Flex Cars and Competition in Ethanol and Gasoline Retail Markets∗ Jo˜ao Paulo Pessoa LSE/CEP [email protected]

Leonardo Rezende PUC-Rio [email protected]

Juliano Assun¸ca˜o PUC-Rio [email protected] November 8, 2011

Abstract The objective of this paper is to investigate how the diffusion of flex cars has affected competition on ethanol and gasoline retail markets. Flex cars have engines that can run on both types of fuel. We present a model of imperfect competition among fuel stations that predicts that, as the share of flex cars in the market grows, prices are expected to fall since fuels become closer substitutes both for consumers and fuel stations. We confirm this prediction using microdata from weekly prices of individual fuel stations in the state of Rio de Janeiro, Brazil. Using panel data methods, we find that a 10% increase in the percentage of flex cars reduces ethanol and gasoline prices by approximately 8 cents and 2 cents, respectively. We also find evidence that the penetration of flex cars increases the price gap between the two fuels, which is consistent with the predictions suggested by the model. We also estimate a structural model to identify differences in demand across fuel types as well as retail pricing responses to shifts in cost. We find that the latter are consistent with predicted responses from oligopoly theory. We also find evidence that while aggregate fuel demand by gasoline and ethanol car owners is inelastic, aggregate demand by flex car owners is elastic. Keywords: Flex-fuel vehicles; Gasoline; Ethanol fuel; price competition; spatial competition; discrete equilibrium price dispersion. JEL Codes: L11, L13, L62, L71. ∗

Preliminary. Comments are most welcome. The authors thank Jo˜ ao Manoel Pinho de Mello and Heleno Pioner for helpful comments, and for Gabriela Rochlin for superb research assistance. All errors are our own.

1

1

Introduction

Flex, or bi-fuel, vehicles can run on any combination of gasoline and ethanol fuel. Flex cars have been made commercially available in Brazil in 2003 and have been rapidly adopted: in 2008, 94% of the new cars registered in this country were flex. In this paper, we examine the consequence of this new technology on the retail fuel market. Our main hypothesis is that the penetration of flex car technology may have lead to lower fuel prices, by turning gasoline and ethanol into closer substitutes and consequently reducing the market power of fuel stations. We confirm this hypothesis empirically, exploring geographical variation in the speed of flex car penetration on a panel dataset of station-specific, weekly fuel prices from the state of Rio de Janeiro, Brazil. Our theoretical argument is similar, but distinct, from the conventional prediction that in an oligopoly where each firm provides a differentiated product, mark-ups are driven by differentiation. In our setting, the same fuel station supplies both gasoline and ethanol. Our point is that a technological change that reduces differentiation between products provided by the same firm, and does not change differentiation across firms, may also reduce mark-ups. To establish this point, we discuss a stylized model of imperfect competition among fuel stations that supply both types of fuel. We suppose there is differentiation across fuel stations, while fuels from the same station are perfect substitutes for flex car owners. We assume that the increase of the flex car fleet does not affect the degree of differentiation across stations or the demand parameters of any group of consumers, and recognize that fuel prices in each location are chosen by the same firm. We find that even in this framework flex car penetration reduces markups in equilibrium. We test this prediction empirically employing a unique panel dataset where we combine weekly data on individual service station retail fuel price with monthly flex car penetration at the municipality level in the state of Rio de Janeiro. Controlling for both time and location fixed effects, we find that flex car penetration leads to a statistically significant decrease in the level of both gasoline and ethanol retail prices. In the analysis, we treat flex car penetration as an exogenous shift in the automotive fuel market. We recognize expectations of future fuel prices can affect automobile demand in the long run and research and development of flex car technology by auto manufacturers in the even longer run. Klier and Linn (2010) provide evidence that automobile prices and sales respond to the price of gasoline in the United States. While flex car adoption is likely to be affected by fuel in Brazil as well, we point out that this is likely to introduce an attenuation bias in our estimates, since flex car sales would be a response to higher, not lower, gasoline prices. Similarly, the increase in correlation between fuel prices is also probably underestimated, since when prices are more correlated flex fuel technology is less useful. 2

Our paper contributes to a growing literature that analyzes the relationship of flex fuel car technology and fuel markets in the US and Brazil. Anderson (2010) examines data from the incipient ethanol fuel market in the US to estimate the demand for this product. He estimates an price elasticity of 2.5–3, and finds evidence of preference heterogeneity, as fuel switching seem to occur at a wide range of relative prices. Corts (2009) analyses empirically how purchases of flex cars by the US government might help to develop the ethanol market in the US. He finds that the adoption of flex cars by the government leads to an increase in the number of ethanol retail stations. Four other papers study how flex car technology has affected the relatively more widespread ethanol market in Brazil. Ferreira, Prado, and Silveira (2009) propose a model of competition between ethanol and gasoline manufacturers that predicts that with flex car penetration the fuel price ratio should converge to the technical rate of substitution of 70%.1 Employing time series methods, they do not find evidence of convergence to that ratio. The theoretical framework proposed by Salvo and Huse (2010b) also predicts that (for a range of international ethanol related market prices and marginal costs of supplying ethanol at the pump) gasoline and ethanol energy-equivalent prices should converge to the technical rate of substitution of 70%. Their model also claims that the two fuel prices should increasingly co-move as the penetration of flex cars rises. Empirically, they find evidence supporting these two theoretical results. Boff (2011) also utilizes time series methods to investigate long run trends on the prices of gasoline, ethanol and sugar, and concludes that the fuel price rate has moved toward the fuel efficiency rate. Contrary to those studies, in the model we develop in this paper, we find that, because retail prices of both fuels are picked by the same firms, they will always find it optimal to post different prices even when flex car penetration approaches 100%. Empirically, we also find that flex car penetration does not reduce the spread between gasoline and ethanol fuels in our sample, one we control for time and location-specific effects. Salvo and Huse (2010a) investigate a survey of flex car owners consumption choice between the two fuels. Consistent with Anderson (2010), they document preference heterogeneity, with a significant share of flex car drivers choosing the most expensive fuel, even when ethanol and gasoline energyequivalent prices differ in 20%. The paper is organized as follows: section 2 provides a brief summary of the general characteristics of the Brazilian fuel market. Section 3 presents a 1

A liter of ethanol has only 70% of the energy content of a liter of gasoline. In principle, a flex car owner concerned only with energy cost should use this ratio to compare the prices of the two fuels.

3

model of oligopolistic competition among fuel stations supplying both types of fuel. Section 4 describes the data we use and also shows some descriptive statistics. We present the empirical results in two section s: In section 5 we employ panel data methods to establish some relationships between flex car penetration and fuel retail pricing. In 6, we exploit these relationships to estimate demand functions for fuel. We make some concluding remarks in section 7.

