Price Competition and Market Segmentation in Retail Gasoline: New Evidence from South Korea* Taehwan Kim† March 2016 Abstract I examine price dispersion in retail gasoline focusing on differentiation along the service dimension (full service vs. self service). Consistent with more intensive search by self-service customers, I find lower price dispersion among densely located self-service stations, but no clear relationship for full-service stations. I also show that higher self-service station density is correlated with lower price dispersion for full-service stations. I observe similar patterns when I segment the market by brand, but the estimates are sensitive to how brands are separated into different types, suggesting that the market is more clearly segmented by service level than by brand. JEL Codes: D22, L11, L81 Keywords: price dispersion, retail gasoline, market segmentation, self service, full service, differentiated sellers, consumer search * Comments welcome. I am extremely grateful to Emek Basker for her guidance and encouragement. I am also thankful to Peter Mueser, Shawn Ni, David Mandy, Cory Koedel, David Kaplan, Kyungsik Nam, Li Tan, and seminar participants at the 33rd Research and Creative Activities Forum at University of Missouri for helpful comments and conversations. † Department of Economics, University of Missouri, Columbia, MO 65201 USA. Email: [email protected] 1. Introduction Price dispersion is ubiquitous even in markets for seemingly homogeneous goods. One common explanation for price dispersion is that consumers are not fully informed about the price distribution because obtaining this information is costly. In addition, markets may be segmented in ways unobservable to the econometrician, so that seemingly homogeneous goods are in fact perceived by consumers as heterogeneous. In this paper, I exploit the recent transition of gasoline stations in Seoul, South Korea, from full service to self service to jointly examine search and the presence of segmentation by service level. I find that self-service stations exhibit greater sensitivity to competition than full-service stations, consistent with higher search intensity among self-service customers. Although I find some evidence of segmentation by brand as well as service level, the evidence for segmentation by brand is weaker and less robust than the evidence for service-level segmentation. Take together, these results highlight the importance of identifying market features important to consumers when testing economic theories. Existing studies of gasoline pricing have often used stations’ brands to identify sellers of different types. For example, Hastings (2004) investigates the price effects of brand contract changes from an independent retailer to ARCO in Southern California, by brand group (high-share, mid-share, or low-share brand); and Lewis (2008) studies panel data from San Diego retail gasoline market and finds that the relationship between number of sellers and price dispersion varies depending on the brand composition of local competitors (high-brand or low-brand stations). These papers assume that demand is segmented by brand type; in other words, that the cross-price elasticity of demand across brands is low. This may not be satisfied in retail gasoline. In addition, brand segmentation may vary with market environments, brand identity may be correlated with observed or unobserved station characteristics that are (more) important to consumers, and there are no clear criteria for sorting brands into different market segments. To identify search and segmentation, I exploit two unique features of the Seoul market from 2009 to 2014. First, during the period of my study, gas prices increased dramatically (over 50% in three years), largely due to wholesale price increases. In normal times, consumers save at most a few cents per gallon from additional search within some bounded areas, limiting consumers’ incentive to search. Second, the introduction of self-service stations, which may have been accelerated by the increase in wholesale prices, also disrupted the market; many stations were converted from full service to self service, and there was also a significant amount of entry and exit. Combined, these features ensure that consumers are very likely to search with increased intensity and this helps overcome the difficulty in observing consumer search directly.1 The variation in service levels generated by the binary choice of service level in the Korean market 1 It is not trivial to measure search intensity from an empirical perspective. Lewis and Marvel (2011) apply a clever method to observe search intensity through the number of visits to a web site that provides gas price information under the assumption that 1 further provides me with identification, under the assumption that self-service customers’ revealed preference for trading effort for lower prices also implies that they search more intensely. Following a specification used by Lewis (2008), I start by examining the relationship between price dispersion and density of local competition. Similar to Lewis, I find significant price dispersion even after controlling for price differences related to station characteristics. The extent of price dispersion varies systematically with the number of sellers, as well as with seller characteristics such as presence of a carwash and/or a repair shop, location, brand type, and – most importantly – service level. Next, I investigate the degree of market segmentation by service level and brand type. I find that the relationship between price dispersion and the local competitive environment differs by service level; specifically, price dispersion among self-service stations is lowest when their density is high, but there is no evidence of such a relationship for full-service stations. Furthermore, regarding competition across service levels, I show that an increase in the number of full-service stations is correlated with higher price dispersion among nearby self-service stations, whereas an increase in the number of self-service stations is correlated with lower price dispersion among full-service stations. I interpret these findings to suggest that stations differentiate on multiple dimensions, many of which are unobservable to the econometrician, and that self-service stations may add differentiating features in the presence of full-service stations. These empirical results driven by differentiation by service level are, to my knowledge, not documented in existing price dispersion literature. Similar patterns are also observed in brand-type segmentations, but the estimates are sensitive to how I segment the market by brand type. These findings support the claim that differentiation by service level is a robust setting less affected by ad-hoc standards compared to differentiation by brand type. There is a vast literature on retail gasoline pricing with respect to various market dimensions (for recent reviews, see Houde, 2010 and Noel, 2016). Focusing on the service dimension, Shephard (1991) estimates the full-service premium in a four-county area in Massachusetts in 1987, and interprets the estimate as evidence for price discrimination over and above cost-based price differences. Png and Reitman (1994) develop a model of service-time competition, and find that consumers pay about 1% more for a 6% reduction in waiting or service time, using the same dataset used by Shephard (1991). Gasoline is also differentiated by product grade and lead content; Borenstein (1991) shows that stations’ markup differs for leaded vs. unleaded gasoline, and reviews some possible but ex-post unsatisfactory explanations for the difference, such as cost-based, purchase-size-based, and paying-method-based explanations. Finally, and perhaps most obviously, stations are differentiated by location. Firgo et al. (2015), using data from Vienna, measure the “centrality” of diesel stations, defined as the number of times a particular station is among the the higher web traffic indicates higher search intensity. Using a reduced-form model, they find asymmetrical searching patterns: consumers search more as prices rise than they do when prices fall. 2 five nearest neighbors of other stations, and shows that prices are more strongly correlated with prices of central competitors. The rest of this paper is organized as follows. I describe the data and provide an overview of the market in Section 2, and show an empirical methodology in Section 3. I present my main results, as well as several robustness tests, in Section 4 and 5, respectively. Section 6 concludes. 