Consumer Search and Automobile Dealer Co-Location

Consumer Search and Automobile Dealer Co-Location Charles Murry and Yiyi Zhou∗ June, 2016

Abstract Retailers co-locate with rivals in order to take advantage of economies of agglomeration when consumers have limited information and engage in costly search, even though co-location implies fiercer price competition. We estimate a structural model of consumer search for spatially differentiated products using rich data on new car transactions. We use the model to separately disentangle the competition and agglomeration effects of retail co-location by simulating retail closures. A full information model that ignores the agglomeration effect would overstate the gains to incumbent rivals and the welfare loss to consumers due to car dealer closures.

Keywords: retail agglomeration, spatial competition, car dealers, retail exit JEL Classification: D83, L13, L62

∗

Charles Murry: Department of Economics, Pennsylvania State University, University Park, PA, 16802, [email protected]. Yiyi Zhou: Department of Economics and College of Business, Stony Brook University, Stony Brook, NY 11794-4384, USA, [email protected]. We thank Simon Anderson, Paul Grieco, Peter Newberry, Regis Renault, Henry Schneider, Steven Stern, Matthijs Wildenbeest, Mo Xiao, and participants at IIOC 2014 in Chicago for useful discussions and comments. The authors are solely responsible for all errors.

1

1

Introduction

Economists have long sought to understand the location decisions of firms and these decisions’ effect on industry profits and consumer welfare. The colocation of firms is especially ubiquitous in many retail industries, and there has been special attention paid as to why some firms tend to locate near each other even though this would typically imply greater price competition.1 A classic explanation for the co-location of retail stores has to do with limited consumer information, see Stahl (1982) and Wolinsky (1983). The basic idea is that if consumers must engage in costly search in order to resolve informational problems before purchase, then consumers are more likely to search areas where there is a concentration of stores in order to limit search costs. This agglomeration effect encourages co-location of stores. However, if stores are close to each other then price competition may be fierce, potentially outweighing benefits to stores from co-location.2 Understanding the agglomeration and competition effects of co-location has important implications for evaluating the consequences of retail closures. On the one hand, the agglomeration effect implies that a nearby rival’s exit would reduce the total attraction of that geographic area and force the incumbent firms to lower their prices in order to continue attracting searching consumers. This would decrease the surplus of incumbent firms and potentially increase consumer welfare. On the other hand, the competition effect implies that a nearby rival’s exit would increase the market power of incumbent firms and lead to higher prices. This would increase the surplus of incumbent firms 1

For example, Hotelling (1929) studied the location decisions of firms selling to geographically disperse consumers and how these decisions influence consumer substitution across products and geography in an attempt to explain the co-location of firms. However, d’Aspremont et al. (1979) showed that Hotelling’s principle of minimum differentiation was invalid and suggested that firms would want to locate far from each other using a variation of the same model. However, the optimal location decisions of firms are very sensitive to changes in the set up of Hotelling’s model. 2 There is a large literature on agglomeration economies that focuses on production driven reasons for co-location of manufacturing, see Rosenthal and Strange (2004) for an overview of empirical evidence from the urban economics literature. We focus on demand driven reasons for co-location because of our focus on retailing, not manufacturing.

2

and reduce consumer welfare. The agglomerative effects of retail closures are particularly salient given the recent U.S. financial crisis of 2007-2009 that saw many retail firm exits due to bankruptcy and other reasons. In this paper we have two main goals. The first goal is to quantify the agglomeration and competition effects of retail co-location and to evaluate how much of these effects are related to limited consumer information. The second goal is to evaluate the welfare effects of retail closures in light of retail agglomeration economies. To accomplish these two goals, we estimate a model of consumer search for spatially differentiated products in the new car retail industry and price competition among new car dealers. This is an ideal setting to examine issues of retail co-location for two reasons. First, retail co-location is ubiquitous in this industry. For example, in Virginia (the geographic region of the data we use in our formal analysis) about 90% of car dealers are located within one-half mile of a competitor (see Figure 1). Second, this industry has been the setting of massive retail closures over the past half century, which was exacerbated by the recent financial crisis. We estimate the model using detailed transactions level data on new car purchases, which includes the population of new car transactions, including the transaction price and distance from the consumers home to the retail transaction location, a key determinant of search costs. The model we present is a parametric version of the optimal portfolio choice problem described in Chade and Smith (2006), very similar to the specification developed in Anderson et al. (1992) (Chapter 7) and recently extended to empirical applications by De los Santos et al. (2012) and Moraga-González et al. (2015).3 In the model, we split the overall market into separate geographic areas, with each area representing a cluster of multiple car dealers. We assume consumers pay a search cost to visit a dealer cluster, and this cost is a function of the distance between the consumers’ home and the cluster. After they pay the cost, consumers are able to inspect all products within a dealer cluster at 3

Moraga-González et al. (2015) also study the new car market. However, they do not study the effects of dealer co-location. Instead, their focus is on the effects of consumer search on prices and market power, and how full information models might lead to biased predictions.

3

no additional cost. Consumers simultaneously decide the set of areas they will search, and conditional on that set, they choose to purchase the best option. As in Stahl (1982) and Wolinsky (1983), the model implies that co-location has two effects, a price competition effect and an agglomeration effect of colocation. To validate the use of the search model, we first present empirical evidence that consumer demand is influenced by clusters of co-located dealers by capturing the effects of co-location in a simpler demand framework. Figure 1: Distance to the Nearest New Car Dealer 500 450 400

Number of Dealers

350 300 250 200 150 100 50 0

0

2

4

6

8

10

Distance to the nearest Dealer (in miles)

Note: This figure plots the distribution of the distance of new car dealers’ closest rival. Data from Virginia Department of Motor Vehicles and includes all new car dealers in the state of Virginia in 2008.

To estimate the model, we use detailed car transactions data that include all new car transactions from all dealers in a single large market, the price of each transaction, and the distance between the dealer and the consumer’s home. Unlike other studies of retail agglomeration, the detailed spatial nature of our data allow us to accurately capture spatial demand substitution patterns which underline the effects of co-location. We estimate a substantially smaller dollar per mile of disutility from traveling to purchase a car than previous 4

studies for this industry, $56 per mile. However, we still find that consumer information, and therefore consumer search, is limited. For example, the model predicts that nearly all consumers search less than four geographic areas when purchasing a new car, and the median consumer searches just one geographic area. These results are in line with survey data from industry reports of new car buying habits. We quantify the importance of search frictions by simulating what equilibrium prices would be if consumer’s search cost were zero. Our simulation results predict that the average retail price would be $422 lower, which in turn suggests that dealers use consumers’ limited information to exercise market power. We conduct counterfactual exercises that simulate the closings of incumbent car dealers. Specifically, we close a single dealer and then re-calculate equilibrium prices and consumer demand. We then do this for every dealer, one at a time, for both our search model and a standard model that assumes full information. We find that for both our search model, and a full information model, dealer closure results in a decrease in consumer surplus because prices rise and consumers have fewer options. We also find that the total surplus of unclosed dealers increases after dealer closure for both models. However, consumer surplus falls by less in the search model compared to the full information model and the total surplus of unclosed dealers increases by less for the search model compared to the full information model. The main reason is that the search model implies a smaller price increase because incumbent dealers in the same geographic area have an extra incentive to keep prices low, that is to continue to attract consumers to their geographic area. Additionally, if the closed dealer was not in a particular consumer’s search/choice set in the first place, closure will have no direct effect on that consumer’s surplus. Our analysis of dealer closures is related to recent literature. Furthermore, our dealer closure counterfactuals are particularly relevant to understanding the effects of massive dealer closures sparked by the financial instability of US car manufacturers over the past decade. Both Benmelech et al. (2014) and Ozturk et al. (forthcoming) study how retail agglomeration effects retail closures, and both papers find evidence of positive agglomeration effects. Benmelech 5

et al. (2014) use data across retail industries to estimate the effect of closures due to chain level financial problems on the closure decisions of close-by incumbent retail outlets. They find that nearby retail outlets are more likely to close after rival’s closure. Ozturk et al. (forthcoming) look at the effect of Chrysler dealer closings on the prices of nearby dealers using a national sample of new car transactions in a differences-in-differences framework. They find that, although prices go up after a closure, the effect of closures on prices moderates with distance. This implies that a co-location agglomeration effect exists, but that it is dominated by a competition effect, a similar result to ours. Most of prior work on retail co-location has focused on inferring the effect of agglomeration economies through firm entry and location decisions. Some of these studies have found closing rivals would have a net negative effect, for example Seim (2006), Jia (2008), and Zhu and Singh (2009). On the other hand, Vitorino (2012) finds evidence of an agglomeration effect of co-location dominates in a shopping mall setting, and Ellickson et al. (2013) find that agglomeration effect is a function of local market size in the big-box retail industry. We distinctly depart from this literature by estimating a structural model of consumer search for spatially differentiated products. By modeling the explicit mechanism of the agglomeration benefit (i.e. how co-location affects consumer demand), we can separately quantify the effects of competition versus agglomeration on firm and consumer behavior. Furthermore, we use the estimated model to evaluate the welfare effects of retail closures, something that is not possible with the optimal entry models in the papers mentioned. In contrast, we must assume that locations are exogenous to demand shocks in the consumer utility function. Our justification, which we explain in more detail later, is that there are many regulations governing the entry and exit of dealers. We also contribute to the growing literature on consumer information, such as Sovinsky Goeree (2008), Hortaçsu and Syverson (2004), and Hong and Shum (2006) among others. Like in those studies, we find evidence that limited consumer information can bias demand results and counterfactuals in full 6

information models. There are many theoretical studies that recognize that limited consumer information and search leads to agglomeration benefits of co-location, such as Stahl (1982), Wolinsky (1983), Wolinsky (1986), Dudey (1990), Fischer and Harrington Jr (1996), among others. However, this idea has not been explored empirically using a consumer search model that captures the demand mechanisms from the theory literature. As such, our paper contributes to more recent literature on the structural estimation of consumer search by explicitly studying the agglomerative benefit of search for firms. In particular, there are numerous recent studies that also nest a simultaneous search framework in a differentiated products demand framework, for example Wildenbeest (2011), De los Santos et al. (2012), Seiler (2013), Honka (2014), and Moraga-González et al. (2015).