2 2.1

Flex Cars and the Automotive Fuel Market in Brazil Ethanol-Powered and Flex Cars

Brazil has a long experience with usage of ethanol as vehicular fuel. In the seventies, in response to the first-oil crisis, the military government launched the Pr´ o-´ alcool program to favor production of ethanol from sugarcane and stimulate the adoption of ethanol-fueled cars. The program included credit subsidies for ethanol production and setting favorable fuel prices at the pump to stimulate adoption of the new technology. Consumers responded to the Pr´ o-´ alcool program — from 1983 to 1989, the majority of new cars was ethanol-fueled. This can be seen in figure 1, that presents shares of new car registrations in Brazil per year by fuel type. In response to a sharp rise of the world price of sugar, that tripled from 1985 to 1990 (USDA (2010),table 3a), domestic ethanol production sharply declined and the ensuing supply crisis lead to a plunge in the sales of ethanolpowered vehicles. Since 1995, sales of ethanol-powered cars correspond only to a small fraction of new vehicle sales in Brazil. However, ethanol-fueled cars still comprise more than 10% of the current fleet. In the first quarter of 2003 the flex cars (or bi-fuel cars) were released in Brazil. Flex cars can run on any mixture of gasoline and ethanol. Since a liter of ethanol contains roughly as much energy as 0.7 liters of gasoline (Marjotta-Maistro and Asai, 2006), a flex car owner can save money if the price ratio drifts away from that threshold.2 As figure 1 shows, flex car penetration has been dramatic: In 2008, 94% of new cars registered in Brazil were flex.

2.2

The Automotive Fuel Market in Brazil

In Brazil, ethanol, or ethyl alcohol, is made from sugarcane. Two kinds of ethanol play a role in the automotive fuel market: anhydrous and hydrated. 2 There are other facts that might influence consumers into choosing one fuel instead of the other. A car running on ethanol is less hazardous to the environment since it does not create net emissions of carbon dioxide. A car powered by gasoline demands less fuel per volume, thus allowing for less frequent refueling.

4

Anhydrous ethanol is mandatorily mixed with gasoline fuel in the proportion of 25%. Hydrated ethanol, a mixture that contains 5% water, is the version of alcohol readily available in drugstores and pumps at fuel stations in Brazil. Brazil is the largest producer of sugarcane in the world, the second largest producer of ethanol, and a net exporter of ethanol. Brazil was a net importer of oil until 2006, but has been a net exporter of gasoline since 1976. Before the Law 9478 from 1997 (“Lei do Petr´oleo”) the Brazilian oil industry was a monopoly in the hands of state-owned Petrobras. The law created the agˆencia nacional do petr´ oleo (ANP), the sector regulatory body, and broke Petrobras’ monopoly on exploration, refining, international trade, and sea transport of oil and its main subproducts. Since January 2002, fuel retail prices are allowed to be freely set by the market. Petrobras continues to be the major player in domestic refining, distribution and retail of gasoline, currently holding market shares in these markets of 96.6%, 28.9%, and 17.8%, respectively (ANP, 2010, tables 2.34, 3.6 and 3.17).

3

Model: Price formation in the retail fuel market in the presence of flex cars

In this section, we present a model of strategic price formation in the retail fuel market and use it to investigate theoretically what is the effect of a larger flex car fleet on equilibrium fuel prices. The model we propose has a number of features that are salient in the automobile fuel market, particularly in areas where multiple fuels are offered simultaneously: i) aggregate demand for fuel (adding up different types of fuel) is proportional to the size of the automobile fleet, and is thus inelastic in the short run; ii) location differences allow fuel stations to have a degree of market power; iii) any fuel station supplies both types of fuel, and selects both prices to maximize joint profits; iv) for flex car owners, ethanol and gasoline are perfect substitutes. In the model, there is (imperfect) competition across fuel stations, but between different types of fuel within a location there is no competition at all: each station provides both ethanol and gasoline, and internalizes the effect of a price change over the other product. The effect of an increase in the flex car fleet is to change the degree of substitutability between fuels, and not across stations. One may tend to think that, since competition and the direct effect of flex cars penetration operate in different dimensions of product differentiation, there would be no effect of the latter on the former. This is not true: we find that in equilibrium, flex car penetration leads to more competition across fuel stations. The model also predicts price dispersion across stations and among fuels within a station. In equilibrium, a gas station generally finds it optimal to charge prices that do not conform to the technical substitution ratio of 70%, 5

even when flex car penetration approaches 100%; thus this theory may help explain why this relationship is not observed in practice. We consider an oligopoly model where N gas stations compete by setting prices for gasoline and ethanol. The population of consumers, that we normalize to 1, is divided in three groups: gasoline-fueled car owners, ethanol-fueled car owners, and flex car owners. We call θ the fraction of flex car owners, and to keep the model symmetric, we assume that the rest of the population is equally divided across gasoline and ethanol. The type of fuel has no effect on consumption, except that for flex car owners, gasoline and ethanol within the same gas station are perfect substitutes. (We measure fuel in terms of energy content, so that flex car owners always buy the cheaper fuel). There is differentiation across fuel stations. For a given fuel price profile, let pif = min{pig , pia } (the price effectively faced by a flex type in station i). Using this notation, we assume that the demand for fuel from station i from a consumer with car type j = g (gasoline-powered), a (alcohol-powered) or f (flex) is qij = α − βpij + γ p¯−ij , where α, β and γ are positive constants, pij is the price of fuel j in station i, and p¯−ij is the average price of fuel j in all stations except i. We adopt the same functional form for all fuel types. This is done for simplicity and to isolate the effect on substitutability across fuels as the car fleet changes. For consumers within each car group, we assume that demand across stations exhibit a simple linear form of symmetric product differentiation. This demand system can be justified by Carlson and McAfee (1983), who model consumer choice by a process of costly search among identical products sold by different firms at (potentially) different prices.3 Carlson and McAffe show that, if the distribution of search costs in the consumer population is uniform, then aggregate demand for firm i exhibits the form postulated above, with α = 1/N and β = γ = (N − 1)/N , where N is the number of firms in the market. We seek to obtain a prediction about Bertrand-Nash equilibrium prices for firms that face the demand arising from this process and have a cost as follows: Ci (qig , qia ) = cig qig + cia qia + Fi , 3

Carlson and McAfee (1983) assume each consumer has a search cost x ≥ 0, drawn from a distribution H, independent of the type of car they own. Consumers know the equilibrium distribution of prices in the market, but initially do not know which firm offers the lowest price. Price search is modeled as follows: Each consumer is randomly assigned to a supplier, and observes the price in that location. She can either make a purchase or decide to search further. If she decides to search, she pays x and samples again from the supplier distribution. For simplicity, Carlson and McAfee (1983) assume sampling with replacement, with all firms equally likely. There is no time discounting (the waste of time in the search process is included in x). This process continues until the consumer makes a purchase.

6

where qig and qia are the quantities sold of gasoline and ethyl alcohol in station i, cig and cia are marginal costs, and Fi a fixed cost component. We assume marginal costs are constant and exogenous, but different across fuels and stations.

3.1

Properties of equilibrium prices

To obtain a characterization of equilibrium prices, we first need to integrate the demand over the mass of consumers with each car type. If pig 6= pia , station i will sell Qig of gasoline and Qia of ethanol, where   1−θ Qig = (α − βpig + γ p¯−ig ) 2 +1I{pig < pia }θ (α − βpig + γ p¯−if ) and 

Qia

 1−θ = (α − βpia + γ p¯−ia ) 2 +1I{pig > pia }θ (α − βpia + γ p¯−if ) .

(1I{A} stands for the indicator function, valued one if A is true and zero otherwise.) If pig = pia , flex car owners are indifferent between the two types of fuel, and we need to specify a sharing rule τ ∈ [0, 1]. Formally, we follow the approach of Simon and Zame (1990) and adopt an endogenous sharing rule, although the specifics of the tie-breaking do not affect the equlibrium determination in this model. The profit of station i is simply πi = (pig − cig )Qig + (pia − cia )Qia . Maximizing this expression with respect to pig and pia yield this firms bestresponse function. Whenever θqif is positive, πi is discontinuous at the point pig = pia (and the profit at the discontinuity point depends on the tie breaking rule τ ). In any pure-strategy equilibrium, we can classify stations into those that decide to charge pig > pia , pig < pia or pig = pia 4 . In the first two cases, profits are continuously differentiable around the chosen prices, and the latter can be characterized by the first order conditions ∂p∂ig πi = 0 and ∂ ∂pia πi

= 0. In the next lemma, we show that the last alternative is never optimal: in equilibrium no fuel station elects to charge pig = pia : Proposition 1 If qij is strictly increasing in p¯−ij and θ > 0 then firm i will post pig 6= pia in equilibrium. 4

That is, in practice, to charge prices within a tenth of a cent of each other.