2. Data and Market Overview 2.1 Data My data cover the period from November 2008 to September 2014. Over this period, I have daily price information on all gasoline stations in Seoul, and I observe entry, exit, address, and service levels, also at a daily frequency. I have created a monthly dataset that includes the average price and the average number of competitors of individual stations. The data come from the Oil Price Information Network (OPINET) in Korea. OPINET is a website operated by Korea National Oil Corporation (KNOC) to provide retail gasoline prices to the public for market transparency since 2008. For stations selecting the “automatic update” option, price information is uploaded automatically six times per day based on transactions data. Otherwise, stations must update their price manually within 24 hours of a change in price. Most of the stations choose the first option. OPINET provides other information, such as each station’s name and station characteristics including level of service. There were a few stations that reported unidentified addresses, leading to geocoding failures. In these cases, I use the stations’ names to fix the addresses since every station reported its name. After carefully matching names with other address resources (such as administrative documents or online map services), I geocode all stations in my dataset.2 A unique market feature in Korea is that and stations are perfectly partitioned by service level at each point in time, with no station simultaneously offering a combination of full-service and self-service pumps, unlike in the U.S. gasoline market. Table 1 describes the data. <Table 1. Summary Statistics> I observe a significant amount of both entry and exit, as well as service-level and brand changes for which a station may close temporarily during a renovation. This feature helps enhance power to estimate the relationship between price dispersion and the number of sellers. To count competitors with the dynamics 2 Three resources on an online map service: Google Inc. (https://www.google.com/maps), Naver Inc. (http://map.naver.com), and Daum Inc. (http://map.daum.net). I also refer to administrative documents available on the website (https://www.data.go.kr) operated by Korean government. 3 properly, in case where entry or exit occurs in the middle of a month, I apply fractions to count the monthly number of stations. For example, if a station enters on the 10th of the month, I count it as 2/3 of a station that month. In addition, my data captures a time period during which gas prices increased very sharply. This feature helps in the identification of the relationship between price dispersion and consumer search because when prices are higher consumers may have greater incentive to roam in search of cheaper gas due to income effects. 2.2 Market Overview The first gas station with self-pumps opened in Seoul in 1992, but by the end of 2007, more than 99% of stations in Seoul still offered only full-service gasoline. Since 2008, the conversion of the market to self service has dramatically accelerated, possibly due to a large spike in crude oil prices. As of September 2014, self-service stations accounted for about 20% of stations in Seoul. One or two new self-service stations have opened every month on average since 2009. In many cases these are a result of switching rather than de novo entry. Self-service customers generally pay 2-4% less per gallon. There are seven brands operating in the market. In descending order by market share, these are: SK Energy, GS Caltex, S-Oil, Hyundai Oilbank, Alddle, Nonghyup Oil, and Namhae Oil.3 <Figure 1. Location of Self-Service Stations> Figure 1 shows the locations of self-service gas stations in the first month (marked as stars) and the last month of my sample (the sum of stars and circles). Seoul is the largest city (234 square miles) in Korea and highly dense (43,600 residents per square mile), and located in a small basin that possibly removes concern about stations near the city boundary. <Figure 2. Price Distributions by Service Level> Figure 2 represents price distributions by service level in four selected months.4 There are three notable phenomena in this market. First, gasoline prices are widely dispersed. For example, the highest price was 25% above the lowest price on November 2008 and this gap was 35% in September 2014. Second, retail prices increased sharply during the sample period in the aftermath of the global financial crisis around 2008. The average price increased 55% during the first three years and then declined, so that prices were 24% 3 In my analysis, I drop out Nonghyup Oil and Namhae Oil. Each brand has only one station in the market, and the former one is a non-private company and the latter one is no longer in the market since June 2012. 4 The selected months are the first and last months of my sample, and two middle points between them. There are no striking jumps in the price distributions. 4 above the initial level at the end of the period. Finally, the shape of price distributions changes in opposite directions for the two service levels: the distribution of self-service prices narrowed while that of fullservice stations widened over time. This suggests that full-service stations may be differentiated from selfservice stations and some full-service stations seem not engaged in price competition. 5 One possible explanation for this phenomenon is that full-service stations differentiate on other, unobserved, dimensions such as extra services (e.g., free carwash, coffee, or raffle tickets). As a preliminary exercise of investigating price dispersion, I trace relative price changes of one arbitrarily selected station (full-service, SK Energy) and its four competitors within a 0.5-mile radius by service level and brand type. Lewis’s (2008) examination is the benchmark of this exercise. The arbitrarily selected station competes with three full-service and one self-service stations; these four stations include two SK Energy, one GS Caltex, and one unbranded station. <Figure 3. Price movements of the five selected stations> Figure 3 shows movements in average prices of the stations by service level and brand type, as well as price quartiles by citywide level and local level. The quartiles are a reference to compare how prices of the stations change relative to citywide or local level. Local quartiles are defined as average prices of stations with a 1.5-mile radius of the arbitrarily selected station; there are twenty-four competitors in that range. There are two useful lessons from this figure. First, price ranking is volatile in the entire period, which makes it costly for consumers to stay well informed about the price distribution.6 More interestingly, when prices increase, the extent of price dispersion remains lower in not only station levels and but also quartile levels. This observation alludes to “rockets and feathers” which is the well-known phenomenon in retail gasoline, whereby stations generally earn lower margins (and exhibit lower price dispersion) when prices increase than they do when prices fall.7 5 Grabowski and Vernon (1992) study price competition after patent expiration in the pharmaceutical-drug market. They document that major companies significantly increased the price of branded products after generic drugs entered the market, consistent with inelastic demand by customers loyal to branded drugs. The gasoline industry obviously has a different market structure, but an increase in the price of full-service stations very close to self-service stations may be rational if full-service stations is unable to beat self-service only with price competition, and use other amenities to attract customers. 