Pricing Dynamics, Leadership, and Misreporting: Evidence from a Mandatory Price-Disclosure Intervention∗ Jorge Lemus† and Fernando Luco‡

August 3, 2017

Abstract Mandatory information-disclosure policies may harm competition by facilitating rivals to monitor each other. We investigate this hypothesis in the context of the Chilean retail-gasoline industry after firms were mandated to post their prices in real-time on a government website. We study pricing dynamics in the 30 months that followed the intervention and find that the fraction of markets with price leaders doubled. Furthermore, firms increased their margins more when a leader was persistent. Finally, we argue that some firms strategically misreported their prices online in areas where consumers searched.

JEL codes: D22, D43, D83, L12, L41 Keywords: Pricing dynamics, information disclosure, price leadership, retail gasoline ∗

We thank the Chilean National Energy Commission (CNE), specially David Pe˜ na and Yamal

Soto, and Itai Ater, George Deltas, and Steven Puller for helpful comments and suggestions. We also thank Sergio Campam´ a for sharing data collected by an app. The usual disclaimers apply. † University of Illinois Urbana-Champaign, Department of Economics. [email protected] ‡ Texas A&M University, Department of Economics. [email protected]

1

1

Introduction

Mandatory information-disclosure policies aim to increase consumer search, competition, market transparency, and to facilitate regulatory oversight. Several countries have recently implemented mandatory price-disclosure policies in the retail-gasoline industry. In particular, Australia, Chile, Greece, and South Korea, have implemented policies that require gas stations to report their prices online.1 The rationale behind these policies is that a reduction of consumers’ search costs should intensify price competition (Schwartz and Wilde, 1978; Varian, 1980; Stahl, 1989). Furthermore, in markets with significant price dispersion, such as retail-gasoline markets (Lewis, 2008), reducing search costs may be particularly beneficial for consumers. The impact of a disclosure policy on welfare, however, depends on whether consumers actually search for low prices and on whether firms respond by competing more fiercely. The main contribution of this paper is to empirically illustrate two mechanisms—tacit collusion and misreporting—by which mandatory information disclosure may harm consumers. Our analysis stems from a mandatory price-disclosure policy implemented in Chile in March of 2012. The policy mandated gas stations to update their prices in real-time on a government website.2 We study three aspects of pricing behavior during the two and a half years that followed the intervention. First, we study pricing dynamics focusing on the number of price changes per week and when these changes take place. Second, we study whether the website facilitated the emergence of price leaders—firms that respond first to changes in wholesale prices in a given market in two consecutive weeks—and whether leadership had any impact on competition. Third, we study how firms that were caught misreporting their prices online to attract consumers that search, changed their behavior after they were sanctioned by the regulator.3 1

Information disclosure policies have also been implemented in other industries including super-

markets (Israel), concrete (Denmark), restaurant hygiene quality (United States). Private pricecomparison platforms (e.g., GasBuddy.com in the U.S.) have also emerged in many industries. 2 For more information, visit http://www.bencinaenlinea.cl/web2/normativa.php. 3 Sanctions are monetary fines or suspension of sales: https://www.leychile.cl/Navegar?idNorma=29819.

2

In our analysis, we exploit unique institutional features of the Chilean gasoline industry: the policy was implemented in the whole country, there is a record of every price that has ever been reported in the system, and wholesale prices change only once per week and the change is publicly announced. The latter feature is especially relevant because wholesale price changes are usually neither decided by a central authority nor publicly announced. We combine these institutional features with a rich dataset that includes the information reported through the website, all the changes in wholesale prices, the set of stations that were sanctioned for misreporting prices (and when they received it), and the location of search requests for gasoline prices executed through a smartphone app powered by data from the government website. Our main findings are related to three dimensions of strategic behavior that could reflect a reduction in competition: price leadership, the timing of price changes, and strategic reporting through the website. First, we show that the first station to change prices in a market in any given week is not randomly selected. For example, if in duopoly markets the first mover was selected at random, the probability of observing the same first mover for 10 consecutive weeks is almost zero whereas in the data is about 25 percent. Furthermore, we show that the probability of observing a price leader was stable during the first months after the policy implementation (around 30 percent) and increased to around 60 percent two years afterward. In addition, relative to markets without leaders, margins are higher in markets that have a leader, and the margin increase is larger the longer a leader has been in place. Second, we show that, over time, stations adjust their prices quicker in response to a change in wholesale prices. On average, five months into the policy implementation, gas stations were changing their prices about 15 hours earlier than what they did immediately after the policy was implemented. Faster adjustments are, however, entirely explained by “follower” stations in oligopoly markets (not first movers) reacting faster to changes in wholesale prices, while the speed of price adjustment of both first movers in oligopoly markets and monopolists did not change.

3

Finally, we argue that the website is a channel through which stations can report lower prices than those they charge at the pump, in an attempt to attract consumers that search for lower prices. However, we show that these attempts are quickly sanctioned by authorities and sanctioned stations stop misreporting after being sanctioned. Our analysis shows that though information-disclosure policies have the potential to intensify competition and increase consumer welfare, their success depends on how the policies are implemented and enforced. Indeed, our evidence is consistent with the website reducing monitoring costs and having an impact only on oligopoly markets. Furthermore, it is in oligopoly markets with leaders where margins increased more. These pieces of evidence suggests that by reducing monitoring costs, the website facilitated tacit coordination and reduced the intensity of competition. Related Literature Our paper contributes to the literature on mandatory disclosure, pricing dynamics, and price leadership. Studies of the effect of price transparency and the reduction of search costs on consumer welfare include the theoretical work of Campbell, Ray and Muhanna (2005), Schultz (2005), Daughety and Reinganum (2008), and St¨ uhmeier (2015). Empirical studies of information-disclosure policies include Mathios (2000) (salad dressing), Jin and Leslie (2003) (restaurant quality), Albæk, Møllgaard and Overgaard (1997) (ready-mixed concrete) and Ater and Rigbi (2017) (supermarkets), among others. Empirical studies on mandatory disclosure policies in the retail-gasoline industry include Rossi and Chintagunta (2015) (highways in Italy), Kim and Lee (2014) and Jang (2014) (South Korea), and Luco (2017) (Chile). These articles, however, do not address the questions we examine in this paper. We contribute to the literature on the use of price leadership as a tacit collusion device (Markham, 1951). Theories of price leadership under capacity constraints and asymmetric information have been developed by Rotemberg and Saloner (1990) and Deneckere and Kovenock (1992), respectively. Empirical evidence of using price leadership as a tacit collusion device has been provided by Kauffman and Wood (2007) 4

in the music-CD and books industry, Busse (2000) in the cellular telephone industry, Lewis (2012) in the US retail-gasoline industry. In this setting, the contemporaneous work of Byrne and de Roos (2017) is the closest to ours. They show that collusive behavior, through the figure of price leaders, began close to a decade after a pricedisclosure policy was implemented in Perth, Australia. We instead document the rise of price leadership soon after disclosure was implemented, and focus our analysis on the impact of leadership on margins and the timing of price adjustments. Additionally, we also provide evidence of strategic misreporting and how sanctions affect post-sanction pricing behavior, which has not been discussed in the literature. Finally, our work is also related to articles that examine different features of gasoline markets including the work of Lewis (2008) (price dispersion and local competition), Deltas (2008) (asymmetric response to changes in wholesale price), Lewis and Noel (2011) (Edgeworth cycles), Lewis (2011) (search), and Clark and Houde (2013, 2014) (collusion). Eckert (2013) surveys this literature.

2 2.1

Industry and Data The Chilean Gasoline Industry

Chile is a net importer of oil. Most of the oil is imported by the state-owned company ENAP that refines it into fuel products. Over the last decade, ENAP has supplied about 90 percent of the Chilean demand for fuel products. Every Wednesday at 7pm, ENAP publicly announces the change in wholesale prices for gasoline and diesel. These prices are effective on Wednesday at midnight and remain fixed until the following Wednesday. This peculiar characteristic enables us to work with a well-defined time period: the week that starts and ends with a public announcement of changes in wholesale prices.

