THREE ESSAYS ON INDUSTRIAL ORGANIZATION AND ENVIRONMENTAL

ECONOMICS

By

ZIYING YANG

A dissertation submitted in partial fulfillment of the requirements for the degree of

DOCTOR OF PHILOSOPHY

WASHINGTON STATE UNIVERSITY School of Economic Sciences

JULY 2017

© Copyright by ZIYING YANG, 2017 All Rights Reserved © Copyright by ZIYING YANG, 2017 All Rights Reserved To the Faculty of Washington State University:

The members of the Committee appointed to examine the dissertation of ZIYING

YANG find it satisfactory and recommend that it be accepted.

Felix Munoz-Garcia, Ph.D., Chair

Ana Espinola-Arredondo, Ph.D.

Jill McCluskey, Ph.D.

Jia Yan, Ph.D.

ii ACKNOWLEDGMENTS

First, I would like to thank my advisor, Felix Munoz-Garcia, for his inspiration, advice and guidance. He helped me know how to be a good researcher and how to be a qualified advisor in the future. It was really nice to work with Dr. Felix Munoz-Garcia. I would also like to express my sincere gratitude to Dr. Ana Espinola-Arredondo, Dr. Jill McCluskey, and Dr. Jia Yan for their insightful feedback, comments, and suggestions. My dissertation was polished with their help.

I am thankful to Xinlong Tan, Haowei Yu, Xiangrui Wang, and Sunny. Thanks for their time to discuss my ideas, listen to my presentation, and to provide valuable suggestions for my research.

It is worth to note that Jaimie Dahl, Tom Dahl, Rich Hoeft, Karla Makus, and other staff in the School of Economic Sciences provided me with great help during my Ph.D. program.

For this, I am deeply grateful.

I would also like to thank my friends in Pullman. Without them, I couldn’t have such a great time in the past four years at Pullman.

Most importantly, I would like to express my deepest gratitude to my wife, Shuhong

Zhang, for her sacrifice and encouragement. I owe much of my achievements during my

Ph.D. studies to her. Without her, I couldn’t get my Ph.D. program done. I acknowledge the support and understanding of my parents, Shengding Yang and Lanxiang Ye, my sister,

Zijiao Yang, and my brother, Helin Yang.

iii THREE ESSAYS ON INDUSTRIAL ORGANIZATION AND ENVIRONMENTAL

ECONOMICS

Abstract

by Ziying Yang, Ph.D. Washington State University July 2017

Chair: Felix Munoz-Garcia

This dissertation consists of three papers on industrial organization and environmental economics. The first paper analyzes a two-stage sequential-move model of location and pricing to identify firm’s location, output, and welfare. We consider two pricing regimes

(mill pricing and spatial price discrimination) and, unlike previous literature, allow in each of them for a non-uniform population density, non-constant location costs (i.e., the setup costs, such rental costs and land prices, differ by firm’s location), and endogenous market boundaries. Under non-constant location costs, we find that welfare is higher (lower) under mill than under discriminatory pricing when transportation rates are low (high, respectively).

The second paper investigates the effect of Beijing’s vehicle lottery system on fleet com- position, fuel consumption, air pollution, and social welfare. We find that the lottery reduced new passenger vehicle sales by 50.15%, fuel consumption by 48.69%, and pollutant emissions by 48.69% in 2012. Also, such lottery shifted new auto purchases towards high-end but less fuel efficient vehicles. In our counterfactual analysis, we show that a progressive tax scheme works better than the lottery system at decreasing fuel consumption and air pollution, and leads to a higher fleet fuel efficiency and less welfare loss.

iv The third paper investigates the effectiveness and welfare consequences of (i) vehicle and vessel usage tax (VVUT) incentives, (ii) fuel-efficient vehicle subsidy program, and (iii) new- energy vehicle (NEV) private purchase subsidy pilot program. The empirical findings suggest that these policies promote the diffusion of fuel-efficient vehicles and NEVs, and improve

fleet fuel efficiency. However, VVUT incentives and fuel-efficient vehicle subsidy program increases oil consumption and CO2 emissions. Although NEV private purchase subsidy pilot program cuts down gas consumption, it raises CO2 emissions. VVUT incentives and fuel- efficient vehicle subsidy program improve social welfare, while NEV private purchase subsidy pilot program causes welfare loss.

v TABLE OF CONTENTS

Page

ACKNOWLEDGMENTS ...... iii

ABSTRACT...... iv

LIST OF TABLES ...... x

LIST OF FIGURES ...... xii

1. Can Banning Spatial Price Discrimination Improve Social Welfare? ...... 1

1.1 Introduction ...... 1 1.2 Model ...... 6 1.3 Second stage: Pricing decisions ...... 8

1.3.1 Mill pricing ...... 8

1.3.2 Discriminatory pricing ...... 10 1.4 First stage: location decisions ...... 12

1.4.1 Equilibrium results ...... 12

1.4.2 Numerical simulation...... 18

1.4.3 Effects of banning spatial price discrimination ...... 23 1.5 Conclusions ...... 24

2. The Effects of Beijing’s Vehicle Lottery System on Fleet Composition and En-

vironment ...... 27

2.1 Introduction ...... 27 2.2 Industry Background, Policy and Data Description ...... 33

2.2.1 Industry Background ...... 33

2.2.2 Policy Description...... 34

vi 2.2.3 Data...... 37

Data Description ...... 37

Stylized Facts ...... 42 2.3 Empirical Model and Estimation ...... 44

2.3.1 Utility Function Specification...... 44

2.3.2 Choice Probability and Aggregate Demand ...... 47

2.3.3 Identification and Estimation ...... 47 2.4 Estimation Results ...... 50

2.4.1 Parameter Estimates from the Structural Demand Model . . . . 50

2.4.2 Robustness Checks ...... 53

2.4.3 Impact on New Vehicle Registration ...... 54 2.5 Counterfactual Analysis ...... 56

2.5.1 Impact on Fleet Composition ...... 57

2.5.2 Impacts on Gasoline Consumption and Pollutant Emissions . . 61

2.5.3 Welfare Analysis ...... 66

2.5.4 Alternative Policies and Comparisons ...... 68 2.6 Conclusion...... 70

3. Welfare Analysis of Government Incentives for Fuel Efficient Vehicles and New

Energy Vehicles in China...... 73

3.1 Introduction ...... 74 3.2 Policy and Data Description ...... 78

3.2.1 Policy Description...... 78

vehicle and vessel usage tax Incentives ...... 78

vii Fuel-efficient Vehicle Subsidy Program ...... 80

NEV Private Purchase Subsidy Pilot Program ...... 82

3.2.2 Data...... 84 3.3 Empirical Model and Estimation ...... 88

3.3.1 Utility Function Specification...... 88

3.3.2 Market Share and Aggregate Demand ...... 91

3.3.3 Identification Method ...... 92

3.3.4 Instruments ...... 94 3.4 Estimation Results ...... 95

3.4.1 Parameter Estimates from the Structural Demand Model . . . . 95

3.4.2 Alternative Specifications ...... 98

3.4.3 Elasticities ...... 100 3.5 Counterfactual Analysis ...... 102

3.5.1 Car Sales ...... 103

3.5.2 Gas Consumption and CO2 Emissions ...... 107

3.5.3 Welfare Analysis ...... 110

3.5.4 Government Support versus Gasoline Tax Comparisons ...... 113 3.6 Conclusion...... 114

APPENDIX ...... 116

A. Proof of Lemma 1...... 116 B. Proof of Lemma 2...... 117 C. Proof of Proposition 1 ...... 118 D. Comparison of equilibrium outcomes in Case 3...... 119 E. Simulation description ...... 121 F. City Characteristics in Beijing, Nanjing, Shenzhen, and Tianjin ...... 124

viii G. Derivation of Consumer Surplus of Household ...... 124

ix LIST OF TABLES

Table Page 1.1 Summary of firm’s optimal locations under different locations costs and population distribution ...... 16 1.2 Simulated equilibrium prices, market radius, optimal locations and profits under mill and discriminatory pricing – non-constant location costs and normal population density ...... 19 1.3 Simulated outputs, consumers’ surpluses, and social welfare under mill and discriminatory pricing – non-constant location costs and normal population density ...... 22 2.1 Summary Statistics of Vehicle Data 2009-2012 ...... 41

2.2 Estimation Results for the Model ...... 51 2.3 Policy Impact on New Passenger Cars Registration in Beijing ...... 55 2.4 First Registration Tax Rate in Hong Kong and Counterfactual Scenario (II) 58

2.5 Consumption Tax Rates in China and Counterfactual Scenario (III)...... 58 2.6 Price, Horsepower, and Fuel Efficiency Summary Statistics under Null Sce- nario and Counterfactual Scenario (I) in Beijing in 2012 ...... 60 2.7 Average Emissions per Liter Gasoline for Passenger Cars ...... 63

2.8 Counterfactual Analysis under Different Scenarios in Beijing 2011Q2-2012Q4 65 3.1 VVUT for Passenger Vehicles in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen based on VVUT Law (2007) ...... 79 3.2 VVUT for Passenger Vehicles based on VVUT Law (2012) ...... 79 3.3 Fuel Consumption Limits Required by Fuel-efficient Vehicle Subsidy Pro- grams Introduced in June 2010 and October 2011 ...... 81

3.4 Summary Statistics of Vehicle Sales, Characteristics, and Consumer Reviews 86 3.5 Estimation Results for the Demand Side ...... 97 3.6 Price Elasticities for Selected Products in Shenzhen in 2012 ...... 101

3.7 Sales Changes in the Five Cities in 2012 under Each Scenario ...... 104 3.8 Counterfactual Analysis under Different Scenarios in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen in 2012 ...... 111

x 0.1 City Characteristics in Beijing, Nanjing, Shenzhen, and Tianjin ...... 125

xi LIST OF FIGURES

Figure Page 1.1 The market line ...... 7

1.2 Simulated MCL(s), MRLm(s) and MRLd(s)...... 17 1.3 Delivery price schedules: mill versus discriminatory pricing ...... 21 2.1 New Vehicle Quarterly Sales in Beijing, Nanjing, Shenzhen, and Tianjin. . . . 38

2.2 Quarterly Sales-weighted Average Prices (1,000 Yuan) 2009-2012 in Four Cities ...... 43

2.3 Quarterly Sales-weighted Average Horsepower (kw) 2009-2012 in Four Cities 43 2.4 Quarterly Sales-weighted Average Fuel Efficiency (L/100km) 2009-2012 in Four Cities ...... 44 2.5 The Cumulative Distribution Functions of Price of New Cars Registered in Beijing in 2012 ...... 61 2.6 The Cumulative Distribution Functions of Horsepower of New Cars Regis- tered in Beijing in 2012 ...... 62

2.7 The Cumulative Distribution Functions of Fuel Efficiency of New Cars Reg- istered in Beijing in 2012...... 63

xii Dedication

I dedicate this dissertation to my wife, Shuhong Zhang.

xiii CHAPTER 1.

CAN BANNING SPATIAL PRICE DISCRIMINATION

IMPROVE SOCIAL WELFARE?

We analyze a two-stage sequential-move model of location and pricing to identify firm’s location, output, and welfare. We consider two pricing regimes (mill pricing and spatial price discrimination) and, unlike previous literature, allow in each of them for a non-uniform population density, non-constant location costs (i.e., the setup costs, such rental costs and land prices, differ by firm’s location), and endogenous market boundaries. Under constan- t location costs, our results show the firm locates at the city center under both mill and discriminatory pricing, and that output is larger under spatial price discrimination. Wel- fare comparisons are, however, ambiguous. Under non-constant location costs, we find the optimal location can move away from the city center, and does not coincide across pricing regimes. Compared with mill pricing, spatial price discrimination generates a higher level of output. We also find that welfare is higher (lower) under mill than under discriminatory pricing when transportation rates are low (high, respectively).

1.1 Introduction

In some industries, such as cement and ready-mixed concrete, spatial price discrimina- tion is possible because firms are geographically differentiated and transportation is costly

(Vogel, 2011; Miller and Osborne, 2014). For example, spatial price discrimination has a

1 long history in the cement industry, where producers privately negotiate contracts with their

customers (Miller and Osborne, 2014).1 While such spatial price discrimination yields larger

profit, its welfare effects have long been the subject of debate. Despite being forbidden in

many countries, like European Union under the Treaty on the Functioning of the European

Union (2007) and China under the Antimonopoly Law (2007),2 many analysts argue that banning spatial price discrimination may harm social welfare; see Greenhut and Ohta (1972) and Holahan (1975). Most previous studies show that price discrimination produces an un- ambiguous welfare improvement, but this paper demonstrates that a welfare reduction can emerge under relatively general conditions.

We analyze a monopolist’s location decision and compare the resulting output and social welfare under two pricing regimes: spatial price discrimination and mill pricing (no discrim- ination).3 Previous studies considered two simplifying assumptions: (1) firm’s location was given; and (2) consumers are uniformly distributed. While some studies relaxed assump- tion (1) by allowing for firm’s location to be endogenous, they assumed that location costs were constant, i.e., suggesting that the firm incurs the same location cost regardless of its distance from the city center.4 We separately relax these assumptions, considering a model

1These contracts specify discounts which depend on the ability of customers to substitute toward cement produced by other firms. Most Portland cement is moved by truck, and purchasers are responsible for its transportation costs, which accounts for a substantial proportion of total costs since this cement is inexpensive relative to its weight. For more institutional details on this industry, see Miller and Osborne (2014). 2In the European Union (EU), the Treaty on the Functioning of the European Union (TFEU) excludes any discrimination between producers or consumers with the Union (see Article 40), where any common price policy should be based on common criteria and uniform methods of calculation. In China, the Antimonopoly Law (2008) prohibits price discrimination in all markets, including final product markets (see Article 17). 3Under mill pricing, the firm charges each consumer a delivery (total) price that is equal to the sum of a mill price and the transportation cost, while under spatial discriminatory pricing the firm sets location- specific delivered prices for consumers. 4Location costs refer to firm’s setup costs, such as rental costs and land prices. In this paper, non-constant

2 of endogenous firm location (relaxing 1); in which consumers are not necessarily uniformly distributed (relaxing 2); and whereby location costs are not necessarily constant. Such a general model allows us to show that output and welfare predictions are critically affected by the assumptions often considered by the previous literature.

As Cheung and Wang (1995) and Hwang and Mai (1990), we solve a two-stage sequential- move model of location and pricing. In particular, the firm chooses its location in the first stage, and prices are chosen in the second stage. We separately identify equilibrium behavior under spatial price discrimination and mill pricing, and provide Monte Carlo simulations for those expressions without explicit solutions.

Our results find that, when location costs are constant, the firm locates at the city center both under mill and discriminatory pricing since most customers concentrate at the city center.5 However, when the location costs are non-constant (as in most industries), we find that optimal location differs across pricing regimes. Under mill pricing, the firm locates closer to the city center than under spatial price discrimination when transportation rates are low; otherwise, the optimal location under mill pricing is further from the city center.

Spatial price discrimination, hence, yields unambiguously a larger market radius, higher profits, and a larger output. Welfare under this pricing regime, however, depends on the transportation cost per unit of distance. For low transportation costs, welfare under mill pricing is higher than under spatial price discrimination; while for high transportation costs the opposite result applies. Therefore, when transportation costs are relatively low (e.g., location costs mean that the location costs differ by firm’s location. 5The market area (i.e., customers served), profit and output are larger under spatial price discrimination than under mill pricing; whereas, the welfare may be larger or lower under discriminatory pricing.

3 roads and railroads are in good condition, or transportation firms facing cheap oil prices),

spatial price discrimination is actually welfare reducing. In these contexts, regulations that

ban spatial price discrimination become welfare improving, while laws that allow or prevent

this type of discrimination under all conditions can entail welfare losses.

Location costs play a crucial role in firm’s location.6 The monopoly spatial price dis-

crimination literature assumes a constant location cost over the market space.7 However,

location costs, such as rental costs and land prices, differ by location (Hinloopen and Martin,

2016). For instance, a one-mile increase in the distance from the city center decreases house

prices in Chicago by 8%, (McMillen, 2003); and, similarly, a 1% increase in the distance from

the city center decreases rentals (land prices) in Shanghai (New York Metropolitan area) by

0.14% (0.95%, respectively), see Wang et al. (2016) and Haughwout et al. (2008).8 Impor- tantly, our results show that, relaxing the assumption of constant location costs changes

firm’s equilibrium location. Specifically, the firm faces a trade-off since locating closer to the city center helps it serve a larger number of customers but entails a higher location cost.

Our above findings show how this trade-off affects firm’s location and, as a consequence, equilibrium output and welfare. Our results can thus be used at urban policies that seek to attract more firms to the city center (e.g., inner cities), as our findings help identify under

6Before a firm starts operation, it must incur some location-dependent costs, such as renting a building as factory. 7See Greenhut and Ohta (1972), Holahan (1975), Gronberg and Meyer (1982), Hobbs (1986), and An- derson et al. (1989) for the case of uniformly distributed population, and Beckmann (1976), Hwang and Mai (1990), and Cheung and Wang (1995) for the case of non-uniformly distributed population. All these papers assume constant location costs. 8In addition, the price differential between city center and suburbs has significantly increased since 2000, as documented by Edlund et al. (2015).

4 which cases these policies can be beneficial.

Related literature. Our paper is related to the literature on monopoly spatial price discrimination. Using a model with uniformly distributed consumers and linear demand,

Greenhut and Ohta (1972) and Holahan (1975) argue that, when the market area is variable, spatial price discrimination results in firms producing larger output, serving larger market areas, and generating larger social welfare than under mill pricing. Beckmann (1976) relaxes the assumption of uniform population density but assumes an exogenous market area.9 He shows that spatial price discrimination yields lower welfare levels than mill pricing; a result that holds for all customer distributions.

Another common assumption in the previous literature is that the monopolist’s location is given and coincides across different pricing policies. Allowing the firm to strategically choose its location based on the pricing regime, however, affects output and welfare, as shown in

Beckmann and Thisse (1987), Hwang and Mai (1990), and Tan (2001). Cheung and Wang

(1995) extend the analysis to non-uniform demands and show that, when the monopolist serves a fixed market area, spatial price discrimination results in the same total output, higher profit, lower consumer surplus and lower total welfare than mill pricing. They also demonstrate that when location is chosen endogenously, output falls, and consumer surplus and total welfare may rise or fall. However, the above studies assume a fixed market. Since market areas served vary with each pricing regime (Greenhut and Ohta, 1972; Holahan, 1975;

Ohta and Wako, 1988), we incorporate the location decision into models of monopolist’s

9Here, exogenous market area (or fixed market area) means that the number of markets is given and all markets are served. Correspondingly, endogenous market area refers to the case where the firm can decide how many markets to serve.

5 spatial price discrimination with endogenous market areas. This general model allows us to identify novel settings under which spatial price discrimination is welfare reducing, thus supporting the arguments of regulators proposing to ban such a pricing practice at least under certain conditions.

This paper, hence, contributes to the monopoly spatial price discrimination literature in two ways. First, although several studies analyze output and welfare effects of spatial price discrimination, few of them simultaneously consider non-uniform population density, endogenous market boundaries, and endogenous plant location. Our setting hence is closer to real market conditions. Second, to our knowledge, this is the first study considering a non-constant location cost in the analysis of monopolist’s spatial price discrimination.

While most studies assume that location costs are constant (i.e., firm incurs the same costs, regardless of its distance from the city center), we allow them to decrease as the firm locates further away from the city center, and show how output and welfare change.

1.2 Model

Consider a setting where a monopolist produces a homogenous good and sells the product to consumers distributed along the market line as shown in Figure 1. The total number of consumers in the market is n. Following Claycombe (1996), we assume the population density is closely approximated using the normal distribution. Let the city center be at point 0. Then the population density at any point x is

2 1 − x φ(x) = √ e 2σ2 (1.1) 2πσ

6 where σ denotes the standard deviation of the population distribution. However, our model

and methodology are not limited to normal density function.

Figure 1.1: The market line

In this paper, we assume that the monopolist produces at a single location s on the market line. The marginal cost of production is assumed constant, and, without loss of generality, can be normalized to zero. We consider two types of location costs: constant and non-constant. When non-constant, location costs are increasing in the market size, and in the firm’s proximity to the city center, where the largest number of customers concentrate.10

In particular, location costs are represented by

2 − s An 2σ2 F (s, n) = √ e F (1.2) 2πσF

where A > 0 and σF is the standard deviation of location cost distribution. If, in contrast, location cost is constant, we consider F (s, n) = An.

10As argued by Berliant and Konishi (2000), the setup costs in a marketplace depend on location and are proportional to the number of consumers in the market, n.

7 Assume each consumer at location x has linear demand function.11

qx = a − b(p + t|x − s|) (1.3)

where a, b > 0, qx is the quantity demanded by consumer at location x, p is the good’s price, t represents transportation cost per unit of distance, and hence t|x − s| denotes the transportation cost that customers at point x face. The monopolist can employ either mill pricing or discriminatory pricing. Under mill (discriminatory) pricing, besides transportation cost, the firm charges the same (different) product price pm (pd) to each individual regardless of (depending on, respectively) his location.

Following models of spatial competition with endogenous location and prices, we assume a sequential-move model, with location chosen in the first stage and prices chosen in the second stage (Hwang and Mai, 1990; Braid, 2008).12 In the following sections, we employ

backward induction to solve for the equilibrium.

1.3 Second stage: Pricing decisions

1.3.1 Mill pricing

Under mill pricing, the monopolist charges each consumer a delivery price, which is equal

to a constant mill price pm plus the transportation cost t|x − s|. With equation (1.1) and

11This assumption conforms to the work by Greenhut and Ohta (1972), Beckmann (1976), Holahan (1975), Guo and Lai (2014), Chen and Hwang (2014), and Andree (2013). 12Here, the time does not matter. We can also solve the profit maximization problem by determining the location and the prices simultaneously. The results do not change.

8 (1.3), we can derive the monopolist’s revenue from selling the good to a customer at location

x

NRm(pm, x) = nφ(x)pmqx (1.4)

= nφ(x)pm[a − b(pm + t|x − s|)] Since production costs are zeros, the market boundaries for the monopolist are customers

at locations x satisfy NRm(pm, x) = 0, that is,

a − bp R = s ± m (1.5) m bt

which means that, when depicted over the market line (Figure 1), the firm’s left boundary

a−bpm a−bpm under mill pricing is LRm = s − bt and the right boundary is RRm = s + bt .

Given the above boundaries, the monopolist’s profit is

a−bpm Z s+ bt πm = nφ(x)pm[a − b(pm + t|x − s|)]dx − F (s, n) (1.6) a−bpm s− bt

13 Taking first order condition with respect to the monopolist’s mill price, pm, yields

Z s+ a−bpm ∂πm bt = nφ(x)[a − 2bpm − bt|x − s|]dx = 0 (1.7) ∂pm a−bpm s− bt

∗ Given that the density function φ(x) follows a normal distribution, we cannot solve for pm in equation (1.7) analytically. It can only be analyzed numerically, as further developed in

2 a−bpm 13 ∂ πm R s+ bt Note that the second-order condition for a maximum is satisfied since ∂p2 = −2b a−bpm nφ(x)dx < m s− bt 0, i.e., profits are concave in the mill price.

9 ∗ ∗ a 14 section 1.4. Let pm solve equation (1.7), where pm ∈ (0, 2b ). We can derive the monopolist’s aggregate output under mill pricing

∗ a−bpm Z s+ bt Q = nφ(x)[a − b(p∗ + t|x − s|)]dx (1.8) m ∗ m a−bpm s− bt which yields profit of

∗ Πm = pmQm − F (s, n) (1.9) consumer’s surplus of

a−bp∗ s+ m ∗ 2 Z bt [a − b(p + t|x − s|)] CS = nφ(x) m dx (1.10) m ∗ a−bpm 2b s− bt

and the social welfare

Wm = Πm + CSm (1.11)

1.3.2 Discriminatory pricing

Under discriminatory pricing, the monopolist is allowed to charge different prices pd for

the good to consumers at different locations. In this case, the monopolist’s net revenue from

customers at location x is15

NRd(pd, x) = nφ(x)pd[a − b(pd + t|x − s|)] (1.12)

14 ∂πm ∂πm Recall that is decreasing in pm. This condition, together with the fact that |p =0 = ∂pm ∂pm m s−a/(bt) s−a/(2bt) R nφ(x)[a − bt|x − s|]dx > 0 and ∂πm | = − R nφ(x)bt|x − s|dx < 0, implies that, s+a/(bt) ∂pm pm=a/(2b) s+a/(2bt) using the mean value theorem, the optimal price p∗ , which is determined by ∂πm = 0, must be unique and m ∂pm a at an interior point of the interval (0, 2b ). 15Similar arguments are made in the work by Holahan (1975) and Cheung and Wang (1995).