6 This practice is called obfuscation in the consumer search literature. See Ellison (2016) for a great overview of price search and obfuscation. 7 Tappata (2009) and Lewis (2011) approach this phenomenon from theoretic perspectives. Hosken et al. (2008), Lewis and Marvel (2011), and Brewer et al. (2014) investigate it empirically. 5 3. Empirical Methodology I calculate price dispersion as a residual from a fixed-effects specification, following Lewis (2008). This approach has the advantage that it controls for price differences related to time-invariant station characteristics, as well as aggregate movements in price due to world oil price shocks.8 The specification is a two-step procedure. In the first step, I estimate a regression of station-level retail prices on station and time fixed effects: (1) 𝑃𝑖𝑡 = 𝛽𝑖 + 𝜙𝑡 + 𝑢𝑖𝑡 𝑤ℎ𝑒𝑟𝑒 𝐸(𝑢𝑖𝑡 ) = 0. The station fixed effects, 𝛽𝑖 , capture price differences due to observed and unobserved time-invariant station characteristics, and the time fixed effects, 𝜙𝑡 , reflect changes in average prices, mostly driven by wholesale gasoline prices. The assumption that station characteristics remain unchanged over time is generally consistent with the data – station characteristics rarely change over time. The one difference from Lewis (2008) is that the total number of stations changes slightly from month to month in this study. This regression gives two important sets of estimates: 𝛽̂𝑖 and 𝑢̂𝑖𝑡 . Each coefficient 𝛽̂𝑖 captures the average price of station i relative to the city average over the entire period.9 The residual 𝑢̂𝑖𝑡 represents the deviation of the price of station i in month t from its average position relative to the city average. In this sense, the estimated variance of 𝑢̂𝑖𝑡 is a measure of price dispersion. In the second step of the procedure, as in Barron et al. (2004) and Lewis (2008), I model price dispersion as a function of the density of local competition and characteristics of sellers. The density of competition is defined as the number of competing stations within 1.5 miles of a station.10 There is one station that has no competitor within the 1.5 miles, so I drop out that station in my analysis. The empirical specification is: 2 ) = 𝛾 ln(𝐷𝑒𝑛𝑠𝑖𝑡𝑦)𝑖𝑡 + 𝜂𝑿𝑖𝑡 + 𝜉𝑡 + 𝑤𝑖𝑡 𝑤ℎ𝑒𝑟𝑒 𝐸(𝑤𝑖𝑡 ) = 0. ln(𝑢̂𝑖𝑡 (2) 𝑤𝑖𝑡 = 𝜌𝑖 𝑤𝑖,𝑡−1 + 𝜇𝑖𝑡 , 𝑤ℎ𝑒𝑟𝑒 𝐸(𝝁𝑡 𝝁′𝑡 ) ≡ 𝚳, 𝐸(𝜇𝑖𝑡 ) = 0, 𝑎𝑛𝑑 𝐸(𝜇𝑖𝑡 𝜇𝑗𝑡 ) ≠ 0 for 𝑖, 𝑗. The coefficient 𝛾 on ln(Density) reveals how price variation not explained by either station or time fixed effects varies with the density, or intensity, of local competition. In some specifications, following Lewis (2008), I also include the variable Same Brand Share, representing the share of stations within a 1.5-mile radius that have the same brand as station i at time t. This variable is a proxy for information on ownership. A fixed-effects approach is commonly used to calculate price dispersion for this purpose in many other fields – Sorensen (2000) in prescription drugs market, Lach (2002) in supermarket, and Hosken et al. (2008) and Lewis (2008) in retail gasoline market. 9 Similar to earlier studies, I also examine the relationship between price dispersion and the coefficient 𝛽 ̂𝑖 , but I exclude explanations and estimations on this relationship hereafter because my results are qualitatively not different from those studies and are a digression from the main focus of this paper. 10 Results are qualitatively similar among one, one and half, and two miles, so that I report the competition of a 1.5-mile radius to be consistent with previous studies. 8 6 Even if brand is not the same as joint ownership, it is intended to control for the case where a local dealer may release or own several stations under the same brand in the market.11 In that case, price dispersion of such stations may be higher or lower depending on their pricing strategies. The vector X includes station characteristics listed in Table 1 to help isolate the effect of ln(Density), as well as to see how they are related to observed price dispersion. I include an indicator of station’s service level, brand type, and the presence of a convenience store, a carwash, or a repair shop. I also include an indicator whether a station is located at an intersection to control for the possibility such station may face higher demand, all else constant. All of these attributes can affect station’s profit, leading to different pricing decisions. However, unlike Barron et al. (2004) and Lewis (2008), I do not include the number of pump, which commonly proxies for a capital stock of stations, due to data limitations. Monthly fixed effects capture differences in the overall level of observed price dispersion over time. I construct the error structure of heteroskedasticity across stations and a station-specific AR(1) process based on data features. 12 The intuitive explanation on the error structure is that I allow the estimated variance of price variation to be robust to arbitrary correlation across stations within time t, and it technically presents that each diagonal elements are not same and off-diagonal elements are non-zero in the variancecovariance matrix 𝐌. All combined, I implement a Feasible Generalized Least Squares procedure to estimate Equation (2). By construction, 𝑢̂𝑖𝑡 reflects citywide price variation. Lewis (2008) reasonably argues that retail gasoline competition is more likely to be localized. To take this argument, I also measure localized price variation according to Lewis’s construction. The specification is as follows: (3) 𝑢̂𝑖𝑡 = λ ̂𝑗𝑡 ) (∑𝑗∈𝐽(𝑡) 𝑢 𝑁𝐽(𝑡) + 𝑣𝑖𝑡 𝑤ℎ𝑒𝑟𝑒 J(t) = {competing stations within 1.5 miles of station 𝑖 at time 𝑡} for each 𝑖. (4) ln(𝑣̂𝑖𝑡2 ) = 𝛾 ln(𝐷𝑒𝑛𝑠𝑖𝑡𝑦)𝑖𝑡 + 𝜂𝑿𝑖𝑡 + 𝜉𝑡 + 𝜔𝑖𝑡 𝑤ℎ𝑒𝑟𝑒 𝐸(𝜔𝑖𝑡 ) = 0. 𝜔𝑖𝑡 = 𝜌𝑖 𝜔𝑖,𝑡−1 + 𝜖𝑖𝑡 , 𝑤ℎ𝑒𝑟𝑒 𝐸(𝝐𝑡 𝝐′𝑡 ) ≡ 𝚬, 𝐸(𝜖𝑖𝑡 ) = 0, 𝑎𝑛𝑑 𝐸(𝜖𝑖𝑡 𝜖𝑗𝑡 ) ≠ 0 for 𝑖, 𝑗. Equation (3) calculates localized price variation for station i at time t by conditioning on the price variation of its nearby competitors in 𝐽, the set of stations within 1.5 miles around station i at time t. As a calculation technique, after having 𝑢̂𝑖𝑡 from Equation (1), I estimate a regression of 𝑢̂𝑖𝑡 on the average of 11 Although I do not have direct information on ownership, I can conjecture the possibility of a joint ownership under the same brand from station’s name. For example, there is a set of the stations naming “Daesung-1st”, “Daesung-2nd”, and up to “Daesung6th.” This may imply that these stations are owned by one dealer, even if they are not located together within 1.5 miles. 