5

In 2012, Chile had four main distributors of fuel products: Copec, Petrobras, Shell, and Terpel. Distributors sell to the industrial market and to the retail market through the gas stations that acquire gasoline from the distributor. A branded station only sells gasoline provided by the brand distributor and it displays the name of that distributor at the pump. Because gasoline is a homogeneous product, gas stations differentiate from each other through services, location, amenities, and by adding chemical products to the gasoline they buy from ENAP.

2.2

Policy Intervention

In March of 2012, the Chilean government began the implementation of a policy mandating all gas stations in the country to post their retail prices on a website.4 The policy was sequentially implemented across the country and it was fully implemented across the whole country, and it began to be enforced, by July 1st, 2012. The policy gives gas stations a 15-minute window to update their prices online after they have changed them at the pump. In addition to price information, the website provides information on stations’ characteristics and a map pinpointing each station.5 Finally, the government enforces the policy by visiting and sanctioning gas stations with online prices different from the prices charged at the pump.6

2.3

Data

In our analysis, we use four datasets collected from different sources. The first dataset, provided by the CNE, contains all the retail price changes reported on the government website as well as station characteristics such as location, brand, etc. We augment 4 5

The website is www.bencinaenlinea.cl. These characteristics include the address, amenities, fuel products, hours of operation, accepted

methods of payment, among others. 6 Consumers can report price mismatches. However, for a station to be sanctioned as a consequence of a consumer report, a government inspector must visit the station and validate the reports.

6

these data with all the public announcements of wholesale prices made by ENAP. There are 1,550 gas stations in the country that sell gasoline of 93 octanes.7 On average, there is one gas station for every 10,000 people in each region of the country.8 Around 90% of the gas stations are branded as either Copec, Shell, Petrobras, or Terpel—the rest are small distributors or independent brands. The second dataset was provided by SEC (Superintendencia de Electricidad y Combustibles), a government agency that, among other duties, monitors gas stations to certify that they meet safety and regulatory standards. From SEC, we obtained data on the owner or operator of 74 percent of the gas stations in the country, which we use to study the relation between pricing strategies and ownership.9 In our data, 78 percent of the owners own a single station and 15 percent own only two stations. The third dataset, also provided by SEC, identifies those stations that were sanctioned because the prices they reported on the website were not the same as those charged at the pump. The data also includes the date of the sanctions. In our sample period, 85 percent of stations were never sanctioned, 13 percent were sanctioned only once, 1.5 percent were sanctioned twice, and 0.4 percent were sanctioned three times. The fourth dataset consists of consumer search data. These data come from the CNE and from an app that uses the information reported on the website. We observe the location and time of search requests made through the app. Figure 1 reports the number of search requests through the government website and the smartphone app. The figure shows that, though there is variation over time, search is scarce—in our data, the website’s median number of daily visits is 947 and the smartphone app’s median number of daily requests is 377.10 7 8

In Online Appendix A we provide more details about the composition of the industry. Chile is administratively divided in regions. Regions are divided in provinces. Provinces are

divided into municipalities, the smallest administrative unit in Chile. 9 The information from SEC is incomplete. We define the owner as the legal representative when reported (most of the time), otherwise we defined ownership using the manager of the station. 10 Conversations with CNE revealed that web scrapping causes the abnormal number of requests.

7

3

Empirical Evidence

In this section, we focus on how three aspects of gas stations’ pricing strategies evolved following the implementation of the policy: the dynamics of price adjustments, the emergence of price leaders, and the post-sanction behavior of stations.

3.1

Patterns of Price Adjustments

Maskin and Tirole (1988) introduce a theory of dynamic competition in homogeneous goods markets. The dynamic of price adjustments, known as Edgeworth cycles, begins with a price increase followed by a series of small price cuts until one firm restarts the cycle with a large price increase. This theory has found empirical support in Noel (2007a,b) and Lewis (2012), among others. In the Chilean retail-gasoline industry, however, we do not find evidence of Edgeworth cycles. Strikingly, most gas stations change their prices only once a week, usually one day after the public announcement of changes in wholesale prices. Figure 2a shows the average number of price changes within each week by station over time—gas stations rarely change more than once a week. On average, the number of price changes is bounded between 9 and 11 changes in 10 weeks. For this reason, we define the relevant period of analysis to be one week—a week that starts and ends with ENAP’s public announcement, every Wednesday at 7pm. Figure 2b shows the timing of price changes, aggregated across gas stations, by week. The figure shows that the bulk of price changes happens between Wednesday at 7pm and Thursday at midnight. Only a few changes take place between Friday of one week and Wednesday before 7pm of the following week, just before a new announcement from ENAP. Finally, Figure 3 shows that stations became faster in responding to changes in wholesale prices since the policy implementation. In fact, Figure 3a shows that, on average, gas stations react to wholesale price changes almost twice as fast toward the end of

8

our sample period compared to the beginning. When we decompose the speed of price adjustment by market structure (Figure 3b), however, we find that the speed of adjustment of monopolies remained almost constant since the full implementation of the system, with the effect entirely driven by oligopolies.11 Thus, following the implementation of the website, markets with more than one gas station began to adjust their prices faster over time.12 The absence of frequent price adjustments may or may not be an indication of the lack of competition in the Chilean retail-gasoline industry. Because wholesale prices change only once a week, the absence of multiple price adjustments could be justified by firms myopically maximizing profits each week and setting prices according to a static Nash equilibrium with full information. However, the evidence suggests that the website facilitated monitoring rivals, because there is a faster speed of adjustment in oligopolies but not monopolies. For this reason, we now turn to study whether the website may have facilitated coordination among gas stations in oligopoly markets, specifically, through price leadership.

3.2

Price Leadership and the Distribution of Margins

Price leadership is one way in which gas stations can tacitly coordinate their prices (Rotemberg and Saloner, 1990; Lewis, 2012; Byrne and de Roos, 2017). In this section, we present evidence showing that price leadership is more pervasive in markets with fewer competitors and that that it became more frequent since the implementation of the price-disclosure policy. We also show that margins—defined as the difference between retail and the wholesale prices—are higher and firms react faster to changes in wholesale-price in markets with price leaders. We begin by formally defining markets, 11

We formally define markets in the next section. Figure 3 reports binned scatterplots. The un-

derlying partitioned regressions control for station fixed effects to take into account station-specific time-invariant characteristics. 12 Figure 2 and Figure 3 are for gasoline of 93 octanes. Figures for other fuel products are similar and we present them in Online Appendix D.

9

Definition 1 A market is a clique in the graph where gas stations are nodes and two stations are connected if their distance is less than some constant d.13

To implement the cliques we use driving times, rather than driving or linear distances, and we use a driving time of d = 5 minutes (motivated by Perdiguero and Borrell 2012).14 Our market definition is similar to other definitions in the retail-gasoline literature in the absence of quantity or volume data. The advantage of our definition is that it reduces the problem of double-counting stations that arises when markets are centered around each station and it avoids setting arbitrary boundaries for when markets are elements of a grid. In Online Appendix E, we study the robustness of our findings to alternative market definitions including cliques with d = 10 minutes, grid partitions, and partitions based on administrative boundaries at the municipality level. Our main results are qualitatively similar across these alternative market definitions. The regular one-week cycles introduced by ENAP’s public announcements allow us to naturally define the first mover in each market-week. This is different to what happens in other countries, where changes in wholesale prices neither happen with a regular frequency nor do they affect all players in the same way and at the same time.

Definition 2 A first mover in market m and week t is the player that changes retail prices before anyone else in that particular market/week.