10 16 Taking first order condition with respect to price pd yields

∂NRd(pd, x) = nφ(x)(a − 2bpd − bt|x − s|) = 0 (1.13) ∂pd

By solving for pd in equation (1.13), we find the price under discriminatory pricing

a − bt|x − s| p (x) = (1.14) d 2b

which is a function of the location of customer x, as opposed to the mill price in expression

(1.7) which was constant for all x. Substituting equation (1.14) into (1.12), revenue becomes

nφ(x)(a − bt|x − s|)2 NR (x) = (1.15) d 4b

Let NRd(x) = 0 in order to obtain the boundaries under discriminatory pricing

a R = s ± (1.16) d bt

a which means that under discriminatory pricing, the monopolist’s left boundary is lRd = s− bt

a and the right boundary is rRd = s + bt . Using the discriminatory price in (1.14) and the boundaries in (1.16), we can find the monopolist’s aggregate output under discriminatory pricing

s+ a Z bt nφ(x)(a − bt|x − s|) Qd = dx (1.17) a 2 s− bt its profit

2 16 ∂ NRd(pd,x) Note that the second-order condition for a maximum is satisfied since 2 = −2bnφ(x) < 0. ∂pd

11 s+ a 2 Z bt nφ(x)(a − bt|x − s|) Πd = dx − F (s, n) (1.18) a 4b s− bt consumers’ surplus

s+ a 2 Z bt nφ(x)(a − bt|x − s|) CSd = dx (1.19) a 8b s− bt and social welfare

Wd = Πd + CSd (1.20)

1.4 First stage: location decisions

In this section, the monopolist chooses the plant location. We consider that location cost

F (s, n) depends on the firm’s distance from the city center, s, and the market size, n, i.e.,

2 − s An 2σ2 F (s, n) = √ e F . When constant location cost, the location cost satisfies F (s, n) = An, 2πσF thus being constant in distance s. Since the population distribution and the location cost are both symmetric with respect to the city center (point 0 in Figure 1), we only need to analyze the case where s ≥ 0. Analogous results apply when s ≤ 0.

1.4.1 Equilibrium results

Mill pricing. Under mill pricing, the monopolist chooses a location to maximize the equilibrium profit in (1.9). Taking first order conditions with respect to distance s yields

12  ∗  s s+ a−bpm 2 Z Z bt − s ∗ Ans 2σ2 btnp φ(x)dx − φ(x)dx = √ e F (1.21) m  ∗  3 a−bpm 2πσ s− bt s F The right-hand side of (1.21) represents the marginal cost that the monopolist bears when locating its plant closer to the city center (since land prices become more expensive as s → 0). The left-hand side, in contrast, indicates the marginal revenue of locating closer to the city center (where a larger mass of customer live). At the optimal location, marginal costs and revenues under mill pricing coincide, i.e., MRLm(s) = MCL(s).

Discriminatory pricing. Under discriminatory pricing, the monopolist chooses a location to maximize the equilibrium profit in (1.18). Taking first order conditions with respect to distance s, we find

s s+ a 2 Z Z bt − s a − bt|x − s| a − bt|x − s| Ans 2σ2 tnφ(x) dx − tnφ(x) dx = √ e F (1.22) a 2 2 2πσ3 s− bt s F

The right-hand side of (1.22) coincides with that of (1.21), indicating that the mo- nopolist’s marginal cost of locating closer to the city center is unaffected by the pricing regime. The marginal revenue (left-hand side) is, however, different from that under mill pricing.17 Similarly as under mill pricing, the monopolist stops approaching the city cen- ter when marginal costs and revenues under discriminatory pricing offset each other, i.e.,

MCLd(s) = MCL(s).

In terms of location costs and population distribution, four cases arise: i) constant loca- tion costs and uniformly distributed customers; ii) non-constant location costs but uniformly

17A direct ranking of the two marginal revenues is infeasible at this general stage of the model; but several numerical simulations are provided at the end of the section.

13 distributed customers; iii) constant location cost and normally distributed customers; and iv) non-constant location costs and normally distributed customers. We next discuss each case.

Case 1: Constant location costs and uniformly distributed customers. Constant location costs mean F (s, n) = An for every distance s. Under the assumption of uniformly distributed customers, the population density is v at each location over the market line. In this context, a

∗ continuum of equilibria emerges for both pricing regimes. That is, equilibrium locations sm,1

∗ and sd,1 can be any points on the market line, where subscript 1 denotes Case 1, since profits coincide at all locations. In addition, spatial price discrimination leads to more markets being served, and generates a higher level of profit, output and greater social welfare than mill pricing; see Holahan (1975) and Greenhut and Ohta (1972). We next present this result.

(All proofs are relegated to the Appendix.)

LEMMA 1: Under uniform population density and constant location cost, price dis- crimination yields a larger output and welfare than mill pricing.

Case 2: Non-constant location costs and uniformly distributed customers. Assume the location costs are distinct at different locations and that the population density is uniform.

Since customers are uniformly distributed in this setting, there are no benefits of locating at the city center, i.e., marginal revenues are zero under both pricing regimes. Marginal costs of location are, however, increasing as the firm approaches the city center, driving it away from the city center, s∗ → ∞. This result applies under both pricing regimes, i.e.,

∗ ∗ ∗ sm,2 = sd,2 = s , where subscript 2 denotes Case 2.

LEMMA 2: Given non-constant location costs and a uniform population distribution,

14 the firm serves a larger market area and yields a higher level of profit, output and social

welfare under spatial price discrimination than under mill pricing.

Case 3: Constant location cost and normally distributed customers. In this case, we as- sume location costs are constant at every location over the market line. In such a setting, the marginal costs of locating closer to the city center are zero, both under mill and dis- criminatory pricing, driving the monopoly to locate as close to the city center as possible in order to benefit from a larger number of customers, as shown in the following proposition.

PROPOSITION 1. When location costs are constant, the firm locates at the city center,

∗ ∗ both under mill and discriminatory pricing, i.e., sm = sd = 0.

In particular, constant location costs entail that, under mill (discriminatory) pricing, the monopolist locates at the median of the population (demand) distribution, thus leading the same number of customers (aggregate demand) to the left- and right-hand side of its location. This is a common result in the literature of spatial discrimination when location costs are constant; see Greenhut and Ohta (1972) and Holahan (1975).

Since the monopolist’s location coincides under both pricing regimes, the market area, profit and total output are larger under spatial price discrimination than under mill pricing.

The welfare, however, may be higher or lower under discriminatory pricing (see Appendix

D. for more details).

Our result on social welfare encompasses that in Cheung and Wang (1995), who show that spatial price discrimination results in lower welfare assuming fixed market area, non- uniform demands, and a given firm’s location. In our model,allowing endogenous market boundaries, a discriminatory pricing monopoly serves a larger market area, thus increasing

15 welfare.

Case 4: Non-constant location costs and normally distributed customers. In this case, the

firm faces two opposing forces in its location decision: on one hand, the firm prefers to locate away from the city center since its location costs are cheaper. On the other hand, it prefers a central location in order to capture a larger amount of customers (normal distribution). As a result of this trade-off, the firm does not locate at the city center (as it did in Case 3) nor as far away from the center as possible (as it did in Case 2), but somewhere in between these two polar locations. Analytical solutions for optimal locations are, however, infeasible because of nonlinearities in equations (1.21) and (1.22), but numerical simulations are provided in section 1.4.2.

Table 1 summarizes firm’s optimal location as a function of location costs (in rows) and population distribution (in column).

Table 1.1: Summary of firm’s optimal locations under different locations costs and population

distribution

Population distribution

Uniform Normal

∗ ∗ ∗ ∗ sm and sd can be any points sm = sd = 0, that is, Constant on the market line the city center

∗ ∗ ∗ ∗ sm = sd = s , where s minimizes ∗ ∗ sm, sd ∈ R Non-constant the location costs. If location costs Location costs (See numerical simulation) follow normal distribution,then s∗ → ∞

16 In order to illustrate the firm’s incentives when choosing its location, Figure 1.2 depicts

18 the marginal cost and revenues of location, MCL(s), MRLm(s) and MRLd(s). Marginal revenues under both pricing regimes are increasing in s first and then decreasing, converging to 0 when s → ∞. For our parameter values, when 0 < s < 1.92 (0 < s < 0.85), the marginal cost of locating closer to the city center, MCL(s), lies above the marginal revenue

MRLm(s)(MRLd(s)), while for s > 1.92 (s > 0.85), MCL(s) lies below the MRLm(s)

(MRLd(s), respectively). MRLd(s) lies above MRLm(s) for all locations.

Figure 1.2: Simulated MCL(s), MRLm(s) and MRLd(s)

As an illustration, we can evaluate the MCL, MRLm and MRLd functions at the spe- cial cases discussed above. As in Case 1, location costs are constant and the population distribution is uniform. In this setting, MCL = MRLm = MRLd = 0 for s ∈ [0, ∞), thus

∗ ∗ yielding a continuum of optimal locations sm and sd. If location costs are non-constant and

18 For simplicity, we set a = b = t = σ = σF = 1 and A = 0.15. More details about the simulation can be found in Appendix E..

17 the population density is uniform, yielding MRLm = MRLd = 0. In addition, MCL lies

above both MRLm and MRLd, leading the firm to choose a plant location as far away from the city center as possible under both pricing systems (as in Case 2). In contrast, when locations costs are constant and the population density is normal, yielding MCL = 0 for

all s ∈ [0, ∞) while the marginal revenue curves MRLm and MRLd lie both above MCL.

As a result, the monopolist chooses the city center as optimal locations under both mill and

discriminatory pricing (as in Case 3).

1.4.2 Numerical simulation

From our previous analysis, we obtained analytical solutions for Cases 1 to 3. The first-

order conditions for optimal locations in Case 4, however, could not be solved analytically.

We next resort to numerical simulation, similar to other studies on monopoly spatial price

discrimination such as Claycombe (1996) and Tan (2001). In particular, consider parameters

19 a = b = σ = σF = 1 and A = 0.15, and market size n of 100,000. Appendix E. provides a

sequential description of our simulation.

Tables 1.2 and 1.3 report the simulated results, which comprise equilibrium prices, lo-

cations, market radius, profits, outputs, consumers’ surplus, and social welfare. Simulation

results are given for transportation costs per unit of distance (transportation rates), t, be-

19Venkatesh and Kamakura (2003) generate a population of 90,000 consumers in their simulation to study bundling strategies and pricing patterns under a monopoly. We use 100,000 consumers in our simulation so that the population sample is closer to a normal distribution. In addition, when n is large enough and the population approximates normal distribution, the value of n does not affect performance comparisons between mill pricing and discriminatory pricing. For example, the sign of Qd − Qm is not affected by the value of n.

18 tween 0.35 and 1.1 with an increment of 0.05.

Table 1.2: Simulated equilibrium prices, market radius, optimal locations and profits under mill

and discriminatory pricing – non-constant location costs and normal population density

∗ ∗ ∗ t pm Radiusm Radiusd sm sd πm πd 0.35 0.3803 1.7705 2.8571 0.0060 0.0080 7372.2304 8113.8006 0.40 0.3721 1.5699 2.5000 0.0073 0.0088 6246.0833 7047.7597 0.45 0.3656 1.4097 2.2222 0.0088 0.0096 5263.6036 6096.3630 0.50 0.3606 1.2788 2.0000 0.0094 0.0100 4404.8691 5247.5841 0.55 0.3566 1.1699 1.8182 0.0121 0.0119 3650.2082 4488.5363 0.60 0.3533 1.0778 1.6667 0.0158 0.0145 2988.8412 3813.6167 0.65 0.3507 0.9989 1.5385 0.0188 0.0162 2405.3920 3210.6041 0.70 0.3488 0.9303 1.4286 0.0293 0.0196 1881.9337 2670.0914 0.75 0.3468 0.8709 1.3333 0.0450 0.0266 1426.2834 2183.6689 0.80 0.3441 0.8199 1.2500 0.2057 0.0371 972.5622 1744.9690 0.85 0.3427 0.7733 1.1765 0.3886 0.0514 634.3469 1348.6521 0.90 0.3340 0.7400 1.1111 1.0192 0.0869 348.8360 988.8451 0.95 0.3246 0.7110 1.0526 1.5004 0.2564 174.0311 666.6155 1.00 0.3154 0.6846 1.0000 1.9215 0.8454 82.6227 403.1798 1.05 0.3077 0.6594 0.9524 2.2796 1.3126 35.9230 228.1564 1.10 0.3015 0.6350 0.9091 2.5719 1.6938 14.8039 121.3127

Prices. From Table 1.2, we can see that the mill price decreases as the transportation

rate t increases. Under discriminatory pricing, the price policy (expression 1.14) also indi-

cates that the good’s price is decreasing in the transportation rate. Intuitively, when the

transportation rate increases, the firm can absorb some transportation cost to sustain sales.

Market radius. Equations 1.5 and 1.16 indicate that the widths of the market area under

both mill and discriminatory pricing increase as the transportation rate t decreases. This point is also confirmed in Table 1.2, since the firm can deliver the products to a more distant area with lower transportation rates. Table 1.2 also shows the market radius is larger under discriminatory pricing than under mill pricing; as shown in Greenhut and Ohta (1972) and

19 Holahan (1975).

This result can be explained by the delivery price DP = p + t|x − s|. Let r represent

∗ each consumer’s distance from the production site. Using pm and pd(x), we can write the

expression for DP under mill pricing and spatial price discrimination as

∗ DPm = pm + tr (1.23)

a tr DP = + (1.24) d 2b 2

Figure 1.3 depicts the delivery price schedules DPm and DPd illustrating that both schedules are linear and positively sloped in the consumer’s distance to the monopolist, r. DPd is,

however, flatter than DPm. No matter which pricing regime the firm adopts, customers at

∗ ∗ a−bpm a−bpm a a distance r ∈ [0, bt ] are served by the firm, while customers at a distance r ∈ ( bt , bt ] are only served under discriminatory pricing.20

Location. Table 1.2 shows that the optimal locations under both mill and discriminatory pricing move away from the city center as the transportation rate increases. For any given location, the marginal revenue of location from approaching the city center decreases as transportation rate, t, increases. As a result, the marginal revenue lines in Figure 1.2,

MRLm and MRLd, become closer to the x axis. While the marginal cost of location, MCL, does not change. Thus, the firm locates further from the city center as transportation become more expensive; a result that holds under both pricing regimes.

∗ 20 1−pm for instance, for the parameter values considered in our simulation, customers at a distance r ∈ [0, t ] ∗ 1−pm are served under both pricing regimes, and customers r ∈ ( t , 1] are only served under discriminatory pricing. In figure 1.3, while intercepts a/b and a/2b become 1 and 1/2 in our parametric example, all ∗ remaining intercepts are still functions of transportation rate t and mill price pm. We could not obtain an analytical expression for such a price, hence we need to rely on numerical simulations.

20 Figure 1.3: Delivery price schedules: mill versus discriminatory pricing

Comparing the optimal locations under each pricing regime, we find that the optimal lo- cation under mill pricing is different from that under price discrimination. For our parameter

∗ ∗ ∗ ∗ 21 values, for t ≤ 0.5, sm < sd, while for t ≥ 0.55, sm > sd.

Profits. As displayed in Table 1.2, higher transportation rates result in lower profits under both mill and discriminatory pricing. This result is intuitive because more of the revenues are used to pay for transportation cost, which in turn lower profits. Table 1.2 show that spatial price discrimination is more profitable than mill pricing at all feasible values of t, i.e.,

πd > πm. Hence, in the absence of regulation, the monopolist would choose a discriminatory pricing policy.

Output. Table 1.3 shows that output under both pricing schedules decreases in trans-

21 ∗ ∗ When t ≈ 0.538, locations essentially coincide under both pricing regimes, sm ≈ sd. This finding is in line with Hwang and Mai (1990) and Cheung and Wang (1995).

21 portation rate.22 Furthermore, as previously discussed, the market area served decreases in

t. Hence, Qm and Qd decrease as t increases, as seen in Table 1.3.

We also find that Qd > Qm for all t. Holahan (1975) argues that spatial price discrim- ination has a ”market expanding effect”, namely a larger market area being served. As a result, spatial price discrimination generates larger output than a mill price policy.

Table 1.3: Simulated outputs, consumers’ surpluses, and social welfare under mill and discrimi-

natory pricing – non-constant location costs and normal population density

t Qm Qd CSm CSd Wm Wd 0.35 35116.08 36058.57 7883.55 7048.66 15255.78 15162.46 0.40 32870.71 34119.29 7368.76 6515.56 13614.84 13563.33 0.45 30761.19 32252.15 6887.84 6039.84 12151.44 12136.20 0.50 28807.95 30476.53 6443.84 5615.37 10848.71 10862.96 0.55 27017.24 28801.99 6038.61 5235.65 9688.82 9724.18 0.60 25390.71 27246.98 5671.47 4897.91 8660.31 8711.52 0.65 23911.07 25807.61 5337.18 4596.12 7742.58 7806.72 0.70 22530.75 24475.73 5025.35 4325.23 6907.29 6995.32 0.75 21312.81 23242.22 4753.08 4080.35 6179.36 6264.02 0.80 19131.77 22097.05 4261.28 3857.69 5233.84 5602.66 0.85 17552.15 21027.00 3908.65 3653.40 4541.00 5002.05 0.90 11686.03 19948.49 2596.70 3450.81 2945.54 4439.65 0.95 6577.27 18196.91 1455.49 3128.26 1629.52 3794.88 1.00 3316.06 13339.52 730.74 2249.41 813.37 2652.59 1.05 1607.33 8435.92 352.51 1386.84 388.43 1615.00 1.10 803.57 4893.26 175.60 785.11 190.41 906.42

Consumer surplus and social welfare. Table 1.3 also shows that consumer surplus and

social welfare decrease in the transportation rate. Intuitively, when transportation rate

increases, the firm can partially absorb the higher transportation cost and pass some of such

22Quantities demanded decrease in delivery prices, which are increasing in the transportation rate (see expressions 1.23 and 1.24).

22 cost to consumers (Martin, 2008; Grg et al., 2010; Baldwin and Harrigan, 2011), ultimately causing a loss in both producer and consumer surplus.

For values of t at or below 0.85, consumer surplus satisfies CSm > CSd, while the opposite ranking applies for higher values of t. Social welfare follows the same pattern, where Wm ≥ Wd for t ≤ 0.45, but Wm < Wd otherwise. Wm > Wd when transportation cost is low, because pd is higher at the city center tan pm even if discriminatory pricing captures more customers. It is also found that output increase is necessary but not sufficient for social welfare to rise under price discrimination, which is in line with Aguirre et al. (2010).

1.4.3 Effects of banning spatial price discrimination

Consider that price strategy is endogenous, allowing the monopoly to choose to adopt either mill pricing or spatial price discrimination. From our above analysis, discriminatory pricing always generates higher profits than mill pricing in Cases 1 to 4, implying that the monopolist would choose a discriminatory pricing strategy.

If the authority bans spatial price discrimination, what would happen to the firm’s output and social welfare? In Case 1 (2), Lemma 1 (2, respectively) indicates that prohibiting discriminatory pricing reduces output and harms social welfare. In Case 3, we also find that banning spatial price discrimination reduces output. However, it may be beneficial or detrimental to social welfare depending on the market expanding effect of discriminatory pricing.

In Case 4, we find that output is larger under spatial price discrimination. Social wel-

23 fare is higher (lower) under mill pricing than that under discriminatory pricing when the transportation cost per mile is low (high, respectively). Thus, for high transportation cost- s, banning discriminatory pricing in Case 4 decreases output and improves social welfare.

Therefore, banning discriminatory pricing reduces welfare if transportation costs are rela- tively high, consumers are not uniformly distributed, and location is more costly close to the city center.

1.5 Conclusions

Using a monopoly spatial model with normal population distribution of consumers, and endogenous market boundaries, the paper analyzes the effects of spatial price discrimination on the firm’s location choices, output and social welfare under both constant and non- constant locations costs along the market line. The main conclusions are the following.

First, when the location costs are constant over the market line, the monopolist locates at the median of the population distribution under both mill and discriminatory pricing.

We also find that the market area, profit and total output are larger under spatial price discrimination than under mill pricing. Intuitively, under spatial price discrimination, the

firm can attract distant consumers by lowering prices, which in turn helps it serve a larger market.

Second, when the location costs are non-constant along the market line, the monopolist locates at different places under mill pricing and spatial price discrimination. Relative to spatial price discrimination, when transportation rate is low, the firm locates closer to the

24 city center under mill pricing, but locates further away from the city center when such a cost increases.23

Spatial price discrimination is banned in many countries because the authorities consider that price discrimination is detrimental to social welfare. However, our findings suggest that spatial price discrimination may raise social welfare. For instance, our results suggest that, in industries with high transportation rate (such as ready-mixed concrete and cement), allowing spatial price discrimination can actually improve social welfare.24 Thus, a blanket prohibition of price discrimination is not socially desirable. As Cheung and Wang (1995) note, a selective regulatory policy would be preferred.

Our model can be extended in several directions. First, we could allow for more com- petitive market structures. Second, we assume that the monopolist produces at a single location. This assumption may be reasonable when the location setup cost is high, but could be relaxed if these costs are low, thus allowing for multiple plant locations. Third, this paper assumes that consumers’ preferences are homogeneous, but preferences could differ depending on consumers’ location due to differences in income or taste. Fourth, Aguirre et al. (2010) show how the curvature of demand determines the welfare effect of monopoly third-degree price discrimination when all markets are served. In our paper, we assume that the firm can endogenously choose the number of markets to serve; and we change the density

23We also find that spatial price discrimination results in a larger market area, higher profit, and larger out- put than mill pricing. However, compared with mill pricing, social welfare under spatial price discrimination is higher (lower) when transportation rate is low (high, respectively). 24While Miller and Osborne (2014) find that banning price discrimination would increase consumer surplus, they do not evaluate profit losses, and thus cannot conclude whether social welfare increases or decreases. Our results, hence, help identify under which contexts price discrimination has welfare improving effects.

25 of the consumers on the market space, which is analogous to changing demand curvature.

Thus, it would be interesting to investigate how demand curvature affects welfare under spatial price discrimination in a model that allows for endogenous market boundaries.

26 CHAPTER 2.

THE EFFECTS OF BEIJING’S VEHICLE LOTTERY SYSTEM

ON FLEET COMPOSITION AND ENVIRONMENT

To control vehicle growth and air pollution, Beijing’s municipal government imposed a vehicle lottery system in 2011, which randomly allocated a quota of licenses to potential buyers. This paper investigates the effect of this policy on fleet composition, fuel consump- tion, air pollution, and social welfare. Using car registration data, we estimate a random coefficient discrete choice model and conduct counterfactual analysis based on the estimated parameters. We find that the lottery reduced new passenger vehicle sales by 50.15%, fuel consumption by 48.69%, and pollutant emissions by 48.69% in 2012. Also, such lottery shift- ed new auto purchases towards high-end but less fuel efficient vehicles. In our counterfactual analysis, we show that a progressive tax scheme works better than the lottery system at decreasing fuel consumption and air pollution, and leads to a higher fleet fuel efficiency and less welfare loss.

2.1 Introduction

A steady increase in vehicle ownership and usage worldwide has been accompanied by tremendous increase in energy consumption, severe air pollution, and health concerns in many cities around the world, especially in developing countries (Unit, 2010; Xiao et al.,

2017). Examining the effectiveness of vehicle policies to reduce fuel consumption and air

27 pollution in emerging markets is increasingly important for at least three reasons. First, air pollution is a particularly acute problem in developing countries where it often exceeds the recommended health limit (Alpert et al., 2012; Greenstone and Hanna, 2014; Li, 2015).