2

Data and Overview of the Market

In this section, we first present a detailed description on the data used in the empirical analysis. Second, we present a set of descriptive statistics documenting consumer travel distance to purchase new cars and the spatial distribution of new car dealers.

2.1

Data

We combine several data sets for our analysis. The first data source provides detailed records of all new vehicle transactions in Richmond, Virginia for four years. The second data source provides general information on characteristics and prices of all vehicles sold during this period, and the third data source provides information on all dealerships. We also use data from the Census for consumer location and demographic characteristics. The primary data are obtained from the Virginia Department of Motor Vehicles, henceforth DMV, and consist of all new vehicle transactions initially registered in Virginia from 2007 through 2010. For each transaction we know the make, model, and transaction price of the car. We also know the iden7

tification number assigned to the dealer by the DMV and the name of the dealer. Finally, the data include either the nine-digit or 5-digit zip code of the purchaser for each transaction.

variable

Table 1: Descriptive Statistics, Car-Dealer Level Mean SD Min Median

Avg. Price HP/Weight*1000 MPG Highway Passengers US Brand Luxury

24,987.25 0.55 2.82 5.20 0.26 0.28

8,490.18 0.09 0.46 0.84 0.44 0.45

11,102.84 0.30 1.50 2.00 0.00 0.00

22,674.95 0.53 2.80 5.00 0.00 0.00

Max 76,866.16 1.36 3.70 10.00 1.00 1.00

Note: New car transactions in Richmond, Virginia from our selected sample from the VA DMV described in the text.

We make a number of sample selection decisions for the raw data in order to focus on the market for new retail cars. We remove all commercial vehicles, motorcycles, trailers, and consumer pickup trucks.4 We also dismiss observations with prices near or at zero, which likely represent something else besides a typical consumer transaction, for example fleet sales, or some error in the data recording process. We also have general information on car characteristics and pricing from intellichoice.com. This includes characteristics of each trim level of each model of car, invoice prices, manufacturer suggested retail price, and other fees assessed at the time of sale. Intellichoice.com also furnished us with a list of all customer incentives provided by manufacturers during the time period, which is crucial to constructing a correct transaction price for a car, as dealers in Virginia report the transaction price less manufacturer rebates to the DMV for tax purposes. In order to focus on our research questions, we limit our study to transactions with buyers and sellers both located in the Richmond metropolitan 4

It is common in the literature to consider pickup trucks a different market. Additionally, some models of pickup trucks have dozens of trim levels that vary widely in price and characteristics, making it problematic to aggregate to the model level.

8

area.5 In total, we are left with 79,097 transactions. Richmond is a relatively isolated market. For example, among all transactions with buyer’s location in Richmond, 93.3% of the associated dealers are located in Richmond area. In other words, Richmond habitants rarely purchase cars from outside the area. Among all transactions with dealer’s location in Richmond, 82.5% of the buyers live in Richmond. We merge the transaction data with a list of all active and recently closed automobile dealers provided to us by the Virginia DMV. To merge the data, we aggregate transactions to the year-model level.6 Our data used in estimation contain 56 dealerships selling 32 different brands and a total of 2,280 unique dealer-model-year combinations. Among them, 22 dealers sell multiple brands, and 6 dealers sell vehicles produced by multiple manufacturers. We complement the data with U.S. Census demographic data on the income and population at the level of “block groups”, which, on average, contain about 1100 people. In total, the Richmond area contains 575 block groups. We present descriptive statistics of the automobile data in Table 1. The average (sales weighted) price of a car in our sample is about $25,000, 26% of the sales are US brand cars, and 28% are cars that are classified as luxury by intellichoice.com.

2.2

Consumer Travel Distance to Purchase

To calculate the distance of each transaction, we geo-code the consumer locations and the dealer addresses into decimal longitude and latitude coordinates. We observe the zip code of each consumer, and using the zip code centroid provided by the US Census, we assign zip codes to the nearest block group using the Census centroid for each block group. Figure 2 illustrates how far consumers travel to purchase a car. The mean travel distance is about 11.75 miles, the median is 9.74 miles, and the standard deviation is 8.21 miles. However, a 5

We will refer to the greater Richmond metropolitan area as Richmond. It would be ideal to use trim information, but unfortunately the transaction is not always recorded in such detail. We assign the base 4-door trim characteristics to each model. 6

9

more telling statistic that shows how important distance is in the consumer’s choice problem is the distance traveled past the nearest dealer. Figure 2 also indicates the distribution of the distances traveled past the nearest dealer. The mean of the extra distance traveled is 6.12 miles, and the median is 5.05 miles. 27% of the buyers bought from the nearest dealer, 52% traveled less than 10 miles, 18% traveled less than 20 miles and only 3% traveled more than 20 miles. Figure 2: Distribution of Buyers’ Travel Distance Purchase Distance Purchase Distance Past Closest Dealer

0.5

Frequency

0.4

0.3

0.2

0.1

0

0

5

10

15

20

25

30

35

40

Distance

Note: Distribution of consumer purchase travel distance. Data from Virginia DMV transactions data. Sample selection described in the text.

2.3

Spatial Distribution of Car Dealers

Dealers in Richmond locate in clusters in a small number of geographic locations, mainly along primary roads and commercial areas. We group the 56 dealers into 9 geographic areas. Figure A.1 in the Appendix A shows the distribution of car dealers in Richmond numerically coded by the geographic areas we assigned each dealer. The areas range in size from a few dealers to over a dozen. The most common type of area is a suburban commercial center common on the outskirts of most U.S. cities. Richmond has many relatively largely populated suburban areas, whose growth is in part driven by the fact 10

that the city and surrounding area is served by three major interstates, I95, I64, and I295. The city of Richmond is about 62 square miles and has a population of a little over 200,000 residents, but the Richmond metropolitan area is home to more than one million residents, and is the 43rd most populous metropolitan area in the United States. We present a very brief description of each dealer area in Table 2. The two largest areas are two suburban centers, Short Pump (a suburb 15 miles west of downtown) and Midlothian (a suburb 15 miles southwest of downtown). Each of these areas has a large mall and other commercial activity along a major thoroughfare surrounded by suburban and exurban residential development. Table 2: Description of Dealer Selling Areas Area

Dealers

Description

1 6 Ashland: Suburb 20 miles north of downtown Richmond on I95 2 24 Short Pump: Suburb 15 miles west of downtown 3 5 Mechanicsville : Historic town 10 miles northwest of downtown 4 22 Midlothian: Large suburb, 15 miles west/southwest of downtown 5 2 Downtown Richmond 6 8 Airport: limited development; 10 miles east of downtown Richmond 7 3 Woodlake: Suburb 20 miles SW of downtown 8 6 Chester: Suburb 15 miles south of downtown 9 5 Colonial Heights: Suburb and military base 25 miles south of downtown Note: A dealer refers to a single franchise location of a brand. For instance, a single location that sells Dodge and Jeep counts as two dealers. Milages are approximate miles using Google Maps driving directions using the geographical centers of each area.

To further illustrate that co-location is a dominant feature of this industry, we display the total sales by dealer US Census Tract in Figure A.2 in Appendix A. In this figure, we shade each Census Tract in the Richmond metro area with a shade of red corresponding to how many car sales originate from dealers in that Census Tract in 2008. Most Census Tracts are not shaded, and therefore not visible on the map. However, the Tracts that are shaded clearly illustrate the geographic clusters of dealers that are present in Richmond. There are clear “pockets” of sales with no dealers in between each shaded pocket. This 11

corresponds to the geographically separated dealer clusters in the Richmond area. Table 3: Sales and Prices by Dealer Area, 2008 US Brands Non-US Brands

Dealer Area 1 2 3 4 5 6 7 8 9

Dealers

Sales/Dealer

Avg. Price

Dealers

Sales/Dealer

Avg. Price

6 7 1 9 2 7 0 2 5

150 168 103 278 344 161 – 320 130

26,426 29,903 24,576 26,437 25,037 26,160 – 26,463 25,820

0 17 4 13 0 1 3 4 0

– 469 526 334 – 374 956 743 –

– 33,134 24,077 31,423 – 20,159 22,078 23,015 –

Note: New car transactions from Richmond, Virginia new car transactions from our selected sample described in the text.