Defining a market leader is often a difficult task. In particular, in most markets, including retail gasoline, wholesale prices change often and by different amounts across markets. Hence, it is not clear how a price leader should be defined. For this reason, the literature on price leadership has proposed various definitions. Lewis (2012) defines a leader as the station that begins the Edgeworth cycle. Seaton and Waterson (2013) define a leader as the supermarket that initiates a price increase that is followed in a 13 14

A clique of an undirected graph is a subgraph in which every node is connected to each other. We compute driving times using the algorithm developed by Huber and Rust (2016).

10

predefined window of time by another player who increases prices by the same amount. Finally, Byrne and de Roos (2017) define a leader as the station that initiates a price increase that is, ex-post, successful. Given that in our setting first movers are well defined, we use the following definition:

Definition 3 A price leader in market m and week t is a first mover in market m and week t that was also a first mover in market m in week t − 1.

The definition of first mover requires us to properly define what a player is. A player can be defined at the level of unique station identifiers, owners of gas stations, or brands. The weakest definition of a leader, the one that is harder to satisfy, is at the level of unique station identifier. However, it is relevant to consider alternative definitions because the absence of price leadership at the station level does not directly imply that stations are not following a leader.15 Figure 4 shows the evolution of leadership in four specific markets that have between six and nine gas stations. In the figures, the horizontal axis represents time (weeks) and the vertical axis represents the (within market) station id. The figures show, for each week, whether a one-week leader (a first mover for two consecutive weeks) is identified. If so, the figures reports the station that was identified as a leader and its brand. The figures shows that, across these markets, leadership became more common over time, though the pattern is different across markets. Figure 4a shows that in a market with six gas stations, only two were leaders at least once (stations with (within market) ids 1 and 4), and that leadership became more common around May 2013. Figure 4b shows a 7-stations market in which leadership was rare before June 2013, but it was concentrated in a single station afterwards. Figure 4c presents a 7-stations market 15

If tacit collusion is at the owner or brand level, different gas stations could be use to signal in the

same market. In this case, we would not see leadership at the station level but at the owner or brand level. At the same time, if a market covers a large geographical area, brand leadership may be more successful than station leadership in inducing tacit collusion.

11

where four stations were leaders at least once. Interestingly, the four stations belong to three different brands. Finally, Figure 4d reports the case of a 9-stations market without a persistent leader until March 2013, when a single station became the leader for the remaining of the sample period. Considering all stations and markets in our data, we find that 50 percent of the stations that were first movers at least once, were never leaders. Before studying whether leadership affected competition, we ask whether the number of weeks with the same first mover or leader observed in the data are consistent with what would be expected if the first mover was chosen at random. To address this question, in Online Appendix B, we derive the distribution of the duration of a gas station as both as first mover and price leader, if the first mover was determined at random. We then compare the random distribution with the empirical one, for different market structures. Figure 5 presents our findings using a 10-week window, aggregating over markets with the same number of competitors.16 Figure 5a shows that, in duopoly markets, price leaders last for 9 weeks (i.e., the same player is the first mover consecutively for 10 weeks) with probability 0.25, whereas the probability of this outcome is only 0.00195 if the first mover was selected at random. In markets with more firms it is harder to observe the same first mover for 10 weeks, but our findings also go in the direction of leadership not being random. For example, Figure 5b reports our findings for markets with 5 competitors. In these markets, if the first mover was chosen at random, the theoretical probability of never observing a leader during a 10-week window is about 15 percent whereas in the data it is about 3 percent. We observe similar patterns for markets with different numbers of stations.17 Finally, we also show that leaders last longer than what it would be expected if the player who chooses prices first each week is random. Figure 6 shows that, for example, in duopoly markets we would expect the same firm to be a leader for significantly less There are 2T −1 possible patterns of first movers. We use T = 10 for computational tractability. 17 Formally, a Kolmogorov-Smirnov test rejects that both distributions are the same. 16

12

than 0.5 consecutive weeks. However, in the data, price leaders last for almost 2.5 consecutive weeks on average. The main finding of this section is, then, that leaders last significantly more than what it would be expected if first movers were chosen at random. In other words, there is strong evidence that both first movers and price leaders in the Chilean gasoline industry are not random.

3.2.1

The Rise of Price Leaders and their Impact on Margins

Figure 7 shows that the probability of observing a price leader increases over time, when leadership is defined at both the station and brand levels. Importantly, leadership was relatively stable until October of 2012, when it started to increase rapidly. This is important because until July 2012, the policy had been implemented only in some parts of the country. Only after the the full implementation in July 2012, sanctions and monetary fines were used to enforce the policy. For this reason, in the remainder of paper, we focus our analysis starting in July 2012. Nonetheless, and something to which we will come back later on, the increase in leadership did not take place until the end of 2012, when the probability of observing a price leader starts increasing from 30 to almost 60 percent at the end of our sample.18 Our interpretation of this finding is that as stations seized the reduction in monitoring costs provided by the website, leadership became pervasive. Our findings so far—the first mover is not selected at random and leadership increased over time—are not direct evidence of an anti-competitive behavior. For instance, as managers of gas stations learned to use the website, some managers may have updated prices online persistently faster than others. However, if this was the case, we should observe no effect of leadership on equilibrium outcomes, e.g., margins. To test this hypothesis, we study whether average margins in markets with a leader differ from 18

Online Appendix C presents the same figures for the different fuel products. The increase in

leadership is common to all fuel products and the magnitudes are similar.

13

margins in markets without one. Thus, we estimate 0 δ + ξj + γm + ηt + εjmt , log(marginjmt ) = β · `jmt + Xmt

(1)

where the indexes j, m, and t, correspond to type of fuel (gasoline of 93, 95, and 97 octanes), market, and week, respectively. log(marginmjt ) is the natural logarithm of the average margin of product j in market m and week t. We define margins as the difference between retail prices and the wholesale price announced by ENAP each week. Hence, our measure of margins includes the margin of the distributor and the station. To control for week-to-week changes in distribution costs, we include the vector Xmt that contains the interaction between the mean distance from the stations in market m to the main pipeline in Santiago and oil prices.19 `jmt is an indicator that takes the value of one if product j in market m has a leader at time t. ξj is a product fixed effect that captures systematic differences in the level of margins for each product (gasoline of 93, 95, and 97 octanes). γm is a market fixed effect that controls for time-invariant market-specific characteristics—such as market structure, location, or income—that may cause systematic differences across markets. Finally, ηt is a week fixed effect that controls for changes in our measure of margins that are common across the country but change from week to week. The coefficient β represents the percentage change in the average margin at the market level if there was a leader in that market in week t. There are three main threats to identification of β. First, stations further away from the wholesale distributor mechanically have higher margins, given that our definition includes transportation costs. Most of this, however, is captured by a market fixed effect. To capture week-to-week variation in distribution costs that are not controlled for by market fixed effects, we explicitly include time-varying market-specific measures of distribution costs (the vector Xmt ). Hence, changes in our measure of margins that may arise from changes in distribution costs should not bias our estimates. Second, whether a market has a leader may be determined by market characteristics. 19

The wholesale price announced by ENAP is the price of each fuel product at this location. In

estimation, we also include two lags of the interaction between distance to the pipeline and oil prices, to control for trends in distribution costs. Our results are robust to including these lags or not.