12 Lewis (2008) assumes a common AR(1) coefficient process to all stations, whereas I assume it with unit-specific 𝜌 and 𝑖 perform robustness checks with a common 𝜌 model. The distribution of the station-specific coefficient estimates is shown in Figure A1 in Appendix A, and the null hypothesis that all stations have common 𝜌 is rejected at a significance level of 0.01. This implies that the serial dependence of individual stations is significantly different across stations in my data. 7 the residuals of station i’s local competitors. Following Lewis (2008), I interpret the new residual 𝑣̂𝑖𝑡 in Equation (4) as a local measure of price dispersion, and I model it as a function of the density and other characteristics as in the specification of citywide price variation in Equation (2) with the same assumption on the error structure. 4. Results 4.1 Price Dispersion with Competition Table 2 shows estimation results for Equation (2). In the baseline regression in Column (1), I find that a 10% increase in the number of competing stations within 1.5 miles is associated with lower price dispersion of 0.91%. The interpretation of this negative coefficient is that the squared price deviation of a station from the city average diminishes when the station faces more competitors (or more competitive environment, generally). The negative coefficient is consistent with previous studies of retail gasoline markets including Marvel (1976), Barron et al. (2004), and Lewis (2008), while counter to Chandra and Tappata (2011) and Lach and Moraga-González (2015). <Table 2. Explaining Citywide Price Dispersion> Next, I test for market segmentation by service level. Customers who buy self-service gasoline are more likely to be price sensitive and have higher search intensity, possibly relating to low dispersion. On the other hand, customers buying full-service gasoline are likely to have lower search intensity and to be less price sensitive, relating to high dispersion. Such illustration basically replies on the intuition mentioned in Salop and Stiglitz (1977; consumer heterogeneity in information costs), and aims to test a search-theoretic inference specifically inspired by Lewis (2008) although Lewis describes this story with respect to the brand dimension. Similarly, I test for segmentation by brand type. Lewis (2008) separates station brands into two categories based on their market shares and average price: high-brand stations (Shell, Chevron, Union 76, and Mobil) and low-brand stations (7-11, Exxon, Texaco, ARCO, and unbranded).13 To take the same approach, I first separate high-brand and low-brand. High-brand stations include SK Energy and GS Caltex, which, together, have an about 70% market share. Similar to the interpretation of service level, low-brand customers are likely to be more price sensitive than high-brand customers. In Korea, people tend to rate SK Energy and GS Caltex as higher quality than the other brands and these stations charge, on average, about 13 Lewis makes an exception for ARCO, arguing that ARCO is thought of as a very competitive brand with lower prices and limited service. 8 3% higher than low-brand stations. 14 Second, I partition brands into a leader brand (SK Energy) and follower brands (all others) since SK Energy alone accounts for almost half (42%) of the entire market. Lastly, I distinguish stations between branded and unbranded, following similar logic. To estimate the relationship between price dispersion and local density depending on the type of sellers, I add an interaction of station type and density of local competition to the baseline specification (Equation 2). The results in Columns (2) of Table 2 show how the relationship varies across sellers of different service levels. I find robust evidence that the extent of price dispersion varies significantly depending on sellers of different types, which is consistent with Lewis (2008) even if I use an alternative dimension of market segmentation. To be specific, the point estimates of Column (2) imply that, ceteris paribus, a self-service station that competes with 10% more stations within 1.5 miles has about 3.6% higher price dispersion but a full-service station with 10% more competing stations has about 2% lower price dispersion (i.e., 0.3610.563). 15 The Wald statistic shows that these coefficient estimates are significantly different from one another and from zero. It is worth noting that one may think that the result is counter-intuitive: self-service stations are likely to engage in more intensive price competition against other competitors, possibly leading to a negative coefficient. However, since the Seoul gasoline market is transitioning from a full-service format to a self-service format, the coefficient’s sign is very likely to change depending on the service composition of local competitors, a possibility not captured by this specification. I return to this issue in next section when I partition the number of competitors by service level. In Columns (3)-(5) I use brand to segment the market; the coefficients on ln(Density) are statistically insignificant and those on the interaction of density and brand type considerably differ by specifications. For example, in the benchmark model with brand classification (high-brand vs. low-brand) in Column (3), the coefficients of interest are both insignificant. This result suggests that either the Seoul market is not segmented by brand or the market is segmented by brand but the brand classification is misspecified. In either case, this finding shows that the estimated coefficients by brand type is sensitive to the criteria used to separate stations.16 The coefficient on same-brand share is positive throughout, indicating higher price volatility in markets with more competitors of the same brand. One possible explanation for this finding is that stations engage in “soft competition” within brands. This positive coefficient is in contrast to Lewis (2008), who estimates a negative but insignificant coefficient on same-brand share. Although it is hard to say why this result is different from Lewis’s, it suggests that different market features across countries may affect results. See the survey on consumer’s shopping pattern on gasoline conducted by Korea Energy Economics Institute (KEEI, 2010). The coefficient on the interaction terms picks up differential sensitivity to the degree of competition when a station is full service (high-brand, leader-brand, or unbranded, respectively). 16 In Column (3)-(4), brand-type specifications, the coefficients on ln(Density) become statistically significant at 5% level in the estimation model with a common coefficient AR(1) process to all stations, whereas the significance levels of the coefficients in the service-level specification stay unchanged. These estimates are not reported in this paper. 14 15 9 In the lower part of Table 2 I show the coefficient estimates on station characteristics. All else equal, stations located at intersections and those with self-service pumps (especially in the service-level specification) exhibit lower price variability, and stations with a carwash facility have higher price variability. It may be that stations with a carwash engage in softer non-price competition and charge higher prices, whereas stations with self-service pumps and those located at intersections compete more heavily on price. <Table 3. Explaining Local Price Dispersion> In addition to citywide variation, I estimate the same specifications using Equation (4), with local price variation of the dependent variable. These results are shown in Table 3. In general, I find similar patterns as observed in the citywide price dispersion. However, in Column (4)-(5) of Table 3, the coefficients on the interaction of density and brand type become statistically insignificant, while the coefficients in the Column (2) have the same sign and significance level as those reported in Table 2. This reflects that the servicelevel specification is robust to the measure of price dispersion, relative to the brand-type specifications. Regardless of the mechanism, using local price dispersion as the LHS variable generally attenuates the coefficients on ln(Density), the interaction terms, and the same-brand share. Overall my results are consistent with previous studies showing that price dispersion varies with heterogeneity of sellers, although the setting is significantly different and my specifications extend the notion of differentiated sellers in retail gasoline (commonly proxied by brand) into service level. The results using service-level market segmentation are more robust across specifications, both across citywide/local competition and with regard to assumptions about the error structure. 