High air pollution has adverse health consequences, such as cardiopulmonary diseases (EPA,

2004), premature deaths (Cropper, 2010; Chen et al., 2013; Wolff, 2014), and infant mor- tality (Chay and Greenstone, 2003; Currie and Neidell, 2005). Second, the fluctuation of world oil prices and instability in the Middle East raise concerns about energy security, es- pecially for some developing countries with surging energy demands.25 The third reason to investigate environmental regulations is that individual countries are responsible to reduce greenhouse gas emission (GHG),with the global increase in GHG emissions been mainly driven by emerging economies, such as China and India.26

China offers a fertile ground for exploring the effectiveness of vehicle environmental reg- ulations. First, China is facing very serious air pollution problems,27 which are particularly serious in big cities, such as Beijing. Twelve of the twenty most polluted cities in the world can be found in China (Bank, 2007). In addition, motor vehicles, specifically cars, are a major source of air pollution in China’s big cities (of Environmental Protection, 2011; Chen and Zhu, 2013; Viard and Fu, 2015). To control vehicle ownership and usage and to ad-

25In 2013, China crude oil consumption was 10.48 million barrels per day, accounting for about 12% of world crude oil consumption according to United State Energy Information Administration. 26In 2013, the United State’s CO2 emissions increased by 2.5% and the EU28’s CO2 emissions decreased by 1.4%, relative to 2012. Other OECD countries also significantly show decreases or minor increases below 2%. In contrast, relative to 2012, CO2 emissions in developing countries mainly increased in 2013, e.g., in China by 4.2%, in India by 4.4%, in Brazil by 6.2% (Agency, 2014). 27Zheng and Kahn (2013) provide a comprehensive review on China’s urban pollution and governments’ policies to address urban pollution externalities. Greenstone and Hanna (2014) demonstrate that ambient particulate matter concentrations in China are seven times the US level.

28 dress related environmental problems, both central and local governments have enacted and

enforced a wide range of policies, including vehicle consumption taxes, fuel taxes, public

transportation subsidies, tightening vehicle emission standards, driving restrictions, vehicle

quota systems, and subsidy schemes on new energy vehicles. As a result, investigating the

efficacy of such policies in emerging economies, such as China, provides insights on designing

environmental regulations to curb fuel consumption and air pollution.

In this paper, we focus on Beijing’s vehicle lottery system (VLS). In 2011, Beijing munici-

pal government imposed a vehicle quota system to control vehicle population growth. Unlike

the vehicle license auction system in Shanghai, about 20,000 new licenses are randomly allo-

cated through monthly non-transferable lotteries in Beijing. Qualified applicants can enter

the lottery at no cost. Only those who win the lottery have the right to register new vehicles

in Beijing.28 However, the effectiveness of the VLS has never been fully investigated. It is urgent to empirically quantify its environmental and welfare consequences, since this is a novel policy no other city has used, and thus is being closely observed and adopted by other large cities in the region.29

This paper investigates the effects of Beijing’s vehicle lottery system (VLS) on fleet

composition, gas consumption, pollutant emissions, and social welfare. To begin with, we

construct and estimate a random coefficient discrete choice model developed by Berry et al.

(1995) using car registration data for Beijing, and three more cities with similar characteris-

28Those residents who scrap or sell their existing cars do not need to enter the lottery. 29Guiyang adopted license lotteries in July 2011. Guangzhou, Tianjin, Hangzhou, and Shenzhen imple- mented a hybrid system combining lottories and auctions in July 2012, December 2013, March 2014, and December 2014, respectively. Other large Chinese cities (Chengdu, Chongqing, Qingdao, and Wuhan) have considered enacting similar control policies.

29 tics that did not implement the VLS, Nanjing, Shenzhen and Tianjin. The model incorpo- rates household preference heterogeneity and unobserved product attributes. To identify the effects of the VLS, we then simulate the outcomes under the counterfactual scenario of no policy and compare them with the observed facts. Moreover, to compare the effectiveness of the VLS with other policies, we conduct another two counterfactual experiments, in which we respectively replace the lottery system with a hypothetical first registration tax and a hypothetical consumption tax.30

Our study provides interesting findings. First, Beijing’s VLS successfully reduced new car sales by 50.15%, gas consumption by 48.69%, and pollutant emissions by 48.69% in

2012. However, the cost of these benefits is a social welfare loss, implying that such policy is costly. Second, the lottery policy changed the fleet composition. We find that the fleet skewed towards high-end and less fuel efficient vehicles. The empirical findings show that the sales-weighted average price of cars registered in Beijing in 2012 under the lottery system is about 42,430 Yuan (US$6,212) higher than under no policy. The fleet fuel efficiency is 8.09

L/100km under the lottery system, relative to 7.85 L/100km under no policy. Finally, our counterfactual analysis shows that progressive tax policies are as effective as Beijing’s VLS in controlling new vehicle sales but achieve better effects in reducing fuel consumption and air pollution. In our study, fuel consumption (or air pollution) were reduced by 56.67% under the hypothetical first registration tax system, by 55.01% under the hypothetical consumption tax

30The first registration tax was introduced in Hong Kong. On the first registration of a motor vehicle, a tax rate proportional to the vehicle class is charged on its taxable value (usually the retail price). In this paper, we revise the tax rates so that new vehicle sales coincide with those under the VLS. Based on the consumption tax in 2008, we construct a hypothetical consumption tax scheme so that it is as effective as the lottery system in controlling for new vehicle sales.

30 scheme, and but only by 48.69% under the VLS in 2012. In addition, these tax policies result in higher fleet fuel efficiency, which could improve from 8.09 liters/100km under the VLS to

7.36 liters/100km under the hypothetical first registration tax system and 7.48 liters/100km under the hypothetical consumption tax scheme. Moreover, welfare could increase by 69.31 billion Yuan (US$10.14 billion) or 79.83 billion Yuan (US$11.68 billion) in 2012 if Beijing were to replace the VLS with the first registration tax system or the consumption tax scheme, respectively.

Related literature. Our paper is related to studies by Xiao and Zhou (2013), Xiao et al. (2017) and Li (2015), who analyze vehicle quota systems in China. Xiao and Zhou

(2013) examine the impacts of Shanghai’s vehicle auction system on vehicle control, fleet efficiency, gas consumption and pollutant emissions. Further, Xiao et al. (2017) investigate the influence of the auction system on the market structure in Shanghai. Our work differs in two ways. First, this paper focuses on Beijing’s VLS which is a non-market based mechanism which allocates the quota through lottery, while Shanghai’s vehicle quota system allocates the license plates by auction, in which households with the highest willingness to pay get the quota. Second, we also provide a welfare measure of the performance of the VLS in

Beijing. Li (2015) finds that, compared with a uniform price auction, Beijing’s VLS led to a welfare loss of nearly 58 billion Yuan (about US$ 9.19 billion) in Beijing in 2012.31 Our paper differs in two respects. First, we mainly focus on the impact of Beijing’s VLS on the

31Under a uniform price auction, the government allocates Q licenses among N bidders, each submitting a one-dimensional bid to obtain a license. Upon receiving the list of bids, each of the Q highest bidders receives one license, and each pays a price equal to the highest rejected bid. For instance, consider an auction with 6 consumers and a quota of 3 licenses to allocate, where consumers 1, 2, 3, 4, 5, and 6 submit bids of $10, $9, $8, $7, $6, and $5, respectively. In this setting, the highest rejected bid is $7, entailing that consumers 1, 2, and 3 win a license, each of them paying $7.

31 automotive fleet’s composition and environment. Our study is the first paper to analyze the fleet composition effect of Beijing’s vehicle lottery system. Second, we also compare the environmental and welfare consequences of the lottery with those of other tax policies, showing that these tax policies are superior to VLS in both environmental benefits and welfare. By comparing our results with Li’s (2015), we find that tax policies work better than a uniform price auction for environmental purpose.

Recently, Yang et al. (2014) use hypothetical Beijing’s gross regional product (GRP) to predict the influence of the lottery system on vehicle growth and hence fuel consumption in 2020. Similarly, Li and Jones (2015) analyze this policy’s effects on vehicle population and CO2 emissions in 2020 based on hypothetical permanent population and GDP in Bei- jing. However, this literature does not consider that vehicle demand is mainly driven by household demographics and vehicle attributes, such as household income, vehicle price and performance.32 Unlike these studies, we derive vehicle demand from household preferences and their choices. Incorporating these features can yield better predictions about the effects of existing policies and generate welfare implications (Chetty, 2015).

Our study also adds to the empirical literature on non-price emission-reduction policies.

Most of the literature focuses on fuel tax (Parry and Small, 2005; Fullerton and Gan, 2005;

Bento et al., 2009; Xiao and Ju, 2014), consumption tax (Xiao and Ju, 2014), congestion fees and road pricing (Small et al., 2005; Eliasson et al., 2009; Gibson and Carnovale, 2015), driving restrictions (Davis, 2008; Gallego et al., 2013; Viard and Fu, 2015), and Low Emission

32Berry et al. (1995), Petrin (2002), Xiao and Ju (2014), and Li et al. (2015) prove the importance of these factors in vehicle demand.

32 Zones (Wolff and Perry, 2010; Wolff, 2014). However, few empirical studies to date have compared the effectiveness of market-based and non-market based policies, except Li (2015), who conducts a welfare comparison between Beijing’s lottery system and a uniform price auction. Economists have raised concerns over non-market based policies because behavioral responses could mitigate net policy benefits (Davis, 2008; Gibson and Carnovale, 2015). For example, households could respond to the vehicle lottery policy by concentrating their budget into a single high-end but less fuel efficient car rather than two low-end cars, which lowers

fleet fuel efficiency. Therefore, broader comparisons between the lottery system with other policy tools such as taxes have important policy implications. Our paper compares the lottery system with other two tax schemes and finds that a progressive tax system could be a good substitute for Beijing’s VLS.

2.2 Industry Background, Policy and Data Description

2.2.1 Industry Background

China’s automobile industry has developed rapidly since 2000. Vehicle population in- creased from 16.09 million units in 2000 to 62.09 million units in 2009 at an average annual rate of 14.5%.33 And the growth rate increased particularly rapidly over the last several years

(17% in 2008, 21.76% in 2009, and 24.36% in 2010). With new vehicle sales of 13.65 million and new vehicle production of 13.79 millions in 2009, China surpassed the U.S. market and

33The average growth rate of vehicle population in the United States from 2000 to 2009 was 1.19%.

33 became the largest auto market in the world in both sales and production.34

In 2004, the Chinese government released the new automotive industry development policy. It specifies that the minimum investment size for new entrants is 2 billion Yuan; the ownership of Chinese partners in the joint venture cannot be lower than 50%; and each foreign firm can form joint ventures with at most two Chinese companies. With these barriers, the number of auto makers has been relatively stable since then. In China, the top ten manufacturers of passenger vehicle accounted for 83.3% of sales in 2008 (Xiao and

Ju, 2014). Among them, joint ventures between local manufacturers and foreign carmakers, such as Shanghai Auto with General Motors and Volkswagen, Beijing Auto with Chrysler and Hyundai, take two thirds of the passenger vehicle market. The rest is taken up by indigenous-brand manufacturers, such Chery and Geely.

2.2.2 Policy Description

Over the last three decades, China’s economy has developed rapidly and household in- come has grown dramatically, especially in some big cities. In Beijing, gross domestic prod- uct (GDP) per capita increased from US $2,915 in 2000 to US $10,910 in 2010,35 increasing vehicle population from 1.04 million in 2000 to 4.50 million in 2010.

However, this fast growth in vehicle population led to traffic congestion and air pollution.

Beijing is often ranked as one of the most congested cities in the world. The average traffic

34In addition, according to China Vehicle Emission Control Annual Report (2010-2013), the percentage of passenger vehicles increased from 8.3% in 1990 to 78% in 2009, 79% in 2010, 80.7% in 2011, and 82.5% in 2012, implying a switch of purchasing from commercial to private purposes. 35Beijing Yearbook, 2013.

34 speed on arterial roads during morning peak hours was 20 km/h in 2010, compared with 30

km/h in 2003 (Center, 2004-2011). Moreover, Beijing was one of the world’s most air polluted

cities in 2013.36 According to the China’s Vehicle Emission Control Annual Report 2010,

emissions from vehicles is the main source of air pollution in China’s major cities. From

2010, Beijing’s average daily concentration of PM2.5 frequently reaches over 250 µg/m3, which is much higher than the recommended daily level of 25 µg/m3 by the World Health

Organization (2006).37

To reduce vehicle congestion and air pollution, Beijing’s municipal government has adopt- ed some policies, such as driving restrictions from July 20, 2008. Viard and Fu (2015) found that the restrictions reduced particulate matter pollution by 7%-19% in the short run. How- ever, some studies indicate that permanent driving restrictions do not successfully reduce air pollution or traffic congestion (Davis, 2008; de Grange and Troncoso, 2011), because households circumvent the restriction by buying second cars.

To further ease traffic congestion and improve air quality, Beijing municipal government issued a plan to control vehicle registrations on December 13, 2010. On December 23, 2010,

Beijing municipal government froze new registrations and announced that, from January

2011, before purchasing a vehicle, residents and corporations need to enter a publicly held lottery and win a license plate, which is necessary to register a vehicle. Each month, there are about 20,000 license plates to allocate, among which, about 88% (or 17,600) for private

36Seven of the world’s 10 most polluted cities are in China, January 14, 2013. 37Air pollution is also linked to some extreme conditions, such as cardiopulmonary diseases, respiratory infections, lung cancer (EPA, 2004), infant mortality (Chay and Greenstone, 2003), and childhood asthma (Neidell, 2004).

35 vehicles and the rest for institutions. The licenses are allocated through random drawings under a monthly lottery-style quota system for private applicants and every two months for businesses.

The lotteries for private licenses are held on the 26th day of each month. Licenses are needed for first-time buyers, second-hand vehicle buyers, and those who accept gifted vehicles or transfer out-of-state registration to Beijing. Those, who destroy, sell, or trade in their existing cars, can retain their license plates to register new vehicles. The eligible participants include Beijing residents and non-residents with temporary residence permits who have been paying social insurance and income tax for at least five years in Beijing. Individuals who have registered vehicles cannot enter the lottery. However, if a household with a car has a second driver, this driver can enter the lottery. To enter the lottery, applicants can fill forms on a government website or apply at a walk-in service center without cost.

Beijing’s Municipal Commission of Transport publishes the lottery results on the lottery system’s website. Each winner can download a certificate online or pick up it at a walk-in service center. The certificate allows the quota holder to purchase a license plate and register a vehicle. The licenses cannot be transferred or sold. Each quota is valid for six months.

If a lottery winner does not register a vehicle during this period, the license will be added to the pool of quotas in the next lottery. Those who allow their quotas to expire cannot participate in the lottery within the next three years.

To strictly enforce the vehicle lottery, additional policies are issued to prevent Beijing residents from registering vehicles in nearby cities and driving in Beijing. Out-of-state vehi-

36 cles need to obtain temporary driving permits to enter the 5th ring road.38 Moreover, these vehicles are banned to travel within the 5th ring road (inclusively) during peak hours.

2.2.3 Data

Data Description

This paper focuses on the effects of vehicle lottery policy in Beijing based on data from

2009 to 2012. To control the effects of other factors, such as trends and tax deduction, we choose Nanjing, Shenzhen, and Tianjin to facilitate identification.39 These cities are the largest cities and they did not have policies on vehicle ownership or vehicle usage during our sample period. The characteristics of these four cities are shown in Appendix Table

0.1.40 Appendix Table 0.1 shows that Shenzhen has the highest average household income,

GDP per capita, and average consumption expenditure per capita, while Tianjin has the lowest. Beijing approximates Nanjing in these dimensions. Generally, these cities are similar in average household income, GDP per capita, and average consumption expenditure per capita. In addition, Li (2015) proves common trend of vehicle sales in Tianjin, Nanjing, and

Beijing.

There are two main data sources for this study. The first data set contains monthly new passenger vehicle registration information in each city from January 2009 to December

38The 5th ring road is about 98.58 km in length and the area within it is about 700 km2. 39In 2009 and 2010, the sales tax was reduced to 5 and 7.5 percent for vehicles with engine displacement no more than 1.6 liter, respectively. From 2011, these tax deductions were canceled. 40In mainland China, Beijing is the second largest city in population, Shenzhen is the fourth, Tianjin is the fifth, and Nanjing is the twelfth.

37 2012, including manufacturer, brand, model year, model, engine displacement, car type, and quantity.41 In this paper, we focus on passenger vehicles and a product is defined as a unique combination of the model year, manufacturer, brand, model, engine displacement, and car type. For example, 2009 Beijing Benz C200 1.8T Sedan, 2010 Beijing Benz C200

1.8T Sedan, and 2009 Beijing Benz C230 2.5L Sedan are different products. In addition, we define a market as a city-year-quarter combination. For example, Beijing-2010-Q1 is a different market from Beijing-2010-Q2. Therefore, we aggregate the monthly data into quarterly levels and use the total sales and average quarterly prices for each quarter to measure their sales and prices. There are 41,006 observations in our sample.

Figure 2.1: New Vehicle Quarterly Sales in Beijing, Nanjing, Shenzhen, and Tianjin

41In China, vehicle registration data are not released to the public. We obtain the data from Dalian Wismar Information Co., Ltd. The data provider required us not to release the data to protect his proprietary information. We adopt the term “sales” to substitute for “the number of newly registered passenger vehicles”. The number of registered vehicles in a city is different from the vehicle sales in that city. For example, consumers in nearby cities may buy vehicles in Beijing and register cars in their own cities. However, vehicle sales in this paper refers in particular to the number of registered vehicles.

38 Figure 3.7 plots quarterly sales of new vehicles in Beijing, Nanjing, Shenzhen, and Tianjin.

Figure 3.7 shows that the lines track each other well before 2011, reflecting a common trend

across these cities. In addition, we can find that there was strong seasonal effects, whereby

sales increase at the end of each year. A possible explanation is that people receive their year-

end bonuses, enabling them to purchase big-items such as cars. Finally, new vehicle sales in

Beijing increase dramatically in the fourth quarter of 2010 and then decrease sharply in the

first quarter of 2011. Since Beijing municipal government issued a plan to control vehicle

registrations on December 13, 2010, consumers may have been afraid to not purchase vehicles

after the quota, moving their purchases into December 2010. To avoid this anticipation effect,

we drop the last quarter in 2010 and the first quarter in 2011 across all cities in our study.42

To complete the data set, we collect vehicle attribute data from the website auto.sohu.com and Car Market Guide. Transaction prices are not available. Following the auto demand literature, we use Manufacturer Suggested Retail Prices (MSRP) in this study. Vehicle prices are computed based on MSRP and the sales tax. Li et al. (2015) argue that, unlike U.S. auto market, promotions in China’s auto markets are not frequent and, hence, MSRP can be good proxy for retail prices. The sales tax is 10% in all of these cities except in 2009 and 2010. In 2009 and 2010, the sales tax was reduced to 5 and 7.5 percent for vehicles with engine displacement below 1.6 liter, respectively. From 2011, these tax deductions were cancelled. The data set also includes horsepower (in kilowatts), car weight (in 1,000

42When the policy was announced, it was possible that consumers in other cities were affected, anticipating that similar policies will be applied in their own cities. As a consequence, we drop the last quarter in 2010 and the first quarter in 2011 in control cities in our main analysis. As a robustness check, we also estimate the model without dropping the data during this period in control cities in section 2.4.2.

39 kilograms), vehicle size (in m2), fuel efficiency (in Liters/100km), and engine displacement

(in Liters).43 In addition, we also obtain gasoline prices from National Development and

Reform Commission of China to construct a fuel economy variable, i.e., kilometers driven for 1 RMB Yuan’s gasoline. All prices are in 2009 RMB Yuan. We use the Consumer Price

Index to deflate.44

Table 3.4 provides the summary statistics of our sample. Both quarterly sales and prices have large variations. The most popular passenger car has a quarter sale of 6,620 units whereas the least popular car had only one sale. The average price is 188,570 Yuan, ranging from 26,040 to 11,243,500 Yuan. The average price is higher than the average household income. Horsepower and engine displacement are correlated with vehicle performance. The most powerful vehicle has a horsepower of 515 kw, while the least has a horsepower of

26.5 kw. Engine displacement varies from 0.8L to 6.8L. Vehicle weight and vehicle size are indicators of comfort and safety (Winston and Yan, 2016). The lightest car weighs 650 kg, whereas the heaviest one weighs 2,950 kg. The largest vehicle is 12.51 m2, while the smallest one is 4.20 m2. Fuel efficiency and fuel economy are used to measure vehicle fuel consumption performance. The least fuel efficient vehicle consumes 26 liters of gasoline for a 100 kilometers’ drive, while the most fuel efficient one consumes only 2.7 liters.

The second data set is the household income distribution in each city and year, which is constructed through Chinese Household Income Survey (2007) and annual statistical year- books of each city.45 Following Li et al. (2015), we use their method to construct household

43Fuel efficiency data are obtained from Ministry of Industry and Information Technology of China. 44Consumer Price index are from National Bureau of Statistics of the People’s Republic of China. 45Chinese Household Income Survey is a national representative survey conducted by University of Michi-

40 Table 2.1: Summary Statistics of Vehicle Data 2009-2012

Variables Mean S.D. Min Max Quarterly sales by product 110.49 292.20 1 6,620 Price (1,000 Yuan) 188.57 225.14 26.04 11,243.50 Weight (1,000 kg) 1.37 0.29 0.65 2.95 Horsepower (kw) 101.88 37.53 26.50 515.00 Vehicle size (m2) 8.03 0.91 4.20 12.51 Fuel efficiency (L/100km) 7.94 1.50 2.70 26.00 Fuel economy (km/Yuan) 2.17 0.40 0.59 7.19 Engine displacement (L) 1.86 0.56 0.80 6.80 Note: All money is in 2009 RMB Yuan. The number of observations is 41,006. The mean of price, weight, horsepower, vehicle size, fuel efficiency, fuel economy, and displacement are sales-weighted means. income distribution from 2009 to 2012 in each city. First, we obtain average household in- come for five income levels or seven income levels from the yearbooks.46 Second, we divide

5,000 observations in the survey into five income levels or seven income levels. Finally, we adjust the household income in the survey proportionally and separately for each income level. After adjustment, the interpolated income distribution from the survey in a given year and city is consistent with income statistics from the yearbook of that city and year.

Following the literature, we assume that the logarithm of household income follows the nor- mal distribution. Since Chinese Household Income Survey (2007) is a national survey, the derived household income distributions may be different from income distributions of vehicle gan and China Institute for Income Distribution. More details can be found on its website. 46In some cities, households are divided into five income levels: low income households (first quintile), medium-low income households (second quintile), medium income households (third quintile), medium-high income households (fourth quintile), and high income households (fifth quintile). Some other cities use seven income levels: lowest income households (first decile), low income households (second decile), medium- low income households (second quintile), medium income households (third quintile), medium-high income households (fourth quintile), high income households (ninth decile), and highest income households (tenth decile).

41 buyers. Hence, we use a survey conducted among vehicle owners in Beijing by the Guanghua

School of Management at Beijing University in 2005 as a robustness check.47

Stylized Facts

The above summary statistics do not show the changes of vehicle characteristics across the cities over time. Figure 2.2 to Figure 2.4 display the quarterly average prices, horsepower, and fuel efficiency, respectively. As shown in Figure 2.2, before 2011, Shenzhen has the highest sales-weighted average price, followed by Beijing, while Tianjin has the lowest.48

However, the sales-weighted average price in Beijing increases from 184,840 Yuan before

January 2011 to 240,410 Yuan after January 2011, representing a 30.06% increase. After the policy was announced, Beijing ranks first in sales-weighted average price. Specifically, during the post policy period, the sales-weighted average price of Beijing is about 13.02% higher than that of Shenzhen, 36.02% higher than that of Nanjing, and 66.66% higher than that of Tianjin. From Figure 2.3 and 2.4, we can find that sales-weighted average horsepower and fuel efficiency follow similar patterns as sales-weighted average prices. These stylized facts suggest that Beijing’s VLS may make buyers switch to high-end, more powerful, but less fuel efficient cars. Moreover, Figure 2.2 to 2.4 indicate that the sales-weighted average price, horsepower, and fuel efficiency are not in steady state in 2011. So we focus on 2012 in our counterfactual analysis.

However, the changes mentioned above could be caused by other factors, such as house-

47Summary statistics of this survey can be found in study by Xiao and Ju (2014). 48In particular, the sales-weighted average price of Beijing is about 8.23% lower than that of Shenzhen, 13.59% higher than that of Nanjing, and 39.21% higher than that of Tianjin.

42 Figure 2.2: Quarterly Sales-weighted Average Prices (1,000 Yuan) 2009-2012 in Four Cities

Figure 2.3: Quarterly Sales-weighted Average Horsepower (kw) 2009-2012 in Four Cities hold income. Our analysis employs a random coefficient discrete choice model to control these factors and identify the effects of the vehicle lottery on fleet composition, fuel consumption, and pollutant emissions in Beijing.