3

Demand Model

We consider a market where differentiated cars sold by many different geographically dispersed dealers are sold to geographically dispersed consumers. We use subscript i to denote consumer, subscript j to denote car model (for example Ford Fusion), subscript f to denote dealer (ie “Bob’s Honda Sales”), and subscript t to denote year. Consumer i makes a discrete choice, either to purchase a new car model from a dealer or to consume an outside option. The indirect utility that consumer i derives from purchasing the car model j from dealer f in year t is uijf t = xjf t β − αi p˜ijf t + ξjf t + ε˜ijf t ,

(1)

where xjf t is a vector of observed product-dealer attributes, p˜ijf t is the price charged by dealer f , ξjf t captures the unobserved product-dealer-year attribute, ε˜ijf t captures an idiosyncratic match value that can be ascertained 12

only upon visiting the dealer, such as the fit and personal comfortability of the car, or a personal image in the car, or the specific way dealer salespeople sell particular cars. αi captures consumer heterogeneity in tastes for price. We assume αi = exp(α0 + α1 hi + α2 ςi ), (2) where hi is the log of household i’s yearly income and ςi follow a standard normal distribution. Note that in the auto industry, prices depend on consumer characteristics and dealer characteristics. As common in the literature on search, we assume that, even though consumers do not know the prices each specific dealer would charge them, they know the average price charged by a specific dealer. This type of information is available on a plethora of car buying websites. Also, advertisements may communicate this information, along with information about dealer specific prices, for example the willingness of each dealer to give price discounts. We assume p˜ijf t = pjf t + ϑijf t , where pjf t is the average price charged by dealer f and ϑijf t is consumer i’s price deviated from the average. Here, ϑijf t captures consumer i’s and dealer f ’s bargaining power that can be ascertained upon visiting the dealer. We assume that consumers must search to find out the exact utility they derive from each car sold by each dealer. To be more specific, we assume that before searching consumers know the product attributes xjf t and ξjf t , and the average price charged by each dealer pjf t . However, consumers do not know the exact values of their own price deviation from the average ϑijf t and their own match value ε˜ijf t . As common in the literature, we assume that that consumers know their distributions before search and costly search reveals the exact values to consumers. Let εijf t = −αi ϑijf t + ε˜ijf t

13

to capture the information that consumer i does not know before search. We assume εijf t follow an EV Type I distribution with location parameter 0 and scale parameter 1.7 Let δ jf t denote the mean utility that is common to all consumers δ jf t = xjf t β + ξjf t ,

(3)

and µijf t denote consumer heterogeneous utility from average price µijf t = −exp(α0 + α1 hi + α2 ςi )pjf t .

(4)

Then, we can write the utility specified in equation (1) as uijf t = δ jf t + µijf t + εijf t .

(5)

Consumers have an outside option, including purchasing from a dealer outside Richmond, or non-purchase, or purchase of a used car. We model the utility from the outside choice as ui0t = εi0t .

3.1

Search Mechanism

Existing theoretical literature typically model consumer search strategies in two ways. One strand of the literature assumes non-sequential, or simultaneous, search strategy, where consumers sample a fixed number of sellers and choose to purchase from the most preferred seller among those they have searched.8 The other strand of the literature assumes sequential search strategy, where after each search consumers choose to purchase from the lowest price observed so far or to make an additional search. Both search strategies have been adopted by empirical researchers. There are two studies we are aware of that compare both strategies in a retail goods setting. De los Santos 7

Although we are specific about how individual prices are related to average prices and uncertainty in the model, we do not explicitly model a bargaining protocol between consumers and dealers. This would severely complicate the model, and it is likely orthogonal to the mechanisms of dealer co-location that is our primary focus in this paper. 8 See Stigler (1961), Burdett and Judd (1983), and Janssen and Moraga-González (2004).

14

et al. (2012) who use detailed data on the browsing and purchasing behavior of a large panel of consumers to empirically test to what extent consumers are indeed using sequential and non-sequential search strategies. They found that in their setting, the non-sequential search strategy outperforms the sequential search model. Honka and Chintagunta (2014) use variation in actual prices in consumers’ observed considerations sets to conclude that simultaneous search better matches their data on the demand for auto insurance. Because we do not observe consumers’ consideration sets in our data, we are unable to let the data tell us which search strategy better represents our empirical setting. In this paper, therefore, we assume that consumers engage in costly simultaneous search in order to learn the exact match values. This would be consistent with a consumer learning about the locations and general offerings of all dealers in the market in some pre-search stage, and then planning a shopping trip that includes all of the geographic areas that make it into the consumers search set. In the model, consumers have the choice to search cars at one or more of nine possible dealer areas. Each dealer area represents a geographic area where dealers are clustered. Consumers pay a fixed cost to search each area, and once the cost is incurred they learn εijf t for every car in the area. The choice set of a particular consumer is made up of only those cars from those areas that she has searched. We normalize the search cost of the outside good to zero, so that every consumer choice set includes the outside option. Consumers simultaneously decide the set of areas they will search, and conditional on that set, they choose the best option. The model is a parametric version of the optimal portfolio choice problem discussed in Chade and Smith (2006), very similar to the specifications of De los Santos et al. (2012) and MoragaGonzález et al. (2015).9 Let Fmt be the set of dealers located in area m, and Jf t be the set of products sold by dealer f . To obtain a closed-form expression for choice prob9

However, we do not adopt the Marginal Improvement Algorithm (MIA) proposed by Chade and Smith (2006) to find the optimal choice set of each consumer, because the conditions of MIA are not satisfied in our context. Instead, we follow De los Santos et al. (2012) and Moraga-González et al. (2015) by analytically computing optimal search set probabilities.

15

abilities from the search model, we follow the literature in assuming that εijf t and εi0t follow a standard type-I extreme value distribution i.i.d. across consumers, car models, dealers and time. Hence, consumer i’s expected gain from visiting a subset of areas S is   Uit (S) = Eε max{ui0t , max uijf t } ∀j∈Jf t ,f ∈Fmt ,m∈S   X = ln 1 + exp(δjf t + µijf t ) . j∈Jf t ,f ∈Fmt ,m∈S

Let cim denote the search cost of visiting area m. We assume that the search cost is linear in the distance from her home address to the dealer location. That is, cim = γdim , where dim is the distance from consumer i’s location to a particular area m. We define the value of visiting a subset S as Vit (S) = Uit (S) −

X

cim + ωiSt ,

m∈S

where ω iSt is an individual and choice set specific term that captures the unobserved search cost shocks, such as the traffic patterns from visiting a particular set of areas. If consumer i does not search any area, we normalize P Uit (Ø) = 0 and m∈Ø cim = 0. To solve for the set of the consumer we obtain a closed-form  optimal search  expression for E maxVit (S) , by assuming that ω iSt follows a standard type S∈S

I extreme value distribution. Then, we can get the probability that consumer i visits a subset S ∗ : PiS ∗ t = P r(Vit (S ∗ ) ≥ Vit (S) f or ∀S) P exp[Uit (S * ) − m∈S * cim ] P , = P S∈S exp[Uit (S) − m∈S cim ] where S is the set of all possible search sets.

16

3.2

Purchase Decision

The probability that consumer i purchases product j from dealer f conditional on a search set S follows the familiar analytical expression: Pijf t|S =

1+

exp(δjf t + µijf t ) , ∀ mf ∈ S exp(δ + µ ) j0f 0t ij0f 0t j0∈Jf t ,f 0 ∈Fmt ,m∈S

P

= 0, ∀ mf ∈ / S, where mf is the area where dealer f is located. Then, the unconditional probability that consumer i purchases product j from dealer f is Pijf t =

X

Pijf t|S PiSt .

(6)

S∈S

Let consumer locations (in our application US Census block-groups) be indexed l = 1..L, and let Fl (·) be the area-specific distribution of consumer income hi . Additionally let nlt be the number of potential consumers in area l. The predicted demand for each vehicle model from each dealer is obtained by aggregating individual choice probabilities over all areas: ˆ qjf t (δt , pt ; θ) =

X

nlt

Pijf t (δt , pt , hi ; θ)dFl (hi ),

(7)

l

where θ = (α0 , α1 , α2 , γ) represents all “non-linear” parameters of the model.

3.3

Dealer Agglomeration and Consumer Demand

The effects of dealer co-location are straightforward in the model. Dealer areas with more dealers will, all else equal, have a greater value of Uit (S). The more products in a search set, the greater the chance that a high utility product is found. This comes through increased variation in the observed characteristics of cars as choice sets increase, along with the fact that more draws of idiosyncratic shock, ijf t , increase the maximum order statistic.10 In turn, the higher 10

This is a well known property of variants of the logit discrete choice model and has the flavor of “love of variety” in representative consumer models. See Anderson et al. (1992) for

17

the value of Uit (S) the more likely the consumer will choose to search that dealer area. Variation in chosen search sets across consumers is generated by different draws of search set specific idiosyncratic cost shocks, ωiSt , and differences in distances to different dealer areas. Holding travel distances constant, dealer areas with more products offered will be searched with greater probability. The size of search costs will ultimately determine how many dealer areas are searched by each consumer. However, as is pointed out by Chade and Smith (2006), the optimal set of dealer areas for each consumer will not necessarily follow a cut off rule of an ordering of Uit (S)’s from highest to lowest.

4

Estimation Methodology

Our estimation follows recent literature that has combined the methods of BLP with micro level data, for example Berry et al. (2004), Petrin (2002), and Sovinsky Goeree (2008), in a General Method of Simulated Moments framework. The method employs the nested fixed point structure of BLP that has an outer and inner loop, but uses both the aggregate moments suggested by BLP as well as moments derived from the individual nature of the data. In the outer loop, we search for all non-linear parameters, θ = (α0 , α1 , α2 , γ), that minimize a GMM objective function. For any guess of θ we solve for δ t (θ) from the contraction mapping suggested in BLP in the inner loop. During the inner loop we recover those fixed effects in the mean utility equation (3) using linear regression. For any candidate value of θ, we calculate the aggregate market share of each product from each dealer: sjf (δt , pt ; θ) = qjf (δt , pt )/Mt ,

(8)

P where Mt is the total number of potential consumers Mt = l nlt . Calculating sjf involves computing the choice probabilities which involves a multi-dimensional integral which cannot be computed analytically. We use details of welfare in discrete choice demand models.

18

simulation methods to approximate the integral, using the empirical distribution of individual characteristics by Census block-group for the case of consumer demographics. We simulate ten consumers from each of the 574 Census block-groups. We then weight each individual based on their block-group population. Let sjf t be the observed market share of product j sold by dealer f in year t. The solution to sjf (δt , pt ; θ) = sjf t exists and is unique (see MoragaGonzález et al. (2015)). The solution is denoted by δ t (θ). We are able to use the contraction mapping suggested by BLP to solve for δt (θ). After inverting demand using the contraction mapping, we solve for the model-dealer-year specific demand unobservables as (9)

ξjf t (θ) = δt (θ) − xjf t β.

