056 [0.334]

Requests within 1 km, squared

-0.000762 [0.002]∗∗∗

Requests within 3 kms

-0.0709 [0.002]∗∗∗

Requests within 3 kms, squared

0.0109 [0.000]∗∗∗

Requests within 5 kms

-0.0958 [0.002]∗∗∗

Requests within 5 kms, squared

0.0163 [0.000]∗∗∗

Fixed Internet connections

Fixed Internet connections, squared

Poverty rate

Rural population

Cost controls

-0.0250

-0.0360

-0.0511

[0.216]

[0.002]∗∗∗

[0.102]

0.0291

0.0354

0.0429

[0.100]

[0.000]∗∗∗

[0.000]∗∗∗

-0.349

-0.509

-0.543

[0.768]

[0.256]

[0.516]

0.00394

0.00377

0.0037

[0.362]

[0.252]

[0.252]

Yes

Yes

Yes

Number of stations within distance threshold

Yes

Yes

Yes

Region

Yes

Yes

Yes

Year and Month

Yes

Yes

Yes

Region-specific trends

Yes

Yes

Yes

N

2278

2278

2278

Mean dependent variable

72.17

R2

0.584

0.588

0.591

Note: The dependent variable in all specifications is log(marginit ). The sample consists only on post-intervention observations, starting in June 2012. All specifications report p-values associated with the Wild bootstrap, in squared brackets, and clustering is at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The

Wild bootstrap was implemented using the cgmwildboot command in Stata, with 1, 000 replications. In all specifications, the number of search requests and fixed Internet connections is standardized. CLP$ stands for Chilean pesos.

vi

Table A.4: Margin dispersion, Income, and Demographics (1)

(2)

(3)

Range

Interquartile range

Std. Deviation

0.121

-0.0235

-0.103

[0.744]

[0.766]

[0.748]

0.708

0.821

0.365

[0.000]∗∗∗

[0.000]∗∗∗

[0.132]

-0.0329

0.0639

-3.880

[0.904]

[0.862]

[0.002]∗∗∗

0.106

-0.00638

0.642

[0.246]

[0.880]

[0.396]

16.13

13.36

2.633

[0.604]

[0.486]

[0.862]

-0.0298

-0.0467

-0.00689

[0.002]∗∗∗

[0.002]∗∗∗

[0.002]∗∗∗

Market FE

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Region-specific trend

Yes

Yes

Yes

N

4683

4683

3555

Mean dependent variable

5.525

4.110

3.759

R2

0.429

0.361

0.337

Disclosure

Disclosure×Income

Fixed Internet Connections

Disclosure×Fixed Internet Connections

Poverty rate

Rural population

Note: The dependent variable in all specifications is measured in Chilean pesos per liter. pvalues, reported in squared brackets, are associated to the Wild bootstrap with clustering at the area of intervention level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The Wild bootstrap

was implemented in Stata using the cgmwildboot command with 1, 000 replications. In all specifications, income and the number of fixed Internet connections is standardized. CLP$ stands for Chilean pesos.

vii

Table A.5: Margin dispersion, Consumer-Search Behavior, and Demographics Range (1) Requests within 1 km

Requests within 1 km, squared

(2)

(6)

Std. Deviation (7) 1.277

(0.543)

(0.210)

(0.545)

0.0457∗

-0.0183

-0.102

(0.0156)

(0.0332)

(0.0482)

(8)

1.822

1.892

1.539

(1.660)

(1.179)

(0.996)

-0.221

-0.255

-0.164

(0.257)

(0.196)

(0.165)

Requests within 5 kms

viii Requests within 5 kms, squared

Poverty rate

(5)

0.734∗∗

Requests within 3 kms, squared

Fixed Internet connections, squared

(4)

0.455

Requests within 3 kms

Fixed Internet connections

Interquartile range (3)

(9)

3.166

3.188

2.514∗

(2.034)

(1.494)

(1.034)

-0.535

-0.596

-0.425

(0.368)

(0.301)

(0.211)

1.356

1.368

1.483

0.0762

0.146

0.290

0.703

1.023∗

1.278∗∗

(0.744)

(0.764)

(0.814)

(0.471)

(0.463)

(0.497)

(0.351)

(0.340)

(0.288)

-0.889

-0.889

-0.913

-0.0720

-0.102

-0.148

-0.742∗∗

-0.907∗∗

-1.035∗∗

(0.510)

(0.515)

(0.508)

(0.329)

(0.317)

(0.313)

(0.200)

(0.239)

(0.200)

15.09

17.32

17.29

6.426

8.657

8.617

-5.849

-3.947

-3.475

(18.67)

(17.48)

(15.28)

(12.48)

(11.43)

(9.740)

(11.08)

(9.614)

(8.244)

-0.0719

-0.0646

-0.0548

-0.113

-0.107

-0.0997

-0.0178

-0.0169

-0.0118

(0.0716)

(0.0697)

(0.0649)

(0.0499)

(0.0480)

(0.0432)

(0.0471)

(0.0457)

(0.0402)

Income

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Income squared

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Number of stations within distance threshold

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Region

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Year and Month FE

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Region-specific trends

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

N

1820

1820

1820

1820

1820

1820

1427

1427

1427

Mean dependent variable

8.017

8.017

8.017

5.887

5.887

5.887

5.140

5.140

5.140

R2

0.324

0.328

0.330

0.270

0.275

0.279

0.323

0.325

0.330

Rural population

Note: The dependent variable in all specifications is measured in Chilean pesos per liter. The sample consists only on post-intervention obser-