14

We take this endogeneity concern into account by including a rich set of market fixed effects in all the regressions that we present below. Third, leadership could have been increasing across the country before the introduction of the website, in which case the cause of the increase in leadership would not be the introduction of mandatory disclosure (even if the website may have facilitated coordination afterwards). For example, if gas stations had been coordinating through phone calls (Clark and Houde, 2014) and we do not take this into account, our estimate of β would assign the impact of explicit collusion to the increase in leadership. There are, however, at least two reasons why we believe this type of argument is not a concern in our setting. One reason is that, as Figure 7 shows, leadership was flat from the moment the system was first introduced in the capital (March 2012) until three months after the full implementation of disclosure across the whole country (October 2012). This suggests that, though some leadership existed, the increase in leadership only happened after disclosure was fully-implemented in July of 2012. This is consistent with stations learning to use the website after the system was fully implemented. Another reason is that, as we will show below, leadership not only affected the level of margins but also the speed at which followers responded to changes in wholesale prices. As we show in detail in the following section, the speed of adjustment only changed among followers in oligopoly markets, who became significantly faster in adjusting their prices, but it did not change among either leaders in oligopoly markets or monopolists. Thus, we argue that because the website reduces the cost of monitoring rivals, it became a facilitator of tacit coordination. We report our estimates of Equation 1 in Table 1. The table shows that margins are higher when a leader is present, regardless of how we identify first movers (station, owner, or brand). The presence of a one-week leader (i.e., two weeks with the same first mover) is associated with approximately 0.6 percent increase in margins per liter sold. However, because leadership is a persistent feature rather than a one-time event, the results reported in Table 1 may underestimate the overall effect of leadership on margins. In other words, the longer a leader is in place, the larger the impact on 15

competition. To take persistence into account, we estimate the following variation of Equation 1 log(marginjmt ) =

T X

0 δ + ξj + γm + ηt + εjmt , `kmjt βk + Xmt

(2)

k=1

where `kmjt is an indicator that takes the value of one if for product j in market m the first mover in week t has been consecutively the first mover for the last k weeks. Thus, when T = 1 this equation corresponds precisely to Equation 1. In Equation 2, the coefficient βk represents the marginal contribution of an extra week with a leader, conditional the same leader for k − 1 weeks. The cumulative effect of consecutive leadership between weeks t − k and t is, then, the sum of the estimated coefficients associated with these k weeks. Because leadership could arise at the brand level rather than at station level (see footnote 15), in Figure 8 we report the cumulative effects on the average margin for the first 10 weeks with a leader, for players identified at both the station and brand level. The figure shows that, regardless of how players are defined, margins are higher in the presence of leaders. The figures, however, also show an interesting difference. In Figure 8a, players are defined at the station level and the figure shows that the effect of leadership on margins follows an inverse-U shape: the relationship is increasing and concave until the eighth week and decreasing afterwards. This suggests that the marginal contribution of an additional week with the same leader is quite significant during the first month of leadership but it declines afterward and even becomes negative after a relatively long period of continuous leadership by a single station. Figure 8b reports our findings for players defined at the brand level. In this case, the figure shows that margins increase persistently until the seventh week of leadership and remain stable afterwards. Overall, we find that the effect of leadership is an increase in margins by 1.5–2 percent during the first eight weeks with the same leader. Our findings also show that leaders not only increase the average margin, but the minimum margin as well. Figure 9 shows that the minimum margin increases by up 16

to 3 percent after eight weeks of continuous leadership, when players are defined using station identifiers. The results are slightly smaller when players are defined using brand identifiers. Overall, this suggests that in the presence of leaders, the whole distribution of margins shifts to the right. In Online Appendix D, we provide additional results for the maximum margin and for different measures of margin dispersion.

3.2.2

Speed of Price Adjustments

Leadership may not only affect the level of margins in a market, but also, for example, when firms make their pricing decisions. For this reason, we now ask: is the speed of price adjustments affected by the presence of leaders? To answer this question, we proceed in two steps. First, Figure 10 reports the average number of hours between when ENAP announces a price change and when stations change their prices, dividing the sample into monopolists, first movers, and followers. The figure shows that, after the system was fully implemented, the timing of prices changes remained stable for monopolists and first movers, but that, over time, followers began to change their prices faster. Second, we estimate a version of Equation 2 where we replace the dependent variable by the average number of hours elapsed between when the first mover changes its price following the announcement by ENAP and when the rest of the firms in the market/week adjust their prices. As before, the effect of time-since-the-policy-implementation on the speed of response, which is common to all gas stations, is captured by week fixed effects. Similarly, any idiosyncratic market condition that affects the speed of response is captured by market fixed effects. We report our findings in Figure 11. The figure shows that stations adjust their prices significantly faster in the presence of leaders: follower stations react, on average, between one to three hours earlier when a leader is present. Finally, in Online Appendix D, we study whether leadership has an asymmetric impact on margins depending on whether wholesale prices increase or decrease. We find 17

that leaders have a greater impact on margins and on the speed of adjustment when wholesale prices increase relative to when they decrease, with margins increasing more and stations reacting faster when wholesale prices increase.

3.3

Sanctions and their Impact on Pricing Strategies

The information that the government provides to consumers through the website is useful only if gas stations report their actual prices in real-time. Mismatches between online prices and those charged at the pump (offline prices) can be classified in two categories. First, managers may either forget to update their prices online after changing prices at the pump or, while the manager may be aware of the requirement of reporting, she may not think it is worth updating prices in a timely manner. This may happen because the manager believes consumers do not use the website or because enforcement is rare. Second, stations may strategically choose to misreport prices online as this may allow them to attract consumers in markets where consumers monitor prices online. To prevent this from happening, SEC encourages consumers to report price mismatches and it also continuously inspects gas stations. If it is determined that a station misreported its price, SEC sanctions it.20 We call skipping the price mismatches events where gas stations do not report price changes on the website in a timely manner; we call misreporting the price mismatches events where gas stations strategically decide report online prices that are different to those charged at the pump. Because skipping or misreporting have different implications for competition, in this section we first study which gas stations are sanctioned and why. We then turn to study how sanctions affect post-sanction pricing behavior. 20

Only between July and October, 2012, SEC made more than 1,700 inspections across the country.

Source (accessed on July 14th, 2017): http://www.sec.cl/portal/page? pageid=33,4883642& dad=portal& schema=PORTAL

18

For this analysis, we obtained from SEC the address of each gas station that was sanctioned, as well as the date of each sanction. In our data, there are 274 sanctions between July 2012 and October 2014. The distribution of sanctions is roughly the same across different regions of the country—about 20% of the stations in any region of the country have received a sanction. Furthermore, the distribution of sanctions over time is also relatively stable, with an average of two sanctions per week. (Figure 12). The only exception is the second week of January 2013, when SEC issued a record number of 78 sanctions. Finally, most stations were either never sanctioned or sanctioned only once. At the brand level, the four largest brands are jointly responsible for 249 out of the 274 sanctions.21 Unfortunately, our data does not allow us to distinguish between skipping or misreporting, so we need to infer it from our data.22

3.3.1

Sanctions and the Probability of Updating Prices Online

Figure 13 shows the percentage of weeks in which a gas station changed prices online, relative to the total number of weeks that the station is observed in our sample. The figure shows that most gas stations report new prices every week—the median percentage of weeks in which stations change prices is 93 percent—but some gas stations report less than 80 percent of the weeks they are observed in the sample. Figure 14 shows the distribution of the number of times stations failed to update their prices online. Although the median number of price skips on a given week is 107, there are some weeks in which a large number of stations do not update their prices. In these weeks, there was a small wholesale price change relative to the previous week, so gas stations kept the same retail price for two consecutive weeks.23 Thus, we can conclude that most stations are quite active in reporting prices online. 21 22

The distribution is Copec (86), Petrobras (73), Shell (59), Terpel (31). SEC informed us that their records do not show why a station was sanctioned, but only that a

sanction exists and they day in which the sanction took place. 23 The dates and the percentage change in wholesale price, relative to the previous week, are: 07/26/2012 (0.045%), 9/27/2012 (0.03%), 12/06/2012 (-0.015%), 1/17/2013 (-0.15%), and 10/17/2013 (0.1%). The average wholesale price change over the whole sample is 1.06%.