43 Figure 2.4: Quarterly Sales-weighted Average Fuel Efficiency (L/100km) 2009-2012 in Four Cities

2.3 Empirical Model and Estimation

2.3.1 Utility Function Specification

Our objective is to investigate the effects of Beijing vehicle lottery. We set up and

estimate a random-coefficient discrete choice model of automobile oligopoly in the spirit of

Berry et al. (1995).

In our analysis, a market t is defined as a city-year-quarter combination, such as Beijing-

2010-Q1. In a market, a product j is defined as a unique combination of the model year, manufacturer, brand, model, engine displacement, and car type, for example, 2009 Beijing

Benz C200 1.8T Sedan. Consider a set of markets, t = 1, ..., T , where a set of products, j = 0, 1, ..., J, is available for each market. We use 0 to denote the outside good (i.e.,

44 the choice of purchasing an electric vehicle or not buying a new vehicle).49 Let i denote a

household. The utility from outside good is normalized to εi0t, which follows i.i.d. type I

extreme value distribution, as in Berry et al. (1995) and Li (2015). In market t, the indirect utility of household i when purchasing product j is given by

uijt =xjtβi + αi ln pjt + λtypedtypej + λf dfirmj + λqdquarter (2.1)

+ λydyear + λcdcity + ξjt + εijt where xjt is a vector of product j’s observed characteristics in market t, including a constant term, logarithm of horsepower, vehicle weight, kilometers driven per Yuan of gasoline (fuel economy), vehicle size, and engine displacement. pjt is the price of product j in market t. dtypej is a vector of car type dummy variables (e.g., sedan, SUV, MPV, station wagon, and coupe)and dfirmj is a vector of firm dummy variables, capturing households’ intrinsic preference for vehicle types and products from different firms. The model also includes quarter dummies, year dummies, and city dummies to capture market-specific effects, such as seasonal effects and city fixed effects. ξjt is the unobserved (to researchers) characteristics of product j in market t, such as product quality. εijt is an independently and identically distributed (across products, households, and markets) idiosyncratic shock that is drawn from the type I extreme value distribution.

Households are heterogenous in their tastes for price and other characteristics. For ex- ample, households with higher income are less price sensitive. Household heterogeneity is captured by random coefficients αi and βi. In particular, αi is household i’s marginal utility

49The electric vehicles accounted for only 0.044% of all newly registered passenger vehicles. Customers of electric vehicle don’t need to enter the lottery to get a license.

45 from income and it is given by

p αi =α ¯ + η ln yi + σpvi (2.2)

p where yi is household income, and vi is unobserved household characteristics that affect household preferences and follow standard normal distribution. Since households with a higher income yi tend to be less price sensitive, we expect η to be positive.

Similarly, βik measures household-specific taste on vehicle characteristic xjtk, which is the kth attribute of product j. Specifically, βik is defined by

¯ k βik = βk + σkvi (2.3)

¯ k k where βk is the average preference across households and σkvi captures random tastes. vi is assumed to have a standard normal distribution.

After combining equations (2.1), (2.2), and (2.3), we obtain

K X ¯ uij = xjkβk +α ¯ ln pj + λtypedtypej + λf dfirmj + λqdquarter k=1 (2.4) K X k p + λydyear + λcdcity + ξj + σkxjkvi + (η ln yi + σpvi ) ln pj + εij k=1 where, for compactness, we suppress the market index t. As in Berry et al. (1995), the above utility function can be decomposed into a mean utility K X ¯ δj = xjkβk +α ¯ ln pj + λtypedtypej + λf dfirmj k=1 (2.5)

+ λqdquarter + λydyear + λcdcity + ξj a household-specific utility (i.e., deviation from the mean utility)

K X k p µij = σkxjkvi + (η ln yi + σpvi ) ln pj (2.6) k=1 and a random taste shock εij.

46 2.3.2 Choice Probability and Aggregate Demand

Household chooses the product that maximizes his utility. Given that εi0t and εijt follow

the i.i.d. type I extreme value distribution, the probability of household i to purchase product

j in market t is exp(δ + µ ) s (p ,Xd, ξ , θ|v , y ) = jt ijt (2.7) ijt t t t i i PJ 1 + m=1 exp(δmt + µimt)

0 d where pt = (p1t, ..., pJt) and Xt includes xjt, car type dummies, firm dummies, quarter dummies, year dummies, and city dummies of the products. θ are the model parameters,

¯ 0 0 0 where θ1 = (¯α, β, λ) , θ2 = (σ, η) , and θ = (θ1, θ2) . Correspondingly, the market share for product j in market t is

Z exp(δ + µ ) s (p ,Xd, ξ , θ) = jt ijt dP (y)dP (v) (2.8) jt t t t PJ 1 + m=1 exp(δmt + µimt)

where P (·) denotes population distribution functions.

d If Nt is the market size in market t, the demand for product j is Ntsjt(pt,Xt , ξt, θ).

Following the literature (Berry et al., 1995; StevenBerry et al., 2004), the measure of market

size is the number of households in the city in a given year.

2.3.3 Identification and Estimation

After the idiosyncratic error term εijt is integrated out analytically, the econometric error term will be the unobserved product characteristics, ξjt, such as prestige and product quality. Prices could be correlated with these product characteristics. For example, vehicles with higher quality generally have higher prices. To address the price endogeneity problem

47 and estimate the parameters in equation (2.1), we employ the GMM estimation method proposed by Berry et al. (1995), which uses the moment condition

E(ξjt|zjt) = 0 (2.9)

where zjt is a vector of instrumental variables described below.

To derive ξjt, we first need to estimate market shares. While the market share in equation

(2.8) does not have a closed form, it can be evaluated by Monte Carlo simulation with ns draws from the distributions of v and y.50 The simulated market shares are calculated as

ns 1 X exp(δjt + µijt) spred(p ,Xd, ξ , θ) = (2.10) jt t t t ns PJ i=1 1 + m=1 exp(δmt + µimt)

Next, we combine the simulated market shares (3.12) with the observed market shares to

0 solve for the mean utility levels δt = (δ1t, ..., δJt) . Theoretically, the vector of mean utilities

δt can be retrieved by equating the estimated market shares with the observed market shares from the data for a given θ2:

obs pred d st = st (pt,Xt , δt; θ2) (2.11)

However, analytical solutions for δt are not available because the system of equations in equation (3.13) is highly nonlinear. In practice, it can be solved numerically by using the contraction mapping proposed by Berry et al. (1995) as follows51

h+1 h obs pred d h δt = δt + ln st − ln st (pt,Xt , δt ; θ2) (2.12)

50To increase computation efficiency and reduce the simulation error, we use Halton sequences to generate the random draws (see Train (2009) for use of Halton sequences). Our results are all based on 150 households in each market. We also checked ns = 250 using the benchmark specification. We found that it made little difference. 51See Berry et al. (1995) for a proof of convergence.

48 h h+1 until the stopping rule ||δt − δt || ≤ in is satisfied, where in is the inner-loop tolerance

−14 52 level. In our analysis, we set in = 10 . Once we find δt, the unobservable attributes ξjt can be solved as

d obs d ξjt(pt,Xt , st , θ) = δjt − (ln pjt,Xjt)θ1 (2.13)

The parameters θ1 in equation (2.13) can be estimated by two-stage least squares (2SLS)

d using instrumental variables (IVs). The demand unobservable ξjt is a function of prices, Xjt, the observed market shares, and parameters. The GMM estimator θˆ solves the problem:

min Q(θ) = min(ξ(θ)0Z)W (Z0ξ(θ)) (2.14) θ θ where W is the weighting matrix. The convergence criterion for the GMM is 10−8.

To address the price endogeneity problem, we need a set of exogenous instrumental variables. Following the literature, we assume that the unobserved product attributes are mean independent of observed product characteristics. Based on this assumption, we use three sets of instrumental variables in our analysis: the observed product characteristics (i.e., constant term, horsepower, vehicle weight, kilometers driven per Yuan of gasoline, vehicle size, and engine displacement), the sum of corresponding characteristics of other products offered by that firm (if the firm produces more than one product), and the sum of the same characteristics of products produced by rival firms. Berry et al. (1995) and Nevo (2000) show that the above instrumental variables are valid for cars and cereals, respectively. We also evaluate the strength of the instruments and the instrument F -statistic in our study is

52See Dube et al. (2012) for the discussion of the importance of a stringent convergence rule. They also provide a new computational algorithm for implementing the BLP estimator, called mathematical program with equilibrium constraints (MPEC). It converges faster than the algorithm that we used here.

49 large, with a p-value almost 0. So the instruments are strong.

2.4 Estimation Results

In this section, we first present parameter estimates for the random coefficient discrete choice model. In the benchmark specification, we estimate the model without using the data from the fourth quarter in 2010 to the first quarter in 2011 across four cities and the post-policy data in Beijing (2011-2012). We also report estimation results from alternative specifications.

2.4.1 Parameter Estimates from the Structural Demand Model

The results of the estimation are presented in Table 3.5. The first panel of the table provides the estimates of the parameters in the mean utility function defined by equation

(3.6). The parameters in the second panel are the estimates of standard deviations of the taste distribution of each attribute. The third panel provides the estimate of the coefficient of the interaction between ln price and ln income.

In the benchmark specification, all coefficients of vehicle attributes in the mean utility function are with the expected signs. The results suggest that households prefer powerful but fuel efficient cars. Vehicles with larger weight and size are more popular, because they are more comfortable and safer. Our findings are consistent with most previous research (Berry et al., 1995; Petrin, 2002; Deng and Ma, 2010; Xiao and Ju, 2014; Li et al., 2015). Moreover,

50 Table 2.2: Estimation Results for the Model

Benchmark Alternative 1 Alternative 2 Variables Coef. S.E. Coef. S.E. Coef. S.E.

Parameters in the mean utility (θ1) Constant -7.4147∗∗∗ 2.7315 -7.3137∗∗∗ 1.1184 -6.6066∗∗∗ 1.3264 Ln(price) -10.2256∗∗∗ 1.7085 -10.3526∗∗∗ 1.6944 -10.5855∗∗∗ 2.8825 Ln(horsepower) 5.0841∗∗ 2.1023 4.7895∗∗∗ 1.5699 4.5350∗∗ 1.8458 Weight 1.8792 1.5866 1.8727 1.3649 2.6624 2.2185 Fuel economy (km/Yuan) 0.1148 0.4457 0.3219 0.4399 0.1674 1.1974 Vehicle size 1.0582∗∗∗ 0.1185 1.0319∗∗∗ 0.1204 1.0476∗∗∗ 0.1465 Displacement 0.6576 0.4853 0.6568∗∗ 0.3279 0.6956∗∗∗ 0.2728

51 Random coefficients (σ) Constant -0.6035 3.2093 -0.0614 0.9824 -0.1223 1.2873 Ln(price) 0.3403 0.3077 0.3665 0.3046 0.3828∗∗∗ 0.1184 Ln(horsepower) 0.4893 0.7227 0.5354∗∗∗ 0.1753 0.7461∗∗∗ 0.1517 Weight 1.0075∗ 0.5277 1.0043∗∗ 0.4463 0.6876 1.0026 Fuel economy (km/Yuan) 0.5178∗∗ 0.1969 0.3773 0.3307 0.5555 0.8956 Vehicle size 0.0035 0.1234 0.0007 0.0609 0.0679 0.0966 Displacement -0.0056 0.9354 0.0023 0.3151 0.0168 0.3821

Interactions with Ln(income) (η) Ln(price) 0.3141∗∗ 0.1488 0.3032∗∗ 0.1360 0.3446∗∗∗ 0.1235 Note: The benchmark model is the preferred model. Alternative specification 1 uses household income distributions derived from a survey conducted among vehicle owners by Beijing University. Alternative specification 2 includes the data from the fourth quarter in 2010 to the first quarter in 2011 in Nanjing, Shenzhen, and Tianjin. All . specifications include car type fixed effects, firm fixed effects, quarter fixed effects, year fixed effects, and city fixed effects. ∗∗∗ significant at 1%; ∗∗ significant at 5%; ∗ significant at 10% the results also imply that households prefer vehicles with larger engine displacement. In the Chinese automotive market, engine displacement is usually correlated with whether a vehicle is high-end or low-end (Deng and Ma, 2010).53 The estimates of vehicle weight, fuel economy and engine displacement are not significant in our benchmark specification.

In the second panel, the estimates for idiosyncratic tastes over price, horsepower, vehicle size and engine displacement are insignificant. This implies that households are rather homogeneous in their preferences on these vehicle attributes. This finding coincides with

Xiao and Ju (2014). However, households do show variation in their preferences on vehicle weight and fuel economy. This adds to the literature on consumer heterogeneity in preference over vehicles.

The coefficient on ln price is negative and significantly different from zero. In addition, the estimate for the standard deviation on the tastes for ln price is not significant. These results indicate that consumers’ preference on vehicle prices is relatively homogeneous and that households dislike high vehicle prices. The estimate on the interaction between ln price and ln income is positive and statistically significant, adding to the literature on household heterogeneity. This suggests that households with higher income are less price sensitive.

With the estimated parameters, we compute price elasticity for each product. The av- erage own-price elasticity is -8.43. The average own-price elasticity is smaller in magnitude than that obtained by Deng and Ma (2010) which is -9.2. The difference can be explained by the increase in household income that occurred from Deng and Ma’s study, which used 1995-

53For example, cars with smaller displacement, such as Alto and Jeely, fall in the low-end category, while cars with larger engine displacement, such Cherokee and Redflag, belong to high-end and luxurious vehicles. Hence, it is intuitive that consumers like high-end cars.

52 2011 data, to our study, using 2009-2012 data. As household income grows, it is reasonable that consumers are less price sensitive to vehicle price changes than before.

2.4.2 Robustness Checks

To verify the sensitivity of our results to other specifications, we perform several robust- ness checks to our empirical analysis. Table 3.5 presents the estimation results of alternative specifications 1 and 2.

Household income. In our benchmark specification, the household income distributions are derived from Chinese Household Income Survey (2007), which is a national representa- tive survey. Hence, the derived household income distributions may be different from the income distributions of vehicle buyers. For example, the average monthly household income is 4,395 Yuan in Beijing 2005, while the average household income of vehicle owners is 8,300

Yuan per month (Xiao and Ju, 2014). In alternative 1, we use data from a survey conduct- ed among vehicle owners in Beijing by Beijing University in 2005 to derive the household income distributions. Generally, the coefficients remain similar to those in the benchmark specification. The average own-price elasticity is -8.20, which is about 2.73% smaller than that from the benchmark specification in magnitude. Since the average household income of vehicle owners is higher than that of the masses, the results suggest that higher income re- duces price sensitivity. As shown in Table 3.5, the estimate of engine displacement becomes significant, implying that high income enables households to purchase high-end cars.

Announcement effect. To avoid the announcement effect of the policy on the control

53 cities, the benchmark specification dropped the data from the fourth quarter in 2010 to the first quarter in 2011 across the control cities. Alternative specification 2 includes these data. Compared with the benchmark model, the coefficient of engine displacement becomes significant and households show variation in tastes for vehicle prices and horsepower. These results may be caused by the announcement effect of the policy. As the policy was announced in Beijing, it could cause panic among households in other cities, who advanced their future purchases. As a consequence, there are two groups of buyers in the market. One contains those who will buy cars even if Beijing did not implemented VLS. The other group contains buyers who enter the market due to Beijing’s VLS. Since consumers in different groups may have different characteristics, we expect a change in consumer’s preference heterogeneity.

2.4.3 Impact on New Vehicle Registration

In this subsection, we examine the effect of Beijing’s VLS on new vehicle registration in

Beijing. With the estimates under benchmark specification, we simulate the quarterly num- ber of newly registered passenger vehicles for each product in Beijing under counterfactual scenario (I) of no policy. As shown in Table 2.3, the counterfactual number is 213,638 in the fourth quarter of 2010, whereas the observed number is 257,489. This increase could be mainly caused by panic buying. With the announcement of license plate restriction in Bei- jing, consumers brought forward their demand for vehicles. Moreover, the new registrations under counterfactual scenario (I) are 757,042 and 1,066,551 in 2011 and 2012, respectively.

Compared with observed registrations in 2011 and 2012 under the policy, the lottery reduced

54 new vehicle registrations in Beijing by 54.84% in 2011 and 50.15% in 2012. This suggests that the policy successfully controlled vehicle growth in Beijing.

Table 2.3: Policy Impact on New Passenger Cars Registration in Beijing

Year and quarter Observed (with vehicle Counterfactual scenario (I)

lottery system) (without lottery system) 2009 548,557 548,557 2010Q1-2010Q3 501,814 501,814 2010Q4 257,489 213,638 2011 341,917 757,042 2012 531,639 1,066,551 Note: The counterfactual outcome is based on benchmark specification.

Our estimates are close to those obtained by Li (2015), where he finds the policy reduced sales by 57% and 54% in 2011 and 2012, respectively. Relative to Li’s (2015) study, the estimated percentage changes in sales were about 2.16 (3.85) lower in 2011 (2012) due to

VLS in our study, respectively. The differences could be attributed to the following reason.

He uses vehicle sales in his analysis, whereas we use vehicle registration data in this paper.

Usually, the number of vehicle sold in a city is not equal to the number of vehicle registered.

For example, given the abundance of dealerships in Beijing, consumers in nearby cities might buy cars in Beijing and register the vehicles in their cities. Since the VLS only applied to these vehicles registered in Beijing, it is more reasonable to exclude those vehicles sold in

Beijing but registered in other cities.

55 2.5 Counterfactual Analysis

The purpose of this section is to evaluate the effects of Beijing’s VLS on fleet composition,

fuel consumption, and pollutant emissions. Moreover, we compare welfare consequences of

vehicle lottery system with other tax-based policies. To examine the effects of the lottery

system, we use the estimates under benchmark specification to simulate market outcomes un-

der counterfactual scenario (I) of no policy and compare the results with observed outcomes

under the lottery system. Scenario (I) will be used as contrast.54

Although vehicle lottery can effectively control vehicle population and reduce gas con- sumption and air pollution in Beijing, it has been subject to criticism (Yang et al., 2014;

Li and Jones, 2015; Li, 2015). In a lottery system, the permits to purchase vehicles are distributed randomly. Those without the highest willingness to pay may get the cars, which in turn causes misallocation and welfare loss.

In this section, we compare the VLS against two alternative market-based policies, which aim at vehicle control. In counterfactual scenario (II), we remove the vehicle lottery and simulate the effect of a “first registration tax” as explained below; and in scenario (III), we replace the vehicle lottery policy with a revised consumption tax scheme. In Hong Kong, the first registration tax system was introduced to encourage the use of environment-friendly petrol private cars with low emissions and high fuel efficiency, which effectively controlled the growth of private cars (Tang and Lo, 2008). Based on the tax system in Hong Kong, we revise

54As shown in Figure 2.2 to 2.4, the sales-weighted average prices, horsepower, and fuel efficiency are not in steady state in 2011. So we conduct the counterfactual analysis using the data in 2012 in Beijing.

56 the tax rates used in our counterfactual scenario (II) so that it is as effective as the vehicle lottery in controlling vehicle sales. The first registration tax systems in Hong Kong and counterfactual scenario (II) are listed in Table 2.4. Similarly, China imposed a consumption tax on vehicles to reduce pollution and save energy.55 Based on the consumption tax rates in 2008, we adjust the tax rates in counterfactual scenario (III) to achieve the same effect as the VLS in controlling vehicle growth in Beijing. Both the current consumption tax rates and the consumption tax rates in counterfactual scenario (III) are shown in Table 2.5. The vehicle sales in Beijing (2011Q2-2012Q4) are 814,708 under scenario (II) and 816,160 under scenario (III), relative to observed sales of 816,001 under the VLS.

In this section, null scenario stands for the observed facts under the VLS; counterfactual scenario (I) stands for a setting with no policy (benchmark); counterfactual scenario (II) stands for the case where we replace the VLS with the hypothetical first registration tax in in Table 2.4; and counterfactual scenario (III) reflects a setting where we replace the VLS with the hypothetical consumption tax in Table 2.5.

2.5.1 Impact on Fleet Composition

With the estimates under the benchmark specification, we simulate the demand of each product in Beijing in 2012 under different scenarios. To estimate the impact of Beijing’s

VLS on fleet composition, we first summarize price, horsepower, and fuel efficiency under null scenario and counterfactual scenario (I) of no policy. Then we depict car price dis-

55In China, three categories of taxes are imposed on a car: the consumption tax, value-added tax (VAT), and vehicle purchase tax; see Xiao and Ju (2014) for more details about these taxes.

57 Table 2.4: First Registration Tax Rate in Hong Kong and Counterfactual Scenario (II)

Tax rates in Hong Kong Tax rates in counterfactual scenario (II) Class of motor vehicle Tax rate Class of motor vehicle Tax rate a. on the first HKD 150,000 40% a. on the first RMB 132,000 8.6% b. on the next HKD 150,000 75% b. on the next RMB 132,000 16% c. on the next HKD 200,000 100% c. on the next RMB 176,000 22% b. on the remainder 115% d. on the remainder 24% Note: Tax rates in Hong Kong come from Source: The Transport Department, The Government of the Hong Kong Special Administrative Region. All money is nominal.

Table 2.5: Consumption Tax Rates in China and Counterfactual Scenario (III)

Engine displacement (Liter) Observed tax rates Tax rates in counterfactual

scenario (III) ≤ 1.0 1% 5.25% 1.0-1.5 3% 11.50% 1.5-2.0 5% 17.75% 2.0-2.5 9% 26.00% 2.5-3.0 12% 33.25% 3.0-4.0 25% 50.50% > 4.0 40% 69.75% Note: The observed consumption tax rates come from Ministry of Finance of the People’s Republic of China.

58 tribution, horsepower distribution, and fuel efficiency distribution under the lottery system and counterfactual scenario (I) of no policy. We compare the distributions under these two scenarios to identify the changes in fleet.

Table 2.6 compares the price, horsepower, and fuel efficiency of the cohort of new pas- senger vehicles registered in Beijing in 2012 under null scenario and counterfactual scenario

(I). As can be seen in Table 2.6, there are significant differences between counterfactual sce- nario (I) and null scenario in price, horsepower, and fuel efficiency. Our results indicate that the VLS caused an increase in the sales-weighted average price in Beijing in 2012 by about

42,430 Yuan, an increase of approximately 22.40%. We also find that the sales-weighted av- erage horsepower under lottery system is about 9.01% higher than that under counterfactual scenario (I) of no policy. Moreover, the fleet becomes less fuel efficient. The sales-weighted average fuel efficiency of new vehicles registered under the lottery is 8.09 liters/100km, rel- ative to 7.85 under no policy. These results are consistent with our discussion above: the lottery system will make car buyers in Beijing switch to high-end and more powerful but less fuel efficient vehicles.

With the derived demand for each product and the total sales, we calculate the cumu- lative density function (CDF) of car prices, horsepower, and fuel efficiency, and plot their distributions under counterfactual scenario (I) and null scenario. We depict the price dis- tribution, horsepower distribution, and fuel efficiency distribution in Figure 2.5, 2.6, and

2.7, respectively. As shown in Figure 2.5, the cumulative distribution of prices under the lottery system lies on the right of that under counterfactual scenario (I). This implies that households would buy more expensive cars under the lottery system. Similarly, Figure 2.6

59 Table 2.6: Price, Horsepower, and Fuel Efficiency Summary Statistics under Null Scenario and

Counterfactual Scenario (I) in Beijing in 2012

Counterfactual Scenario (I) Null Scenario Variables Mean S.D. Mean S.D. Difference Price (1,000 Yuan) 189.42 214.80 231.85 259.75 42.43∗∗∗ Horsepower (kw) 105.20 38.17 114.68 41.30 9.48∗∗∗ Fuel Efficiency(L/100km) 7.85 1.62 8.09 1.60 0.24∗∗∗ Note: All money is in 2009 RMB Yuan. The number of observations is 2,964. The mean of price, horsepower, and fuel efficiency are sales-weighted means. The new vehicle sales are 1,066,551 under counterfactual scenario (I) and 531,639 under null scenario in 2012. *** significant at 1%; ** significant at 5%; * significant at 10%. and 2.7 indicate that the cumulative distribution of horsepower and fuel efficiency shift to the left after removing the lottery system. That is, the lottery system would lead buyers to purchase more powerful but less fuel efficient cars.