7

Proof: If some of the other fuel stations charge a different price for each fuel, p¯−if < p¯−ig or p¯−ia . Without loss of generality, suppose p¯−if < p¯−ig . The profit firm i, as a function of pig , exhibits a kink at the point pig = pia . Let x and y be the right and left derivatives at that point: x=

∂ + π = (1 − θ/2)[q(pig , p¯−ig )) − β(pig − cig )] ∂pig i

and y=

∂ − π = x + (θ)[q(pig , p¯−if )) − β(pig − cig )] ∂pig i

For firm i to find it optimal to charge pig = pia , it must be the case that y ≥ 0 ≥ x. But this is impossible, since x ≤ 0 ⇒ y < 0.  Considering the two possible first order conditions that must be satisfied by pij , we obtain the following expressions:   1 γ α θ γ pig = cig + p¯−ig + − 1I{pig < pia } (¯ p−ig − p¯−if ) 2 β β 1+θβ and pia =

  1 θ γ γ α cia + p¯−ia + − 1I{pig > pia } (¯ p−ia − p¯−if ) 2 β β 1+θβ

Notice that p¯−if ≤ p¯−ig and p¯−if ≤ p¯−ia , so the right-hand sides of the expressions above are decreasing in θ. Since prices across stations are strategic complements, we conclude that prices are decreasing in θ. This result is summarized in the proposition below. Proposition 2 Ethanol and gasoline prices are decreasing with respect to the fraction of flex cars. A larger fleet of flex cars pulls prices down because flex cars provide an option value to their owners: if fuel prices are dispersed, flex car owners expect to find lower prices than other drivers since they can always pick the cheapest alternative. Because of this, they are willing to pay less, and fuel stations respond to a lower demand by lowering prices. Let’s turn to the analysis of the difference between gasoline and ethanol prices. Salvo and Huse (2010b) suggest that the increase of the flex fleet has reduced the spread between gasoline and ethanol prices. In our model this is not the case. The effect of flex car penetration on the difference between

8

gasoline and ethanol prices is ambiguous and depend on the competition pressures from other station in the market. More precisely, we have that   1 γ pig − pia = cig − cia + (¯ p−ig − p¯−ia ) − 2 β θ γ − [1I{pig < pia }(¯ p−ig − p¯−if ) − 1I{pia < pig }(¯ p−ia − p¯−if )] . 1+θβ Thus, the price difference pig − pia does not necessarily decrease with θ.

4

Data

4.1

Data sources

This study combines data from different sources. The first source is the Levantamento de Pre¸cos e de Margens de Comercializa¸c˜ ao de Combust´ıveis, a weekly survey carried out by the national petroleum agency (ANP), the Brazilian regulatory agency of the oil and gas industry. ANP collects data on retail prices for ethanol and gasoline prices, as well as prices paid to fuel distributors, at individual fuel stations in 10% of the municipalities in Brazil. There is also information on the brand of the station (or if it has no brand), the date prices were collected and the address of the station. Our sample contains weekly prices from January, 2002 to March, 2008 for stations located in 36 municipalities in the Rio de Janeiro state. Not all fuel stations are surveyed every week. Coverage is 100% in small municipalities, while for larger markets the survey adopts a rotating sample (with random selection) that eventually covers all fuel stations in the location. Table 1 provides information on the number of stations sampled, as a proportion of the overall population, in the 38 municipalities used in this study. The second data source is a monthly dataset on the number of cars with license plates from each municipality in the state of Rio de Janeiro, classified according to fuel type (gasoline, ethanol, flex, gasoline + CNG 5 , ethanol + CNG, flex + CNG). The time period is the same considered in the ANP’s survey (from January, 2002 to March, 2008). This data set was provided by the department of motor vehicles of the state of Rio de Janeiro (Detran - RJ). Although ANP verifies the price charged by fuel stations in all Brazilian states, we are not able to expand our analysis to the entire Brazilian territory because we do not have access to data on the number of cars by fuel type in other states. Combining the two data sources, we focus on the 38 municipalities from the state of Rio de Janeiro sampled by ANP that are listed in Table1. We included annual municipal GDP per capita as a regressor (obtained from the Brazilian institute of geography and statistics - IBGE). We also 5

CNG stands for compressed natural gas. In Brazil it is possible to convert vehicles to run on natural gas.

9

included annual data on number of hotels per square km in each municipality (provided by the data and information centre of Rio de Janeiro - CIDE RJ) as a proxy for the amount of drivers that are out of town. The former series is available from 2002 to 2006 and the latter from 2002 to 2004. Since we are analysing the period between 2002 and 2008, in both cases the missing years were replaced by the most recent available ones.

4.2

Descriptive statistics

Table 1 shows how the percentage of flex cars changed between 2004 and 2007 in the 38 municipalities in our sample. Figure 2 presents a map of Rio de Janeiro state, with a bullet for each municipality in our sample. The size of the bullet is proportional to (the log of) the local car fleet, and the color is coded according to flex car penetration in 2007. As we can see, the flex car fleet grew considerably across time in all locations, reaching a maximum value of 14.1% in the city of Mangaratiba in 2007. There is variation both in the time series and the cross-section/geographic dimensions. Figure 2 strongly suggests that flex car adoption is closely related to local income, occurring in the capital (Rio de Janeiro city), tourist resort towns (Arma¸c˜ ao de B´ uzios, Parati, Angra dos Reis) and rapidly developing areas (Maca´e, Mangaratiba). The municipalities with the lowest adoption rates are located in the northeastern part of the state, which is the traditional area of sugarcane production in Rio de Janeiro. Table 2 provides basic statistics about wholesale prices, retail prices and margins in our data. Table 4 presents the evolution of retail prices for each type of fuel. Ethanol average retail prices have been below the 70% threshold of gasoline average retail prices, except for the year of 2006. Thus, it is reasonable to infer that most owners of flex cars in our sample are choosing to fill their tanks with ethanol. Although we do not have access to data on gasoline and ethanol sales at stations, ANP publishes the distributors’ consolidated sales volume in the state of Rio de Janeiro, as shown in table 3. In this table, we also present the aggregate size of the fleet in our dataset. We have also computed sales per car, in cubic meters, in the last two columns. (For ethanol, we added the fleet of ethanol-powered and flex cars in the denominator, under the assumption that in this period flex car owners were mostly buying this fuel.)

5

Empirical Results

In this section, we document three effects of flex car penetration on the distribution of fuel prices. First, we estimate the effect of the penetration of the flex cars θ on the level of retail prices and margins, for both ethanol and gasoline. Second, we investigate, and fail to find, evidence that the price

10

spread between gasoline and ethanol is reduced by flex car penetration. Finally, we document that flex car penetration has made gasoline and ethanol prices more correlated within each station.