To control for the correlation of price pjf t with the unobserved product attribute ξ jf t , we use the competing product characteristics as the instrumental variables. Hence, the first set of empirical moment conditions are given by (1)

GN (θ) =

1 X (1) ξjf t (θ)Zjf t , N1 j,f,t

(10) (1)

where N1 is the number of model-dealer-year-level observations and Zjf t is a set of instrument variables. Notice that because of the car style, location, and year fixed effects, the moment conditions are expressed over the transitory component of the unobserved quality. The second set of moment conditions matches the predicted consumer purchase distance with the observed purchase distance from our individual level choice data. Because we include individual heterogeneity, we also use simulation at this step. We integrate out the random shock and income heterogeneity in αi by taking ten draws over the corresponding distributions for each individual in our sample. In particular, let yijf t equal to one if consumer i bought the vehicle j from dealer f in year t and equal to zero otherwise. We compute

19

the following moment conditions: " # X 1 P (θ) (2) Pijf t G2N2 (θ) = − yijf t Zijf t , N2 i,j,f,t 1 − j,f Pijf t (θ)

(11)

(2)

where Zijf t includes the distance traveled by consumer i to buy the vehicle j from dealer f in year t, and Pijf t (θ) is the simulated choice probability for each consumer. The number N2 is the number of consumer-model-dealer-year-level observations in the sample.11 We stack the micro and macro moment conditions and then minimize their weighted distance from zero by choosing θ. The estimates of θ are given by θˆ = arg min G(θ)0 W G(θ), θ

(1)

(12)

(2)

where G(θ) = [GN (θ) GN2 (θ)]0 and W is a block diagonal weighting matrix.12

4.1

Identification

In this section we provide an informal discussion on the identification of the key parameters of the model, the price coefficients α and the transportation cost parameter γ. To identify the price coefficient, we need to find relevant and valid instruments to solve the simultaneity problem between price and unobserved characteristics, ξ jf t . Here, the simultaneity problem arises because dealers and consumers observe the unobserved attribute ξ jf t when making their decisions and so prices will adjust in the short-run to changes in ξ jf t . Following the literature, we use the own product characteristics and the average exogenous characteristics of competing products as instruments for price. Here, we have rich variation in these instruments because we define competing firms as those 11

In practice we construct the micro-moments using a randomly selected 20% sample of the individual transactions each time period. 12 We use a two stage procedure. In the first stage we use the 2SLS weight matrix for the first set of moments, and the identity matrix for the second set of moments. Then we use the first stage parameters to calculate the optimal weight matrix for each type of moment and re-estimate the model.

20

dealers within 5 miles. In addition, we use the number of car models with the same body style and the the number of all car models within 5 miles. The specific search mechanism that we model is not identified per se. Because we do not observe choice sets or search behavior, we cannot reject another search model, or non-search model, in favor of our model. Our analysis in section 2.4 is designed to validate our choice of search model. In turn, the search cost parameters are identified conditional on our parametric assumptions about search. The parameter on distance, γ, is identified from variation in distance and choice probabilities in the data. The identification of the search cost parameters relies on the assumption that dealer entry, exit and location are not correlated with the unobserved demand shocks, ξ jf t . Since the utility function includes dealers’ location fixed effect and time fixed effect, ξ jf t captures only transitory local demand shocks. Hence, this assumption is valid if dealers’ entry, exit or location decisions are based on the long-run local conditions that are allowed to be contained in the location fixed effects and aggregate economic shocks that are captured by the time fixed effects, but not on the realization of the transitory shock ξ jf t . This assumption is reasonable in our context for the following reasons. First, the sunk cost involving the entry, exit and location change of a dealer is large. This is partly due to regulations that limit entry and exit. Hence, there is very little entry of dealers in the industry, and when there is entry it is often a new brand entering at an existing dealer location. Also, to the extent the local demographics and population change over time, initial decisions about entry may not reflect current demographics, population, or other transitory factors. Third, forced exit of dealers by the manufacturer is very difficult in this industry because of state laws requiring payments to dealers for termination of franchise contracts. Lastly, there are other state laws that make it difficult for entry and exit, including mandated exclusive territories for brands. For a discussion of the regulatory environment see Lafontaine and Morton (2010) and Murry and Schnieder (2015). If our argument is invalid and dealers endogenously make their location choices based on the realization of the transitory shocks ξ jf t , our estimates of search cost parameter would be upward biased. 21

5

Discussion of Results

In this section we discuss the parameter estimates of the demand model. First we present results from a simplified model, and then we present the results from the search model described in Section 4. We also simulate the estimated model to learn about the extent of consumer information the model implies. Lastly, we present a model of pricing, and use the model to infer marginal costs, which will be important for counterfactual simulations.

5.1

Evidence of the Effect of Dealer Co-location

To better understand the covariation in the data, we estimate a simple version of our model where we make the following restrictions: (i) consumers have full information and zero search costs and (ii) consumer heterogeneity comes through the idiosyncratic shock only, if jt (see Berry, 1994). We present the results in Table 4. Generally the results look sensible and are in line with other studies of demand for new cars. For example, the results imply an average demand elasticity of either -3.27 or -4.43, consumers like acceleration (HP/Weight), more passenger seats, and dislike US brand cars. The coefficients on miles per gallon (MPG) and luxury are not significant.13 We include an additional variable in consumer utility specifications in Table 4 to understand how dealer co-location affects the consumers choice decision. Specifically, we add two product “characteristics,” separately in two specifications, to the consumer’s indirect utility function: the number of available cars within two miles of the product (Njf t ), and the number of cars availstyle able within two miles of the same body style (Njf t ). In a typical consumer choice problem the existence of nearby competitors should not directly affect the consumer’s utility. Nearby competitors would effect the consumer choice probability only though the denominator of the standard logit choice probability formula. However, we find these two variables have positive and significant style coefficients. Furthermore, the coefficient on Njf t is larger than the coefficient 13

The result on MPG is not an uncommon finding. For example see BLP and Petrin (2002).

22

on N , which makes sense because consumers likely care more about searching for similar cars. We take this as evidence that dealer co-location effects the consumer choice problem in this market. In other words, there is a positive effect of nearby competitors in consumer utility.14 Table 4: Simple Demand Parameter Estimates (1) (2) Variable Estimate s.e. Estimate Price ($10,000) Constant HP/Weight*100 MPG/10 (hwy) Passenger Seats U.S. Brand Luxury N N style

-1.160 -10.134 3.453 -0.185 0.233 -0.503 -0.359 0.006 –

Mean Elasticity

-3.27

(0.57) (1.68) (2.16) (0.77) (0.09) (0.12) (0.38) (0.003) –

-1.571 -9.030 4.987 -0.707 0.291 -0.590 -0.391 – 0.012

s.e. (0.69) (2.00) (2.63) (0.93) (0.11) (0.15) (0.38) – (0.005)

-4.43

Note: The sample includes 79,097 transactions and 2,280 model/dealer/year observations. We use the sum of attributes of all other products of the same types of cars within 5 miles as instruments for the price. N is the number of products offered within 2 miles and N style is the number within 2 miles of the same style. Both specificaitons include body style dummies (ie SUV, sedan etc.), dealer area dummies, and year dummies.

5.2

Search Model Estimates

The parameter estimates for the search model are presented in Table 5. The estimates for horsepower/size, MPG, seats, and US brand are similar to the 14

Endogeneity of these two variables might be a concern if the unobserved demand shock, ξjf t , is correlated with firms location decisions. However, we include dealer area dummies (where the areas are those defined in Section 2.3) in the model to capture location specific unobserved heterogeneity. Identification of these parameters comes from changes in product sets across time within a 2 mile radius of each dealer. In general, this would be entry and exit of dealers that is not correlated with ξ, and changes to the products that manufacturers offer, which is likely not a function of contemporaneous demand shocks in a single market because it is a national decision by the manufacturer.

23

simpler specifications. Recall that price coefficients enter the consumers’ utility function in a non-linear way, see equation (2). Hence, a positive estimate of α1 implies that price has a negative impact on consumers’ mean utility, and a negative estimate of α2 implies that consumers with higher income are less price sensitive. The implied consumer-model-dealer-year level own-price elasticities of demand are between -9.46 and -2.12, with a sales weighted average of -4.10. This suggests that consumers are price sensitive on average, but there is substantial heterogeneity. Overall, our estimates of price elasticities are generally consistent with previous studies of automobile demand. For example, the average own-price elasticity is equal to -4.1 in Albuquerque and Brooenberg (2012), -5.3 in Murry (2015), and -3.14 in Nurski and Verboven (2015). The distance search cost parameter is large in magnitude and precisely estimated. It implies that consumers are very sensitive to distance. Next, we explore the implications of the estimated search costs in detail. Table 5: Demand Parameter Estimates from the Search Model Variable Distance (100 miles) Price, Mean ($10,000) Price*log(inc) Price*random Horsepower/Size MPG/10 (hwy) Passenger Seats U.S. Brand Constant

Parameter γ α0 α1 α2 β1 β2 β3 β4 β0

Coefficient 16.949 4.5657 -0.3505 0.1161 4.8244 -0.7091 0.2594 -0.5783 -4.9825

Std. Error (0.5689) (0.5867) (0.0380) (0.0248) (0.2892) (0.1116) (0.0348) (0.0778) (0.4912)

Note: The sample includes 79,097 transactions and 2,280 model/dealer/year observations. We use the mean of attributes of all other products of the same types of cars within 5 miles as instruments for the price. We include body style dummies (ie SUV, sedan etc.), dealer area dummies, and year dummies. Standard errors are computed directly.