Our results imply that households in Beijing switch to high-end, more powerful, but less fuel efficient vehicles due to the policy. Here we provide a possible explanation. Most households in China use their savings to purchase a vehicle (Xiao et al., 2017). Households who do not win the lottery can set aside additional money for vehicle purchase each period.

As a result, they have more savings and can afford more expensive cars when they have the right to buy cars. In addition, the households would concentrate their transportation investment on high-end and more expensive vehicles, shifting the fleet toward less fuel efficient vehicles (Yang et al., 2014). It is worthwhile to investigate the underlying mechanism for such change in households’ vehicle choices. We leave this question for future research.

60 Figure 2.5: The Cumulative Distribution Functions of Price of New Cars Registered in Beijing in

2012

2.5.2 Impacts on Gasoline Consumption and Pollutant Emissions

Although the vehicle lottery policy is effective to control vehicle population growth in

Beijing, the cost of the vehicle restriction is a less fuel-efficient vehicle fleet. Here, we estimate the gas consumption and pollutant emissions under both null scenario and counterfactual scenario (I). The pollutants include carbon dioxide (CO2), particulate matter (PM10 and

56 PM2.5), nitrogen oxides (NOx), and carbon monoxide (CO). Moreover, we also calculate the sales-weighted average fuel efficiency since better fuel economy is associated with lower vehicle emissions (Harrington, 1997).

56 PM10 and PM2.5 are particulate matters with diameter less than 10 micrometers and 2.5 micrometers, respectively.

61 Figure 2.6: The Cumulative Distribution Functions of Horsepower of New Cars Registered in

Beijing in 2012

To quantify gasoline consumption and pollutant emissions, we first need to specify the household’s average annual vehicle miles traveled (VMT) in Beijing and the lifetime of the vehicles. We acquire the average annual VMT data from the 2010 Beijing Household Travel

Survey conducted by Beijing Transportation Research Center. The average VMT (100 km) is 161 in Beijing. In addition, we follow the literature and assume that the lifetime of a new vehicle is 15 years (Beresteanu and Li, 2011).57

The total gasoline consumption in market t is given by

z X z GASt = qtj × VMT × FEj × 15 (2.15) j∈Jt

57While some studies choose 10 years (Li, 2015), different lifetime horizons do not affect the qualitative comparison.

62 Figure 2.7: The Cumulative Distribution Functions of Fuel Efficiency of New Cars Registered in

Beijing in 2012 where the superscripts z = 0, 1, 2, 3 are respectively used to index null scenario and coun-

z terfactual scenario (I), (II), (III), qtj is the number of car j at the z scenario, VMT is the household’s average annual vehicle miles traveled, and FEj is the fuel efficiency (liter- s/100km) of product j.

Table 2.7: Average Emissions per Liter Gasoline for Passenger Cars

CO2 PM10 PM2.5 NOx CO g/liter 2345.649 0.028 0.026 4.412 59.851

58 Let el be the emission volume of pollutant l per gallon of gasoline listed in Table 2.7,

58When estimating pollutant emissions of vehicles, some studies employ emission per kilometer driven (Xiao and Zhou, 2013; Li and Jones, 2015). Instead, we use emission factors on a per gallon of gasoline basis

63 obtained el from Environmental Protection Agency (EPA) (2008). The total emissions of

pollutant l in market t are computed by

z X z EMt = qtj × VMT × FEj × el × 15 (2.16) j∈Jt

Table 3.8 presents the lifetime gas consumption and pollutant emissions under null and counterfactual scenarios. They are calculated for the new passenger cars registered in Beijing in 2012. Previous analysis shows that new vehicle sales would have been 1,066,551 units in the absence of the lottery policy, compared to the observed 531,639 under the lottery system.

This implies that new vehicles sales decreased by 50.15% due to the policy. Row 2 shows that the vehicle quota led to a lower fleet fuel efficiency. Row 3 suggests that Beijing’s vehicle lottery policy reduced fuel consumption by 0.98 × 1010 liters, a decline of approximately

48.69%. Similarly, the emissions of each pollutant were reduced by 48.69% with vehicle restriction.

An interesting finding is that the percentage reductions in gasoline consumption and pollutant emissions are not as much as that in new vehicle sales. This is because of the controversial effects of the policy. As shown in above analysis, the policy effectively decreased new vehicle sales in Beijing, which in turn reduced fuel consumption. However, it also shifted the fleet toward less fuel efficient vehicles.

0 1 Here, we use fuel consumption as illustration. Let Qt and Qt be the total sales product

0 1 j in market t at the null and counterfactual scenario (I), respectively. rstj and rstj are the relative market shares of product j in market t under null and counterfactual scenario (I).

since more gasoline combusted contributes to higher emissions of pollutants (Harrington, 1997).

64 Table 2.8: Counterfactual Analysis under Different Scenarios in Beijing 2011Q2-2012Q4

Counterfactual Counterfactual Counterfactual scenario (I) Null scenario scenario (II) scenario (III) (Baseline) Sales of new passenger cars 1,066,551 531,639 493,148 503,892 Fleet fuel efficiency (Liters/100km) 7.85 8.09 7.36 7.48 Gasoline consumption (liters) 2.02×1010 1.04×1010 0.88 × 1010 0.91 × 1010

65 7 7 7 7 CO2 emissions in tons 4.75 × 10 2.43 × 10 2.06 × 10 2.13 × 10 PM10 emissions in tons 566.41 290.62 245.45 254.85 PM2.5 emissions in tons 525.95 269.89 227.92 236.65 4 4 4 4 NOx emissions in tons 8.93 × 10 4.58 × 10 3.87 × 10 4.02 × 10 CO emissions in tons 12.11 × 105 6.21 × 105 5.25 × 105 5.45 × 105 Change in external costs (billion Yuan) - -54.70 -63.63 -61.96 Change in consumer welfare (billion Yuan) - -151.66 -101.52 -90.23 Change in government revenue (billion Yuan) - 0 10.24 11.14 Change in social welfare (billion Yuan) - -96.96 -27.65 -17.13 Note: All money is in 2009 Yuan. The external costs are calculated based on a discount rate of 5% during 15 years. In market t, the change in gas consumption is given by

X 0 X 1 ∆GASt = qtj · VMT · FEj · 15 − qtj · VMT · FEj · 15 j∈Jt j∈Jt   X 0 0 X 1 1 = 15 · VMT ·  Qt · rstj · FEj − Qt · rstj · FEj j∈Jt j∈Jt     1 0 X 1 0 X 0 1 = −15 · VMT · (Qt − Qt ) · rstj · FEj + 15 · VMT · Qt · (rstj − rstj) · FEj j∈Jt j∈Jt

where the first term represents the effect on gasoline consumption due to sales decrease alone,

whereas the second term measures the impact on gasoline consumption because of fleet fuel

efficiency changes. In our analysis, the license plate restriction would have reduced the gas

consumption by 50.09% in Beijing holding fleet fuel efficiency constant, changes in fleet fuel

efficiency increased fuel consumption by only 1.40%.

2.5.3 Welfare Analysis

In this subsection, we estimate the welfare effect of the policy in 2012. In our study, we

define the social welfare as consumer welfare plus government revenue minus external costs

related to vehicle usage.

In this paper, we use compensating variation (CV) as a measure of the change in consumer

welfare due to a policy. Since the marginal utility of income is nonlinear in our specification,

we cannot calculate CV with the approach used by Nevo (2000), Nevo (2003), and Xiao and

Ju (2014).59 Given our utility function (2.1), we follow Small and Rosen (1981) and Herriges

59 In those studies, consumer i’s indirect utility function from product j in market t is uijt = Vijt + εijt, where εijt follows i.i.d. extreme value distribution. For a logit discrete choice model, if the marginal utility of income is constant, then the expected compensating variation of household i due to a policy is

66 and Kling (1999) and derive our estimation of the expected compensating variation. As in

previous studies, let wo and w denote the case without policy versus the case with a policy.

The expected CV of household i in market t due to a policy is given by60

  wo w  Uilt − Uimt E(CVit|yi, vi) = Eε yi − exp ln yi + w (2.17) η ln pmt

w w wo wo where Uilt = max uijt and Uimt = max uijt . yi is household income, and vi is a vector j=1,...,Jt j=1,...,Jt of unobservable household characteristics. Then the measure of total consumer welfare in

market t can be calculated as

ns Z Qwo X CV = Qwo E(CV |y , v )dP (y)dP (v) = t E(CV |y , v ) (2.18) t t it i i ns it i i i=1

wo where Qt is the number of vehicles sold in market t under no policy. In this paper, we

follow the simulated method developed by Herriges and Kling (1999) for consumer welfare

computation. We first generate ns = 150 households from the distributions of y and v by

Halton sequences for each quarter in Beijing. For each household, we simulate 200 draws

from the distribution of εijt for each product. With simulated εijt, we can compute each

household’s expected CV in equation (2.17). Finally, the consumer welfare in market t can

be calculated by equation (2.18).61

We now turn to external costs associated with vehicle usage. Parry et al. (2007) argue that

vehicle usage could cause various externalities, such as air pollution and traffic congestion.

Creutzig and He (2009) estimate the external costs of one gallon of gasoline consumption to

PJt w PJt wo ln[ j=1 exp(Vijt)]−ln[ j=1 exp(Vijt )] given by E(CVit) = MU , where w and wo denote with and without the policy, respectively. Here MU is the constant marginal utility of income. 60In Appendix G., we give details of the derivation. 61 We also tried 300 draws of εijt but that made little difference in our estimations.

67 be 9.7 Yuan in Beijing, i.e., $6.02 per gallon of gasoline.62 Given the fact that the gasoline tax in China includes 1 Yuan per liter to deal with externalities, we follow Li (2015) and use

8.7 Yuan (in 2012 term) per gallon of gasoline as the external costs in our analysis. The total external costs are calculated based on an annual discount rate of 5% during the vehicles’ lifetime.

Table 3.8 presents our empirical results on welfare. On the benefits side, the vehicle lottery significantly reduces gas consumption and pollutant emissions in Beijing. The ex- ternality reduction from the lottery system is estimated to be 54.70 billion Yuan during a

15-year time span. On the cost side, the vehicle quota system leads to about 151.66 billion

Yuan loss in consumer welfare. Consumer welfare loss from the lottery policy is due to:

(1) those households who have demand for private vehicles but cannot purchase cars since they do not win the lottery; and (2) the lottery allocates vehicles to households who do not necessarily have the highest willingness-to-pay. Overall, the reduction in external costs is dominated by the consumer welfare loss and the vehicle lottery causes a social welfare loss of 96.96 billion Yuan.

2.5.4 Alternative Policies and Comparisons

Generally, market-based mechanisms (e.g., taxes) achieve better allocative efficiency than non-market based mechanisms (e.g., lottery). Therefore, we check the effectiveness of the two alternative market-based policies discussed at the beginning of Section 2.5 and in Table

62Similarly, Parry and Timilsina (2009) find that the external costs from vehicle usage is $6.06 per gallon of gasoline consumed in Mexico city, which is a comparable city to Beijing.

68 2.4 and 2.5, i.e., a first registration tax system and a consumption tax scheme. The results are shown in Table 3.8.

In counterfactual scenario (II), we apply a first registration tax system to Beijing. As shown in Table 3.8, the fleet efficiency under the registration tax scheme is better compared with vehicle lottery, . Moreover, the gas consumption and pollutant emissions are lower in counterfactual scenario (II). The consumer welfare loss is much smaller under this tax system because tax policies achieve better efficiency by allocating the resource to those with the highest willingness to pay. Specifically, the social welfare would increase by 69.31 billion

Yuan in Beijing in 2012 if Beijing municipal government replaced the vehicle lottery with the first registration tax system. Also, such a tax policy could generate 10.24 billion Yuan tax revenue for Beijing government, which makes it likely to gain political support.

Column 5 in Table 3.8 presents the market outcomes under scenario (III) with a hypo- thetical consumption tax scheme.63 We find that the fleet is more fuel efficient under the consumption tax system. This result implies that such consumption tax can further lower gas consumption by 1.3 billion liters. We also find that the consumption tax policy generates less pollutant emissions relative to the vehicle quota system. In terms of social welfare, such a tax policy only leads to 17.13 billion Yuan loss in social welfare in Beijing in 2012, relative to welfare loss of 96.96 billion Yuan under the VLS. That is to say, Beijing loses 79.83 billion

Yuan in welfare in 2012 by using the lottery system rather than a consumption tax to control vehicle registrations.

63The tax rates were shown in Table 2.5. The tax rates are designed so that the fleet size is the same under both vehicle lottery policy and the consumption tax regime.

69 In a related paper, Li (2015) compares Beijing’s lottery system with a uniform-price auction to study the welfare consequences of these mechanisms. He finds that the social welfare (consumer welfare plus government revenue minus external costs) loss from the lottery system is about 58 billion Yuan in Beijing in 2012, compared with a uniform price auction.

Our study indicates that the lottery system leads to welfare loss of nearly 69.31 and 79.83 billion Yuan compared with the first registration tax scheme and the consumption tax system, respectively. Thus, a uniform price auction is superior in terms of improving social welfare.

However, Li (2015) also finds that fleet efficiency of new vehicles sold under the auction system is lower than that under the VLS. In contrast, our paper reveals that our tax systems shift new vehicle purchases toward more fuel efficient cars and generate a more fuel efficient

fleet relative to the VLS. Thus, these results suggest that the tax schemes work better than the auction system for environmental purposes.

From the above analysis, we find that the vehicle lottery system is not the first choice to control vehicle growth and air pollution. Beijing municipal government could consider some market-based policies, such as tax policies, which can be more effective in slowing down vehicle increase and reducing pollutant emissions.

2.6 Conclusion

With the rapid growth in vehicle population, problems such congestion, energy shortages, air pollution and its health consequences have become a major concern in Beijing and many other cities worldwide. To control vehicle growth and thereby address related environmental

70 issues, Beijing municipal government imposed a vehicle quota system and allocated the quota through lottery. In this paper, we investigate the impacts of such novel policy on

fleet composition, fuel consumption, pollutant emissions, and social welfare. To do this, we estimate a random coefficient discrete choice model of automotive oligopoly using registration data of new passenger vehicles in Beijing, Nanjing, Shenzhen, and Tianjin. To identify the effects of the lottery and compare with other policies, we then conduct counterfactual analysis based on the model estimates.

Our main results suggest that Beijing’s vehicle lottery is effective in limiting new vehicle sales and reducing gasoline consumption and air pollution. However, vehicle fleet composi- tion changes due to the policy. The vehicle quota system shifts the demand for new vehicles toward less fuel efficient cars because of households’ behavioral responses, such as concen- trating their transportation investment into single high-end but less fuel efficient vehicles.

This change can undermine the potential benefits. In addition, our analysis shows that this system leads to a welfare loss since it arbitrarily reduces demand for vehicles regardless of the willingness to pay for new vehicles.

The vehicle lottery system in Beijing has been emulated in Guiyang, Guangzhou, Tianjin,

Hangzhou, and Shenzhen, and similar programs are being considered for other Chinese cities, such as Chengdu and Wuhan. While this vehicle quota system may seem a reasonable approach for resolving vehicular environmental issues, our counterfactual analysis shows that a progressive tax system works better than the vehicle lottery policy in reducing fuel consumption and air pollution. Moreover, we find that, relative to the lottery system, such a tax scheme can achieve the same effect in controlling vehicle growth but improving fleet fuel

71 efficiency and producing a smaller welfare loss. This implies that other cities could consider other options to the lottery system in an effort to control vehicle, gas consumption, and air pollution.

There are several worthwhile directions for future research. First, our conclusion is based on the assumption that the lottery does not change driving patterns. Examining how the lottery system affects driving patterns would be helpful to accurately assess the effects of the policy. Second, our results suggest that Beijing’s vehicle lottery system shifts households’ purchases toward high-end vehicles, which in turn could affect the local automotive market structure. Exploring the impact of this system on the market structure has important implications for the industry. Finally, future research could consider a dynamic model to investigate how households choose among available products or wait to purchase in the future, which helps to understand the policy effects on consumers’ behavior.

72 CHAPTER 3.

WELFARE ANALYSIS OF GOVERNMENT INCENTIVES FOR

FUEL EFFICIENT VEHICLES AND NEW ENERGY

VEHICLES IN CHINA

To solve energy security and environmental problems, the Chinese government announces

various incentives for fuel-efficient vehicles and new-energy vehicles (NEVs). This paper

investigates the effectiveness and welfare consequences of (i) vehicle and vessel usage tax

(VVUT) incentives, (ii) fuel-efficient vehicle subsidy program, and (iii) NEV private pur-

chase subsidy pilot program. The empirical findings suggest that these policies promote

the diffusion of fuel-efficient vehicles and NEVs, and improve fleet fuel efficiency. Howev-

er, VVUT incentives and fuel-efficient vehicle subsidy program increases oil consumption

and CO2 emissions. Although NEV private purchase subsidy pilot program cuts down gas

consumption, it raises CO2 emissions. VVUT incentives and fuel-efficient vehicle subsidy

program improve social welfare, while NEV private purchase subsidy pilot program causes

welfare loss. In addition, we examine how results varies with rising gasoline tax. We find

that increasing gasoline tax is more effective in reducing oil consumption and CO2 emissions, and improving welfare, but at the expense of adoption of fuel-efficient vehicles and NEVs.

73 3.1 Introduction

China’s automobile industry has developed rapidly since 2000. Vehicle population in-

creased from 16.09 million units in 2000 to 62.09 million units in 2009 at an average annual

rate of 14.5%.64 With new vehicle sales of 13.65 million and new vehicle production of 13.79 millions in 2009, China surpassed the U.S. market and became the largest auto market in the world in both sales and production. However, the rapid growth in vehicle ownership and usage have created some serious problems, such as energy security, air pollution, and health concerns in China. To solve these problems, the Chinese government issued various fiscal incentives to improve the fuel economy of new vehicles and promote the diffusion of new- energy vehicles (NEVs).65 For example, the Chinese government offered vehicle and vessel

usage tax (VVUT) incentives to energy-saving vehicles and NEVs from January 2012, with

50% tax reductions for energy-efficient vehicles and tax exemptions for NEVs. In June, 2010,

the central government issued a subsidy program which provides consumers with a subsidy

of 3,000 yuan per car to purchase officially approved fuel efficient vehicles. In addition,

both central and local governments would subsidize private purchase of NEVs in Shanghai,

Changchun, Hangzhou, Hefei, and Shenzhen during 2010 and 2012.

To our knowledge, no studies investigate the welfare effects of these government incen-

tives and their effectiveness on solving energy security and environmental problems. In this

paper, we the examine the effectiveness of various government incentives on gas consumption

64The average growth rate of vehicle population in the United States from 2000 to 2009 was 1.19%. 65New energy vehicles mainly refer to electric vehicle, fuel cell battery vehicle, and plug-in hybrid electric vehicle.

74 and CO2 emissions as well as the welfare consequences of these policies. These governmen- t support programs includes VVUT incentives, fuel-efficient vehicle subsidy program, and

NEVs private purchase subsidy pilot program. We construct and estimate a random coeffi- cient discrete choice model developed by Berry et al. (1995) using passenger car registration data in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen from 2011 to 2012. Based on the parameter estimates, we conduct several counterfactual experiments to identify the effects and welfare consequences of VVUT incentives, fuel-efficient vehicle subsidy program, and NEVs private purchase subsidy pilot program. Moreover, to compare the effectiveness of these government support programs with that of a higher gasoline tax, we simulate the outcomes under another three counterfactual scenarios of an increase in the gasoline tax of

0.1 Yuan, 0.2 Yuan, and 0.3 Yuan per liter.

Taking advantage of customer reviews data on a vehicle website, we reduce the model’s dependence on the the exogeneity assumption in the literature that observed product char- acteristics are uncorrelated with unobserved product characteristics. We exploit customer reviews on car models to obtain some product attributes (e.g., vehicle interior features and perception of fuel efficiency), which affect consumers’ purchase decision and may be cor- related with observed product attributes. For example, horsepower and car weight may be negative correlated with consumers’ perception of fuel efficiency (Allcott and Wozny,

2014). Leaving this information from customers reviews would render the instrument vari- ables invalid and biases parameter estimates. So our study uses online customer reviews and mitigates the effect of the exogeneity assumption on model estimation.

Our study provides interesting findings. First, these government incentives would in-

75 crease sales of fuel-efficient vehicles and NEVs, and improve fleet fuel efficiency. However, our counterfactual analysis shows that VVUT incentives and fuel-efficient vehicle subsidy program increase oil consumption and CO2 emissions. Although NEVs private purchase pilot program reduce gas consumption, it increases CO2 emissions because most of NEVs consume electricity. The welfare analysis indicates that VVUT incentives and fuel-efficient vehicle subsidy program respectively improve social welfare by 0.0815 and 0.4230 billion yuan in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen in 2012. While NEVs pri- vate purchase pilot program causes a total welfare loss of 0.0151 billion yuan in these cities in 2012. Comparing these government incentives with increasing gasoline tax, we find that these government support programs are superior to an increase in gasoline tax in improving

fleet fuel efficiency and promoting the adoption of fuel-efficient vehicles and NEVs. Whereas the gasoline tax works better than the governments incentives with respect to reducing gas consumption, cutting down CO2 emissions, and raising social welfare.

Government incentives are widely used around the world to improve fleet fuel economy.

For example, there are federal income tax deductions and federal income tax credits for hybrid vehicles in U.S.. And several papers have examined the effectiveness of these poli- cies. Gallagher and Muehlegger (2011) study the efficacy of state sales tax waivers, income tax credits, and rising gasoline prices on adoption of hybrid vehicles and find they are all important in U.S.. Beresteanu and Li (2011) investigate how gasoline prices and income tax incentives affect demand for hybrid vehicles in U.S.. They find that both increases in gasoline prices and federal income tax incentives are of importance to grow the market share hybrid vehicles, while the effects of government support on reducing gasoline and CO2 emissions

76 are inconsequential due to the small share of hybrid vehicles. However, little is known about the effectiveness of government support programs in China. Our study contributes to the strand of literature.

Our paper is also related to research on various policies in Chinese auto market to reduce energy consumption and pollutant emissions. For example, Viard and Fu (2015) evalu- ate the pollution reductions from Beijing’s driving restrictions and find that pollutant falls

21% during driving restrictions. Xiao and Ju (2014) explore the environmental effects of consumption-tax and fuel tax adjustments and find that fuel tax is effective in reducing fuel consumption while consumption tax does not. Li (2015), Xiao et al. (2017), and Yang et al.

(2017) examine the effectiveness of vehicle quota systems in China. However, little research has been devoted to study the environmental effects and welfare consequences of government incentives for fuel-efficient vehicles and NEVs. Our analysis fill the gap.

The rest of the paper is organized as follows. Section 3.2 briefly introduces VVUT incentives, fuel-efficient vehicle subsidy program, and NEVs private purchase subsidy pilot program in China, and discuss our data. Section 3.3 describes the empirical model and the estimation strategies. Section 3.4 reports our estimation results. We provide counterfactual analysis in Section 3.5. Section 3.6 concludes.

77 3.2 Policy and Data Description

3.2.1 Policy Description

Since we only focus on private passenger vehicles in this study, we only introduce policy content related to private passenger cars. The articles for business-use or government-use vehicles are not presented here. vehicle and vessel usage tax Incentives

The owner or managers of vehicles and/or vessels within China need to pay vehicle and vessel usage tax (hereafter VVUT). It is behavioral and property tax. The provisional regulations of China on VVUT were first promulgated by the State Councile on Septermber

15, 1986, and took effect on October 1, 1986. The provisional regulations were revised on

December 27, 2006, and came into effect on January 1, 2007. According to the regulations, the annual amount of VVUT ranges from 60 to 660 Yuan per car for passenger vehicles.

The specific tax amounts applicable to vehicles are determined by provincial governments.

In Table 3.1, we give taxable items and tax amounts for passenger vehicles in Shanghai,

Changchun, Hangzhou, Hefei, and Shenzhen based on VVUT law (2007).