5.1

Effect of flex cars on retail prices and margins

According to proposition 2, the penetration of flex cars tightens the competition in fuel market, pushing down gasoline and ethanol prices and margins. Tables 5 and 6 present a reduced form analysis of the impact of flex cars on prices and margins. Columns (1)-(4) of both tables consider fuel prices, while columns (5)-(8) consider fuel margins. Each column adopts a different set of controls: monthly dummies, individual station fixed effects and brand fixed effects. This allows us to flexibly control for both aggregate effects over time and site-specific effects. In specifications (4) and (8) we also include gdp per capita, number of stations per car and number of hotels per squared kilometer to control for the effect of income growth, variation in relative fuel station scarcity and intensity of local tourist activity. Table 5 analyzes the effect of flex cars on gasoline market. Column (1) suggests that, contrary to the prediction of the model, the estimated coefficient is positive. However, this is a result drawn by a common upward trend of prices and flex car penetration. In fact, when we introduce time dummies, in column (2), the coefficient becomes negative. The point estimate implies that an increase of 10 percentage points in flex car penetration reduces gasoline price by 3.84/0.7 = 5.49 cents (of Brazilian Reais) per liter. Considering station and flag/brand fixed effects, this impact reduces to 2.74/0.7 = 3.91 cents per liter. Finally, controlling for gdp per capita, stations per car and hotel density, it becomes 2.05/0.7 = 2.93 cents per liter. The effect on gasoline margin presents the same pattern and have similar magnitude. Table 6 presents the same analysis for the ethanol case. Findings for the ethanol prices are similar, but the effect is larger. In our preferred specification (column 4), a 10 percentage point increase in flex car penetration reduces ethanol price by 7.8 cents (of Brazilian Reais) per liter. We estimate a negative, but significantly smaller, effect on the ethanol retail margins: a reduction of 3.7 cents per liter. This finding suggests that unlike the gasoline case, flex car penetration may have affected mark-ups in the ethanol upstream markets. In order to put the empirical relevance of the results above in perspective, it is important to consider the velocity in which the composition of the Brazilian fleet is changing over time. According to ANFAVEA (Brazilian Association of Automobile Producers), more than 90% of new cars sold in the country is flex since 2007. This may change substantially the figures exhibited above over time, with the renewal of the fleet.

11

5.2

Effect of flex cars on the spread between gasoline and ethanol prices

Salvo and Huse (2010b) suggest that the flex cars reduce the spread between gasoline and ethanol prices. In our model, on the other hand, the competitive changes created by flex cars in fuel markets determine an ambiguous effect on the equilibrium spread between gasoline and ethanol prices. We now investigate the empirical relationship between flex cars (θ) and the spread between gasoline and ethanol prices (pg − pa ). Figure 3 shows the scatter plot between the spread pg − pa and θ, with a spline trend estimated with bandwidth 8. The relationship depicted in the figure is non-monotonic. It decreases to 0 with θ until flex cars reach a fraction of 6%. Then, it increases and stays away from 0 for larger values of θ.6 Although figure 3 suggests a non monotonic effect of the flex cars on spread, it might be contaminated by other undesired sources of variation. In order to take this into account, table ?? presents regressions of the the absolute value of the spread |pg − pa | on θ, controlling for different sets of fixed-effects and other variables. In columns (1) and (2), the coefficient on the penetration of flex cars is negative and statistically significant. However, after eliminating the crosssection variation, the coefficient becomes positive. Notice that, as in the previous analysis, the empirical environment emulated in columns (3) and (4) is closer to the one described in the model, where a given market experience an increase of the flex car fleet. This positive coefficient, although compatible with our model, is against the results found by Salvo and Huse (2010b).

5.3

Fuel price correlation

While we do not find evidence that the spread of fuel prices has decreased, we have found that in markets with more flex penetration, fuel prices tend to be more correlated, which is consistent with the hypothesis that fuel stations make the pricing decisions of both fuels jointly. In this section, we provide evidence that the correlation between fuel prices has increased with flex car penetration using two different methods. In the first method, we assume that fuel prices are jointly distributed with flex car penetration affecting both the expectation and the covariance of prices. Put another way, we assume that the equations estimated in tables 5 and 6 form a system of seemingly unrelated regressions (SUR), and the residuals are heteroskedastic, with the residuals covariance increasing 6 Figure 4 plots the spread against time. There are substantial seasonal effects that affect the affect relative prices of the fuels, but there is no evidence of a reduction of the spread ofver time.

12

with flex car penetration. A simple way to test this hypothesis is to regress the product of the residuals from regressions in tables 5 and 6 in flex car penetration, which is the same as the second stage in the standard feasible GLS procedure to estimate a SUR model. Coefficients in this regression show how the (conditional) covariance of fuel prices depends on the regressors. Table 8 presents the results of this regression. We consider three specifications: the first one is a regression involving flex car penetration and station fixed effects; in the second, brand effects are added, and in the third, month effects are added. In all specifications, we find that flex car penetration has a positive and significant effect on the price covariance. In the second method, we introduce interactions of flex fuel penetration with the price of the other fuel on the the original price regressions. If our hypothesis is correct, in markets with more flex fuel cars the price of a given fuel should respond more sharply to shifts in the price of the other fuel — which would lead to a positive coefficient for this interaction term. The results of the second method are presented in table 9. We find that all interaction terms are positive as predicted, and all but one are statistically significant.

6

Structural Estimation of the Model

In this section, we employ the theory of price formation that we proposed in section 3 in order to estimate empirically properties of fuel demand from price data alone. In order to bring the theory to the data, we change two aspects of the basic model presented in section 3: First, we recognize that the gasoline fleet is much larger than the ethanol fleet in our sample: let the fleet of gasoline, ethanol and flex vehicles be θg , θa and θf , respectively. Second, we recognize that demand for fuel may differ systematically with car type, in response to differences in usage and fleet composition: let the demand for fuel from a vehicle of type j from station i in market m is qij = αmj − βj pij + γj p¯ij , where αmj = αm + αj , the composition of a market-specific fixed effect and a fuel-specific intercept. Retaking the same analysis performed in section 3, we obtain the following first-order conditions:   αmg γg 1 pig = cig + 1I{pig > pia } + p¯ig 2 2βg 2βg   θg αmg + θf αf θf γ f θg γ g +1I{pig < pia } + p¯ig + p¯if , 2(θg βg + θf βf ) 2(θg βg + θf βf ) 2(θg βg + θf βf ) 13

pia

  1 αma γa = + p¯ia cia + 1I{pia > pig } 2 2βa 2βa   θf γ f θa αma + θf αf θ a γa +1I{pia < pig } + p¯ia + p¯if . 2(θa βa + θf βf ) 2(θa βa + θf βf ) 2(θa βa + θf βf )

Our objective in this section is to seek to identify the demand parameters by estimating price best response functions (Pinkse, Slade, and Brett, 2002). Since we do not observe quantities directly, we cannot identify the absolute scale of the demand coefficients only from pricing responses, and a normalization must be made. For convenience, we normalize βf = 1. If we further assume in that βg = βa = βf = 1, we obtain best responses that are linear in the remaining parameters: hα γg i 1 mg pig = cig + 1I{pig > pia } + p¯ig 2 2 2   θg αmg + θf αmf θf γ f θg γ g +1I{pig < pia } + p¯ig + p¯if , 2(θg + θf ) 2(θg + θf ) 2(θg + θf )

pia =

hα γa i 1 ma cia + 1I{pia > pig } + p¯ia 2 2 2   θa αma + θf αmf θf γ f θa γ a +1I{pia < pig } + p¯ia + p¯if . 2(θa + θf ) 2(θa + θf ) 2(θa + θf )