4.2 Relative Effects by Seller Type Next, I partition the number of competitors to examine relative effects of density by service level and brand type to test whether stations’ pricing behavior depends on their competitors’ types. As differentiation by service level, I estimate Equation (5) under the same error structure that I assume in the baseline specification. (5) 2 ) = 𝛾1 ln(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝑆𝑒𝑙𝑓 ) + 𝛾2 ln(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝑆𝑒𝑙𝑓 ) ∗ 𝐹𝑢𝑙𝑙 𝑆𝑒𝑟𝑣𝑖𝑐𝑒𝑖𝑡 ln(𝑢̂𝑖𝑡 𝑖𝑡 𝑖𝑡 + 𝛾3 ln(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐹𝑢𝑙𝑙 )𝑖𝑡 + 𝛾4 ln(𝐷𝑒𝑛𝑠𝑖𝑡𝑦𝐹𝑢𝑙𝑙 )𝑖𝑡 ∗ 𝐹𝑢𝑙𝑙 𝑆𝑒𝑟𝑣𝑖𝑐𝑒𝑖𝑡 + 𝜂𝑿𝑖𝑡 + 𝜉𝑡 + 𝜀𝑖𝑡 In Equation (5), by interacting the density of each service level with the service level indicator, I can identify four relative price effects: the effect of full-service density on full-service stations, the effect of 10 full-service density on self-service stations, the effect of self-service density on full-service stations, and the effect of self-service density on self-service stations. I specify the empirical models of brand classifications (high-brand vs. low-brand, leader-brand vs. follower-brand, and branded vs. unbranded) similarly. By partitioning the density into service level, I drop some stations from this analysis that have no selfservice competitor within 1.5 miles. To ensure that the estimation is robust to the missing stations, I also estimate Equation (5) using a 3.5-mile radius, which is large enough to make almost all stations have at least one full-service and one self-service competitor nearby. Results using 1.5 miles or 3.5 miles are qualitatively similar. Therefore, I use 1.5 miles to be consistent with the previous results as well as to compare my findings with Lewis (2008).17 <Table 4. Relative Effects of Different Sellers, by Service Level> Table 4 contains the estimates by service level. Starting with citywide variation, in Column (1) of Table 4, within the same service level of competition, I find that the effect of self-service density on self-service dispersion is negative (-0.19), and the effect of full-service density on full-service dispersion is also negative but twice higher (-0.08) (i.e., 0.397-0.477) than competition among self-service stations. These results are consistent with the inference that self-service customers are more likely to have high search intensity relative to full-service customers, although the Wald statistic on this difference is not large enough to distinguish them statistically. I observe the similar and statistically significant pattern in local price dispersion in Column (2). I find that self-service station density is negatively correlated with price dispersion for self-service stations (-0.32). Full-service station density is also negatively correlated with dispersion among full-service stations, but this correlation is three times smaller (-0.11) (i.e., 0.421-0.543). Again, this difference is consistent with a search-based story: high search intensity is correlated with low price dispersion, whereby self-service customers exhibit higher search intensity than full-service customers. My result on competition within service level contrasts with Lewis’s (2008) result within brand type, which is that price dispersion among low-brand stations is higher than that among high-brand stations. In addition, I note that the relationship between local density and price dispersion reveals an asymmetry across service levels in both citywide and local variation. For example, in Column (2) of Table 4, I show an apparently positive relationship between the full-service station density and price dispersion for self- 17 Table A1 in Appendix A shows the 3.5-mile results. Additionally, using a logistic regression with odds ratio, I find that the dropped stations are about twice as likely to be full-service and 1.8 times more likely to have a convenience store, but are about one-third times less likely in other characteristics such as the presence of a carwash or repair facility and location at an intersection. These results are shown in Table A2 in Appendix A. 11 service stations (0.42), but there is no evidence of such a clear relationship between the self-service density and dispersion for full-service stations (0.06) (i.e., 0.263-0.320). I observe the same pattern in citywide variation too. This result consistently indicates that self-service competitors are highly correlated with low price dispersion. Recall that in Column (2) of Table 2 and 3 the positive coefficient for self service with respect to number of competitors was seemingly counter-intuitive. Now, these relative results present that the type of competitors that self-service stations face is dominated by full service in my data. Lewis (2008) finds the opposite magnitude to mine: high-brand station density is correlated with much smaller dispersion for low-brand stations compared to the effect of low-brand station density on dispersion for high-brand stations. The difference between Lewis’s (2008) results and mine is primarily seen in the coefficient on the density of low-type sellers in local variations (self-service in this paper and low-brand in Lewis’s paper) which reflects the effect of low-type station density on price dispersion for low-type stations. Whereas Lewis found a small and statistically insignificant coefficient, I find a significant effect an order of magnitude much larger (-0.32). There are two possible reasons for this difference. First, the conversion of many full-service stations to self-service stations may have increased price competition among self-service stations, indicating that the increased competition is negatively correlated with dispersion for self-service stations. Second, the difference between full-service and self-service prices may be large enough to divide consumers by their search intensity, while in Lewis’s setting price differences between brands are smaller and may not fully partition consumers by their search intensity. <Table 5. Relative Effects of Different Sellers, by Brand Type> Next, I do the same type of the estimation with respect to different brand types, which is high-brand and low-brand or leader-brand and follower-brand. In these classifications, almost all stations have both brand types of competitors within 1.5 miles. Table 5 shows the estimates by the brand classifications. In Column (3)-(4) and (6)-(7), I generally observe similar (and even clearer) patterns to those I report using the service-level setting, indicating that brand type is a reasonable proxy to divide demand by different search intensity. However, as observed in Column (3)-(4) of Table 2 and 3, the magnitude of the coefficient on density and the interaction of density and its brand-type indicator considerably differ between two specifications. This suggests that service level is an obvious dimension on which consumer may sort and free from the selection of brand sorting. Regarding the branded and unbranded setting in Column (5) and (8), I should mention that the results are generally insignificant and may be imprecise because I drop one-third of observations (many have no 12 unbranded stations within 1.5 miles).18 Nonetheless, the relative magnitude of the coefficients between unbranded-station competition and branded-station competition are similar to the previous results that I document in the other brand settings. 