On February 25, 2011, the Standing Committee of China’s National People’s Congress passed the Vehicle and Vessel Tax Law of the People’s Republic of China (PRC Presidential

Decree No. 43. The new law will replace the 2007 interim law and come into force on

January 1, 2012. VVT for Passenger cars with small engine size does not change much, but the tax for cars with large engine size increases a lot under the new law relative to the 2007

78 Table 3.1: VVUT for Passenger Vehicles in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen

based on VVUT Law (2007)

Annual VVUT per car (Yuan) Taxable Items Shanghai Changchun Hangzhou Hefei Shenzhen Large vehicles 540 600 540 540 600 Intermediate vehicles 510 540 480 450 480 Small vehicles 450 480 360 360 420 Mini vehicles 300 300 240 240 240 Note: Large vehicles: passenger capacity is greater than or equal to 20 persons. Intermediate vehicles: passenger capacity is greater than 9 but less than 20. Small vehicles: passenger capacity is not greater than 9. Mini vehicles: engine size is not greater than 1.0L. interim law. The State Council regulates a range of the annual tax rates for various types of vehicles. And the local governments determine the specific tax amounts. Table 3.2 reports the tax range by the State Council and the tax schemes in Shanghai, Changchun, Hangzhou,

Hefei, and Shenzhen since January 1, 2012.

Table 3.2: VVUT for Passenger Vehicles based on VVUT Law (2012)

Annual VVUT per car (Yuan) Engine Size State Council Shanghai Changchun Hangzhou Hefei Shenzhen ≤1.0 60-360 180 240 180 180 180 1.0-1.6 300-540 360 420 300 300 360 1.6-2.0 360-660 450 480 360 360 420 2.0-2.5 660-1,200 720 900 660 660 720 2.5-3.0 1,200-2,400 1,500 1,800 1,500 1,200 1,800 3.0-4.0 2,400-3,600 3,000 3,000 3,000 2,700 3,000 >4.0 3,600-5,400 4,500 4,500 4,500 3,900 4,500 Note: The tax rates are for passenger cars with seating for no more than nine passengers.

As China becomes more concerned over environmental issues, the new law offers tax

79 incentives to energy-saving vehicles and new-energy cars. The law stipulates that VVUT may be reduced or exempted with respect to vehicles that contributes to energy conservation or use new energies. As a result, new-energy vehicles, including pure electric vehicles, fuel cell vehicles and plug-in bybrid vehicles, can enjoy VVUT exemptions. And other types of energy-saving passenger vehicles are subject to half of the regular tax rates. But these vehicles should meet some standards set by Ministry of Finance (MOF), State Administration of Taxation, and Ministry of Industry and Information Technology (MIIT). In 2012, these departments jointly issued two groups of energy-saving and new-energy vehicles eligible for vehicle and vessel tax reduction and exemption in 2012.66

Fuel-efficient Vehicle Subsidy Program

To save energy and reduce greenhouse gas (GHG) emissions, Chinese policymakers real- ized the importance of increasing market share of fuel-efficient vehicles. On June 30, 2010,

China established a national subsidy program for fuel-efficient vehicles that provided a one- time direct deduction of 3,000 Yuan (about $470 U.S. dollars) per vehicle from the retail prices to consumers who purchased eligible fuel efficient vehicles. Eligible cars should be with 1.6L or smaller engine size and meet certain fuel consumption limits as shown in Ta- ble 3.3. Firms should submit application for their vehicles to be involved in the program.

MIIT, MOF, and National Development and Reform Commission (NDRC) will review the applications and announce the list of eligible vehicles.

The policy was renewed in October 2011 with more stringent fuel consumption limits, as

66The catalogue of the vehicles can be found on the autohome’s website.

80 Table 3.3: Fuel Consumption Limits Required by Fuel-efficient Vehicle Subsidy Programs Introduced in June 2010 and October

2011

Fuel-efficient Vehicle Fuel-efficient Vehicle Subsidy Program (June 2010) Subsidy Program (October 2012) Fuel Consumption Fuel Consumption Fuel Consumption Fuel Consumption Curb Weight Curb Weight Limits for Categories Limits for Categories Limits for Categories Limits for Categories (CW) (CW) (i) and (ii) (iii) and (iv) (i) and (ii) (iii) and (iv) kg L/100km L/100km kg L/100km L/100km 750

81 980

From June 2010 to December 2012, the government released 8 official lists of car models eligible for the subsidy program.

NEV Private Purchase Subsidy Pilot Program

New energy vehicles (NEVs) refer to vehicles using non-traditional fuel (ethanol, biogas, biodiesel), electric vehicles, fuel cell vehicles, and various hybrids of these (Liu and Kokko,

2013). They are particularly important for China since they can effectively relieve energy security and environmental pollution problems. As a result, the NEV industry gains special incentives from the Chinese government.

Government’s subsidy to support NEV industry development can be dated back to around

2000. The Chinese 9th Five-Year Plan (1995-2000) and 10th Five-Year Plan (2001-2005) highlight the importance of NEVs. Consequently, the “National High Technology Research and Development Program” (the 863 Program) started to fund NEVs related R&D from

2001.

To raise the demand for NEVs, the government planned to increase the adoption of NEVs in public sector and government agencies in large cities. In 2009, the central government published a “Notice on New Energy Vehicle Demonstration and Extension Work”, which announced to subsidize the public sector’s purchases of NEVs. 13 cities are selected for the

82 demonstration and promotion of NEVs.

Given that the percentage of private passenger vehicles was 79% in 2009 according to

China Emission Control Annual Report (2010), stimulating the purchases of NEVs for public

transportation is not enough. To further promote the development of the NEV industry and

reduce oil consumption and pollutant emissions, the central government issued a “Notice on

Subsidies for Private Purchases of New Energy Vehicles”. We called it NEV private purchase

subsidy pilot program. The program identified Shanghai, Changchun, Hangzhou, Hefei, and

Shenzhen as the pilot cities for the subsidies from 2010 to 2012. It is worth to note that only

pure electric vehicles (EVs or BEVs) and plug-in hybrid electric vehicles (PHEVs) are eligible

for the program. For EVs, the minimum battery capacity is 15 kWh. Eligible PHEVs should

have a minimum battery capacity of 10 kWh and a minimum all-electric range of 50km.

The subsidies to NEVs private consumers are based on the battery capacity and the tech-

nology. In these five pilot cities, qualified NEVs are subsidized with 3,000 Yuan (about $490

U.S. dollars) per kWh. The amount of subsidy is subject to a cap of 50,000 or 60,000 Yuan

(U.S. $7,320 or $8,784 in June 2010) for PHEVs and EVs, respectively. The subsidies are paid to carmakers instead of consumers, but the retail prices will be reduced correspondingly.

Moreover, the program stipulated that the pilot cities should also issue their own subsidy plans (2010-2012). So private consumers in these cities could obtain both central and local governments’ subsidies for NEVs purchases. In Hefei, the subsidies for an EV and a PHEV are 10,000 and 15,000 Yuan, respectively. The Shenzhen government subsidizes PHEVs with

2,000 Yuan per kWh for the first 10,000 sales, 1,500 Yuan per kWh for the second 10,000 sales, and 1,000 Yuan per kWh for the rest. The maximum subsidy for PHEVs is 30,000

83 Yuan per car. EVs are subsidized with 3,000 Yuan per kWh for battery capacity above 20 kWh with a maximum subsidy of 60,000 Yuan per car. In Hangzhou, purchases of PHEVs would be subsidized with 2,000 Yuan per kWh for battery capacity above 10 kWh with a cap of 30,000, and purchases of EVs would be subsidized with 3,000 Yuan per kWh for battery capacity above 20 kWh with a cap of 60,000 Yuan. In Shanghai, the amount of local subsidy was indexed to the vehicle battery capacity (2,000 Yuan per kWh) and subject to a cap of

20,000 or 40,000 Yuan for PHEVs and EVs, respectively. Changchun government provided subsidies for NEVs of 20% of retail prices. The subsidies were up to 40,000 Yuan for PHEVs and up to 45,000 Yuan for EVs.

3.2.2 Data

There are two main data sources for this study. The first data set provides use with the monthly vehicle registration information in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen from January 2011 to December 2012.67 It contains new passenger vehicles’ registration city (e.g., Shanghai), registration time (e.g., July 2011), vehicle nameplate- vintage, model number (only for domestic vehicles), manufacturer, brand, model name, engine size, bodystyle, passenger capacity, transmission, country of origin, purpose of usage

(private or business), and sales.

To complete the data set, we further collect vehicle attributes from autohome’s website and Car Market Guide. We define a car model by the vehicle nameplate-vintage (e.g., 2011

67In China, vehicle registration data are not released to the public. We purchase the data from Beijing CyberRay Co., Ltd.

84 Benz C200), manufacturer, and vehicle attributes (i.e., MSRP, horsepower, curb weight, length, width, height, engine size, transmission, bodystyle).68 We also obtain fuel con- sumption per 100 km of each car model from China’s Ministry of Industry and Information

Technology. Gasoline prices and electricity prices are from National Development and Re- form Commission of China. We then construct energy expenditure variable (yuan/100km), which would affect consumers’ vehicle purchase decision.

It is worth to point out that we use Manufacturer Suggested Retail Prices (MSRP) in this study. MSRPS are set by manufacturers and are generally constant across locations and within a model year. Theoretically, transaction prices would be better. But it is not available. Li et al. (2015) argue that, unlike U.S. auto market, promotions in China’s auto markets are not frequent and, hence, MSRP can be good proxy for retail prices. Vehicle prices are computed base on MSRP, taxes, license fee, and government subsidies.69 In Shanghai, consumer who need a new license need to pay the license fee, while the license fee is zero in other city. Given that cars and licenses are perfect complementary goods, we include the fee in the vehicle prices. We obtain the average monthly license price in Shanghai online.

We also crawl data of customer reviews on each car model from the autohome’s web- site. We collect 336,886 online comments, including consumers’ ratings over each product’s interior room, ease to handle, fuel efficiency, comfort, appearance, interior features, and price-performance ratio. The ratings are based on a 5-point Likert scale (1 means ”Ex-

68Vehicles with same nameplate-vintage but different attributes will be defined as different car models. For example, 2011 Beijing Benz C200 1.8T Sedan and 2011 Beijing Benz C200 2.0L Sedan are different models. 69The taxes includes consumption tax, value-added tax, vehicle purchase tax, and vehicle and vessel usage tax. See Xiao and Ju (2014) for more details about the first three taxes.

85 Table 3.4: Summary Statistics of Vehicle Sales, Characteristics, and Consumer Reviews

Variables Obs Mean S.D. Min Max Monthly sales by model 80,001 21.4088 47.2635 1.00 1,383.00 Price (1,000 Yuan) 80,001 247.3482 262.2664 20.07 14,508.08 Weight (1,000 kg) 80,001 1.4288 0.2800 0.65 2.95 Horsepower (kw) 80,001 107.5145 38.3229 12.00 515.00 Car size (m2) 80,001 8.1626 0.8187 4.20 12.21 Fuel consumption (L/100km) 80,001 7.9687 1.4692 0.00 20.00 Energy expenditure (Yuan/100km) 80,001 60.0433 11.0751 5.40 154.51 Sedan dummy 80,001 0.6348 0.4815 0.00 1.00 Hatch back vehicle dummy 80,001 0.1370 0.3438 0.00 1.00 Multi-purpose vehicle dummy 80,001 0.0512 0.2204 0.00 1.00 SUV dummy 80,001 0.1762 0.3810 0.00 1.00 Station wagon dummy 80,001 0.0008 0.0285 0.00 1.00 MT dummy 80,001 0.3054 0.4606 0.00 1.00 AT dummy 80,001 0.5202 0.4996 0.00 1.00 CVT dummy 80,001 0.0894 0.2854 0.00 1.00 DCT dummy 80,001 0.0850 0.2788 0.00 1.00 Rating on interior room 2,217 4.3241 0.5117 1.00 5.00 Rating on ease to handle 2,217 4.3395 0.3167 1.00 5.00 Rating on fuel efficiency 2,217 4.1485 0.4241 1.00 5.00 Rating on comfort 2,217 3.9335 0.4168 1.00 5.00 Rating on appearance 2,217 4.5140 0.3066 2.00 5.00 Rating on interior features 2,217 3.9936 0.4368 1.00 5.00 Rating on price-performance ratio 2,217 4.3955 0.2831 1.00 5.00 Note: All money is in 2012 RMB Yuan. MT stands for Manual transmission, AT for Automatic transmission, CVT for Continuously variable transmission, and DCT for Dual clutch transmission. The mean of price, weight, horsepower, car size, fuel consumption, energy expenditure, bodystyle dummies, and transmission dummies are sales-weighted means. The mean of ratings are weighted by number of customer reviews.

86 tremely bad”, 2 ”Bad”, 3 ”Neither bad nor good”, 4 ”Good”, and 5 ”Extremely good”).

Exploiting online customer reviews is to capture some product features which may affect households’ vehicle choice but are unobserved to researchers.

We drop observations registered for business use due to the difference between private consumers and business consumers. There are 80,001 observations with 2,217 different car models in our sample.

Table 3.4 provides the summary statistics of our sample. Both monthly sales and prices have large variations. The most popular passenger car has a monthly sale of 1383 units, while the lease popular car has only one sale. The sales-weighted average price is 247,348 Yuan with a range of 20,070 to 14,508,080 Yuan. The average price is higher than the average household income in these cities according to the yearbooks. The lightest car weighs 650 kg, whereas the heaviest one weighs 2950 kg. The most powerful vehicle has a horsepower of 515 kw, while the lease has a horsepower of 12 kw. The largest vehicle is 12.21 m2, but the smallest one is 4.2 m2. In term of fuel consumption performance, the least fuel efficient vehicle consumes 20 liters of gasoline for a 10 kilometers’ drive, while the most fuel efficient one consumes only 0 liters since electric vehicle consume electricity. The average expenditure is 60.0433 Yuan per 100 km, ranging from 5.4 to 154.51 Yuan/100km. Among the vehicle sales, sedan accounts for 63.48%. Cars with manual transmission and automatic transmission count 30.54% and 52.02%, respectively. Consumers ratings tend to be positive on average.

The second data set is the household income distribution in each city and year. China’s

National Bureau of Statistics does not release household-level income data to the public.

87 However, we can obtain the annual average income of each quantile from the statistical yearbook of these and some income information from Chinese Household Income Survey

(2007). Following Li et al. (2015) and Yang et al. (2017), we construct the income distribution based on these two data sources. We first obtain each city’s average household income for each income level. Then we interpolate the data using the survey data. Finally, we adjust the household income so that the derived income statistics are consistent with those from the yearbooks. Following the literature, we assume that household income follows lognormal distribution. With the interpolated household income data, we can estimate the mean and standard deviation of the lognormal distribution.

3.3 Empirical Model and Estimation

3.3.1 Utility Function Specification

Let m = {1, 2, 3, 4, 5} denote a market (i.e., Changchun, Hangzhou, Hefei, Shanghai, and Shenzhen) and t denote time (half year by year, e.g., the first half year of 2011 and the second half year of 2011). In market m at time t, a set of products, j = 1, ..., J, is available. It is worth to note that, as stated in data description section, a product in our analysis is defined as a unique combination of the vehicle nameplate-vintage, manufacturer, and vehicle attributes. Let i denote a household. Then the indirect utility of household i from purchasing product j in market m at time t is given by

uijmt = Vijmt(pjmt,XX jmt, ξjmt, yimt,vvimt) + εijmt (3.1)

88 where pjmt is the price of product j. X jmt is a vector of observed product attributes (other than price), including a constant term, logarithm of horsepower, vehicle curb weight, vehicle size, energy expenditure per kilometer drive. ξjmt is unobserved (by the researcher) product characteristic, such as product quality. yimt is the income of household i and vimt is a vector of unobserved household demographics. With yimt and vimt, we assume that households are heterogenous in income and idiosyncratic tastes. εijmt is an independently and identically distributed (across products, households, markets, and time) idiosyncratic shock, which follows the type I extreme value distribution.

Following the literature (Nevo, 2000, 2001; Li et al., 2015; Yang et al., 2017), we use some dummy variables to capture some components of the unobserved characteristics. That is, we model ξjmt = ξm + ξj + ξt + ∆ξjmt. We can capture ξm and ξt by city dummies and time dummies. ξj includes firm dummies, vehicle bodystyle dummies (i.e., NB, HB, M-

PV, SUV, Station Wagon), and transmission dummies (i.e., manual transmission, automatic transmission, continuously variable transmission, and dual clutch transmission), capturing households’ intrinsic preference for products from different firms, vehicle bodystyle, and car transmission. Unlike previous studies, we use customer reviews from a vehicle website to capture some other components of the unobserved characteristics. Some product character- istics (like vehicle appearance) are unobserved to researchers, but some consumers would indicate them in their reviews. Let Rj denote a vector of product j’s consumer ratings over vehicle interior room, ease to handle, fuel efficiency, comfort, appearance, interior features,

89 and price-performance ratio. The specification of the indirect utility is

uijmt =αi ln pjmt + X jmtβ i + Rjγ + Dcitymλ1 + Dtimetλ2 + Dfirmjλ3 (3.2)

+ Dbodystylejλ4 + Dtransmissionjλ5 + ∆ξjmt + εijmt where ∆ξjmt is the econometric error term. We use 0 to denote the outside good (i.e., the choice of purchasing a used car or not buying a vehicle). The utility from outside good is normalized to εi0mt, which follows i.i.d. type I extreme value distribution.

The model allows household heterogeneity in their tastes for price and other characteris- tics. For example, households with higher income are less price sensitive. And the random coefficients αi and β i capture the household heterogeneity. In particular, αi measures house- hold i’s preference for price and it is given by

p αi =α ¯ + η ln yi + σpvi (3.3)

p where yi is household income, and vi is unobserved household characteristics that affect household preferences and follow standard normal distribution. Since households with a higher income yi tend to be less price sensitive, we expect η to be positive.

Similarly, βik measures household-specific taste on vehicle characteristic xjmtk, which is the kth attribute of product j. Specifically, βik is defined by

¯ k βik = βk + σkvi (3.4)

¯ k k where βk is the average preference across households and σkvi captures random tastes. vi is assumed to have a standard normal distribution.

0 Let Djmt be the vector of dummy variables in equation (3.2) and λ = (λ1,λλ2,λλ3,λλ4,λλ5) .

90 By substituting equation (3.3) and (3.4) into (3.2), we obtain

K X ¯ uijmt = xjmtkβk +α ¯ ln pjmt + Rjγ + Djmtλ + ∆ξjmt k=1 (3.5) K X k p + σkxjmtkvi + (η ln yi + σpvi ) ln pjmt + εijmt k=1 The above utility function can be decomposed into a mean utility of product j in market m at time t, which is common to all households

K X ¯ δjmt = xjmtkβk +α ¯ ln pjmt + Rjγ + Djmtλ + ∆ξjmt (3.6) k=1 a household-specific utility (i.e., deviation from the mean utility), which captures household’s heterogenous tastes for observed product characteristics

K X k p µijmt = σkxjmtkvi + (η ln yi + σpvi ) ln pjmt (3.7) k=1 and a random taste shock εijmt. For compactness, denote the parameters in the mean utility

0 (3.6) as θ1, the parameters in the household-specific utility (3.7) as θ2, and θ = (θ1, θ2) .

3.3.2 Market Share and Aggregate Demand

A household chooses the product that maximizes his utility. Then the set of household attributes that induce the choice of product j in market m at time t is defined as

Ajmt(X ·mt, p·mt, δ·mt; θ2) = {(yi,vvi, εi0mt, ..., εiJmt)|uijmt ≥ uilmt, for l = 0, 1, ..., J} (3.8)

0 0 0 where X ·mt = (X 1mt, ...,XX Jmt) , p·mt = (p1mt, ..., pJmt) , and δ·mt = (δ1mt, ..., δJmt) are ob- served vehicle attributes, prices, and mean utilities of all products in market m at time t,

91 respectively. By integrating over all consumers in the set Ajmt, we find the market share of product j in market m at time t as follows

Z sjmt(X ·mt, p·mt, δ·mt; θ2) = dP (y)dP (v)dP (ε) (3.9) Ajmt where P (·) denotes population distribution functions. Given that εi0t and εijt follow the i.i.d. type I extreme value distribution, the market share of product j can be written as

Z exp(δ + µ ) s (X , p , δ ; θ ) = jmt ijmt dP (y)dP (v) (3.10) jmt ·mt ·mt ·mt 2 PJ 1 + r=1 exp(δrmt + µirmt)

Based on the market share function, we then can derive the aggregate demand function.

Let Nmt be the market size in market m at time t. Then the demand for product j is

Nmtsjmt(X ·mt, p·mt, δ·mt; θ2). Following the literature (Berry et al., 1995; StevenBerry et al.,

2004; Yang et al., 2017), the measure of market size is the number of households in the city in a given year.70

3.3.3 Identification Method

After the idiosyncratic error term εijmt is integrated out analytically, the econometric error term will be the unobserved product characteristics, ∆ξjmt, such as prestige and product quality. Prices could be correlated with these product characteristics. For example, vehicles with higher quality generally have higher prices. To address the price endogeneity problem and estimate the parameters in equation (3.2), we employ the GMM estimation method

70Choosing a different market size measure does not change the estimates except the parameter of the constant term. Because it only causes a change in the market share of each product compared to the outside goods, but does not change the relative market shares between products (Xiao and Ju, 2014).

92 proposed by Berry et al. (1995), which uses the moment condition

E(∆ξjmt|z jmt) = 0 (3.11)

where z jmt is a vector of instrumental variables described below.

To derive ∆ξjt, we first need to estimate market shares. While the market share in equation (3.10) does not have a closed form, it can be evaluated by Monte Carlo simulation with ns draws from the distributions of v and y. To increase computation efficiency and reduce the simulation error, we use Halton sequences to generate the random draws (see

Train (2009) for use of Halton sequences). Following Li (2015) and Yang et al. (2017), our results are based on 150 households in each market.71 The simulated market shares are calculated as

ns 1 X exp(δjmt + µijmt) spred(X , p , δ ; θ ) = (3.12) jmt ·mt ·mt ·mt 2 ns PJ i=1 1 + r=1 exp(δrmt + µirmt)

Next, we combine the simulated market shares (3.12) with the observed market shares

0 to solve for the mean utility levels δmt = (δ1mt, ..., δJmt) . Theoretically, the vector of mean utilities δmt can be retrieved by equating the estimated market shares with the observed market shares from the data for a given θ2:

obs pred smt = smt (X ·mt, p·mt, δ·mt; θ2) (3.13)

However, analytical solutions for δmt are not available because the system of equations in equation (3.13) is highly nonlinear. In practice, it can be solved numerically by using the

71We also checked ns = 200. We found that it made little difference.

93 contraction mapping proposed by Berry et al. (1995) as follows72

h+1 h obs pred δmt = δmt + ln smt − ln smt (X ·mt, p·mt, δ·mt; θ2) (3.14)

h h+1 until the stopping rule ||δmt − δmt || ≤ in is satisfied, where in is the inner-loop tolerance

−14 73 level. In our analysis, we set in = 10 . Once we find δmt, the unobservable attributes

∆ξjmt can be solved as

obs ∆ξjmt(X ·mt, p·mt,RRR,DD·mt, s·mt; θ) = δjmt − (X jmt, ln pjmt,RRj,DDjmt)θ1 (3.15)

The parameters θ1 in equation (3.15) can be estimated by two-stage least squares (2SLS) using instrumental variables (IVs). The demand unobservable ∆ξjmt is a function of prices, observed vehicle attributes, consumer reviews, dummies, the observed market shares, and parameters. The GMM estimator θˆ solves the problem:

min Q(θ) = min(∆ξ(θ)0Z)W (Z0∆ξ(θ)) (3.16) θ θ where W is the weighting matrix. The convergence criterion for the GMM is 10−8.

3.3.4 Instruments

To address the price endogeneity problem, we need instrumental variables. The first set of instrumental variables that comes to mind are: the observed product characteristics (i.e., constant term, logarithm of horsepower, vehicle weight, energy expenditure per kilometer,

72See Berry et al. (1995) for a proof of convergence. 73See Dube et al. (2012) for the discussion of the importance of a stringent convergence rule. They also provide a new computational algorithm for implementing the BLP estimator, called mathematical program with equilibrium constraints (MPEC). It converges faster than the algorithm that we used here.

94 and vehicle size), the sum of corresponding characteristics of other products offered by that

firm (if the firm produces more than one product), and the sum of the same characteristics of products produced by rival firms, following Berry et al. (1995), Nevo (2000), and Xiao and

Ju (2014). We also evaluate the strength of the instruments and the instrument F -statistic in our study is large, with a p-value almost 0. So the instruments are strong.

3.4 Estimation Results

In this section, we present estimation results for four specifications. We first present the coefficient estimates from the benchmark model and then compare the results from other specifications. Finally, we discuss price elasticities based on parameter estimates in the benchmark model.