We report estimates for both the case where the β coefficients are assumed to be identical and for the more general, nonlinear case. In order to estimate these models, we must deal with two aditional issues. First, we need to define the relevant market for each fuel station. For this sake, we define a market to consist of all stations that share the same fivedigit zip code (CEP). For the smaller cities the use of the five-digit zip code is equivalent to definining the market as the whole city. In larger cities, on the other hand, five-digit zip codes roughly correspond to specific neighborhoods. The second challenge is how to deal with endogeneity. If there are stochastic unobserved components in the demand function or in the marginal cost, since all prices are determined in equilibrium, all terms in the right hand side of the form p¯ij or 1I{pij < pik } are endogeneous. (Throughout, we still maintain the assumptions that the fleet composition and the marginal costs are exogenous.) To handle this problem, we follow Pinkse, Slade, and Brett (2002) and instrument each endogenous regressor by the analogous term involving costs: that is, we substitute 1I{cij < cik } for 1I{pij < pik }, etc. In addition, we estimate the model with and without market fixed effects, to account for unobserved variation across markets. In the linear case, we 14

can add fixed effects for each market (zip code) separately. In the nonlinear case, we use municipality-specific fixed effects. Finally, we can estimate price reaction functions separately for each fuel type (which will yield two different sets of estimates for the demand from flex car owners) or stack the data to impose the restriction that αf and γf must be the same in both regressions. Tables 10 and 11 present the results of estimating the linear model separately and jointly, respectively. Table 12 presents the results of the nonlinear model. Columns 1, 2 and 3 presents estimates of the nonlinear model without imposing the theoretical restriction that the coefficients on cost should be 0.5. A remarkable finding is that, even though we do not impose the restriction that the coefficient on cij is 0.5, our estimates are near that figure. The coefficient of cia is not statistically different from 0.5 in column 3. This fact suggests that pricing in this market does indeed comply with the logic of a price-setting oligopoly game. This coefficient is also of independent interest; the fact that is less than unity means that demand is cost-absorbing; according to Weyl and Fabinger (2009), a number of comparative statics predictions can be derived from that fact. For example, entry of a new station will necessarily reduce prices, and a merger (without synergies) will raise prices of all firms (Weyl and Fabinger, 2009, theorem 4). In column 4, we present estimates of the model obtained when we impose the theoretical restriction that the coefficients on costs should be 0.5. This is our preferred specification. We find estimates for βG and βA of around 0.95, not substantially different from the values assumed in the linear model. We also find that the other coefficients do not differ substantially from those obtained in tables 10 and 11. Due to our normalization, the magnitude of the γj coefficients represent the ratio between the slopes of demand with respect to i’s price and the average market price. If γj = βj (= 1), as in the Carlson and McAfee (1982), aggregate demand for fuel is inelastic. Our estimates suggest that aggregate demand for gasoline and ethanol car owners is nearly inelastic, whereas for flex car owners, demand is more elastic (estimates for γf are imprecise, but tend to be subtantially smaller than one). Finally, column (5) in table 12 presents the results obtained when we widen our definition of the local market: here we specify as a market an areas with the same first 5 digits of the CEP code. The number of usable observations increases considerably as fewer markets now contain a single observation; however, the estimated coefficients do not change substantially.

6.1

Exploring the demand estimates

In this section, we report some counterfactual calculations based on our demand estimates presented in the previous section. Since we have used 15

price best responses to estimate these parameters, our procedure does not identify the scale of the coefficients. To obtain quantity predictions from our demand estimates, we need additional information on the volume sold of fuel. We only have information on quantities aggregated at the state × year level. From the ANP yearbook (ANP, 2010) we know that 1635 thousand cubic meters of gasoline and 359400 cubic meters of ethanol were sold in the Rio de Janeiro state in 20077 , an average of 15000 liters of gasoline and 3299 liters of ethanol per fuel station per week. We can use these figures to find the appropriate scale for the demand coefficients. The multipliers that match the observed average quantities sold in 2007 are 25.21 and 25.35 respectively for gasoline and ethanol. Using these multipliers, we obtain the coefficients for demand in our model presented in table 13. At observed prices, on average in our sample demand per car per station per week is 5.52 liters for ethanol-only cars, 5.59 liters for gasoline-only cars, and 11.65 liters of ethanol and 6.35 liters of gasoline for flex cars. We believe these large differences in consumption reflect the fact that the flex car and more intensively used. Also, the model predicts that flex cars owners are using mostly (but not exclusively) ethanol since this is the cheapest fuel most of the time. In figure 5 we present these figures and sketch how they would change if one would shift the average price of ethanol, holding the prices of gasoline and the distribution of ethanol prices as observed in the data. We show the effect of ethanol price on the ethanol market in the top panel, and the effect on the gasoline market on the bottom panel. In the top panel, the demand by ethanol-only vehicles is linear (by assumption) and estimated to be nearly inelastic. (In the bottom panel, the demand by gasoline-only vehicles is not affected by ethanol prices, by assumption.) On the other hand, demand by flex car owners is subtantially more elastic, as relative price shifts lead to substitution across fuels. However, this substitution is not complete: due to local price heterogeneity, even when the price of ethanol on average 40 cents lower, there is still positive demand for gasoline by flex cars owners in markets where cheap ethanol is unavailable.

7

Concluding Remarks

In this work we investigate how the penetration of flex cars have affected the fuel retail market in the state of Rio de Janeiro. Our main hypothesis was that flex car penetration has increased the degree of substitution between gasoline and ethanol and that fuel stations would respond strategically to this shock reducing retail prices in equilibrium. 7

C.f. table 3.

16

Our estimates suggest that the model prediction is correct and as the percentage of flex cars increase by 10%, ethanol and gasoline energy equivalent prices per liter fall approximately by 8 cents and 2 cents, respectively. Considering the volume of sales and the size of the flex car fleet in 2007, we estimate consumer savings on the order of 70 million Reais in Rio de Janeiro state that year. Our estimates also show that the price gap between the two fuels increased, and price correlation has increased, with flex car penetration. We also propose a method to structurally identify fuel demand parameters from the estimation of best reply price response functions. We find that the coeeficient of pass-through from costs to prices in fuel retail is near 0.5, as predicted by oligopoly theory. In future work we intend to develop further the theoretical model to obtain an estimable expression for the best reply pricing functions that account more realistically for geographical differentiation among fuel stations. We believe that this analysis, combined with detailed geographic information from our unique dataset, can lead to further insights into how competition works in this market.