5. Robustness Checks I perform several robustness checks on all regressions above: citywide variation, local variation, and relative effects of both variations by service level and brand type. In particular, I attach the results of the robustness checks on relative effects by service level as Appendix B. 5.1 Shorter sample period An important assumption in the fixed-effects specification is that station fixed effects control for price differences driven by station characteristics. My sample is long (71 months, or almost six years), a period long enough for significant changes in station characteristics. In Table B1, I use a subsample of 25 months (from September 2012 to September 2014) to test robustness of the results. The observed patterns of estimates are generally robust to the subsample regression, even though I lose some observations on the stations that switch from full service to self service and on the periods when prices increased sharply, weakening statistical significance. 5.2 Asymmetric effects By squaring the residuals obtained in Equation (1) to have the estimated variance of price variation, I should necessarily impose a symmetric relationship between local density and price dispersion, regardless of station’s pricing position relative to a city (or local) average. To investigate whether positive and negative residuals are differentially correlated with competitive conditions, I separate my sample into two subsamples, one containing only observations for which residuals are non-negative, and the other containing the negative residuals. In Table B2, I confirm that my results are robust. Results for both samples are symmetric overall in the citywide and local variation. 5.3 Modification in the first-step procedure As with entry and exit, stations’ conversion to self service gives me space to modify the empirical methodology and check the robustness of it, especially Equation (1). In particular, I replace station fixed effects with the interaction of station fixed effects and service level to confirm the robustness of current 18 I reestimate the same specification using a 4-mile radius in which a fair number of stations has both branded and unbranded competitors nearby. Although such competition boundary seems too wide and some coefficients are insignificant, the results are generally similar to other brand-type specifications. This result is shown in Table A3 in Appendix A. 13 estimates to switching stations. Intuitively, this modification implies when a station changes from full service to self service, I treat the switching station as a new competitor. Repeating all analyses, I conclude that my estimates remain unchanged on that concern, which is shown in Table B3. 5. Concluding Remarks This paper argues on ex ante grounds that demand for gasoline is segmented by service level, and demonstrates that price dispersion patterns are consistent with this segmentation. Gas stations sell nearly identical quality of gasoline and driving is costly to consumers, so that spatial differentiation is probably of first order and the most obvious dimension on which the market is segmented. Lewis (2008) shows the importance of brand as another dimension on which the gasoline market is segmented. In addition to those dimensions of differentiation, I show that the service dimension – full service vs. self service – is important, and I argue that it is more intuitive and robust than brand segmentation. My setting is in line with search-theoretic models that attempt to explain the presence of price dispersion with heterogeneity among consumers and/or producers, including Reinganum (1979), Salop and Stiglitz (1977), and, especially, Carlson and McAfee (1983).19 Reinganum generates equilibrium price dispersion by introducing firm heterogeneity in marginal costs; Salop and Stiglitz show the existence of price dispersion by introducing consumer heterogeneity in search costs; Carlson and McAfee adopt both. All of these models are highly stylized and therefore not well-suited to direct empirical testing, but their settings are conceptually close to this study in that self-service stations have different marginal costs from fullservice stations (generating seller heterogeneity) and that self-service customers may have different search costs from full-service customers (generating consumer heterogeneity). 20 Although it is hard to decide which search models match better the observations, the empirical results of this paper extend the notion of price dispersion into a service dimension which has been overlooked in context of consumer search. 19 The assumptions of search-based models differ with respect to consumer search behavior (e.g., sequential search vs. fixed sample), as well as assumptions about firm heterogeneity, consumer heterogeneity, and market structures. Mixed together, model predictions vary with these settings – a positive sign (e.g., Reinganum, 1979), a negative sign (e.g., MacMinn, 1980), or a nonmonotonic sign (e.g., Varian, 1980). See Baye et al. (2006) for an excellent review on a wide range of search models from both theoretic and empirical perspectives. 20 Basker et al. (2015) study the effects of the adoption of self-service pumps on employment between 1977 and 1992 in the US retail gasoline market, and find about 0.4 fewer workers per pump using fuzzy regression-discontinuity and instrument variable identification strategies. 14 References Barron, J. M., Taylor, B. A., & Umbeck, J. R. (2004). Number of Sellers, Average Prices, and Price Dispersion. International Journal of Industrial Organization, 22(8), 1041-1066. Basker, E., Foster, L., & Klimek, S. D. (2015). Customer-Labor Substitution: Evidence from Gasoline Stations. US Census Bureau Center for Economic Studies (CES) Working Paper 15-45. Baye, M. R., Morgan, J., & Scholten, P. (2006). In T. Hendershott (Ed.), Information, Search, and Price Dispersion. Handbook on Economics and Information Systems, Elsevier Press, 323-371. Borenstein, S. (1991). Selling Costs and Switching Costs: Explaining Retail Gasoline Margins. RAND Journal of Economics, 22(3), 354-369. Brewer, J., Nelson, D. M., & Overstreet, G. (2014). The Economic Significance of Gasoline Wholesale Price Volatility to Retailers. Energy Economics, 43(3), 274-283. Carlson, J. A., & McAfee, R. P. (1983). Discrete Equilibrium Price Dispersion. Journal of Political Economy, 91(3), 480-493. Chandra, A., & Tappata, M. (2011). Consumer Search and Dynamic Price Dispersion: An Application to Gasoline Markets. RAND Journal of Economics, 42(4), 681-704. DAUM (2016), Daum Maps, available at http://map.daum.net (accessed 14 March 2016). Ellison, S. F. (2016). Price Search and Obfuscation: An Overview of the Theory and Empirics. In E. Basker (Ed.), Handbook on the Economics of Retailing and Distribution, Edward Edgar, 287-305. Firgo, M., Pennerstorfer, D., & Weiss, C. R. (2015). Centrality and Pricing in Spatially Differentiated Markets: The Case of Gasoline. International Journal of Industrial Organization, 40(1), 81-90. Grabowski, H. G., & Vernon, J. M. (1992). Brand Loyalty, Entry, and Price Competition in Pharmaceuticals After the 1984 Drug Act. Journal of Law and Economics, 35(2), 331-350. Google (2016), Google Maps, available at http://maps.google.com/ (accessed 14 March 2016). Hastings, J. S. (2004). Vertical Relationships and Competition in Retail Gasoline Markets: Empirical Evidence from Contract Changes in Southern California. American Economic Review, 94(1), 317-328. Hosken, D. S., McMillan, R. S., & Taylor, C. T. (2008). Retail Gasoline Pricing: What Do We Know?. International Journal of Industrial Organization, 26(6), 1425-1436. Houde, J.