24

5.3

Search Cost and Search Frequency

To get a sense of the economic magnitude of the parameters, it is useful to consider how much a product’s price should be reduced to compensate consumers if they have to travel one more mile. Our results suggest that the average value is $56 per mile. We also report the dollar-per-mile for each brand in the last column of Table A.1 in Appendix A.15 The estimated travel cost is substantially lower than those reported in other studies that estimate consumer distance costs in the industry. For example, Moraga-González et al. (2015), the closest paper to ours, estimate a median travel cost of €107 per kilometer. This difference could come from different reasons. First, Moraga-González et al. (2015) consider the market for cars in the Netherlands. It is reasonable to think that search costs are much higher in Europe for multiple reasons, including congestion and the price of fuel. Second, the unit of observation in our data is different. We use individual purchase data whereas use transaction data aggregated to a relatively fine geography. Our parameter estimates could reflect important micro level information captured in the co-variation of distance and purchase probabilities. Consider that a consumer bought a Toyota Camry. Moraga-González et al. (2015) do not observe which dealer she bought from, so they rely on the model to tell them what dealer she likely purchased from. However, we incorporate exact purchase distance information, which may imply that consumers purchase from further away dealer than their model predicts. Our search cost is also much lower than studies of consumer demand for cars that include distance in the indirect utility function, like Nurski and Verboven (2015), Murry (2015), and Albuquerque and Brooenberg (2012). For example Nurski and Verboven (2015) obtain an average travel cost of €112 per kilometer. It is possible search frictions may help rationalize the purchase patterns with lower costs of distance. Next, we consider what the estimates imply about how much consumers 15

To calculate the cost per mile we first calculate the own distance elasticities, which tells us how choice probabilities change if a dealer were to move one mile away from every consumer. The average distance elasticity is -0.92%. Next, we use the own price elasticity to calculate the price decrease that would increase choice probabilities by the same amount. We average this across all products to arrive at our average dollar per mile of $56.

25

search by using the results of the search model to simulate purchase behavior. In Figure 3, we plot the histogram of the number of searches per consumer from this simulation. In general, search is limited. We predict that conditional on positive search, 49% of consumers search only one area and less than 1% search four areas. This results are generally consist with industry reports and previous studies. For example, in a survey by DME Automotive, an industry consulting group, 47% of all new car buyers report visiting a single dealer before purchasing a car. Moraga-González et al. (2015) report that 47% of survey respondents in their data searched one dealer. Although our model does not have empirical content regarding specifically how many dealerships are searched because search happens at the area level, we can at least say that searching two or more areas implies searching at least two dealers, and so our predictions compare favorably to these other sources.16

Figure 3: Predicted Density of Search Intensity 0.6

0.5

Density

0.4

0.3

0.2

0.1

0

1

2

3

4

5

6

7

8

9

Number of Searched Areas

Note: Model predicted distribution of the number of searches conditional on positive search.

16

See http://www.dmeautomotive.com/announcements/1-in-6-car-buyers-skips-testdrive-nearly-half-visit-just-one-or-no-dealership-prior-to-purchase. In principle, we could use this information to add a moment inequality to our estimation objective function. However, we prefer to use survey data such as this to think about model validation.

26

6

Retail Markups

Although we do not use a model of car retail supply to estimate the demand parameters, we do use a supply model in counterfactuals, so it is important to report the implications of the supply model given the estimated demand parameters. Below we describe our supply-side model.

6.1

A Supply Model

The profit of dealer f is defined as π f (pt ) =

X

(pjf t − mcjf t )qjf (pt ) − F Cf t ,

j∈Jf t

where mcjf t is the constant marginal cost of product j sold by dealer f in year t, and F Cf t represents fixed cost. Dealers simultaneously maximize profit by setting price, taking into account prices and attributes of competing dealers.17 The first order condition for a particular dealer that defines a Nash Equilibrium in prices is qjf (pt ) +

X

(pjf t − mcjf t )

j∈Jf t

∂qjf (pt ) = 0. ∂pjf t

(13)

Let 4 denote the price derivative matrix with the row j column k element 4jk

 ´ P ∂qj (pt )  l nlt (α0 + α1 hi )Pijt (1 − Pijt )dFl (hi ), if j = k = = − P n ´ (α + α h )P P dF (h ), ∂pkt if j 6= k 0 1 i ijt ikt l i l lt

We define an ownership matrix Ω∗ , with Ω(j, k)∗ = 1 if product j and k are sold by the same dealer and zero otherwise. Let Ω = Ω∗ × 4(p) . Then, equation (13) can be written in matrix notation as the following markup equation p − mc = Ω−1 q(p). 17

(14)

Price here is the same concept as in the demand model: the average price for each model at each dealer.

27

6.2

Recovering Price-Cost Markups

Given the estimated demand parameters, we compute the price cost markups from equation (14). The weighted average price-cost markup, defined as the difference between price and marginal cost, is $6,015 and the median is $6,051. , in Figure 4. The weighted We display the distribution of dealer margins, p−mc p average markup is 25.88% and the median is 23.75%. These are in line with other studies of the automobile industry, for example, 24% in Berry et al. (1995), 17% in Petrin (2002), $6,220 in Albuquerque and Brooenberg (2012), $5,238 in Murry (2015), and 43% in Nurski and Verboven (2015) and MoragaGonzález et al. (2015). The third column of Table A.1 in Appendix A shows the average markup in dollars for all brands. Among all brands, BMW’s markup is the highest and Suzuki’s markup is the lowest.

Figure 4: Dealer Margins (%) 0.1

0.08

Density

0.06

0.04

0.02

0

0

10

20

30

40

50

Dealer Margin (%)

Note: Dealer margins predicted by the model. Dealer margins are defined as

6.3

p−mc p .

Contribution of Search Frictions

The estimated model provides a way for us to understand the sources of dealer margins. The standard source of market power in discrete choice models of de28

mand is attributed to product differentiation. Additionally, our model implies that market power is also attributed to search frictions. Search frictions create incomplete choice sets for consumers as is seen in Figure 3, where the implied search intensity of consumers is quite low – the median searching consumer searches only one geographic area. Search frictions can have two opposite effects on price. On one hand, incomplete choice sets can lead to market power for firms because consumers are more captive to firms given they have fewer choice options. Firms have the ability to raise prices because, conditional on a choice set, consumers do not have as many options available as they would if search costs were zero. This is the “competition effect.” Due to this effect, higher search cost would lead to higher price. On the other hand, because consumers decide their search set of geographic areas by making a tradeoff between the expected utility of a search set and the search costs associated with that set, higher search cost would encourage dealers to lower prices to attract more consumers to visit their area. This is the “agglomeration effect.” Due to this effect, higher search cost would lead to lower price. Whether the competition effect dominates the agglomeration effect is an empirical question, depending on the model primitives.18 To quantify the impact of search cost on market power, we consider a scenario in which each consumer has full information and considers all products, the standard assumption in the literature. In Table 6 we present comparisons of the price predicted by the full information model to the prices predicted by the search model.19 The weighted average observed price (weighted by market share) is $24,940 and the weighted average simulated price from the full information model (weighted by market share) is $24,518. Our result implies that search frictions contribute $422 to the price of a car on average.20 18

Sovinsky Goeree (2008) examines a slightly different scenario where limited choice sets affect the markup of firms. In her setting, firms advertise the existence of products to consumers, which creates limited choice sets. She finds that her model implies markups about three times higher than a full information model. She does not have the second effect that if prices rise too high then consumers will not include a product in their choice set. 19 In estimation we match predicted shares to the data exactly, so prices in the data are identical to model predictions for the search model. 20 Moraga-González et al. (2015) perform a similar counterfactual. They find average

29

Table 6: Price Impacts of Search Frictions Variables 2007 2008 2009 2010 all years average price search $28,035 $27,897 $28,449 $28,521 $28,216 full information $27,610 $27,462 $28,120 $28,066 $27,803 difference $425 $434 $329 $455 $413 sales weighted average price search $26,008 $23,966 $24,890 $24,798 $24,940 full information $25,578 $23,538 $24,563 $24,353 $24,518 difference $430 $428 $327 $445 $422 Note: price difference = price predicted by the search model minus price predicted by full information model. Averages are weighted by market share.

Table A.1 in Appendix A compares the weighted average markups for each brand in the case of our estimated search cost to those in the case of zero search cost. This allows us to examine the additional markup firms earn as a result of search frictions. The weighted average markup predicted by the search model is $6,015 and the weighted average markup predicted by the full information model is $5,593. Our results imply that about 8% of markups are attributed to search frictions in this market. The results suggest that the “competition effect” dominates the “agglomeration effect” so that the prices and hence markup predicted by the search model are in general higher than those predicted by the full information model. Currently the idiosyncratic shocks in our model generate product differentiation and thus market power. Although we are unable to estimate the variance of these shocks, if they had 2 less variance than what we assume, π6 , then markups would be even more attributable to search frictions, so we consider our finding conservative.

prices between their search model and a full information model very similar to ours. The key difference in the two models is that in their model, consumers search at the dealer level as opposed to a geographic area. They also predict changes to manufacturer surplus after dealer re-organization and find that manufacturers would prefer to consolidate brands under one dealership.

30

7

Retail Closures

Using our structural model we simulate dealer closures in order to tease out the competition and agglomeration effect of dealer co-location, and to understand what biases a full information model might have when considering such a counterfactual. A dealership closure generates two effects to the remaining dealers located in the same area. First, closing a dealer reduces the total attraction of the whole area, and thus forces the dealers in this area to reduce their prices. This is the “agglomeration effect”. Second, closing a dealer directly reduces the price competition among dealers in the same area, and hence pushes the price higher. This is the “competition effect”.21

7.1

Background

The agglomerative effects of retail closures are particularly salient given the recent U.S. financial crisis of 2007-2009 that saw many retail firm exits due to bankruptcy and other financial issues. Benmelech et al. (2014) document massive retail exits due to financial reasons, such as bankruptcy, during the financial crisis.22 They find evidence of agglomeration effects of closures of bankrupt firms’ stores on non-bankrupt incumbent stores using data on the location and closures of multiple retail chains across the US. New car dealers also saw a large swath of retail closures during and immediately following the financial crisis. For example, Chrysler, General Motors, and Ford all closed dealers for financial reasons in 2009 and 2010. However, as described in Lafontaine and Morton (2010), this industry has experienced large numbers of 21

In complementary work, Ozturk et al. (forthcoming) examine the agglomeration versus competition effects of dealer closures. They use a national sample of car sales to infer a treatment effect of Chrysler dealership closures on rivals’ prices. They find that an average, rivals’ prices increase after a Chrysler closure, but prices of nearby dealers increase much less than distant dealer. They interpret their results as providing evidence that the competition effect dominates the agglomeration effect overall but the agglomeration effect is present because nearby dealers experience a lower price increase than far away dealer after the closures. 22 For example, they document the complete liquidation of multiple large retailers, including Circuit City, Linens ’N Things, and The Sharper Image. Other large retail chains that experienced massive closings due to financial trouble include Kmart and Sears.