References Anderson, S. (2010): “The demand for ethanol as a gasoline substitute,” Discussion paper, National Bureau of Economic Research. ANP (2010): “Anu´ ario Estat´ıstico Brasileiro do Petr´oleo e do G´as Natural 2010,” http://www.anp.gov.br/?dw=33213, accessed on October 20th, 2010. Boff, H. P. (2011): “Modeling the Brazilian Ethanol Market: How FlexFuel Vehicles are Shaping the Long Run Equilibrium,” China-USA Business Review, 10(4), 245–264. Carlson, J., and R. McAfee (1982): “Discrete equilibrium price dispersion: Extensions and technical details,” Discussion paper, Institute for Research in the Behavioral, Economic, and Management Sciences, Krannert Graduate School of Management, Purdue University. (1983): “Discrete equilibrium price dispersion,” The Journal of Political Economy, 91(3), 480–493. Corts, K. S. (2009): “Building out alternative fuel retail infrastructure: Government fleet spillovers in E85,” Discussion paper, Center for the Study of Energy Markets. Ferreira, A., F. Prado, and J. Silveira (2009): “Flex cars and the alcohol price,” Energy Economics, 31(3), 382–394. 17

Klier, T., and J. Linn (2010): “The Price of Gasoline and New Vehicle Fuel Economy: Evidence from Monthly Sales Data,” American Economic Journal: Economic Policy, 2, 134–153, http://www.aeaweb.org/articles.php?doi=10.1257/pol.2.3.134. ´ Marjotta-Maistro, M. C., and G. A. Asai (2006): “Alcool combust´ıvel: do carro a ´ alcool ao carro flex,” mimeo. Pinkse, J., M. Slade, and C. Brett (2002): “Spatial price competition: A semiparametric approach,” Econometrica, 70(3), 1111–1153. Salvo, A., and C. Huse (2010a): “Consumer Choice between Gasoline and Sugarcane Ethanol,” Northwestern University and Stockholm School of Economics. (2010b): “Is Arbitrage Tying the Price of Ethanol to that of Gasoline? Evidence from the Uptake of Flexible-Fuel Technology,” Northwestern University and Stockholm School of Economics; forthcoming, Energy Journal. Simon, L., and W. Zame (1990): “Discontinuous Games and Endogenous Sharing Rules,” Econometrica, 58, 861–872. USDA (2010): “Sugar and Sweeteners Yearbook Tables,” http://www.ers.usda.gov/Briefing/Sugar/data.htm#yearbook, accessed on October 20th, 2010. Weyl, E. G., and M. Fabinger (2009): “Passthrough as an Economic Tool,” Harvard University.

Registration of new cars by fuel type

Figure 1: Registration of new cars by fuel type — Share in percentage.

(share in  percentage)

100 80 60 40 20 0

% gasoline 

% ethanol 

18

% flex 

Table 1: Stock of Vehicles, Percentage of Flex Cars and Number of Fuel Stations, by City

Mangaratiba Niter´ oi Maca´e Rio de Janeiro Arma¸ca ˜o de B´ uzios Angra dos Reis Parati Maric´ a Resende Cabo Frio Nil´ opolis Trˆes Rios Araruama Saquarema Barra Mansa Vassouras Volta Redonda Para´ıba do Sul S˜ ao Gon¸calo Petr´ opolis Sapucaia Belford Roxo Teres´ opolis Queimados Nova Igua¸cu Mag´e Duque de Caxias Nova Friburgo Itagua´ı Valen¸ca S˜ ao Jo˜ ao de Meriti Rio Bonito Itabora´ı Barra do Pira´ı Santo Antˆ onio de P. Itaperuna Campos dos G. S˜ ao Francisco de I. Total

2004 Number of % vehicles 3.090 176.647 41.048 1.800.614 4.340 21.503 2.937 14.193 26.085 32.315 24.305 14.792 20.290 9.955 31.788 7.297 68.441 6.101 107.977 86.912 1.381 26.574 42.920 9.278 115.523 21.034 128.426 60.539 28.016 10.297 64.607 22.251 27.800 18.436 9.234 18.810 91.651 3.395 3.200.802

of flex cars 1,2% 1,0% 1,1% 0,8% 1,4% 0,9% 0,6% 0,7% 0,8% 1,0% 0,5% 0,9% 0,6% 0,6% 0,8% 1,0% 0,7% 0,8% 0,4% 0,6% 0,9% 0,3% 0,5% 0,3% 0,4% 0,4% 0,3% 0,4% 0,2% 0,3% 0,2% 0,5% 0,3% 0,5% 0,7% 0,6% 0,3% 0,3% 0,7%

19

2007 Number of % of vehicles 4.812 194.511 57.261 1.969.128 6.703 27.030 3.675 24.120 33.368 46.371 29.579 18.246 28.037 13.771 36.420 8.319 81.203 7.427 134.860 97.177 1.508 37.744 50.787 13.612 137.337 27.495 151.257 68.588 31.609 11.540 76.891 35.453 37.075 21.239 10.540 22.123 109.388 4.204 3.670.408

of flex cars 14,1% 12,4% 11,1% 10,5% 10,5% 10,0% 9,1% 9,0% 8,9% 8,6% 7,4% 7,3% 7,2% 7,1% 6,9% 6,9% 6,7% 6,5% 6,5% 6,3% 6,2% 6,1% 5,6% 5,6% 5,5% 5,5% 5,3% 5,3% 4,6% 4,5% 4,4% 4,4% 4,3% 4,3% 4,2% 3,8% 2,6% 2,3% 8,8%

Number of fuel stations In the Total % sampled sample (2010) weekly 10 11 91% 34 85 40% 14 23 61% 190 805 24% 7 7 100% 13 23 57% 7 11 64% 11 22 50% 14 26 54% 10 23 43% 9 10 90% 10 15 67% 16 23 70% 9 13 69% 21 31 68% 6 10 60% 21 30 70% 7 7 100% 35 88 40% 18 54 33% 10 13 77% 11 19 58% 23 31 74% 7 7 100% 25 65 38% 10 13 77% 30 88 34% 21 38 55% 8 13 62% 10 15 67% 20 33 61% 11 17 65% 20 34 59% 10 16 63% 10 16 63% 16 20 80% 38 109 35% 8 10 80% 750 1874 40%

Table 2: Fuel wholesale and retail prices and margins (in R$)

Ethanol retail price Ethanol wholesale price Ethanol margin Gasoline retail price Gasoline wholesale price Gasoline margin

Observations 266,030 154,185 154,185 286,126 198,866 198,866

Mean (R$) 1.510 1.253 0.249 2.279 2.015 0.258

St. Deviation 0.331 0.338 0.133 0.332 0.298 0.104

Min 0.600 0.278 -0.927 1.299 1.140 -0.747

Max 2.879 2.527 0.964 3.119 2.846 1.469

Table 3: Distributors’ Fuel Sales, in cubic meters, and size of fleet in the State of Rio de Janeiro

2001 2002 2003 2004 2005 2006 2007

Gasoline Sales (a) 1,772,337 1,971,934 1,764,595 1,848,172 1,739,319 1,660,803 1,635,152

Ethanol Sales (b) 155,572 157,567 98,178 109,817 180,528 224,255 359,404

Gasoline Fleet (c) 2,424,674 2,606,238 2,775,071 2,879,902 2,958,560 3,016,335 3,098,499

Ethanol Fleet (d) 471,053 473,434 478,060 476,632 475,307 473,880 472,670

Flex Fleet (e) 23,561 84,297 188,271 335,629

(a)/(c)

(b)/[(d)+(e)]

0.73 0.76 0.64 0.64 0.59 0.55 0.53

0.33 0.33 0.21 0.22 0.32 0.34 0.44

Table 4: Gasoline and Ethanol Average Prices by Year in the State of Rio de Janeiro (in R$)

2002 2003 2004 2005 2006 2007

Ethanol Price 1.06 1.40 1.30 1.56 1.88 1.70

Gasoline Price 1.71 2.12 2.11 2.37 2.60 2.57

20

Ethanol Price/ Gasoline Price 0.62 0.66 0.61 0.66 0.72 0.66

Frequency of Ethanol pr. < 0.7 Gasoline pr. 0.78 0.60 0.76 0.71 0.37 0.66

21

Gasoline price (pg ) (2) (3) ∗∗∗ -0.384 -0.274∗∗∗ (0.0744) (0.0263)