-F. (2010). Gasoline Markets. In S. Durlauf and L. Blume (Eds), The New Palgrave Dictionary of Economics, available at http://www.dictionaryofeconomics.com/article?id=pde2010_G000218&edition=current&q=gasoline%20market&topicid=&result_number=1 (accessed 14 March 2016). Kim, D.-W., & Kim, J.-H. (2011). The Impact of the Entry of Self-Service Stations in the Korean Retail Gasoline Market: Evidence from the Difference-in-Differences Methods. Korean Journal of Economic Studies, 59(2), 77-99. Korea Energy Economics Institute (KEEI) (2010). 소비자의 주유소 구매행태 설문조사 (Survey on Consumer’s Shopping Pattern on Gasoline), available at http://www.keei.re.kr/web_keei (accessed 14 March 2016). Korea Ministry of Trade, Industry and Energy (2014). 전국 주유소 등록현황 (The Current Registration of Gasoline Stations in the Entire Country), available at http://www.data.go.kr (accessed 14 March 2016). 15 Lach, S. (2002). Existence and Persistence of Price Dispersion: An Empirical Analysis. Review of Economics and Statistics, 84(3), 433-444. Lach, S., & Moraga-González, J. L. (2015). Asymmetric Price Effects of Competition. Centre for Economic Policy Research (CEPR) Discussion Papers 10456. Lewis, M. (2008). Price Dispersion and Competition with Differentiated Sellers. Journal of Industrial Economics, 56(3), 654-678. Lewis, M. 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, M. S., & Marvel, H. P. (2011). When Do Consumers Search? Journal of Industrial Economics, 59(3), 457-483. MacMinn, R. D. (1980). Search and Market Equilibrium. Journal of Political Economy, 88(2), 308-327. Marvel, H. P. (1976). The Economics of Information and Retail Gasoline Price Behavior: An Empirical Analysis. Journal of Political Economy, 84(5), 1033-1060. NAVER (2016), Naver Map, available at http://map.naver.com (accessed 14 March 2016). Noel, M. D. (2016). Retail Gasoline Markets. In E. Basker (Ed.), Handbook on the Economics of Retailing and Distribution, Edward Edgar, 392-412. Png, I. P., & Reitman, D. (1994). Service Time Competition. RAND Journal of Economics, 25(4), 619-634. Reinganum, J. F. (1979). A Simple Model of Equilibrium Price Dispersion. Journal of Political Economy, 87(4), 851-858. Salop, S., & Stiglitz, J. (1977). Bargains and Ripoffs: A Model of Monopolistically Competitive Price Dispersion. Review of Economic Studies, 44(3), 493-510. Shepard, A. (1991). Price Discrimination and Retail Configuration. Journal of Political Economy, 99(1) 30-53. Sorensen, A. T. (2000). Equilibrium Price Dispersion in Retail Markets for Prescription Drugs. Journal of Political Economy, 108(4), 833-850. Tappata, M. (2009). Rockets and Feathers: Understanding Asymmetric Pricing. RAND Journal of Economics, 40(4), 673-687. Varian, H. R. (1980). A Model of Sales. American Economic Review, 70(4), 651-659. 16 <Figures and Tables> Figure 1. Expansion of Self-Service Stations: Existing Stations (Star) as of November 2008; New Stations (Circle) as of September 2014 Note: The layout is GoogleMap inserted by WTMS in ArcGIS 10.3. 17 Figure 2. Price Distribution by Service Level in Four Selected Months Note: The first and last of months in the sample, and a twenty-one month interval between them. 18 Figure 3. Price Movements of Five Selected Stations and Their Rivals Relative to City and Local Quartiles 19 Table 1. Summary Statistics, by Competition and Station Characteristics Variable Description P Station price (Korean won/liter) Competition Variables Density Number of stations within a 1.5-mile radius around station i ln(Density) Log of Density Same Brand Share Share of same brand stations within a 1.5-mile radius as station i Station Characteristics(Xi)a Self-Service Station serves as self-service Intersection Station is located at an intersection of a road Store Station has a convenience store Carwash Station has an automatic carwash machine Repair Station has an auto-repair center SK Station brand: SK Energy GS Station brand: GS Caltex Hyundai Station brand: Hyundai Oilbank S-Oil Station brand: S-Oil Alddle Station brand: Alddle Unbranded Station brand: Unbranded High-Brand Station brand group: SK Energy and GS Caltex Follower-Brand Station brand group: All brands expect for SK Energy b Station Average number of stations Observationc Number of observations Mean 1881.18 Std 201.06 18.93 2.852 0.278 7.29 0.459 0.165 0.146 0.134 0.104 0.692 0.280 0.421 0.278 0.134 0.121 0.015 0.026 0.700 0.578 629.94 44,146 0.353 0.341 0.305 0.461 0.449 0.493 0.448 0.341 0.326 0.123 0.160 0.458 0.493 22.89 a. Rare variations in brands and attributes, and one or two stations switch to self-service each month on average. b. Number of stations observed between November 2008 and September 2014 (71 months). c. Number of observations slightly changes across specifications in which there is no certain type of competitors nearby 1.5 miles in month t. 20 Table 2. Explaining Citywide Price Dispersion: 2 Dependent Variable: ln(𝑢̂𝑖𝑡 ) Competition Variables ln(Density) (1) (2) (3) (4) (5) -0.091** (0.039) 0.361*** (0.087) -0.563*** (0.097) -0.028 (0.071) 0.010 (0.050) 0.313 (0.208) ln(Density)*Full-Service -0.085 (0.083) ln(Density)*High-Brand -0.235*** (0.077) ln(Density)*Leader-Brand 0.885*** (0.171) 0.952*** (0.172) 0.875*** (0.172) 0.834*** (0.172) -0.416** (0.211) 0.877*** (0.170) 0.043 (0.057) -0.371*** (0.059) -0.015 (0.066) 0.374*** (0.043) 0.031 (0.046) 0.831*** (0.113) 0.454*** (0.105) 0.189* (0.105) 0.483*** (0.104) 1.046*** (0.164) -1.538*** (0.279) -0.375*** (0.058) -0.038 (0.066) 0.369*** (0.043) 0.038 (0.046) 0.787*** (0.113) 0.419*** (0.105) 0.170 (0.104) 0.433*** (0.104) 0.949*** (0.164) 0.040 (0.054) -0.370*** (0.058) -0.016 (0.066) 0.374*** (0.044) 0.030 (0.043) 1.093*** (0.281) 0.715*** (0.276) 0.206* (0.105) 0.497*** (0.105) 1.086*** (0.166) 0.036 (0.055) -0.367*** (0.058) -0.018 (0.066) 0.375*** (0.043) 0.026 (0.046) 1.543*** (0.258) 0.495*** (0.106) 0.221** (0.105) 0.510*** (0.104) 1.116*** (0.164) 0.042 (0.054) -0.369*** (0.058) -0.008 (0.067) 0.372*** (0.043) 0.029 (0.046) 2.102*** (0.652) 1.723*** (0.652) 1.459*** (0.651) 1.752*** (0.650) 2.304*** (0.658) 26.24 [0.000] 44064 6.197 2.482 0.15 [0.707] 44064 6.197 2.482 4.46 [0.035] 44064 6.197 2.482 3.04 [0.082] 44064 6.197 2.482 ln(Density)*Unbranded Same Brand Share Station Characteristics Self-Service Intersection Store Carwash Repair SK GS Hyundai S-Oil Alddle Wald statistic [p-value]a - Num. Observation 44064 Mean of Dep. Var 6.197 S.D of Dep. Var 2.484 All specifications also include month fixed effects. Denote significance ***, **, * at 1%, 5%, 10% level, respectively. Unbranded is omitted. a. Wald statistic results from the coefficient test hypothesis that ln(Density) = ln(Density)*Type-Indicator. 