19

If the stations that do not update their prices online after changing them at the pump did so because they forgot to do so, rather than because they strategically misreported their prices, then we should observe that after being sanctioned, stations start reporting prices online more actively. To determine whether this is the case, we estimate the following equation Pr(skipst ) = β · sanctionst + γs + ηt + εst ,

(3)

where s indicates a station and t indicates a week. skipst is an indicator that takes the value of one if gas station s did not report a new price on week t.24 If gas station s was sanctioned at time t¯s , the variable sanctionst takes the value of zero for t ≤ t¯s and one for t > t¯s . γs is a gas station fixed effect and ηt is a time fixed effect. Table 2 reports the estimate of β in Equation 3 and shows that gas stations that are sanctioned are less likely to skip reporting after the sanction. Thus, sanctions have been effective in inducing firms to report their prices in a timely manner more frequently.

3.3.2

Sanctions and Consumer-Search Behavior

Having shown that sanctions induce stations to post new prices online more often, we now turn to study whether sanctions are enough to persuade gas stations to truthfully report prices online. What are the gas stations’ incentives to misreport? If consumers do not use the website to search, it is dominant for gas stations to truthfully report prices online. This is so because misreporting prices online would not attract additional consumers but it would increase the risk of being sanctioned. However, if consumers decide where to purchase gasoline based on the online prices reported on the website, stations face a trade-off between reporting low prices online to attract consumers that search and increasing the probability of being sanctioned. If sanctions are not harsh or if the probability of detection is low, gas stations may find 24

The website reports prices as of the last time prices were updated by a station manager. Hence,

if a manager does not update prices in a specific week, the website still reports a price for that station but the price will not coincide with the price charged at the pump.

20

it advantageous to misreport prices to attract those consumers who search.25 However, the probability of detection may increase in areas with high consumer search, because consumers can inform SEC about mismatches between online and offline prices. At the same time, gas stations also have higher incentives to misreport in markets that have both more stations and higher search intensity. Consumers, on the other hand, may benefit from searching more in markets with more gas stations. Therefore, misreporting prices and consumer search create a coordination game: consumers only want to search when firms report truthfully, but when consumers search, firms have incentives to misreport. We investigate the relationship between the probability of receiving a sanction and market characteristics such as number of firms and search intensity. We measure search intensity as the number of search requests through the smartphone app. We estimate several versions of the following equation Pr(sanctions ) = β · requests 3kms + γ · n3kms + Brands + εs ,

(4)

where s indicates a station and sanctions is an indicator that takes the value of one if station s was ever sanctioned. requests 3kms is the (standardized) total number of search requests executed within a radius of 3 kilometers from gas station s. n3kms is the number of gas stations located within 3 kilometers from station s (as before, we use driving distances). Brands is the brand of gas station s. Columns 1 and 2 in Table 3 report the estimate of β and show that stations located in areas with higher consumer search are more likely to be sanctioned. The average marginal effect implies that a one standard-deviation increase in the number of requests executed within 3 kilometers of a station, increases the probability of a station being sanctioned at least once by 10 percent. As mentioned above, these findings may be explained by two (complementary) reasons. First, in areas where consumers search more, authorities 25

The average fine until November 2014 was of $113,000 Chilean pesos, or around $170 U.S. dollars

of May 2017. Source (accessed on July 31st, 2017): http://www.sernac.cl/sernac-sec-y-cne-vigilaran-que-estaciones-de-servicio-bajen-precios-anunciadospor-gobierno/.

21

may receive a larger number of complaints which could lead to more sanctions. Second, gas stations located in areas with high rates of consumer search (and with more rivals) face greater incentives to misreport prices than those stations located in areas with little search. Equation 4 studies the connection between the total number of requests executed around a station over our whole sample, and the likelihood of being sanctioned. Equation 4, however, does not allow for week-to-week variation in search. To capture this demand from searchers around the date of the sanction, we estimate Pr(sanctionst ) = β · requests 3kmst + γ · n3kms + Brands + εst ,

(5)

where sanctionst is an indicator that takes the value of one if station s was sanctioned in week t. requests 3kmst is the number of search requests during week t within a radius of 3 kilometers from the gas station. n3kms is the number of gas stations located within 3 kilometers from station s. Brands is the brand of gas station s. Columns 3 and 4 in Table 3 reports the estimate of β. Again, the estimates are positive but only the estimate in column 4 is significant. The results show that a one standarddeviation increase in the number of search requests executed within the week and within 3 kilometers of a station, increases the probability of a sanction.

3.3.3

Impact of Sanctions on Pricing Behavior

We finish this section studying how sanctions affect the pricing behavior of sanctioned stations. We use a difference-in-difference approach to compare the pricing behavior of sanctioned stations versus stations in the same market that were not sanctioned. Formally, we estimate log(margin)jmt = α + βSmt · Ij + γt + ξj + εjmt ,

(6)

where log(margin)jmt corresponds to the natural logarithm of the margin of station j in market m in week t, Smt is an indicator that takes the value of one after the first 22

sanction in market m and zero otherwise, Ij is an indicator that takes the value of one station j was ever sanctioned, γt is a week fixed effect, and ξj is a station fixed effect. The coefficient β measures the differential impact of a sanction on sanctioned stations relative to stations that were not sanctioned. We consider the first sanction of each station as the event of interest, because most stations are sanctioned only once. As discussed in Section 3.3.1, some stations were sanctioned most likely because they skipped posting new prices, rather than because they strategically misreported their prices. To focus on sanctions due to misreporting, we estimate Equation 6 for different subsets of gas stations characterized by how active they were in reporting prices, according to their position in the distribution presented in Figure 13. We cluster standard errors at the market level. Table 4 presents the estimates of Equation 6. Panel A uses the whole sample of weeks whereas Panel B excludes the week of 1/17/2013—when an abnormally large number of stations were sanctioned. Columns 1 and 7 consider all the observations. Columns 2 and 8 consider only those gas stations that had reported prices for two consecutive weeks before the sanction (to avoid considering cases in which sanctions may have happened because a station’s manager forgot to update prices online). In the other columns, we only include stations according to the percentage of weeks in which they reported price changes: more than 25 percent (columns 3 and 9), more than 50 percent (columns 4 and 10), more than 75 percent (columns 5 and 11), and more than 90 percent (columns 6 and 12). The results in the table show that after the first sanction in a market, sanctioned stations increase their margins relative to those non-sanctioned stations in the same market. The increase in margins is about 3 percent, except for the most active stations (columns 6 and 12) where the effect is 1.8 percent. Given that we only observe reported prices, these results should not be interpreted as an increase in the profitability of the sanctioned stations after the sanction, but rather that their reported prices increase relative to those of non-sanctioned stations after a sanction. On average, we observe that sanctioned stations report higher prices (and

23

we observe higher margins) after the sanction, relative to non-sanctioned stations. If stations report truthfully after the sanction, and this is reflected in higher observed margins, we infer that the reason of the sanction was that gas stations must have under-reported their prices online before they were sanctioned. Finally, Figure 15 shows the estimated impact of sanctions on margins, decomposing the effect in the six weeks before and after a sanction. The figure is based on the sample used in columns 2 and 8 in Table 4 and excludes the week of 1/17/2013.26 From the figure, we extract three insights. First, one week before a station is sanctioned, it reports lower prices than stations in the same market that are not sanctioned in the week that follows. This is evidence that stations are sanctioned because they under-reported prices on the website. Interestingly, this effect shows up only in the week before the sanction and not earlier, which suggests that misreporting is quickly observed and sanctioned by SEC. If this was not the case, we would observe negative estimates in more of the weeks before a station is sanctioned. This suggests that SEC is able to quickly detect stations that misreport prices. Second, we observe that stations that are sanctioned charge higher prices than their non-sanctioned rivals both in the week when they are sanctioned and in the week that follows the sanction. These effects could be explained by stations changing prices after the sanction to pass the cost of the sanction on to the consumers.27 Finally, and interesting from a policy perspective, the figure reveals that after the first week following a sanction, stations revert to charging the same prices as their rivals that were not sanctioned. This suggests that sanctions are effective in inducing stations that have mis-reported prices, to report their prices truthfully. 26 27

The omitted dummy corresponds to six weeks before the sanction took place. In our data, 43 percent of gas stations change their prices shortly after the sanction, before a new

week starts.