(4) % flex cars (θ) -0.205∗∗∗ (0.0294) gdp per capita 0.00109∗∗∗ (0.000138) stations/car -0.000211∗∗∗ (0.0000213) hotels/km2 0.0545∗ (0.0243) ∗∗∗ ∗∗∗ ∗∗∗ constant 1.480 1.103 1.097 1.100∗∗∗ (0.00431) (0.00679) (0.00551) (0.00736) monthly dummies No Yes Yes Yes station fixed effects No No Yes Yes flag/brand fixed effects No No Yes Yes N 280827 280827 280798 272243 R2 0.388 0.876 0.975 0.975 Robust standard errors in parentheses, clustered by city-week. Regressions include month, brand and station fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

(1) 5.270∗∗∗ (0.162)

(5) 0.458∗∗∗ (0.0379)

Gasoline margin (pg − cg ) (6) (7) (8) ∗∗∗ ∗∗∗ -0.428 -0.305 -0.279∗∗∗ (0.0518) (0.0216) (0.0262) 0.000411∗∗∗ (0.000124) -0.000105∗∗∗ (0.0000167) 0.0434∗ (0.0206) ∗∗∗ ∗∗∗ ∗∗∗ 0.171 0.143 0.141 0.140∗∗∗ (0.00105) (0.00407) (0.00328) (0.00578) No Yes Yes Yes No No Yes Yes No No Yes Yes 195528 195528 195501 189942 0.031 0.120 0.701 0.697

Table 5: Effect of flex car penetration on gasoline price and margin

22

Ethanol price (pa ) (2) (3) ∗∗∗ 0.396 -0.795∗∗∗ (0.118) (0.0756)

(4) % flex cars (θ) -0.780∗∗∗ (0.0901) gdp per capita -0.00183∗∗∗ (0.000365) stations/car -0.000175∗∗∗ (0.0000364) hotels/km2 0.0637 (0.0455) constant 1.381∗∗∗ 1.036∗∗∗ 1.033∗∗∗ 1.056∗∗∗ (0.00624) (0.00898) (0.00530) (0.0127) monthly dummies No Yes Yes Yes station fixed effects No No Yes Yes flag/brand fixed effects No No Yes Yes N 261048 261048 261023 252695 R2 0.236 0.795 0.928 0.928 Robust standard errors in parentheses, clustered by city-week. Regressions include month, brand and station fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

(1) 5.847∗∗∗ (0.242)

(5) 0.0608 (0.0413)

Ethanol margin (pa − ca ) (6) (7) (8) ∗∗∗ -0.000272 -0.351 -0.370∗∗∗ (0.0808) (0.0632) (0.0782) 0.000601∗ (0.000290) 0.000162∗∗∗ (0.0000351) -0.0532 (0.0372) 0.248∗∗∗ 0.209∗∗∗ 0.218∗∗∗ 0.205∗∗∗ (0.00116) (0.00572) (0.00531) (0.0110) No Yes Yes Yes No No Yes Yes No No Yes Yes 151687 151687 151665 146841 0.000 0.067 0.358 0.357

Table 6: Effect of flex car penetration on ethanol price and margin

Table 7: Effect of flex car penetration on spread between gasoline and ethanol prices Absolute spread (|pg − pa |) (1) (2) (3) (4) % flex cars (θ) -0.351∗∗∗ -0.509∗∗∗ 0.192∗∗∗ 0.183∗∗∗ (0.0602) (0.0524) (0.0542) (0.0543) gdp per capita -0.00214∗∗∗ (0.000536) stations/car -0.000102∗∗∗ (0.0000271) hotels/km2 -0.0598 (0.0346) Constant 0.160∗∗∗ 0.0974∗∗∗ 0.0855∗∗∗ 0.129∗∗∗ (0.00190) (0.00333) (0.00394) (0.00945) monthly dummies No Yes Yes Yes station fixed effects No No Yes Yes flag/brand fixed effects No No Yes Yes N 260472 260472 260447 259098 R2 0.007 0.394 0.529 0.529 Standard errors in parentheses ∗

p < 0.05,

∗∗

p < 0.01,

∗∗∗

p < 0.001

Table 8: The effect of flex cars penetration on fuel prices correlation, I product of residuals from tables 5 and 6, specification (4) (1) (2) (3) ∗∗∗ ∗∗∗ % flex cars 0.0000552 0.0000561 0.0000712∗∗ (0.0000115) (0.0000115) (0.0000266) ∗∗∗ 0.00163 0.00142∗∗∗ 0.00137∗∗∗ constant (0.0000380) (0.000151) (0.000323) Station Fixed-Effects Yes Yes Yes Brand Fixed-Effects No Yes Yes Month Fixed-Effects No No Yes Robust standard errors in parentheses, clustered by city-week. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

23

24

Gasoline price (2) -0.0198∗∗∗ (0.00132) 0.0103∗∗∗ (0.000702) 0.165∗∗∗ (0.00320) 0.297∗∗∗ (0.00751) (3) -0.0174∗∗∗ (0.00134) 0.00955∗∗∗ (0.000699) 0.166∗∗∗ (0.00320) 0.297∗∗∗ (0.00750) -0.00447∗∗∗ (0.00110)

gdp per capita

0.00187∗∗∗ 0.00136∗∗∗ 0.00135∗∗∗ (0.000183) (0.000170) (0.000171) stations/car -0.000241∗∗∗ -0.000210∗∗∗ -0.000182∗∗∗ (0.0000272) (0.0000254) (0.0000250) hotels/km2 0.107∗∗∗ 0.101∗∗∗ 0.122∗∗∗ (0.0321) (0.0275) (0.0302) ∗∗∗ ∗∗∗ cons 1.364 0.979 0.992∗∗∗ (0.0106) (0.0138) (0.0138) N 252145 175424 175424 Robust standard errors in parentheses, clustered by city-week. Regressions include month, brand and station fixed effects. ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001

Ethanol fleet

Ethanol wholesale price

Gas price

% Flex×Gasoline price

Gasoline fleet

Gas wholesale price

Ethanol price

% Flex×Ethanol price

% Flex

(1) -0.0243∗∗∗ (0.00135) 0.0133∗∗∗ (0.000713) 0.186∗∗∗ (0.00331)

-0.00277∗∗∗ (0.000335) -0.00000814∗ (0.00000381) -0.0139 (0.0390) 0.0997∗∗∗ (0.0206) 252145

0.000220 (0.00209) 0.617∗∗∗ (0.0115)

(4) -0.00614 (0.00557)

-0.00197∗∗∗ (0.000286) 0.0000105∗∗ (0.00000393) -0.0407 (0.0301) 0.0132 (0.0175) 146543

0.0185∗∗∗ (0.00184) 0.502∗∗∗ (0.00968) 0.316∗∗∗ (0.00418)

Ethanol price (5) -0.0519∗∗∗ (0.00490)

Table 9: The effect of flex cars penetration on fuel price correlation, II

0.00910∗∗∗ (0.00186) 0.522∗∗∗ (0.00996) 0.315∗∗∗ (0.00411) -0.0329∗∗∗ (0.00336) -0.00173∗∗∗ (0.000284) 0.0000188∗∗∗ (0.00000395) 0.00654 (0.0312) -0.00481 (0.0179) 146543

(6) -0.0217∗∗∗ (0.00511)

25

Gasoline price (2) (3) OLS IV 0.615∗∗∗ 0.646∗∗∗ (0.00329) (0.00508) 4.663∗∗∗ -0.770 (0.236) (0.812) 0.845∗∗∗ 0.836∗∗∗ (0.00602) (0.00918) -1.522∗∗∗ 1.338∗∗ (0.132) (0.442)

∗

p < 0.05,

∗∗

p < 0.01,

∗∗∗

p < 0.001

Standard errors in parentheses

Constant

0.00765∗∗∗ 0.0535∗∗∗ 0.0198∗∗∗ (0.00134) (0.00132) (0.00171) Market Fixed-Effects No Yes No N 93169 89124 83402 R2 0.947 0.949 0.943 Estimation method subsumes that βG = βA = βF (= 1).