31

retail closures for many decades, from a high of around 60,000 outlets in the US in the 1940’s, to less than 20,000 by 2010. Most recently a number of entire brands have dissolved, such as Oldsmobile, Pontiac, Hummer, and Saab, among others, making the issue of the agglomerative effects of retail closures especially important in new car retailing. Recent dealer closures stemmed from two primary causes. First, American manufacturers discontinued a number of brands in the mid to late 2000s, starting with Oldsmobile in 2004, and continuing with Saturn and Pontiac in 2009, Mercury in 2010, and Saab in 2011. 23 These brands had seen steady declines in sales, and were reported as being unpopular and out of touch with consumer needs in media and industry reports.24 For the case of GM owned Pontiac and Saab, there was also pressure from the U.S. government, who provided a large loan to GM in 2009 under the Troubled Asset Relief Program (TARP), to make the company “leaner” and more focused on core products that had a history of satisfactory sales and performance. The second cause of the dealer closures had to do with the financial crisis more directly. GM and Chrysler received TARP U.S. government loans in 2009, and because of their subsequent reorganization were allowed to terminate dealers without answering to state auto franchise laws. Both companies, along with Ford Motors, had a clear policy to create smaller dealer networks, but were generally unable to do so because state regulations prohibit dealer franchise contract termination by manufacturers in the automobile industry. 25 Dealers lobbied against dealer closures, citing existing state regulations that prohibit closures. Many of the proposed closures (from both of the reasons stated above) went into legal arbitration. For example, when GM closed the Oldsmobile brand, they reportedly paid over $1 billion to their dealers. In this section we examine the effects of dealer closures, and offer an explanation of why even unclosed dealers might prefer other dealers not to close.26 Our results also suggest that the 23

Saab, the major Swedish produced car brand, was owned by General Motors until 2011. After 2011, the company re-organized, and started producing cars again in 2014. 24 http://tinyurl.com/cotzn9 25 For example of popular press coverage of dealer closures, see http://tinyurl.com/p7zvgys. 26 For an alternative explanation having to do with promotional activities by the manu-

32

loss to consumer welfare from a dealer closure, and the gain of rival dealers, is exaggerated by the standard full information model of demand.

7.2

Competition Effect and Agglomeration Effect

To separately examine the agglomeration effect and the competition effect we calculate the equilibrium outcomes in the following two scenarios. In the first scenario, after the dealer closure we allow the consumers to adjust their probability of searching (PiS ), but we hold constant the probability of purchase conditional on each search set (Pijf t|S ). This quantifies the agglomeration effect: firms will adjust their equilibrium prices only for the reason to attract new searchers to their dealer area, not to price compete against a rival in the same geographic area. In the second scenario, we hold constant the probability of searching (PiS ) but adjust the probability of purchasing condition on each search set (Pijf t|S ). This quantifies the competition effect: firms will adjust their price to compete with local rivals, not to attract more consumers to search their geographic area. In both scenarios, we measure the separate effects on the dealers by comparing the difference in prices, sales and markups. Lastly, we allow the consumer to adjust both her search decisions and choice probabilities conditional on search to quantify the total effect. To measure the welfare effects of dealer closure, we follow the literature (see Petrin (2002) and Fan (2013)) and define consumer welfare change as the compensating variation. The compensating variation for consumer i is given by V 1 − Vilt0 , CVilt = ilt αi where αi < 0 is the negative of the consumer i’s marginal value of income, and Vilt0 − αi and Vilt1 − αi are the expected maximum utility for consumer i before and after a dealer closure. Specifically, " Vilt0 = ln 1 +

!# X

0 exp UiSt −

S

X m∈S

facturer, see Murry (2015).

33

cim

i h P 0 where UiSt = ln 1 + j∈Jf t ,f ∈Fmt ,m∈S exp(u0ijlt − ijlt ) . The post-closure util0 1 with UiSt and u0ijlt with u1ijlt . ity Vilt1 is analogously defined by replacing UiSt Given the compensating variation for a specific consumer, the change in the average consumer surplus in zip code l in year t is given by ∆CSlt = Eςi (CVilt ). The total consumer welfare change is the sum of the welfare changes in all zip P codes, 4CS = lt nlt ∆CSlt , where nlt is the number of households in zip code l in year t. The change in average per-consumer consumer surplus is 4CS 4CS = P . nlt l

7.3

A Case Study

First, we simulate new equilibria as described above for the closing of a small dealer, henceforth Small. Small terminated operation in 2009, and previous to closure was one of the worst performing dealers in terms of sales in our sample. Small sold 94 units of cars in 2007 and 69 units of cars in 2008. We consider a scenario in which this dealer was closed in 2007. Table 7 presents the agglomeration effect, competition effect and the total effect of closing this particular dealer. In the first two rows, we present the results when we only allow adjustment of the probability of searching (i.e., PiS ). This is the agglomeration effect of closing a dealer. We can see that other dealers in the same area would reduce their price by $3 and their total sales in 2007 would suffer a loss of around 7 units. Dealers located in other areas would increase their price by $1 and their total sales would increase by 4 units. In the third and fourth rows, we present results when we only allow adjustment of the probability of buying each product given a search set (i.e., Pijf t|S ). This is the competition effect of closing a dealer. We can see that other dealers’ price would increase by $4 on average and total sales in 2007 would increase by around 7 units. Dealers located in other areas would increase their price by $1 and total sales would increase by 10 units. In the last two rows, we present results when we allow adjustment of both PiS and Pijf t|S . This is the total effect of closing a dealer. We can see that 34

other dealers’ average price in the same area stay almost the same and total sales in 2007 increase by very slightly. This is because the agglomeration effect and competition effect offset with each other. Dealers located in other areas would increase their price by $2 on average and total sales would increase by around 13 units. Table 7: Effects of Dealers Closing Dealer “Small” before Agglomeration Effect Same area 1,831 Other areas 22,798 Competition Effect Same area 1,831 Other areas 22,798 Total Effect Same area 1,831 Other areas 22,798

Sales after

Average Price before after change

change

1,824 22,802

-7 +4

$26,153 $28,197

$26,150 $28,198

-$3 +$1

1,838 22,808

+7 +10

$26,153 $28,197

$26,157 $28,198

+$4 +$1

1,833 22,811

+2 +13

$26,153 $28,197

$26,153 $28,198

$0 +$1

Note: Dealers in the same area are referred to those dealers located in the same area as dealer "Small". Dealers in other areas are referred to those dealers located in different area as dealer "Small". Sales refers to the total sales across all dealers in the relevant geography for 2007.

The welfare effects of closing Small is reported in Table 8. We compare the effects implied by the search model to a model with full information. The first row shows that the welfare change for the search model. Overall, the consumer surplus declines by 0.027% of the pre-closure consumer surplus and the total surplus of all dealers decreases by 0.294% of the pre-closure dealer surplus. Among these dealer surplus change, the total surplus of unclosed dealers in the same area increases by 0.104% whereas the total surplus of all unclosed dealers in other areas increases by 0.067%. The second row of Table 8 shows that the consumer surplus change for full information model is -0.281%, and the total surplus of unclosed dealers in the same area increases by 0.211% 35

whereas the total surplus of all unclosed dealers in other areas increases by 0.213%. The full information model over-predicts the consumer welfare loss. The main reason is that in the full information model, a dealer closure lessens the total choices available to all consumers. But in the search model, only a fraction of consumers had the closed dealer in their choice set in the first place, so the closure does not impact the surplus of as many consumers. Additionally, the price changes are less severe in the search model because higher prices have the added negative effect of decreasing search probabilities. The full information model also over-predicts the surplus gain for unclosed dealers after a dealer closure. The reason is that agglomeration effect a lower rise in prices to offset the decreased attraction of the area due to the closure of a dealer in the area. Ignoring this effect would overestimate the consumer welfare loss and overestimate the benefits that other dealers can get from a closure of a competing dealer.

Table 8: Welfare Effects of Closing Dealer “Small”

Search Model: change (m $) percentage change (%) Full-Information Model: change (m $) percentage change (%)

Consumer Surplus

all dealers

Dealer Surplus same area other areas

-0.578 -0.027

-0.435 -0.294

0.011 0.104

0.092 0.067

-2.548 -0.281

-1.088 -0.190

0.095 0.211

1.120 0.213

Note: "all areas" refers to the total surplus change of all dealers in the sample including the closed dealer, "same area" refers to the total surplus change of all other dealers located in the same geographic area as the closed dealer, and "other areas" refers to the total surplus change of all dealers located in all geographic areas other than the area of the closed dealer.