24

4

Discussion

Mandatory information-disclosure policies are meant to encourage consumer search and to intensify competition, but in practice these objectives are not always achieved. Following a price-disclosure policy in the Chilean retail-gasoline industry, we study firms’ pricing strategies during the two and a half years that followed the policy implementation. We focus on three dimensions of pricing strategies: the timing and frequency of price changes, the rise of price leadership, and the strategic misreporting of prices through the disclosure platform. We find no evidence indicating that the policy intensified competition. On the contrary, our findings indicate a decreased the intensity of competition. Specifically, we find that first movers are not random and that price leaders are more likely to appear and persist in markets with fewer firms. Furthermore, price leaders became more pervasive over time and their presence resulted in higher margins and faster reactions to changes in wholesale prices. We also argue that price disclosure allowed stations to post lower prices online than those charged at the pump, in an attempt to attract consumers that search for low prices. However, we find that enforcement measures meant to guarantee compliance with the disclosure policy have had a positive effect. Indeed, gas stations sanctioned for misreporting abandoned the practice soon after being sanctioned. Finally, we show that gas stations that were sanctioned for not update their prices online consistently became more complaint after they were sanctioned. The main contribution of this paper is to shed light on some of the potential anticompetitive effects of information-disclosure policies, such as price leadership and strategic misreporting. Because mandatory information-disclosure policies have the potential to both benefit and harm consumers, the implementation of these policies should carefully consider how consumers and firms will respond to the disclosure of information, and strict enforcement measures are necessary to prevent firms from using the policy in their own benefit.

25

References Albæk, Svend, Peter Møllgaard, and Per B. Overgaard. 1997. “GovernmentAssisted Oligopoly Coordination? A Concrete Case.” The Journal of Industrial Economics, 45(4): 429–443. Ater, Itai, and Oren Rigbi. 2017. “The Effects of Mandatory Price Disclosure of Supermarket Prices.” working paper. Busse, Meghan R. 2000. “Multimarket contact and price coordination in the cellular telephone industry.” Journal of Economics & Management Strategy, 9(3): 287–320. Byrne, David P, and Nicolas de Roos. 2017. “Learning to coordinate: A study in retail gasoline.” Campbell, Colin, Gautam Ray, and Waleed A. Muhanna. 2005. “Search and Collusion in Electronic Markets.” Management Science, 51(3): pp. 497–507. Clark, Robert, and Jean-Fran¸cois Houde. 2013. “Collusion with Asymmetric Retailers: Evidence from a Gasoline Price-Fixing Case.” American Economic Journal: Microeconomics, 5(3): 97–123. Clark, Robert, and Jean-Fran¸cois Houde. 2014. “The Effect of Explicit Communication on pricing: Evidence from the Collapse of a Gasoline Cartel.” The Journal of Industrial Economics, 62(2): 191–228. Daughety, Andrew F, and Jennifer F Reinganum. 2008. “Communicating quality: a unified model of disclosure and signalling.” The RAND Journal of Economics, 39(4): 973–989. Deltas, George. 2008. “Retail gasoline price dynamics and local market power.” The Journal of Industrial Economics, 56(3): 613–628. Deneckere, Raymond J, and Dan Kovenock. 1992. “Price leadership.” The Review of Economic Studies, 59(1): 143–162.

26

Eckert, Andrew. 2013. “Empirical studies of gasoline retailing: A guide to the literature.” Journal of Economic Surveys, 27(1): 140–166. Huber, Stephan, and Christoph Rust. 2016. “Calculate travel time and distance with OpenStreetMap data using the Open Source Routing Machine (OSRM).” Stata Journal, 16(2): 416–423. Jang, Youngjun. 2014. “How Do Technological Advances Affect Consumer Search: Evidence from the Korean Gasoline Market.” Jin, Ginger Zhe, and Phillip Leslie. 2003. “The Effect of Information on Product Quality: Evidence from Restaurant Hygiene Grade Cards.” The Quarterly Journal of Economic, 118(2): 409–451. Kauffman, Robert J, and Charles A Wood. 2007. “Follow the leader: price change timing in internet-based selling.” Managerial and Decision Economics, 679–700. Kim, Donghun, and Jiyon Lee. 2014. “Spatial Price Competition in the Korean Retail Gasoline Market.” Environmental and Resource Economics Review, 23(4): 553– 581. Lewis, M. 2008. “Price Dispersion and Competition with Differentiated Sellers.” The Journal of Industrial Economics, LVI(3): 654–679. Lewis, Matthew, and Michael Noel. 2011. “The Speed of Gasoline Price Response in Markets With and Without Edgeworth Cycles.” The Review of Economics and Statistics, 93(May): 672–682. Lewis, Matthew S. 2011. “Asymmetric price adjustment and consumer search: An examination of the retail gasoline market.” Journal of Economics & Management Strategy, 20(2): 409–449. Lewis, Matthew S. 2012. “Price leadership and coordination in retail gasoline markets with price cycles.” International Journal of Industrial Organization, 30(4): 342– 351. 27

Luco, Fernando. 2017. “Who Benefits from Information Disclosure? The Case of Retail Gasoline.” Texas A&M University, working paper. Markham, Jesse W. 1951. “The nature and significance of price leadership.” The American Economic Review, 891–905. Maskin, Eric, and Jean Tirole. 1988. “A theory of dynamic oligopoly, II: Price competition, kinked demand curves, and Edgeworth cycles.” Econometrica: Journal of the Econometric Society, 571–599. Mathios, Alan D. 2000. “The Impact of Mandatory Disclosure Laws on Product Choices: An Analysis of the Salad Dressing Market.” Journal of Law & Economics, XLIII: 651–678. Noel, Michael D. 2007a. “Edgeworth price cycles, cost-based pricing, and sticky pricing in retail gasoline markets.” The Review of Economics and Statistics, 89(2): 324– 334. Noel, Michael D. 2007b. “Edgeworth price cycles: Evidence from the Toronto retail gasoline market.” The Journal of Industrial Economics, 55(1): 69–92. Perdiguero, Jordi, and Joan-Ramon Borrell. 2012. “Driving competition in local gasoline markets.” Rossi, Federico, and Pradeep K Chintagunta. 2015. “Price Transparency and Retail Prices: Evidence from Fuel Price Signs in the Italian Motorway.” Journal of Marketing Research. Rotemberg, Julio J, and Garth Saloner. 1990. “Collusive price leadership.” The Journal of Industrial Economics, 93–111. Schultz, Christian. 2005. “Transparency on the consumer side and tacit collusion.” European Economic Review, 49(2): 279–297. Schwartz, Alan, and Louis L Wilde. 1978. “Intervening in markets on the basis of imperfect information: A legal and economic analysis.” U. Pa. L. Rev., 127: 630. 28

Seaton, Jonathan S, and Michael Waterson. 2013. “Identifying and characterising price leadership in British supermarkets.” International Journal of Industrial Organization, 31(5): 392–403. Stahl, Dale O. 1989. “Oligopolistic Pricing with Sequential Consumer Search.” The American Economic Review, 79(4): pp. 700–712. St¨ uhmeier, Torben. 2015. “Price disclosure rules and consumer price comparison.” The BE Journal of Economic Analysis & Policy, 15(2): 815–835. Varian, Hal R. 1980. “A model of sales.” The American Economic Review, 651–659.

29

5

Figures Figure 1: Number of daily requests over time 20000

Number of search requests per day

1000

Number of users per day

15000

10000

5000

0

500

0

01jan2013

01jul2013

01jan2014 01jul2014 Date The horizontal dashed line reports the median number of visits per day: 947

01jan2015

01jul2012

01jan2013 01jul2013 01jan2014 01jul2014 Date The horizontal dashed line reports the median number of requests per day: 377

(a) CNE data

01jan2015

(b) App data

Note: The figure reports the number of search requests executed directly through the website (upper panel) or through a smartphone app (lower panel) over time.