3.4.1 Parameter Estimates from the Structural Demand Model

The results of the estimation are presented in Table 3.5. The first panel of the table provides the estimates of the parameters in the mean utility function defined by equation

(3.6). The second panel provides the estimates of the coefficients in the household-specific utility function defined by equation (3.7).

We focus on the benchmark specification in this subsection. The benchmark specification is the model introduced in Section 3.3. This specification takes into account (i) the endo- geneity problem caused by the correlation between unobserved product characteristics and

95 price and (ii) household heterogenous tastes over product characteristics due to observed and observed household demographics.

In the benchmark specification, the coefficient on ln(price) negative and significantly different from zero, suggesting that households dislike high vehicle prices. Horsepower and energy expenditure are measures of vehicle performance and fuel cost, respectively. Vehicle size is a measure of itself, comfort, and safety. The coefficients of these vehicle attributes in the mean utility are significant and with the expected signs. Consumers prefer vehicles of higher horsepower, larger size, but less fuel cost. Moreover, Vehicles with larger weight are more comfortable and safer. Although the estimate for car weight is not significant, it has positive effect on consumers’ utility. So vehicles with larger weight are preferred. These

findings are consistent with most previous research (Berry et al., 1995; Deng and Ma, 2010;

Xiao and Ju, 2014; Li et al., 2015; Yang et al., 2017).

In terms of customer reviews, ratings over interior room, fuel efficiency, comfort, and price-performance ratio do not have significant effects on consumers. This is because other variables in the model already capture most of the vehicle attributes. For example, ln(price) and ln(horsepower) are good measures of price and performance. So the left unobserved car characteristics of price-performance are insignificant. As shown in Table 3.5, the coefficient on ease to handle is negative and significant, indicating that consumers dislike ease to handle but prefer challenge. The positive coefficient estimates on appearance and interior features suggest that a vehicle with better exterior look and more beautiful interior features is valued more.

The second panel in Table 3.5 also reports the estimates of parameters in the household-

96 Table 3.5: Estimation Results for the Demand Side

Benchmark Specification 1 Specification 2 Specification 3 BLP-with OLS Logit IV Logit BLP-without Variables customer reviews customer reviews Coef. S.E. Coef. S.E. Coef. S.E. Coef. S.E.

Parameters in the mean utility (θ1) Constant -4.9713∗∗∗ 1.7717 -7.1835∗∗∗ 0.3640 -5.5824∗∗∗ 0.5180 -5.0921∗ 2.1179 Ln(price) -8.1140∗∗∗ 1.4044 -1.7647∗∗∗ 0.0643 -4.9780∗∗∗ 0.6962 -7.5769∗∗∗ 1.3667 Ln(horsepower) 2.1955∗ 1.0957 -0.1540 0.0908 2.2884∗∗∗ 0.5354 1.6301 1.0446 Weight (1,000kg) 0.7818 0.9338 -0.2970∗ 0.1297 2.2079∗∗∗ 0.5574 0.3074 1.0105 Energy Expenditure (Yuan/km) -2.6087∗∗∗ 0.5308 -1.9843∗∗∗ 0.1902 -1.0627∗∗∗ 0.2832 -2.5942∗∗∗ 0.4958 Vehicle size 0.8242∗∗∗ 0.0789 0.8115∗∗∗ 0.0367 0.7823∗∗∗ 0.0394 0.8462∗∗∗ 0.0832 Interior room -0.0565 0.0341 -0.0297 0.0295 -0.0347 0.0313 Ease to handle -0.1909∗∗∗ 0.0563 -0.2981∗∗∗ 0.0416 -0.2198∗∗∗ 0.0473 Fuel efficiency -0.0960 0.0992 -0.2208∗∗∗ 0.0296 -0.0277 0.0521 97 Comfort 0.0155 0.0606 0.0502 0.0417 -0.0293 0.0475 Appearance 0.8056∗∗∗ 0.0822 0.6338∗∗∗ 0.0427 0.7962∗∗∗ 0.0572 Interior features 0.2968∗∗∗ 0.1052 0.1232∗∗∗ 0.0379 0.3310∗∗∗ 0.0602 Price-performance ratio -0.2418 0.1526 0.0826∗ 0.0371 -0.2616∗∗∗ 0.0840 City fixed effects Yes Yes Yes Yes Time fixed effects Yes Yes Yes Yes Bodystyle fixed effects Yes Yes Yes Yes Transmission fixed effects Yes Yes Yes Yes Parameters in household-specific utility (θ2) Constant 2.3352 1.3103 2.9302∗ 1.3121 Ln(price) -0.7000∗∗∗ 0.2300 -0.796∗∗ 0.2598 Ln(horsepower) 0.1037 0.0575 0.1406∗∗ 0.0518 Weight (1,000kg) -1.1623∗∗∗ 0.3340 -1.1482∗∗∗ 0.2982 Energy Expenditure (Yuan/km) 2.0517∗∗∗ 0.6112 1.9097∗∗∗ 0.6292 Vehicle size -0.1718 0.0951 -0.2045∗ 0.1043 Ln(price)*Ln(income) 0.4928∗ 0.2288 0.5348∗ 0.2150 GMM Objective Function Value 17.1915 21.3613 Note: ∗∗∗ significant at 0.5%; ∗∗ significant at 1%; ∗ significant at 5%. specific utility as defined by equation (3.7). These parameters capture household heteroge- nous preferences on vehicle attributes. The standard deviation of the taste parameter for constant reflects the variation in the utility difference between a product and the outside good. It is not statistically significant. The estimates for idiosyncratic tastes over horsepow- er and car size are insignificant, implying that households are rather homogeneous in their preferences on horsepower and vehicle size. The random coefficients on price, car weight, and energy expenditure are statistically significant, suggesting consumer heterogeneity due to unobserved household demographics. To better interpret the preference parameters, here we use vehicle weight as an example. The preference parameter on vehicle weight follows a normal distribution with mean 0.7818 and standard deviation 1.1623. This suggests that over 74% of households have positive preference parameters on vehicle weight. Moreover, the estimate on the interaction term between price and income is positive and significant, implying that households with higher income are less price sensitive.

3.4.2 Alternative Specifications

Specification 1 is standard logit model using ordinary least squares regression of ln(sj) − ln(s0) on logarithm of price, product attributes, consumer ratings, and dummy variables.

Specification 2 maintains the functional form of specification 1, but allows the unobserved product characteristics to be correlated with price. To address endogeneity problem, we estimate the logit model by two stage least squares, using the instruments discussed in

Section 3.3.4. As shown in Table 3.5, several of the parameter estimates from OLS logit are

98 quite different from those obtained by IV logit. The price coefficient in IV logit model is about three times bigger in absolute value than that in OLS logit model. The interpretation of this finding is intuitive: the unobserve product attributes, such as product quality, are positive correlated with price, and hence bias the coefficient toward zero without controlling the endogeneity. These results indicate the importance to control for the endogeneity of prices.

The difference between benchmark and specification 2 is that benchmark incorporates household heterogenous preferences on vehicle attributes. We can find that benchmark gen- erates some changes in some parameter estimates because it controls for the impacts of both observed and unobserved household demographics on the preferences. For example, house- hold with higher income is less price sensitive. Families with children, which is unobserved in our study, may prefer larger vehicles. Including household heterogenous tastes induced by observed and unobserved household characteristics makes the model more reasonable.

Hence, The results from benchmark specification allow for a more flexible set of substitution patterns and provide reasonable implications for how the demand will change with prices and product characteristics. We will discuss this later.

Unlike benchmark, specification 3 does not incorporate customer reviews in estimation.

Compared to results from benchmark, several of the parameter estimates in specification

3 change substantially. Specification 3 is based on the assumption that observed product attributes are uncorrelated with the error term. Here, the error term includes consumers’ ratings over products and other unobserved product characteristics. The correlation between observed product attributes and product online ratings may render instruments invalid in

99 specification 3. For example, vehicle weight and horsepower may be negatively correlated with ratings over fuel efficiency, since fuel efficiency is mechanically determined by weight and horsepower (Atkinson and Halvorsen, 1984; Allcott and Wozny, 2014). Generally, heav- ier vehicles burn more fuel. These negative correlations would lead to underestimation of coefficients of weight and horsepower. As shown in Table 3.5, the coefficient estimates on weight and logarithm of horsepower in benchmark are larger than those in specification 3.

This finding suggests that omitting customers reviews in the estimation may violate the exogeneity assumption about observed product characteristics.

Based on previous discussion, the benchmark specification is superior to other specifica- tions. So we will use the estimates from benchmark specification for further analysis.

3.4.3 Elasticities

We first calculate own-price elasticities for each product using parameter estimates from

IV logit. The sales-weighted average own-price elasticity is -4.9770 with a range of -4.9780 to

iv -4.9659. These own price elasticities are computed by ejmt = αiv(1 − sjmt), which are nearly constant because the market share are small. This contradicts to the fact that products with lower prices have lower own-elasticities (in absolute value).

Our benchmark specification allows for more flexible own-price and cross-price elasticities because the model incorporates household heterogenous tastes. Based on the estimates in benchmark specification, we compute price elasticity for each product. The sales-weighted average own-price elasticity is -4.7991 and the range is from -5.7276 to -3.2542. Table 3.6

100 Table 3.6: Price Elasticities for Selected Products in Shenzhen in 2012

MSRP Own-price Cross-price Elasticity ID Manufacturer Nameplate (Yuan)Sales Elasticity 1 2 3 4 1 BYD 2012 BYD F0 1.0L MT 36,900 183 -5.1889 - 0.0021 0.0121 0.0203 2 Changan Suzuki 2012 Suzuki Alto 1.0L MT 39,900 128 -5.1651 0.0030 - 0.0124 0.0209 3 BYD 2012 BYD F3 1.5L MT 50,900 719 -5.1194 0.0031 0.0022 - 0.0244 101 4 BYD 2012 BYD L3 1.5L MT 53,900 1,215 -5.0996 0.0031 0.0022 0.0144 - 5 Changan Ford 2012 Ford Focus 1.6L MT 119,900 380 -4.9493 0.0029 0.0021 0.0139 0.0237 6 Dongfeng Honda 2012 Honda Civic 1.8L AT 139,800 877 -4.8796 0.0028 0.0021 0.0136 0.0232 7 FAW-Toyota 2012 Toyota Corolla 1.6L AT 140,800 2,643 -4.8442 0.0028 0.0021 0.0137 0.0233 8 Dongfeng Honda 2012 Honda CR-V 2.0L AT 193,800 995 -4.8148 0.0027 0.0020 0.0136 0.0232 9 SAIC-VW 2012 VW Tiguan 1.8T MT 199,800 188 -4.8265 0.0026 0.0020 0.0135 0.0230 10 Dongfeng Honda 2012 Honda Elysion 2.4L AT 283,800 153 -4.7830 0.0022 0.0016 0.0126 0.0214 11 Ford 2012 Ford Edge 2.0L AT 292,800 128 -4.7694 0.0020 0.0015 0.0120 0.0205 Note: Columns labeled 1-4 correspond to the 4 products. These price elasticities are calculated using the estimates from the benchmark specification. reports a sample of own- and cross-price elasticities for selected products in Shenzhen in 2012.

The own-price elasticities are larger in magnitude among products with lower prices. The interpretation for this finding is: households who buy more expensive vehicles are generally less price sensitive because they are more likely to have higher income. Moveover, we note that substitution patterns are more flexible. Table 3.6 shows that cross-price elasticities are larger for cars with similar attributes. For example, the cross-price elasticity for 2012 BYD

F0 1.0L MT implies that an increase in its price is most likely to drive households to switch to 2012 BYD F3 1.5L MT, 2012 BYD L3 1.5L MT, and 2012 Suzuki Alto 1.0L MT among the 10 products in the table.

3.5 Counterfactual Analysis

In this section, we examine the effectiveness and welfare effects of vehicle and vessel usage tax incentives, fuel-efficient vehicle subsidy program, and NEV private purchase subsidy pilot program. We conduct the counterfactual experiments using data in 2012 in the five cities.

We assume that the market structure and firms’ pricing strategies do not change, since the total sales of these five cities account for a small portion of the sales in China (5.99% in

2012).

We first conduct four counterfactual experiments. Scenario (I): The vehicle and vessel us- age tax incentives, fuel-efficient vehicle subsidy program, and NEV private purchase subsidy pilot program were not introduced. Scenario (II): Only vehicle and vessel usage tax incentives were introduced. Scenario (III): Only fuel-efficient vehicle subsidy program was introduced.

102 Scenario (IV): Only NEV private purchase subsidy pilot program was introduced. Scenario

(I) will be used as contrast.

In addition, gasoline tax could effectively promote the diffusion of fuel-efficient vehicles and green cars. As a result, gasoline tax has drawn the attention of both policymakers and researchers (Austin and Dinan, 2005; Bento et al., 2009; Parry and Small, 2005; Fullerton and Gan, 2005; Beresteanu and Li, 2011; Xiao and Ju, 2014). So we also compare the above policies against gasoline tax. We stimulate the market outcomes in additional three counterfactual scenarios. Scenario (V): The gasoline tax increases by 0.1 Yuan per liter, while the above tax incentives and subsidy programs are removed. Unlike Scenario (V),

Scenario (VI) and (VII) increase gasoline tax by 0.2 Yuan per liter and 0.3 Yuan per liter, respectively.

To generate the simulations, we first calculate the mean utility under each scenario based on the estimates from the benchmark specification. Then we use Halton sequences to make

150 random draws from the income distribution and unobserved household characteristics distributions. With the parameter estimates and mean utility, we can get each household’s choice probability for each product. Finally, we derive the market shares and sales for each product under each scenario.

3.5.1 Car Sales

In this subsection, we study how the new vehicle sales will change under different policies.

Table 3.7 reports vehicle sales changes under each scenario. Generally, government supports

103 Table 3.7: Sales Changes in the Five Cities in 2012 under Each Scenario

Sales over Fuel Consumption Sales over Engine Size Sales of vehicles with (liters/100km) (L) Scenarios Total Sales government Supports (0,6) (6,8) (8,10) (10,max) (0,1.6) (1.6,2.0) (2.0,3.0) (3.0,max) (1) (2) (4) (5) (6) (7) (8) (9) (10) (11) (I) 922,022 102,996 34,681 501,721 325,432 60,188 442,177 312,767 152,052 15,026 Sales changes under different scenarios

104 (II) 729 1,221 362 583 -191 -25 988 -185 -71 -3 (0.08%) (1.19%) (1.04%) (0.12%) (-0.06%) (-0.04%) (0.22%) (-0.06%) (-0.05%) (-0.02%) (III) 3,695 6,724 1,406 3,679 -1,209 -181 5,450 -1,213 -509 -33 (0.40%) (6.53%) (4.05%) (0.73%) (-0.37%) (-0.30%) (1.23%) (-0.39%) (-0.33%) (-0.22%) (IV) 1,566 2,292 2,396 -532 -267 -31 1,947 -273 -102 -6 (0.17%) (2.23%) (6.91%) (-0.11%) (-0.08%) (-0.05%) (0.44%) (-0.09%) (-0.07%) (-0.04%) (V) -4,309 -444 -135 -2,188 -1,690 -296 -1,981 -1,503 -775 -50 (-0.47%) (-0.43%) (-0.39%) (-0.44%) (-0.52%) (-0.49%) (-0.45%) (-0.48%) (-0.51%) (-0.33%) (VI) -8,839 -892 -266 -4,487 -3,430 -656 -4,037 -3,063 -1,601 -138 (-0.96%) (-0.87%) (-0.77%) (-0.89%) (-1.05%) (-1.09%) (-0.91%) (-0.98%) (-1.05%) (-0.92%) (VII) -13,469 -1,304 -407 -6,804 -5,224 -1,034 -6,126 -4,648 -2,467 -228 (-1.46%) (-1.27%) (-1.17%) (-1.36%) (-1.61%) (-1.72%) (-1.39%) (-1.49%) (-1.62%) (-1.52%) Note: The percentage changes are in parentheses. would increase the utility of vehicles with supports relative to outside good and cars without supports, which in turn would stimulates new purchases and leads consumers to switch from other car models to those with government supports. In contrast, increasing gasoline tax would drive consumers to substitute less efficient vehicles with more fuel efficient ones and even discourage vehicle purchases. Because higher gasoline tax raises vehicle usage costs, which in turn decreases utility of buying vehicles, especially for less fuel efficient vehicles.

Column (1) in Table 3.7 reports changes in total sales under each scenario. The vehicle and vessel usage tax incentives (scenario (II)), fuel-efficient vehicle subsidy program (scenario

(III)), and NEV private purchase subsidy pilot program (scenario (IV)) will increase car sales, while increasing gasoline tax decreases vehicle purchases. For example, the vehicle and vessel usage tax incentives could increase new passenger car sales by 729 units in Shanghai,

Changchun, Hangzhou, Hefei, and Shenzhen in 2012, while total sales drop by 4,309 units with an increase in gasoline tax of 0.1 Yuan/liter.

Column 2 in Table 3.7 presents sales changes of vehicles with government supports.

Generally, these vehicles ave new energy vehicles and fuel efficient vehicles with engine size no greater than 1.6L. The increases in sales for this kind of vehicles are larger than the changes in total sales under scenario (II), (III), and (IV), as shown in column 1. This finding suggests that tax incentives and subsidy programs could lead consumers to switch from other vehicle models. For instance, 3,029 buyers switch to vehicles with government support due to fuel-efficient vehicle subsidy program. On the other hand, column 2 in Table 3.7 shows that sales of vehicles with government supports, i.e., new energy vehicles and fuel efficient vehicles, drop as gasoline tax increases.

105 It is worth noting that the tax incentives and subsidy programs aim to improve fleet fuel efficiency, save energy, and reduce emissions. Since the government often uses fuel consumption and engine size to decide whether a vehicle meets the requirements of the incentives and programs, we categorize cars into various groups by fuel consumption and engine size.74 We then calculate the sales changes for each group under different policies.

Columns 4 to 11 in Table 3.7 show the sales changes over fuel consumption and engine size under each policy. Under scenario (II), (III), and (IV), sales of more fuel-efficient passenger cars and vehicles with smaller engine size will increase, while less fuel-efficient vehicles and cars with large engine size become less popular. Moveover, for those vehicles which do not meet the requirements of government support, the more they close to the requirements, their sales are more likely to drop. For example, the fuel-efficient vehicle subsidy program requires engine size to be no greater than 1.6L. The program causes a larger decrease in group of 1.6-

2.0 than than in group of 2.0-3.0. This result can be explained by the fact that products with similar attributes have larger substitution effects. In addition, columns 4 to 11 in Table 3.7 suggest that the sales decrease nonproportionally among groups due to gasoline tax increase.

Because usage costs of fuel-inefficient cars raise more relative to vehicles that burn less fuel, as gasoline is more expensive.

74For example, the fuel-efficient vehicle subsidy program requires that the engine sizes of the cars are not greater than 1.6L.

106 3.5.2 Gas Consumption and CO2 Emissions

A steady increase in vehicle ownership and usage has been accompanied by large en-

75 ergy consumption and CO2 emissions. So improving the fuel economy of new vehicles

and promoting new-energy vehicles are important to save energy and reduce greenhouse gas

emissions. Here, we estimate the gas consumption and CO2 emissions under each scenari-

o. Moreover, we also calculate the sales-weighted average fuel efficiency since better fuel

economy is associated with lower vehicle emissions (Harrington, 1997).

 PJ The average fuel consumption is calculated by fuel consumptionj ∗ qˆj k=1 qk. We also calculate total gas consumption for the new vehicles sold in these five cities in 2012 during the lifetime of these vehicles. Following Li (2015), we assume that the lifetime of a new vehicle is 10 years. We also assume that the average annual vehicle miles traveled (VMT) is 16,100 km according to 2010 Beijing Household Travel Survey conducted by Beijing Transportation

Research Center. Then the total gasoline consumption in market m at time t under each scenario is given by

X GASmt = qjmt × VMT × FCj × 15 (3.17)

j∈Jmt where qjmt is the total sales of car j, and FCj is the fuel consumption per 100km (liter-

s/100km) of vehicle j. Similarly, we can also derive total electricity consumption.

To compute CO2 emissions, we first need to specify the emission factors of gasoline

75The daily motor gasoline consumption is 1.61 million barrels in China according to United State Energy Information Administration.

107 and electricity. According to Environmental Protection Agency (EPA) 2008, consuming one gallon of gasoline generates 2345.649 gram CO2. And the CO2 emission factor of electricity is 997g/kWh based on the power supply structure in China (Tang et al., 2013). With the total fuel consumption and electricity consumption, we can estimate CO2 emissions for new vehicles sold in 2012 under each scenario.

Table 3.8 presents the impacts of different policies on fleet fuel efficiency, total gasoline consumption, and total CO2 emissions for the new vehicles sold in the five cities in 2012.

Column 1 shows the sales-weighted average fuel consumption (liters/100km) of new vehicles.

The mean fuel consumption would have been 7.8849 without any government supports, compared to 7.8824 with vehicle and vehicle tax incentives, 7.8733 with fuel-efficient vehicle subsidy program, and 7.8648 with NEV private purchase subsidy pilot program. So among the three government support policies, NEV private purchase subsidy pilot program is most effective in improving fleet fuel efficiency. Comparing the effectiveness of these policies with that of increasing gasoline tax, it is shown that an increase of 0.3 Yuan in gasoline taxes would have similar effect on fleet fuel efficiency as that generated by vehicle and vessel usage tax incentives. 76

Columns 2 and 3 in Table 3.8 report the changes in gasoline consumption and CO2 emis- sions over the lifetime of new vehicles.77 Our results show that vehicle and vessel usage tax

76To generate the same impacts on mean fuel consumption as fuel-efficient vehicle subsidy program, the gasoline tax should increase by about 1.5 Yuan. 77The changes are not significant relative to their amounts. There are two reasons. First, we only focus on the new passenger cars sold in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen in 2012. Second, the number of car models eligible for government supports is not large. For example, there are only 9 new energy vehicle models available in our sample.

108 incentives (scenario (II)) and fuel-efficient vehicle subsidy program (scenario (III)) do not

reduce gasoline consumption and CO2 emissions. By comparison, Beresteanu and Li (2011)

find that government incentives, such as income tax credit program for hybrid vehicles, could

reduce gasoline consumption and CO2 emissions in the United States. The interpretation

for our finding is: the government supports have two opposite effects: (i) the gasoline con-

sumption cutting down effect due to consumers switching from fuel-inefficient vehicles and

(ii) gasoline consumption increasing effect due to consumers switching from outside good.

As shown in Table 3.7, numbers of new buyers are larger than the number of switchers under

scenario (II) and (III). Hence, the gasoline consumption cutting down effect is dominated by

the gasoline consumption increasing effect.

In scenario (IV), our analysis shows that NEV private purchase subsidy pilot program

would reduce gasoline consumption. This finding is intuitive because most new energy

vehicles consume electricity instead of gasoline. However, we find that CO2 emissions increase

due to this program. Although NEVs do not generate CO2 directly, they indirectly produce

CO2 by consuming electricity. Generally, a NEV emits less CO2 than a gasoline vehicle for a same distance drive. But the large amount of subsidy stimulates many new purchases, which in turn increases total CO2 emissions. Our estimate indicates that the government revenue decreases by 0.1941 billion Yuan and the reduction in gasoline consumption is 0.01 billion liters in scenario (IV). This suggests that the cost of gasoline consumption reduction using

NEV private purchase subsidy pilot program is 19.41 Yuan per liter. Our results also show that NEV private purchase subsidy pilot program is more effective than the fuel-efficient vehicle subsidy program in controlling gasoline consumption.

109 Columns 2 and 3 in Table 3.8 also show that a higher gasoline tax works better in reducing

gasoline consumption and CO2 emissions than the tax incentives and subsidy programs.

For example, an increase in the gasoline tax of 0.1 Yuan per liter would cut down gasoline

consumption and CO2 emissions by 0.0554 billion liters and 0.1301 million tons, respectively.

3.5.3 Welfare Analysis

We compare the welfare consequences under different policies. In this paper, we define the social welfare as consumer surplus plus government net revenue minus external costs associated with vehicle usage.