36

7.4

Average Effect of Dealer Closures

The previous section shows how the framework in this paper can be used to study the effects of a single dealer closure in one specific market. It gives insight into the particular mechanisms at play in our model. However, it was just a case study, and this dealer may not be representative of all dealers in the market we consider. Next, we investigate the general pattern of how the welfare effect of dealer closure varies with dealer characteristics. To do this, we compute the welfare effects when we close each dealer, one at a time. Table 9: Welfare Effects of Dealer Closure Mean

SD

Min

Median

Max

-2.430 -0.115 -1.832 -1.567 0.108 0.486 0.305 0.299

2.300 0.108 1.727 1.365 0.144 0.437 0.311 0.274

-0.076 -0.004 -0.065 -0.094 2.43 ∗ 10−4 0.016 0.005 0.010

-1.683 -0.079 -1.247 -1.033 0.052 0.344 0.211 0.201

-12.298 -0.576 -9.072 -6.265 0.905 2.451 1.647 1.465

Full Information Model: 4CSf t (m$) -10.295 (%) -1.393 4DSf t (m$) -4.918 (%) -1.033 4DSfsame (m$) 0.614 t (%) 0.887 4DSfother (m$) 3.784 t (%) 0.882

9.892 1.277 4.786 1.014 0.708 0.822 3.855 0.831

-0.364 -0.718 -0.232 -0.052 0.007 0.035 0.083 0.034

-6.852 -0.984 -3.423 -0.691 0.386 0.608 2.473 0.597

-46.553 -6.296 -22.235 -5.376 3.820 3.786 19.001 3.875

13 0.468

307 9.980

2131 81.105

Search Model: 4CSf t (m$) (%) 4DSf t (m$) (%) 4DSfsame (m$) t (%) other 4DSf t (m$) (%)

Characteristics of Closed Dealer: Qf t 424.255 394.271 AreaSharef t (%) 16.981 16.697

Note: Descriptive statistics from the results of counterfactual exercise of closing every dealer, one at a time. Variable definitions in the text.

37

Table 9 provides summary statistics on the distribution of welfare changes for our search model and the full information model. Let 4CSf t denote the denote the percentpercentage change in average consumer surplus, 4DSfsame t age change in total surplus of all unclosed dealers located in the same area as the closed dealer, 4DSfother denote the percentage change in total surplus of t all unclosed dealers located in the other areas, 4DSf t denote the percentage change in total surplus of all dealers, Qf t denote the sales of the closed dealer, and AreaSharef t denote the market share of the closed dealer in its geographic area. The search model predicts that consumer surplus declines by 0.115% on average, across dealer closures, and the total surplus of all unclosed dealers in the same area increases by around 0.486% on average. The full information model predicts that the mean consumer welfare loss is 1.393% and the total surplus gain of all unclosed dealers in the same area is 0.887% on average. Therefore, ignoring the search frictions and agglomeration effect would overestimate the consumer welfare loss and overestimate the benefits that other dealers can get from a closure of a competing dealer. Looking at the percentiles of the welfare change distributions, the search model predicts very skewed distributions. This corresponds to the skewed distribution of the size of the closed dealer and the skewed distribution of the market share of the closed dealer in its area. As seen in Table 9, dealer surplus for incumbent dealers increases after a single dealer closure (rows three and four in each panel). However, notable is that the full information model overstates the gain in dealer surplus after a closure for those dealers in the same geographic area as the closed dealer. This is because incumbent dealers do not raise prices by too much after a rival closure in order to keep their area attractive to search, whereas in the no-search model this mechanism does not exist. Change in total producer surplus, ∆DS, is strictly negative under the full information model, but this is not the case for the search model. This is due to the different effects depending on the size of the closed dealer and the total sales and search intensity of the geographic area of the closed dealer. To understand the relationship between the welfare effects and the char38

acteristics of the closed dealer, we run regressions of welfare effects on the characteristics of the closed dealer, in particular, the size of the closed dealer and the market share of the closed dealer in its area. Notice that this regression captures a correlation pattern rather than a causal effect. However, it is a useful way to summarize the results from the counterfactual simulations and to broadly confirm our intuition about the different mechanisms in the model. The results of the regressions are reported in Table 10. The welfare effect of dealer closure depends on the size of the closed dealer. If a large dealer is closed, after controlling for its share in its area, this would increase the market power of its neighboring competitors. As a result, other unclosed dealers in the same area would increase their prices. Higher prices in that area would also lead to higher prices of other areas due to the competition among areas. This explains why both the total surplus of unclosed dealers in the same area (4DSfsame ) and the total surplus of unclosed dealers in t other other areas (4DSf t ) increase in the size of the closed dealer. Due to the higher prices, consumers are worse off; in other words, consumer welfare change (4CSf t ) decreases in the size of the closed dealer. Another factor is the market share of the closed dealer in its area. This measures how important the dealer is to its area. If a dealer with a large share in its area is closed, after controlling for its size, this would significantly decrease the attraction of that area. As a result, other unclosed dealers in the same area would lose a substantial amount of consumer visits and thus sales. To offset the declined attraction of the whole area, they have to reduce their prices. Lower price in that area would also lead to lower prices of other areas due to the competition among areas. This explain why both the total surplus of unclosed dealers in the same area and the total surplus of unclosed dealers in other areas decrease in the area share of the closed dealer. Due to the lower prices, consumers are better off ; in other words, 4CSf t increases in the area share of the closed dealer.

39

Table 10: Regression: Welfare Effects and Area Characteristics Qf t AreaSharef t Constant

4CSf t 4DSfsame t -0.098 0.378 (0.005) (0.021) 0.016 0.075 (0.026) (0.111) 0.442 -1.665 (0.029) (0.122)

4DSfother t 0.257 (0.013) -0.106 (0.067) -1.144 (0.074)

Note: Estimates from regressions of the change in welfare from dealer closures versus area characteristics. Unit of observation is a single closure dealer closure in a given time period. ∆CSf t is the change in consumer surplus from the closure of dealer f in time t; ∆DSfsame t is the change in total surplus of the remaing dealers in the same area as the closed dealer; ∆DSfother t is the total change in surplus for dealers in other areas. Standard errors in parentheses.

Next, we examine the sources that drive the fact that the full information models overestimates the consumer welfare loss and the surplus gain of unclosed dealers due to a dealer closing. We define the bias as the difference between the percentage change predicted by the full information model and the percentage change predicted by the search model. We regress these bias terms on dealer and market characteristics. The two characteristics we use are the total sales of the dealer, and the share of dealer sales in its area before closure. We use both of these variables because we are interested in teasing out the contribution of dealer size conditional on the size of the dealer cluster, and also the relative importance in a cluster conditional on gross dealer size. The results of the regressions are reported in Table 11. Both factors are important contributors to the bias between the full information model and the search model because these two factors are both positively related to how large the agglomeration effect is implied by the model in a particular dealer cluster. The “larger” a dealer, both in raw terms and relative terms to nearby dealers, the more the full information model will overstate the effects of dealer closures because it does not capture the countervailing agglomeration effect.

40

Table 11: Regression: Welfare Effects Bias and Area Characteristics Qf t AreaSharef t Constant

4CS Biasf t 0.921 (0.058) 1.848 (0.306) -4.634 (0.336)

4DS Biassame ft 0.266 (0.027) 0.940 (0.140) -1.341 (0.154

4DS Biasother ft 0.378 (0.035) 1.216 0.184) -1.828 (0.203)

Note: Estimates from regressions of the bias in the change in welfare from dealer closures versus area characteristics. The bias is defined as the difference in the change between the search model and the full information model. The unit of observation is a single closure dealer closure in a given time period. ∆CSBiasf t is the bias in the change in consumer surplus from the closure of dealer f in time t; ∆DSBiassame is the bias in the change in ft total surplus of the remaing dealers in the same area as the closed dealer; ∆DSBiasother t f is the bias in the total change in surplus for dealers in other areas. Standard errors in parentheses.

8

Conclusion

In this paper, we estimate a structural model of consumer search for spatially differentiated products in the new car market. Our approach contributes to the literature on consumer demand with limited information and the literature on retail agglomeration by formally modeling consumer search for spatially differentiated products where the co-location of retail stores effects the consumers’ search and purchase decisions. We estimate a substantially lower dollar per mile of traveling to purchase than other studies, both search and not search, for the automobile industry. Compared to the standard full-information model, our model implies greater market power for car dealers. After a dealer closure, our model predicts smaller consumer welfare loss and smaller dealer surplus gain (for incumbent dealers) after a single dealer closure than the a full information model. Our results suggest that both the competition and agglomeration effects matter after a dealer closure. Our results are consistent with Ozturk et al. (forthcoming), but are at odds with the findings of Benmelech et al. (2014), who find a negative overall effect on nearby rivals after closure. However, in the case of 41

Benmelech et al. (2014), the competition effect would be expected to be small because they examine the effect across unrelated products. We also validate our modeling assumptions by showing, in a simpler framework, that consumer purchasing decisions are partly a function of dealer agglomeration. We do this by estimating a simple model of demand where we include the number of nearby available products as a “characteristic” in consumer utility. Given the amount of co-location in different retail industries, this is an important finding on its own because it provides evidence of demand based reasons for retail agglomeration, as opposed to cost side reasons. To be sure, our analysis relies on particular assumptions, and although we are confident that our model captures the major features of this industry, some caveats are worth mentioning. First, although we feel like the evidence we present strongly suggests dealer agglomeration is a consumer consideration during the car buying process, the search process may be more complicated than our model presents. In particular, the recent proliferation of car-buying websites aimed at alleviating consumer information has likely started to change the way consumers search for cars. However, cars are likely an experience good, so websites could never fully inform a consumer completely about the utility from purchase like personal interaction can. Second, consumers may search in a more complicated way, nesting geographical concerns with the search for a dealer (as in Moraga-González et al. (2015)) and the search for a car type. Thirdly, although we present a demand driven reason for dealers to colocate, there are likely cost driven reasons, for example land prices, zoning, and management convenience for multi-dealership dealer conglomerates, among others. Our analysis can not be used to balance all the tradeoffs associated with the optimal location decision, only to quantify the demand side incentives.