30

Average number of price changes per week

Figure 2: Number of price changes per week and by day of the week 1.5

1.3

1.1

.9

.7

.5 0

50

100

150

Week in sample since March, 2012

(a) Number of changes per week

Number of price changes

1000

800

600

400

200

0 0

50

100

150

Week in sample since March 1st, 2012 Wed 19:00−Thursday

Friday to Wednesday 18:59

The vertical red line marks the day when the system was implemented across the whole country (July 1st, 2012).

(b) Number of price changes by day of the week

Note: Figure 2a reports a local polynomial of the number of price changes per station and week, over the whole sample. the horizontal red lines mark the interval that bounds the number of price changes per week. Figure 2b reports local polynomials of the number of price changes by day of the week, and classifies the data in two groups: price changes that happen between Wednesday at 7pm (just after ENAP’s announcement) and midnight on Thursday, and price changes that happen between Friday and Wednesday before 7pm in the following calendar week. In both figures, the vertical red lines mark the day when the system was implemented and enforced across the whole country.

31

Figure 3: Timing of price changes 60

Hours after ENAP’s announcement

Hours after ENAP’s announcement

60

50

40

30

20

50

40

30

Monopolists Oligopoly markets

20 Jan12

Jul12

Feb13

Aug13

Mar14

Date

Jan12 Oct14

Jul12

Feb13

Aug13

Mar14

Oct14

Date

Quadratic fit including station fixed effects.

Linear fit including station fixed effects.

(a) All price changes

(b) Monopolies and Oligopolies

Note: Figure 3a reports the number of hours elapsed between ENAP announces a change in wholesale prices and when stations change their prices. Figure 3b replicates Figure 3a separating the sample into monopoly and oligopoly markets, according to the definition introduced in section 3.2. The figures are binned scatterplots with associated linear fits that follow from regressions that includes station fixed effects to control for station-specific time-invariant characteristics. In both figures, the vertical red lines mark the day when the system was implemented and enforced across the whole country.

32

Figure 4: Weeks with a Leader 6

7

6

5

Station ID

Station ID

5 4

3

4

3 Copec

2

Copec 2

Petrobras Shell Terpel

1 July12

January13

July13

January14

Petrobras Shell Terpel

1

July14

July12

January13

July13

Date Note: 6−station market.

July14

Note: 7−station market.

(a) Market 468

(b) Market 483

9

9

8

8

7

7

6

6

Station ID

Station ID

January14

Date

5 4

5 4

3

3 Copec

Copec

Petrobras

2 1 July12

January13

July13

January14

Petrobras

2

Shell Terpel

Shell Terpel

1

July14

July12

January13

July13

Date

January14

July14

Date

Note: 7−station market.

Note: 9−station market.

(c) Market 646

(d) Market 705

Note: The figures show leadership events for four different markets with different market structure. In each plot, time (weeks) is in the horizontal axis and each gas station is represented in the vertical axis. For each week, a marker shows which station (if any) was a leader. Different shapes in the market represent different brands. For instance, Copec is represented by a circle whereas Petrobras by a square marker. Figure 4a shows a market with six stations where only two of them were ever leaders. In this figure, leadership switched from gas station 1 (Copec brand) to gas station 4 (also Copec) around May 2013. Similarly, Figure 4b shows that in a seven-station market, leadership became more common starting in June 2013, and it was driven by a single Copec station. Figure 4c shows something different. In this case, the market has seven stations and four of these, of three different brands, were leaders at different moments in time. Finally, Figure 4d shows the case of a market with nine stations that did not have a clear leader until March 2013, when a Copec station became the leader for the rest of our sample period.

33

Figure 5: Distribution of weeks with leaders conditional on the number n of firms in a market. .25

.3

.2

Probability

Probability

.2 .15

.1

.1 .05 Empirical Random

0 0

2

4

6

8

10

Empirical Random

0 0

Consecutive weeks as a leader

2

4

6

8

Consecutive weeks as a leader

(a) n = 2

(b) n = 5

Note: The figure reports the distribution of the number of consecutive weeks with a leader for markets with different market structure. The figure reports both the theoretical distribution derived assuming that leadership is random and the empirical distribution from the data. The figure shows that the distribution are not the same, which is confirmed by a Kolmogorov-Smirnov test.

34

10

Figure 6: Expected duration of a leader (defined at the station id level)

Consecutive weeks as a leaders

2.5

2

1.5

1

.5 Realized (data) Expected (random)

0 0

5

10

15

20

Firms in the Market

Note: The figure reports the expected and realized number of consecutive weeks with a leader, for markets with different market structure. The expected number of weeks is derived assuming that leadership is random.

Figure 7: Leadership by product and definition .8

Markets and leadership

.7

.6

.5

Leadership based on Brand

.4

Station January12

July12

February13

August13

March14

October14

Local polynomial smoothing with a 95% confidence interval. The vertical red line marks the day when the system was implemented and enforced in the whole country. We focus on gasoline of 93 octanes because for all other fuel products the figures look qualitatively identical (we report these in the Online Appendix).

35

2

2

1.5

1.5

Percentage change

Percentage change

Figure 8: Cumulative effect of leadership on mean margins at the market level

1

1

.5

.5

0

0 1

2

3

4

5

6

7

8

9

10

1

Weeks with same leader Regressions exclude monopoly markets

2

3

4

5

6

7

8

9

10

Weeks with same leader Regressions exclude monopoly markets

(a) Leadership at Station level

(b) Leadership at Brand level

Note: In Figure 8a, price leaders are based on station identifiers. In Figure 8b, price leaders are based on brand identifiers. The plots correspond to the sum of the estimated coefficients (confidence intervals adjusted accordingly). In Online Appendix D, we include the impact of leadership for price leaders defined at the owner level as well. The underlying regressions include week, market, and product fixed effects.

Figure 9: Cumulative effect of leadership on the minimum margin at the market level 3

Percentage change

Percentage change

3

2

1

2

1

0

0 1

2

3

4

5

6

7

8

9

10

Weeks with same leader Regressions exclude monopoly markets

1

2

3

4

5

6

7

8

9

10

Weeks with same leader Regressions exclude monopoly markets

(a) Leadership at Station level

(b) Leadership at Brand level

Note: In Figure 9a, price leaders are based on station identifiers. In Figure 9b, price leaders are based on brand identifiers. The plots correspond to the sum of the estimated coefficients (confidence intervals adjusted accordingly). In Online Appendix D, we include the impact of leadership on the distribution of margins as well. The underlying regressions include week, market, and product fixed effects.

36

Figure 10: Time elapsed between the change in wholesale prices and when station change their prices, by type of station Hours after ENAP’s announcement

60

50

40

30 Monopolists First movers 20

Followers Jan12

Jul12

Feb13

Aug13

Mar14

Oct14

Date Linear fit including station fixed effects.

Note: The figure reports the number of hours elapsed between ENAP’s announcement of a new wholesale price (Wednesday at 7pm) and when stations change their retail prices, as a function of station type (first mover or follower) in oligopolies, and also for monopolies, over time. The figure is a binned scatterplot with the associated linear fit that follows from a regression that includes station fixed effects to control for station-specific time-invariant characteristics. In the figure, the vertical red line marks the day when the system was implemented and enforced across the whole country.

Figure 11: Cumulative effect of leadership on the speed of price adjustments (minutes) 0

0

−50

Minutes

Minutes

−50

−100

−100 −150

−150

−200 1

2

3

4

5

6

7

8

9

10

Weeks with same leader Regressions exclude monopoly markets

1

2

3

4

5

6

7

8

9

10

Weeks with same leader Regressions exclude monopoly markets

(a) Brand leadership

(b) Station leadership

Note: The plots correspond to the sum of the estimated coefficients (confidence intervals adjusted accordingly). The estimates are available from the authors.