γF

γA

αF

ca

γF

γG

αF

cg

(1) OLS 0.413∗∗∗ (0.00278) -0.283 (0.257) 1.261∗∗∗ (0.00490) 1.612∗∗∗ (0.143)

0.0441∗∗∗ (0.00169) Yes 79357

(4) IV 0.530∗∗∗ (0.00524) 8.594∗∗∗ (0.772) 1.007∗∗∗ (0.00963) -3.507∗∗∗ (0.419) 0.512∗∗∗ (0.00248) 0.624∗∗∗ (0.0638) 0.979∗∗∗ (0.00533) 0.548∗∗∗ (0.0369) 0.132∗∗∗ (0.00201) No 68858 0.903

(5) OLS

0.554∗∗∗ (0.00364) 1.363∗∗∗ (0.128) 0.842∗∗∗ (0.00833) 0.181∗ (0.0767) 0.167∗∗∗ (0.00268) No 57494 0.898

Ethanol price (6) (7) OLS IV

0.543∗∗∗ (0.00251) 1.379∗∗∗ (0.0661) 0.871∗∗∗ (0.00553) -0.0192 (0.0385) 0.173∗∗∗ (0.00212) Yes 65671 0.892

Table 10: Structural regressions, separate equations

0.538∗∗∗ (0.00351) 1.356∗∗∗ (0.136) 0.856∗∗∗ (0.00821) 0.178∗ (0.0818) 0.178∗∗∗ (0.00277) Yes 54307

(8) IV

Table 11: Structural regressions, stacked equations (1) (2) (3) OLS OLS IV cg 0.466∗∗∗ 0.627∗∗∗ 0.640∗∗∗ (0.00346) (0.00391) (0.00637) ca 0.502∗∗∗ 0.531∗∗∗ 0.553∗∗∗ (0.00187) (0.00182) (0.00275) αF 0.548∗∗∗ 1.430∗∗∗ 1.302∗∗∗ (0.0483) (0.0476) (0.0968) ∗∗∗ γF 0.625 -0.0102 0.222∗∗∗ (0.0278) (0.0276) (0.0578) ∗∗∗ ∗∗∗ γG 1.190 0.832 0.846∗∗∗ (0.00640) (0.00742) (0.0124) γA 1.000∗∗∗ 0.909∗∗∗ 0.846∗∗∗ (0.00402) (0.00396) (0.00630) Constant -0.144∗∗∗ -0.0592∗∗∗ -0.127∗∗∗ (0.00413) (0.00404) (0.00493) Stack (1 if gas, 2 if eth) 0.136∗∗∗ 0.109∗∗∗ 0.147∗∗∗ (0.00246) (0.00235) (0.00305) Market Fixed-Effects No Yes No N 161784 154561 140668 R2 0.922 0.917 0.920 Estimation method subsumes that βG = βA = βF (= 1). Standard errors in parentheses ∗

p < 0.05,

∗∗

p < 0.01,

∗∗∗

p < 0.001

26

(4) IV 0.583∗∗∗ (0.00724) 0.532∗∗∗ (0.00259) 1.486∗∗∗ (0.0974) 0.115∗ (0.0582) 0.921∗∗∗ (0.0143) 0.873∗∗∗ (0.00597) -0.0872∗∗∗ (0.00490) 0.129∗∗∗ (0.00292) Yes 133445

27

Table 12: Structural regressions, Nonlinear model (1) (2) (3) (4) Gasoline price Ethanol price Joint equations Joint equations IV IV IV IV cig .571∗∗∗ 0.572∗∗∗ .5 (.00901) (0.00799) cia .534∗∗∗ 0.503∗∗∗ .5 (.00455) (0.00380) αG .0778 0.0378∗∗∗ 0.0288∗∗∗ (.105) (0.00662) (0.00557) αA .215∗∗ 0.300∗∗∗ 0.303∗∗∗ (.0806) (0.0436) (0.0539) αF 1.489 1.553∗∗∗ 1.602∗∗∗ 1.617∗∗∗ (2.262) (.393) (0.208) (0.238) βG .981 0.974∗∗∗ 0.940∗∗∗ (1.307) (0.128) (0.161) ∗∗ ∗∗∗ βA .976 0.955 0.955∗∗∗ (.370) (0.138) (0.171) βF 1 1 1 1 γG .940 0.947∗∗∗ 1.041∗∗∗ (1.252) (0.120) (0.179) ∗∗ ∗∗∗ γA .907 0.907 0.908∗∗∗ (.347) (0.132) (0.163) γF .120 .231 0.314∗∗ 0.341∗∗ (1.290) (.223) (0.114) (0.126) Market definition 8 digit CEP 8 digit CEP 8 digit CEP 8 digit CEP City Fixed-Effects Yes Yes Yes Yes N 55304 55304 55304 55304 Standard errors in parentheses ∗ p < 0.05, ∗∗ p < 0.01, ∗∗∗ p < 0.001 0.0210∗∗∗ (0.00417) 0.284∗∗∗ (0.0493) 1.600∗∗∗ (0.202) 0.938∗∗∗ (0.156) 0.952∗∗∗ (0.165) 1 1.042∗∗∗ (0.174) 0.922∗∗∗ (0.160) 0.313∗∗ (0.108) 5 digit CEP Yes 94152

.5

(5) Joint equations IV .5

28

9.1%

10.0%

4.6%

> 10%

7% -10%

5% - 7%

2% - 5%

6.9%

14.1%

6.7%

4.3%

4.6%

4.50%

5.6%

6.9%

10.5%

7.4%

5.5%

6.5%

4.4%

6.1%

5.3%

Percentage of Flex Cars 2007

12.4%

5.5%

6.3%

7.3%

6.5%

4.3%

5.6%

9.0%

6.2%

4.4%

7.1%

5.3%

7.2%

11.1%

4.2%

8.6%

3.8%

10.5%

Figure 2: Geographic distribution of sample across Rio de Janeiro state.

2.6%

2.3%

Table 13: Unnormalized demand coefficients

αG βG γG αA βA γA αF βF γF

Original Estimates 0.03 0.94 1.04 0.30 0.96 0.91 1.62 1.00 0.34

Scaling to match: Gasoline sales Ethanol sales 0.73 0.73 23.70 23.83 26.24 26.39 7.64 7.68 24.08 24.21 22.89 23.02 40.76 40.99 25.21 25.35 8.60 8.64

-1

-.5

Spread (pg-pa) 0

.5

1

Figure 3: Flex cars and the spread between gasoline and ethanol prices

0

.05

.1 !

29

.15

Figure 4: Spread between gasoline and ethanol prices over time

30

Figure 5: Effect of average ethanol price change in demand for fuel

2 by ethanol only cars by flex cars

price of ethanol R$/liter

1.8

1.6

1.4

1.2

1 0

5

10

15

20

demand for ethanol in liters, per car per station per week

2

price of ethanol R$/liter

1.8

1.6

1.4

1.2 by gasoline only cars by flex cars 1 0

5

10 demand for gasoline in liters, per car per station per week

31

15

20