References Albuquerque, P. and B. Brooenberg, “Measuring the Impact of Negative Demand Shocks on Car Dealer Networks,” Marketing Science 31 (2012), 4–23. 42

Anderson, S. P., A. D. Palma and J. F. Thisse, Discrete choice theory of product differentiation (MIT press, 1992). Benmelech, E., N. Bergman, A. Milanez and V. Mukharlyamov, “The Agglomeration of Bankruptcy,” Technical Report, National Bureau of Economic Research, 2014. Berry, S., J. Levinsohn and A. Pakes, “Automobile Prices in Market Equilibrium,” Econometrica 63 (July 1995), 841–890. ———, “Differentiated Products Demand Systems from a Combination of Micro and Macro Data: The New Car Market,” Journal of Political Economy 112 (February 2004), 68–105. Berry, S. T., “Estimating discrete-choice models of product differentiation,” The RAND Journal of Economics (1994), 242–262. Burdett, K. and K. L. Judd, “Equilibrium price dispersion,” Econometrica: Journal of the Econometric Society (1983), 955–969. Chade, H. and L. Smith, “Simultaneous search,” Econometrica 74 (2006), 1293–1307. d’Aspremont, C., J. J. Gabszewicz and J.-F. Thisse, “On Hotelling’s" Stability in competition",” Econometrica: Journal of the Econometric Society (1979), 1145–1150. De los Santos, B., A. Hortacsu and M. R. Wildenbeest, “Testing models of consumer search using data on web browsing and purchasing behavior,” The American Economic Review 102 (2012), 2955–2980. Dudey, M., “Competition by choice: the effect of consumer search on firm location decisions,” American Economic Review 80 (1990), 1092–1104. Ellickson, P. B., S. Houghton and C. Timmins, “Estimating network economies in retail chains: a revealed preference approach,” The RAND Journal of Economics 44 (2013), 169–193. Fan, Y., “Ownership Consolidation and Product Characteristics: A Study of the US Daily Newspaper Market,” American Economic Review 103 (Auguest 2013), 1598–1628. Fischer, J. H. and J. E. Harrington Jr, “Product variety and firm agglomeration,” The RAND Journal of Economics (1996), 281–309. 43

Hong, H. and M. Shum, “Using price distributions to estimate search costs,” The RAND Journal of Economics 37 (2006), 257–275. Honka, E., “Quantifying search and switching costs in the US auto insurance industry,” The Rand Journal of Economics 45 (2014), 847–884. Honka, E. and P. K. Chintagunta, “Simultaneous or sequential? search strategies in the us auto insurance industry,” Search Strategies in the US Auto Insurance Industry (April 29, 2014) (2014). Hortaçsu, A. and C. Syverson, “Product Differentiation, Search Costs, and Competition in the Mutual Fund Industry: A Case Study of S&P 500 Index Funds,” The Quarterly Journal of Economics (2004), 403–456. Hotelling, H., “Stability in Competition,” The Economic Journal 39 (1929), 41–57. Janssen, M. and J. L. Moraga-González, “Strategic Pricing, Consumer Search and the Number of Firms,” Review of Economic Studies 71 (2004), 1089–1118. Jia, P., “What Happens When Wal-Mart Comes to Town: An Empirical Analysis of the Discount Retailing Industry,” Econometrica 76 (2008), 1263– 1316. Lafontaine, F. and F. S. Morton, “Markets: State Franchise Laws, Dealer Terminations, and the Auto Crisis,” Journal of Economic Perspectives 24 (Summer 2010), 233–50. Moraga-González, J. L., Z. Sándor and M. R. Wildenbeest, “Consumer search and prices in the automobile market,” unpublished manuscript (2015). Murry, C., “Advertising in Vertical Relationships: An Equilibrium Model of the Automobile Industry,” unpublished manuscript, Penn State University (2015). Murry, C. and H. Schnieder, “The Economics of Retail Markets for New and Used Cars,” unpublished manuscript, Cornell University (2015). Nurski, L. and F. Verboven, “Exclusive Dealing as a Barrier to Entry? Evidence from Automobiles,” Review of Economic Studies forthcoming (2015).

44

Ozturk, O. C., S. Venkataraman and P. K. Chintagunta, “Price Reactions to Rivals’ Local Channel Exits,” Marketing Science (forthcoming). Petrin, A., “Quantifying the Benefits of New Products: The Case of the Minivan,” Journal of Political Economy 110 (2002), 705–729. Rosenthal, S. S. and W. C. Strange, “Evidence on the nature and sources of agglomeration economies,” Handbook of regional and urban economics 4 (2004), 2119–2171. Seiler, S., “The impact of search costs on consumer behavior: A dynamic approach,” Quantitative Marketing and Economics 11 (2013), 155–203. Seim, K., “An Empirical Model of Firm Entry with Endogenous Product-Type Choices,” The RAND Journal of Economics 37 (Autumn 2006), 619–640. Sovinsky Goeree, M., “Limited Information and Advertising in the US Personal Computer Industry,” Econometrica 76 (September 2008), 1017– 1074. Stahl, K., “Location and spatial pricing theory with nonconvex transportation cost schedules,” The Bell Journal of Economics (1982), 575–582. Stigler, G. J., “The Economics of Information,” Journal of Political Economy 69 (1961), 213–225. Vitorino, M. A., “Empirical entry games with complementarities: An application to the shopping center industry,” Journal of Marketing Research 49 (2012), 175–191. Wildenbeest, M. R., “An Empirical Model of Search with Vertically Differentiated Products,” RAND Journal of Economics 42 (2011), 729–757. Wolinsky, A., “Retail trade concentration due to consumers’ imperfect information,” The Bell Journal of Economics (1983), 275–282. ———, “True Monopolistic Competition as a Result of Imperfect Competition,” Quarterly Journal of Economics 101 (1986), 493–511. Zhu, T. and V. Singh, “Spatial competition with endogenous location choices: An application to discount retailing,” Quantitative Marketing and Economics 7 (2009), 1–35.

45

Appendix A: Additional Tables and Figures Table A.1: Results by Brand

average price ($) average markup ($) price elasticity dollar per observed full info diff search full info diff (%) search full info mile ($) Acura 34,493 33,927 566 6,595 6,029 9.38 -5.23 -5.63 60.23 Audi 39,851 39,280 571 6,879 6,309 9.04 -5.76 -6.19 62.78 BMW 45,942 45,361 581 7,195 6,613 8.79 -6.40 -6.88 65.18 Buick 30,125 29,637 488 6,301 5,813 8.39 -4.77 -5.10 57.94 Cadillac 40,994 40,404 590 6,919 6,329 9.33 -5.93 -6.38 63.31 Chevrolet 23,395 23,030 365 5,829 5,464 6.68 -3.93 -4.14 53.82 Chrysler 27,298 26,865 433 6,129 5,696 7.60 -4.45 -4.73 56.03 Dodge 23,420 23,013 407 5,872 5,465 7.46 -3.98 -4.21 54.35 Ford 23,467 23,100 367 5,871 5,504 6.67 -3.98 -4.19 54.10 GMC 38,190 37,708 482 5,582 6,201 7.76 -5.70 -6.07 60.93 Honda 22,445 22,071 374 5,884 5,511 6.78 -3.82 -4.05 53.92 Hyundai 19,699 19,259 440 5,739 5,399 8.30 -3.42 -3.63 52.17 Infiniti 41,191 40,607 584 6,895 6,311 9.25 -5.96 -6.42 63.04 Jeep 26,068 25,607 461 6,117 5,655 8.16 -4.26 -4.55 55.78 Kia 18,674 18,301 373 5,604 5,231 7.14 -3.33 -3.51 50.89 Lexus 42,024 41,460 564 6,994 6,430 8.77 -6.02 -6.46 62.74 Lincoln 38,166 37,668 498 6,740 6,242 7.98 -5.67 -6.06 61.22 Mazda 22,490 22,001 489 5,936 5,447 8.98 -3.78 -4.05 53.93 Mercedes-Benz 45,190 44,612 578 7,108 6,530 8.84 -6.32 -6.78 64.51 Mercury 24,355 23,869 486 6,029 5,544 8.77 -4.05 -4.33 54.64 Mini 24,721 24,196 525 6,099 5,573 9.43 -4.07 -4.36 55.67 Mitsubishi 22,177 21,646 531 5,896 5,364 9.90 -3.76 -4.03 53.27 Nissan 22,736 22,291 445 5,910 5,466 8.13 -3.83 -4.07 53.71 Pontiac 21,217 20,760 457 5,807 5,349 8.55 -3.65 -3.88 53.00 Saab 32,625 32,077 548 6,393 5,846 9.36 -5.11 -5.50 58.43 Saturn 24,063 23,534 529 5,994 5,465 9.68 -4.00 -4.29 54.13 Scion 17,246 16,892 354 5,574 5,219 6.79 -3.13 -3.31 50.86 Subaru 22,672 22,166 506 5,974 5,468 9.25 -3.81 -4.08 54.14 Suzuki 18,253 17,520 733 5,633 5,100 10.44 -3.18 -3.41 50.89 Toyota 23,341 22,973 368 5,941 5,573 6.60 -3.93 -4.16 54.28 Volkswagen 22,136 21,621 515 5,929 5,415 9.50 -3.73 -3.99 53.87 Volvo 34,019 33,456 563 6,561 5,998 9.38 -5.17 -5.57 59.78 All 24,940 24,518 422 6,015 5,593 7.55 -4.10 -4.35 56.59 Note: All averages are weighted by the market shares. Difference = observed - predicted by full information model. The own price elasticities are measured in the absolute values.

46

Figure A.1: Dealers and Dealer Areas in Richmond, VA

Note: This maps displays the locaitons of dealers across the Richmond metropolitan area, color coded by assigned co-location area. Areas 1 and 2 are both shaded light purple and area 9, not shown, is about 10 miles south of area 8 near Petersburg, VA.

47

Figure A.2: Sales by Dealer’s Census Tract, 2008

Note: This maps displays the sales of new car dealers in the Richmond metropolitan area for the year 2008. Total sales are shaded by the US Census Tract of the dealer location. Darker red represents more total sales from that tract. As can be seen from Figure A.1, there are typically multiple dealers per tract.

48