21 Table 3. Explaining Local Price Dispersion: Dependent Variable: ln(𝑣̂𝑖𝑡2 ) Competition Variables ln(Density) (1) (2) (3) (4) (5) -0.043 (0.039) 0.299*** (0.103) -0.396*** (0.110) -0.079 (0.064) -0.093* (0.050) 0.264 (0.212) ln(Density)*Full-Service 0.059 (0.080) ln(Density)*High-Brand 0.131* (0.079) ln(Density)*Leader-Brand ln(Density)*Unbranded Same Brand Share 0.182 (0.172) 0.218 (0.172) 0.195 (0.173) 0.212 (0.173) -0.316 (0.215) 0.177 (0.172) Wald statistic [p-value]a 11.01 1.02 3.70 1.86 [0.000] [0.311] [0.054] [0.172] Num. Observation 44064 44064 44064 44064 44064 Mean of Dep. Var 6.008 6.008 6.008 6.008 6.008 S.D of Dep. Var 2.500 2.500 2.500 2.500 2.500 All specifications also include station characteristics, station characteristics, and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. Omitted brand is unbranded. a. Wald statistic results from the coefficient test hypothesis that ln(Density) = ln(Density)*Type-Indicator. - Table 4. Relative Effects of Different Sellers, by Service Level Full-Service vs. Self-Service ln(DensitySS) ln(DensitySS)*Full-Service ln(DensityFS) ln(DensityFS)*Full-Service 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (1) 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (2) -0.185*** (0.069) 0.072 (0.075) 0.397*** (0.072) -0.477*** (0.084) -0.320*** (0.083) 0.263*** (0.089) 0.421*** (0.089) -0.534*** (0.099) Wald statistics [p-value] 1.57 4.57 [0.209] [0.032] 41.53 25.01 Competition across service levelsb [0.000] [0.000] Num. Observation 37322 37322 Mean of Dep. Var 6.128 5.971 S.D of Dep. Var 2.479 2.521 All specifications also include same share brand, station characteristics, and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. a. Wald statistic on the coefficient test hypothesis for competition within the same level: ln(DensitySS) = ln(DensityFS)+ln(DensityFS)*Full-Service. b. Wald statistic on the coefficient test hypothesis for competition across levels: ln(DensityFS) = ln(DensitySS)+ln(DensitySS)*Full-Service. Competition within service levela 22 Table 5. Relative Effects of Different Sellers, by Brand Type High-Brand vs. Low-Brand ln(DensityLB) ln(DensityLB)*High-Brand ln(DensityHB) ln(DensityHB)*High-Brand 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (3) (4) (5) 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (6) (7) (8) -0.107** (0.049) -0.393*** (0.064) 0.151* (0.088) 0.267*** (0.100) -0.169*** (0.053) -0.256*** (0.067) 0.182*** (0.082) 0.320*** (0.097) Leader-Brand vs. Follower-Brand ln(DensityFB) -0.258*** (0.058) -0.326*** (0.113) 0.256*** (0.060) 0.134 (0.102) ln(DensityFB)*Leader-Brand ln(DensityLB) ln(DensityLB)*Leader-Brand Branded vs. Unbranded ln(DensityB) -0.400*** (0.055) 0.035 (0.112) 0.394*** (0.056) 0.034 (0.104) 0.034 (0.119) -0.034 (0.121) 0.352 (0.392) -0.215 (0.397) ln(DensityB)*Branded ln(DensityUB) ln(DensityUB)*Branded -0.019 (0.137) 0.008 (0.138) 1.441*** (0.278) -1.205*** (0.280) Wald statistics [p-value] Competition within brand typea Competition across brand typesb Num. Observation Mean of Dep. Var S.D of Dep. Var 24.73 [0.000] 35.45 [0.000] 42981 6.191 2.483 52.22 [0.000] 72.68 [0.000] 43851 6.198 2.484 0.54 [0.461] 0.81 [0.369] 17980 5.859 2.517 48.36 [0.000] 35.91 [0.000] 42981 6.008 2.507 85.58 [0.000] 59.24 [0.000] 43851 6.007 2.502 2.69 [0.101] 27.04 [0.000] 17980 5.697 2.564 All specifications also include same share brand, station characteristics, and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. High-Type (HT: high, leader, and branded) and Low-Type (LT: low, follower, and unbranded), respectively: a. Wald statistic on the coefficient test hypothesis for competition within the same type: ln(DensityLT) = ln(DensityHT)+ln(DensityHT)*High-Type. b. Wald statistic on the coefficient test hypothesis for competition across types: ln(DensityHT) = ln(DensityLT)+ln(DensityLT)*High-Type. 23 <Appendix A: Supplementary Results> Figure A1. The distribution of estimated coefficients on an AR (1) process in each station. Table A1. Relative Effects of Different Sellers By Service Level Using 3.5 Miles Full-Service vs. Self-Service ln(DensitySS) ln(DensitySS)*Full-Service ln(DensityFS) ln(DensityFS)*Full-Service Num. Observation Mean of Dep. Var S.D of Dep. Var 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (1) 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (2) -0.452*** (0.078) -0.224*** (0.085) 0.500*** (0.098) 0.125 (0.112) 43894 6.197 2.482 -0.690*** (0.081) 0.216** (0.088) 0.917*** (0.102) -0.382*** (0.115) 43894 6.113 2.453 All specifications also include same share brand, station characteristics and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. Wald statistics are omitted. 24 Table A2. Logistic Regression of Dropped Stations on Station Characteristics Dropped Stations: No Self-Service Competitors within 1.5 miles Odds Ratio 2.241*** (0.109) 0.660*** (0.029) 1.839*** (0.076) 0.825*** (0.024) 0.700*** (0.023) 44146 0.154 0.361 Full-Service Intersection Store Carwash Repair Num. Observation Mean of Dep. Var S.D of Dep. Var Denote significance *** at 1% level. Table A3. Relative Effects of Different Sellers By Brand Type Using 4 Miles Branded vs. Unbranded ln(DensityB) ln(DensityB)*Branded ln(DensityUB) ln(DensityUB)*Branded Num. Observation Mean of Dep. Var S.D of Dep. Var 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (1) 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (2) -0.269 (0.182) 0.115 (0.185) 0.952*** (0.365) -0.425 (0.368) 37844 6.121 2.494 -0.367** (0.178) 0.174 (0.180) 0.696* (0.358) -0.296 (0.362) 37844 6.027 2.513 All specifications also include same share brand, station characteristics and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. Wald statistics are omitted. 25 <Appendix B: Robustness Checks> Table B1. Robustness Check: Subsample (2012.09 – 2014.09) Relative Effects of Different Sellers, by Service Level Full-Service vs. Self-Service ln(DensitySS) ln(DensitySS)*Full-Service ln(DensityFS) ln(DensityFS)*Full-Service Num. Observation Mean of Dep. Var S.D of Dep. Var 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (1) 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (2) -0.179* (0.096) -0.329*** (0.106) 0.343** (0.097) -0.288** (0.121) 12952 6.533 2.443 -0.135 (0.101) -0.307*** (0.103) 0.133*** (0.103) -0.036 (0.125) 12952 6.242 2.585 All specifications also include same share brand, station characteristics, and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. Wald statistics are omitted. Table B2. Robustness Check: Asymmetric Effects Relative Effects of Different Sellers, by Service Level Full-Service vs. Self-Service ln(DensitySS) ln(DensitySS)*Full-Service ln(DensityFS) ln(DensityFS)*Full-Service Num. Observation Mean of Dep. Var S.D of Dep. Var 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (1) (2) 𝑢̂𝑖𝑡 >=0 𝑢̂𝑖𝑡 <0 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (3) (4) 𝑢̂𝑖𝑡 >=0 𝑢̂𝑖𝑡 <0 -0.094 (0.096) 0.031 (0.105) 0.481*** (0.094) -0.558*** (0.112) 17615 6.133 2.576 -0.165* (0.080) 0.258*** (0.096) 0.344*** (0.109) -0.534*** (0.122) 18485 5.985 2.593 -0.170** (0.075) 0.011 (0.082) 0.447*** (0.089) -0.425*** (0.103) 19699 6.123 2.389 -0.351*** (0.103) 0.218** (0.106) 0.245** (0.110) -0.315** (0.122) 18827 5.958 2.454 All specifications also include same share brand, station characteristics, and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. Wald statistics are omitted. 26 Table B3. Robustness Check: Modification in the First-Step Procedure Relative Effects of Different Sellers, by Service Level Full-Service vs. Self-Service ln(DensitySS) ln(DensitySS)*Full-Service ln(DensityFS) ln(DensityFS)*Full-Service Num. Observation Mean of Dep. Var S.D of Dep. Var 2 Citywide variation: ln(𝑢̂𝑖𝑡 ) (1) 2 Local Variation: ln(𝑣̂𝑖𝑡 ) (2) -0.198*** (0.068) 0.113 (0.075) 0.243*** (0.073) -0.330*** (0.085) 37322 6.128 2.479 -0.372*** (0.079) 0.357*** (0.085) 0.282*** (0.088) -0.426*** (0.099) 37322 5.971 2.521 All specifications also include same share brand, station characteristics, and month fixed effects. Denote significance ***, **,* at 1%, 5%, 10% level, respectively. Wald statistics are omitted. 27
© Copyright 2026 Paperzz