37

Figure 12: Number of stations sanctioned per week

Number of Sanctions

80

60

40

20

0 20

40

60

80

100

120

Week in sample since July 1st, 2012

Note: The figure reports the number of stations sanctioned per week, for weeks in which at least one station was sanctioned. The unconditional average number of sanctions per week in the sample is two and the median is zero.

Figure 13: Distribution of the number of weeks each station changed prices relative to the number of weeks the station was present in the data

Percentage of stations

.3

.2

.1

0 0

.2 .4 .6 .8 Percentage of weeks with price changes relative to the total number of weeks in sample

1

Note: The figure reports a measure of how active each firm is in reporting new prices in the website. This measure is computed as the ratio between the number of weeks in which a station changed prices at least once and the number of weeks elapsed between the first and last week in which that station was observed in the dataset.

38

Figure 14: Number of Stations skipping a price change for a given week.

Number of price skips

1000

800

600

400

200

0 0

50

100

Week in sample since July 1st, 2012

Note: The figure reports the number of stations that skip updating their prices for each week. The five observations at the top of the figure correspond to weeks in which the change in wholesale prices was negligible relative to the mean change over the sample period.

Figure 15: Impact of sanctions on margins 2

Impact on Margins Percentage

1

0

−1

−2 −5

−4

−3

−2

−1 0 1 2 Week relative to sanction

3

4

5

6

The figure reports the estimates of a linear regression around the time of a sanction. For each market, we define 0 as the week of a sanction and estimate the impact of a sanction on sanctioned stations over the six weeks before and after the sanction, relative to stations that were not sanctioned. The omitted category corresponds to six weeks before the sanction takes place. The regression only considers stations that did not skip a week at the moment of the sanction, and also exclude the week of 1/17/2013 because of the abnormal number of sanctions that took place that week. Standard errors are clustered at the market level.

39

6

Tables Table 1: Impact of Leadership on Mean Margins (1)

(2)

(3)

Brand

Owner

Station

0.686∗∗∗

0.586∗∗∗

0.575∗∗∗

(0.0900)

(0.0730)

(0.0738)

Market FE

Yes

Yes

Yes

Week FE

Yes

Yes

Yes

Product FE

Yes

Yes

Yes

65.15

65.15

65.15

N

377487

377487

377487

R2

0.746

0.746

0.746

Leader

Mean margins (CLP$ per liter)

Note: Standard errors, in parentheses, are clustered at the market level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. The dependent vari-

able in all regressions is the average margin at the market level, with margins measured in Chilean pesos per liter. The difference across specifications is how leaders are identified (i.e., brand, owner, or station identifiers).

40

Table 2: Effect of Sanctions on the Probability of Skipping a Week Posting a New Price (1)

(2)

(3)

(4)

-1.612∗∗∗

-1.441∗∗∗

-1.685∗∗∗

-1.238∗∗∗

(0.1543)

(0.1596)

(0.1703)

(0.1838)

Station FE

No

No

Yes

Yes

Week FE

No

Yes

No

Yes

175150

172050

173229

173229

-55539.75

-46824.35

-46381.29

-36451.711

-0.076

-0.067

-0.344

-0.0049

Sanction

Observations log(likelihood) Average marginal effect

Note: Standard errors, in parentheses.

∗

p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01.

Columns 1 and 2 correspond to Logit regressions. Columns 3 and 4 correspond to Fixed-Effect Logit regressions. All columns estimated by Maximum Likelihood. In all specifications, the dependent variable is equal to one if a station skips a week and zero otherwise. The variable sanction takes the value of one starting when a station is sanction and remains equal to one afterwards. An observation is a station-week combination.

41

Table 3: Effect of the Number of Search Requests within 3 Kilometers on the Probability of Probit Regression on Sanction

Requests within 3k

(1)

(2)

0.5059∗∗∗

0.4560∗∗∗

(0.0743)

(0.0754)

Requests within 3k by week

Number of rival stations within 3 k

log(likelihood) Average marginal effect Note: Standard errors, in parentheses.

(4)

0.0163

0.0568∗∗∗

(0.0204)

(0.0179)

0.0111

0.0134

0.0207∗∗∗

0.01934∗∗∗

(0.0084)

(0.0084)

(0.0038)

(0.0039)

No

Yes

No

Yes

1454

1313

136685

131409

-551.08

-1797.46

-1939.89

-448.925

0.105

0.101

0.00009

0.0002747

Brand FE Observations

(3)

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01. All

specifications correspond to Probit regressions estimated by Maximum Likelihood. In columns 1 and 2, the dependent variable is equal to one if a station was ever sanctioned, and the dependent variables corresponds to the total number of search requests that were ever executed within 3 kilometers of a station and the number of rival stations within the same driving distance. An observation is a station. In columns 3 and 4, the dependent variable is equal to one if a station was sanctioned in a particular week and zero otherwise. Search requests are also measured at the week level. An observation is a station-week combination.

42

Table 4: Impact of Sanctions on Pricing Behavior log(margin) Panel A (all weeks) Interaction

(1)

(2)

(3)

(4)

(5)

(6)

0.0297∗∗∗

0.0286∗∗∗

0.0297∗∗∗

0.0298∗∗∗

0.0298∗∗∗

0.0179∗∗∗

(0.00322)

(0.00322)

(0.00322)

(0.00323)

(0.00324)

(0.00409)

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Week FE

Yes

Yes

Yes

Yes

Yes

Yes

Station FE

Yes

Yes

Yes

Yes

Yes

Yes

58.83

58.77

58.83

58.82

58.68

59.21

N

1395110

1382414

1395074

1394364

1355286

910519

R2

0.828

0.826

0.828

0.828

0.828

0.826

(10)

(11)

(12)

Mean margins (CLP$ per liter)

log(margin) Panel B: Excludes week of 1/17/2013 Interaction

(7)

(8) ∗∗∗

(9) ∗∗∗

∗∗∗

∗∗∗

∗∗∗

0.0301

0.0288

0.0301

0.0302

0.0302

0.0181∗∗∗

(0.00326)

(0.00325)

(0.00326)

(0.00327)

(0.00328)

(0.00414)

Cost controls

Yes

Yes

Yes

Yes

Yes

Yes

Week FE

Yes

Yes

Yes

Yes

Yes

Yes

Station FE

Yes

Yes

Yes

Yes

Yes

Yes

Mean margins (CLP$ per liter)

58.86

58.70

58.86

58.85

58.71

59.24

N

1384236

1372738

1384200

1383498

1344556

903558

R2

0.828

0.826

0.828

0.828

0.828

0.826

Note: Standard errors, in parentheses, are clustered at the market level.

∗

p < 0.10,

∗∗

p < 0.05,

∗∗∗

p < 0.01.

Across columns, the dependent variable is the natural logarithm of the margin of a station for gasoline of 93 octanes. An observation is a station-market-week triple. The panels differ on which weeks are included in the sample. Panel A includes all weeks, while Panel B excludes the week of 1/17/2013, during which an abnormal number of stations were sanctioned. Specifications differ on the sample of stations considered according to how active the stations were in reporting (new) prices to the website, as measured in Figure 13. Columns 1 and 7 consider all stations. Columns 2 and 8 exclude stations that reported new prices less than 25 percent of the weeks in which they could have reported new prices to the website. Columns 3 and 9 exclude stations that reported new prices less than 50 percent of the weeks when they could have done so. Columns 4 and 10 exclude stations that reported new prices less than 75 percent of the weeks when they could have reported new prices. Columns 5 and 11 exclude stations that reported new prices less than 90 percent of the weeks when they could have reported new prices. Finally, columns 6 and 12 exclude stations that skip posting a price in the week previous to the sanction.

43