Based on parameter estimates from the benchmark specification in Table 3.5, the expected consumer surplus (CS) for household i from the most preferred vehicle is given by

  δj + µij + εij E(CSi) = Eε max (3.18) j=1,...,J MUij

where MUij is marginal utility of money. Since our utility function defined by equation (3.2)

is nonlinear in price, we follow Herriges and Kling (1999) and Li (2015) to use the simulation

method to calculate the estimated CS from the most preferred car.78

Vehicle usage could cause various externalities, such as air pollution and traffic congestion

(Parry et al., 2007; Creutzig and He, 2009; Parry and Timilsina, 2009). Li (2015) argues that

the external costs from automobile usage are 8.7 Yuan (in 2012 term) per gallon of gasoline.

78In each market at time t, we generate 150 households characterized by a vector of random draws from distributions of income and unobserved household characteristics using Halton sequences. For each house- hold, we then simulate 200 draws from the distribution of εij for each product. The CS estimation results are based on 20 replications.

110 Table 3.8: Counterfactual Analysis under Different Scenarios in Shanghai, Changchun, Hangzhou, Hefei, and Shenzhen in 2012

(1) (2) (3) (4) (5) (6) (7)=(4)+(5)-(6) (8) (9) (10)=(7)+(8)-(9) Vehicle- Gasoline Government Consumer External Mean Fuel Total Gas Total CO related Subsidy Social Welfare 2 Tax Revenue Surplus Costs Consumption Consumption Emissions Taxes Scenario (liters/100km)(billion liters) (million tons) (billion Yuan) Panel 1: Counterfactual estimation results under each scenario (I) 7.8849 11.7048 27.4594 59.0447 9.3934 - 68.4381 207.2914 81.7348 193.9948 (II) 7.8824 11.7104 27.4724 59.0153 9.3979 - 68.4132 207.4366 81.7735 194.0763 (III) 7.8733 11.7344 27.5286 58.9668 9.4171 0.1409 68.2431 208.1155 81.9408 194.4178 (IV) 7.8648 11.6948 27.4905 59.0566 9.3854 0.1979 68.2440 207.5630 81.8273 193.9797 (V) 7.8844 11.6494 27.3293 58.7790 10.2838 - 69.0628 206.6999 81.3474 194.4153 (VI) 7.8836 11.5907 27.1918 58.4809 11.1622 - 69.6431 206.0635 80.9381 194.7686

111 (VII) 7.8828 11.5307 27.0510 58.1709 12.0298 - 70.2007 205.4054 80.5191 195.0869

Panel 2: Changes under different policies (Scenario (I) as contrast) (II) -0.0025 0.0056 0.0130 -0.0294 0.0045 - -0.0249 0.1452 0.0387 0.0815 (III) -0.0116 0.0296 0.0692 -0.0779 0.0237 0.1409 -0.1950 0.8241 0.2060 0.4230 (IV) -0.0201 -0.0100 0.0311 0.0119 -0.0080 0.1979 -0.1941 0.2716 0.0925 -0.0151 (V) -0.0005 -0.0554 -0.1301 -0.2657 0.8904 - 0.6247 -0.5915 -0.3874 0.4205 (VI) -0.0013 -0.1141 -0.2676 -0.5638 1.7688 - 1.2050 -1.2279 -0.7967 0.7738 (VII) -0.0021 -0.1741 -0.4084 -0.8738 2.6364 - 1.7626 -1.8860 -1.2157 1.0921 Note: Vehicle-related taxes in column (4) includes value-added tax, purchase tax, consumption tax, and vehicle and vessel usage tax. All money is 2012 RMB Yuan. Scenario (I) assumes that the vehicle and vessel usage tax incentives, fuel-efficient vehicle subsidy program, and NEV private purchase subsidy pilot program were not introduced. Scenario (II) is the case where only vehicle and vessel usage tax incentives were introduced. Scenario (III) is the case where only fuel-efficient vehicle subsidy program was introduced. Scenario (IV) is the case where only NEV private purchase subsidy pilot program was introduced. Scenario (V), (VI), and (VII) are cases where no tax incentives or subsidy programs were introduced but the gasoline tax increases by 0.1 Yuan, 0.2 Yuan, and 0.3 Yuan per liter, respectively. We assume vehicle lifetime of 10 years. The total gas consumption, total CO2 emission, gasoline tax, and external costs are calculated for the new vehicles during their lifetime. The gasoline tax and external costs are calculated based on a discount rate of 5% during the vehicle lifetime. Using the CO2 emission factor of gasoline, we translate the external cost to 3708.9948 Yuan per ton of CO2 emissions. The total external costs are calculated based on an annual rate of 5% during the vehicle’s lifetime.

In China, vehicles are exposed to value-added tax, consumption tax, vehicle purchase tax, and vehicle and vehicle tax. Moreover, vehicle usage can also generate gasoline tax for the government. Total gasoline tax is calculated for the new vehicles sold in the five cities in 2012 during the lifetime of these vehicles based on an annual rate of 5%. In addition, if the government subsidizes purchases of fuel-efficient vehicles or new-energy cars, government expense will increase.

Columns 4 to 10 in Table 3.8 give our empirical results on social welfare. Our estimates show that all the tax incentives and subsidy programs are beneficial to consumers, but at the cost of government revenue decrease and external costs increase. The estimated welfare gain from vehicle and vessel usage tax incentives is 0.0815 billion Yuan in Shanghai, Changchun,

Hangzhou, Hefei, and Shenzhen in 2012. Although the fuel-efficient vehicle subsidy program and NEV private purchase subsidy pilot program decrease government revenue about the same, the fuel-efficient vehicle subsidy program increases social welfare by 0.4230 billion

Yuan but NEV private purchase subsidy pilot program causes a social welfare loss of 0.0151 billion Yuan in 2012. So from the point of view of social welfare, fuel-efficient vehicle subsidy program works best among the three government support policies.

Our analysis shows that an increase in the gasoline taxes could improve social welfare.

Increasing gasoline tax by 0.1 Yuan would generate a welfare gain of 0.4205 billion Yuan, which is nearly the same as fuel-efficient vehicle subsidy program. An increase in the gasoline

112 tax of 0.2 Yuan would improve social welfare by 0.7738 billion Yuan. This suggests that an increase of 0.2 Yuan in gasoline is superior to the above three government support policies in terms of social welfare. Given that tax policy could generate revenue for the government, if the revenue is used to improve public transportation, it would further enhance the welfare effect of gasoline tax.

3.5.4 Government Support versus Gasoline Tax Comparisons

For policymakers, they care more about the impacts of each policy and which policy to choose. Hence, we compare the above government support policies with gasoline tax on various aspects.

First, social welfare. Considering welfare effect, NEV private purchase subsidy pilot program is not a wise choice because it causes welfare loss. Increasing gasoline tax could lead to a higher social welfare and generate revenue for the government to support public transit. As a result, it will be more likely to get political support.

Second, energy security and environmental effect. Our results indicate that vehicle and vessel usage tax incentives and fuel-efficient vehicle subsidy program would increase gas consumption and CO2 emissions, whereas gas tax would achieve the opposite. Although,

NEV private purchase subsidy pilot program would cut down gasoline consumption, it results in an increase in CO2 emissions. To solve energy security and climate change problems, a higher gasoline tax would be a good choice.

Third, technology advance and green auto industry development. As discussed previ-

113 ously, sales of fuel efficient vehicles and new energy vehicles increase under the cases with government supports, while their sales drop as gasoline tax becomes higher. That is, tax in- centives and subsidy programs could stimulate demand for green cars and promote diffusion of green technologies, such as hybrid technology and electric vehicle technology. As sales accumulate, firms can achieve economy of scale and reduce cost through learning by doing, which in turn would lead price to decline.79 And future demand increase could be realized.

Given movement in employees and spillover effect of knowledge, the knowledge accumulated through learning by doing and R&D could benefit the whole green auto industry.

3.6 Conclusion

During the past decades, the Chinese government has become much concerned over en- ergy security and environmental issues. To save energy and reduce pollutant emissions, the

Chinese government offers various incentives for fuel-efficient vehicles and new energy vehi- cles. In this paper, we investigate the effectiveness as well as welfare impacts of (i) vehicle and vessel usage tax (VVUT) incentives, (ii) fuel-efficient vehicle subsidy program, and (iii)

NEVs private purchase subsidy pilot program.

Our empirical results indicate that all these three government support program promote the demand for fuel-efficient vehicles and NEVs, and hence improve fleet fuel economy.

However, we find that VVUT incentives and fuel-efficient vehicle subsidy program increase oil consumption and CO2 emissions. Because the reductions in gasoline consumption and

79Li et al. (2015) find that learning by doing is a main driver for the dramatic decline in China’s automobile market.

114 CO2 emissions due to consumers switching from other fuel-inefficient vehicles are less than

the increases due to consumers switching from outside goods (i.e., new buyers). NEVs private

purchase subsidy pilot program decreases gasoline consumption but raises CO2 emissions,

since most of NEVs consume electricity and produce CO2 indirectly. Our results also suggest

that VVUT incentives and fuel-efficient vehicle subsidy program lead to social welfare gain,

while NEVs private purchase subsidy pilot program cause social welfare loss in the five pilot

cities in 2012.

We also estimate the effects of rising gasoline tax and compare the results with those

from the government support programs. It is found that, relative to increasing gasoline

tax, government incentives is more effective to promote the diffusion of fuel-efficient vehicles

and NEVs. The incentives also works better from the perspective of improving fleet fuel

efficiency. However, an increase in gasoline tax achieves better effects in cutting down oil

consumption and CO2 emissions, and generates higher social welfare, but has an adverse effect on consumer surplus.

115 APPENDIX

A. Proof of Lemma 1

Consider a uniform population density, φ(x) = v at each location x. In Case 1, location costs are constant for all s, i.e., F (s, n) = An.

Second stage: Pricing decisions. Mill pricing. In this case, the first order condition

a−bpm R s+ bt for optimal price under mill pricing (expression (1.7)) becomes 2 0 nv[a − 2bpm −

∗ ∗ a bt|x − s|]dx = 0. Solving for pm, we obtain pm = 3b . Thus, under mill pricing, we obtain

2a 4nva2 4nva3 the market boundaries Rm = s ± 3bt , aggregate output Qm = 9bt , profit Πm = 27b2t − An,

nva3 and social welfare Wd = 4b2t − An.

Discriminatory pricing. Given φ(x) = v and F (s, n) = An, we find market boundaries

a nva2 nva3 Rd = s ± bt , aggregate output Qd = bt , profit Πd = 6b2t − An, and social welfare Wd =

nva3 4b2t − An.

4nva3 First stage: location decisions. Under mill pricing, profit function is Πm = 27b2t −An,

which is independent on s.

nva3 Under discriminatory pricing, profit function is Πd = 6b2t −An, which is also independent

on s.

Hence, both first order conditions hold for all s, indicating the monopolist obtains the

same profits at any location. The market radius, output, and social welfare are not affected

by s, either. Finally, we can easy show that Qd > Qm and Wd > Wm.

116 B. Proof of Lemma 2

Second stage: Pricing decisions Mill pricing. Under a uniformly distributed pop-

ulation density φ(x) = v, the first order condition for optimal price under mill pricing

(expression (1.7)) becomes

a−bpm Z s+ bt 2 nv[a − 2bpm − bt|x − s|]dx = 0 0

∗ ∗ a Solving for pm, we obtain pm = 3b . Thus, under mill pricing, the market boundaries (expres- sion (1.5)), aggregate output (expression (1.8)), profit (expression (1.9)) and social welfare

(expression (1.11)) become 2a R = s ± m 3bt 4nva2 Q = m 9bt 4nva3 Π = − F (s, n) m 27b2t 20nva3 W = − F (s, n) m 81b2t Discriminatory pricing. Still under a uniformly distributed population, the expressions for market boundaries, aggregate output, profit, and social welfare (equations (1.16)-(1.20)) become a R = s ± d bt nva2 Q = d bt nva3 Π = − F (s, n) d 6b2t nva3 W = − F (s, n) d 4b2t 4nva2 First stage: location decisions. Under mill pricing, profit function is Πm = 27b2t −

117 F (s, n). Taking first order condition with respect to s, we obtain

∂Π ∂F (s, n) m = − = 0 ∂s ∂s

nva2 Under discriminatory pricing, profit function is Πd = 6b2t − F (s, n). Taking first order condition with respect to s, we find

∂Π ∂F (s, n) d = − = 0 ∂s ∂s

Hence, both first order conditions indicate that the monopolist will locate at the location

∗ ∗ ∗ where the location cost is minimum, which implies sd = sm = s .

C. Proof of Proposition 1

Under mill pricing, the profit function is shown in expression (1.9). Taking derivative

with respect to firm’s location, we get

 ∗  s+ a−bpm s ∂Π Z bt Z m = btnp∗ φ(x)dx − φ(x)dx m  ∗  ∂s a−bpm s s− bt

∂Πm Given the normal density function φ(x) in equation (1.1), we can find ∂s = 0 for s = 0

∂Πm and ∂s < 0 for s > 0. Thus, the unique optimal location under mill pricing and constant

∗ location cost is the city center, sm = 0.

Under discriminatory pricing, the profit function is shown in expression (1.18), taking

derivative with respect to s, we obtain

s+ a s ∂Π Z bt a − bt(x − s) Z a − bt(s − x) d = tnφ(x) dx − tnφ(x) dx ∂s 2 a 2 s s− bt

118 a Now let r = |x − s|, where r is the distance from the firm’s location and r ∈ [0, bt ]. Each

a−btr ∂Πd consumer at distance r has a demand qr = 2 ≥ 0 under price policy (1.14). Then ∂s becomes Z a ∂Πd bt = t n[φ(s + r) − φ(s − r)]qrdr ∂s 0

∂Πd ∂Πd Given the normal density function φ(x), it follows ∂s = 0 for s = 0 and ∂s < 0 for s > 0. Thus, similarly to mill pricing, the unique optimal location under discriminatory

∗ pricing and constant location cost is the city center, sd = 0.

D. Comparison of equilibrium outcomes in Case 3

∗ ∗ Based on Proposition 1, we know sm = sd = 0. We next compare market radius, profits, output, and welfare in mill and discriminatory pricing.

Market radius. Using (1.5) and (1.16), the market radius under mill pricing and spatial

∗ ∗ a−bpm ∗ a price discrimination are radiusm = |Rm −sm| = bt and radiusd = |Rd −sd| = bt . Because

∗ a a pm ∈ (0, 2b ), it follows that radiusm < bt , so the market area is larger under discriminatory

pricing than under mill pricing when the firm’s location is Given.

Profits. Under discriminatory pricing, the market area can be divided into three regions

∗ ∗ ∗ ∗ a a−bpm a−bpm a−bpm a−bpm a x ∈ [− bt , − bt ), x ∈ [− bt , bt ], and x ∈ ( bt , bt ]. For any market x in the market

∗ ∗ a a−bpm a−bpm a interval [− bt , − bt ) and ( bt , bt ], the firm can make positive net revenue above location

cost under discriminatory pricing, while zero net revenue above location cost under mill

pricing since the demand in this market interval is zero. Under mill pricing, for any market

∗ ∗ a−bpm a−bpm a−bt|x| x in [− bt , bt ], the net revenue above location cost (1.4) is maximized at pm = 2b

119 nφ(x)(a−bt|x|)2 with a value of 4b , which is the optimal net revenue above location cost (1.12)

∗ ∗ a−bt|x| under price discrimination. Since pm is a constant mill price, pm cannot be equal to 2b

∗ ∗ a−bpm a−bpm and maximize the net revenue for every market x ∈ [− bt , bt ]. Thus, discriminatory

pricing yields higher aggregate revenue than under mill pricing. Given the same location

cost of the given location, discriminatory pricing is more profitable than mill pricing.

∗ a−bpm ∗ R bt Output. Since p solves the first order condition (1.7), this means ∗ nφ(x)(a − m a−bpm − bt ∗ 2bpm − bt|x|)dx = 0. Using (1.8) and (1.17), we can calculate the output difference between the two pricing systems:

∗ − a−bpm a Z bt nφ(x)(a − bt|x|) Z bt nφ(x)(a − bt|x|) Q − Q = dx + dx d m ∗ a 2 a−bpm 2 − bt bt ∗ a−bpm 1 Z bt − nφ(x)[a − 2bp∗ − bt|x|]dx ∗ m 2 a−bpm − bt ∗ − a−bpm a Z bt nφ(x)(a − bt|x|) Z bt nφ(x)(a − bt|x|) = dx + dx ∗ a 2 a−bpm 2 − bt bt

> 0

Thus, the output of the monopolist is higher under spatial price discrimination than mill pricing. From above equation, we can clearly see that the output difference between discriminatory and mill pricing is equal to the output gain from the extra market area under spatial price discrimination.

Social welfare. Using (1.11) and (1.20), we can calculate the social welfare difference between the two pricing regimes:

120 a−bp∗ − m 2 a 2 Z bt 3nφ(x)(a − bt|x|) Z bt 3nφ(x)(a − bt|x|) W − W = dx + dx d m ∗ a 8b a−bpm 8b − bt bt

| {z ∗ ∗ } a a−bpm a−bpm a >0,Welfare gain from extra market regions [− bt , − bt ) and ( bt , bt ] 1 + (Π a−bp∗ a−bp∗ − Π a−bp∗ a−bp∗ ) m|x∈[− m , m ] d|x∈[− m , m ] 2 bt bt bt bt

| {z ∗ }∗ a−bpm a−bpm <0,Welfare loss in the nearby market interval [− bt , bt ]

T 0

∗ ∗ ∗ ∗ where Π a−bpm a−bpm and Π a−bpm a−bpm are the profits under mill pricing and dis- m|x∈[− bt , bt ] d|x∈[− bt , bt ]

∗ ∗ a−bpm a−bpm criminatory pricing in market area [− bt , bt ], respectively. As we argued previously,

a−bp∗ a−bp∗ ∗ ∗ ∗ ∗ m m Π a−bpm a−bpm < Π a−bpm a−bpm . This implies that in market area [− bt , bt ], m|x∈[− bt , bt ] d|x∈[− bt , bt ] spatial price discrimination reduces welfare. Relative to mill pricing, discriminatory pricing

∗ ∗ a a−bpm a−bpm a regime serves extra market regions [− bt , − bt ) and ( bt , bt ], where the welfare increases.

∗ ∗ a a−bpm a−bpm a The sign of Wd −Wm depends on the welfare gain from [− bt , − bt ) and ( bt , bt ] and the

∗ ∗ a−bpm a−bpm welfare loss from [− bt , bt ]. Thus, the welfare therefore may be higher or lower under

discrimination than under mill pricing.

E. Simulation description

Under mill pricing, the two-stage model can be formulated as constrained optimization

problem

a−bpm Z s+ bt max πm(pm, s) = nφ(x)pm[a − b(pm + t|x − s|)]dx − F (s, n) pm,s s− a−bpm bt (0.1) a−bpm Z s+ bt s.t. nφ(x)[a − 2bpm − bt|x − s|]dx = 0 a−bpm s− bt

121 The two-stage model under discriminatory pricing can be formulated as constrained op-

timization problem

a Z s+ bt max πd(pd, s) = nφ(x)pd[a − b(pd + t|x − s|)]dx − F (s, n) pd,s s− a bt (0.2) a − bt|x − s| s.t. p = d 2b The integrals can be approximated with Monte Carlo Simulation. Generally, suppose q(x) is density function of x and that we want to compute R g(x)q(x)dx. We can simulate N

−1 PN R draws (x1, ..., xN ) from q(x), and let N i=1 g(xi) be the approximation of g(x)q(x)dx.

In practice, many researchers adopt this technique to approximate integral in their studies

(Berry et al., 1995; Dube et al., 2012; Lee and Seo, 2015).

We set a = b = σ = σF = 1, and A = 0.15. Now we simulate n = 100, 000 artificial

consumers drawn from φ(x). We only analyze the case where s ≥ 0. Analogous results apply

a when s ≤ 0. In footnote section 1.3.1, we also show that pm ∈ (0, 2b ). Thus, the constrained

optimization problem under mill pricing becomes

n 1 X a − bpm a − bpm max πm(pm, s) = (1(s − ≤ xi ≤ s + )npm pm,s n bt bt i=1

[a − b(pm + t|xi − s|)]) − F (s, n) n 1 X a − bpm a − bpm s.t. (1(s − ≤ x ≤ s + )n[a − 2bp (0.3) n bt i bt m i=1

− bt|xi − s|]) = 0 a 0 < p < , s ≥ 0 m 2b

a−bpm a−bpm where indicator function 1(s − bt ≤ xi ≤ s + bt ) takes 1 if xi is in the interval

a−bpm a−bpm (s − bt , s + bt ) and 0 otherwise.

122 Under discriminatory pricing, the constrained optimization problem becomes

 2  1 X a a n(a − bt|xi − s|) max πd(s) = n 1(s − ≤ xi ≤ s + ) − F (s, n) s n bt bt 4b i=1 (0.4) s.t. s ≥ 0

a a a a where 1(s − bt ≤ xi ≤ s + bt ) takes 1 if xi is in the interval (s − bt , s + bt ) and 0 otherwise.

In this paper, we solve the Mathematical Program with Equilibrium Constraints (MPEC) with KNITRO optimization solver (Su and Judd, 2012; Dube et al., 2012). After we find the

∗ ∗ ∗ equilibrium pm, sm and sd, we can use Monte Carlo approximation to get the equilibrium

profits, outputs, consumer surplus, and welfare under both pricing regimes. For example,

∗ ∗ 1 Pn ∗ a−bpm equilibrium output under mill pricing can be approximated by Qm = n i=1(1(sm − bt ≤

∗ ∗ a−bpm ∗ ∗ xi ≤ sm + bt )[a − b(pm + t|xi − sm|)]). We replicate the Monte Carlo simulation 1000

times and find the mean of each variable.

To get Figure 1.2, we first generate a sequence of location (s1, ..., sns) and calculate MCL,

MRLm, and MRLd at each location. For a given location sk, the marginal cost of location

s2 − k Ans 2σ2 √ k F can be obtained by MCL(sk) = 3 e . 2πσF

Under discriminatory pricing, the marginal revenue of location at sk can be approximated

by n   1 X a nt(a − bt|xi − sk|) MRL (s ) = 1(s − ≤ x ≤ s ) d k n k bt i k 2 i=1 (0.5) n   1 X a nt(a − bt|xi − sk|) − 1(s ≤ x ≤ s + ) n k i k bt 2 i=1

Under mill pricing, we need to find the optimal price at location sk. To achieve this, we

123 solve n 1 X a − bpm a − bpm max πm(pm, sk) = (1(sk − ≤ xi ≤ sk + )npm pm n bt bt i=1

[a − b(pm + t|xi − sk|)]) − F (sk, n) a s.t. 0 < p < m 2b ∗ By solving above problem with KNITRO, we get the optimal price pm. Then we can ap- proximate marginal revenue of location at sk under mill pricing by Monte Carlo simulation.

n 1 X  a − bp∗  MRL (s ) = 1(s − m ≤ x ≤ s )nbtp∗ m k n k bt i k m i=1 n (0.6) 1 X  a − bp∗  − 1(s ≤ x ≤ s + m )nbtp∗ n k i k bt m i=1

Finally, we can plot MCL, MRLm, and MRLd and obtain figure 2.

F. City Characteristics in Beijing, Nanjing, Shenzhen, and Tian- jin

G. Derivation of Consumer Surplus of Household

In this appendix, we will derive the compensating variation of household i. In our paper, the utility function is defined by

K X k p uijt(yi, vi, pjt, xjt, δjt, εijt) = δjt + σkxjktvi + (η ln yi + σpvi ) ln pjt + εijt (0.7) k=1

Let wo and w denote the cases without and with policy, respectively. In case without policy, the household chooses product l, i.e.,

wo wo wo wo Uilt (yi, vi, plt , xlt , δlt , εilt) = max uijt(yi, vi, pjt, xjt, δjt, εijt) j=1,..,Jt

124 Table 0.1: City Characteristics in Beijing, Nanjing, Shenzhen, and Tianjin

No. of Average GDP per Average Consumption Year City Households Household Income Capita Expenditure per (10,000) (in RMB Yuan) (Yuan/person) Capita (Yuan/person) 2009 Beijing 636.29 74,866.40 66,940 17,885 2010 Beijing 668.10 81,404.40 73,856 19,929 2011 Beijing 687.86 88,838.10 81,658 21,973 2012 Beijing 704.89 98,466.30 87,475 23,980 2009 Nanjing 230.84 68,351.04 55,290 16,339 2010 Nanjing 237.00 75,876.16 63,771 18,156 2011 Nanjing 240.08 86,940.00 76,263 20,763 2012 Nanjing 241.62 96,979.74 88,525 23,493 2009 Shenzhen 307.10 94,752.24 84,147 21,526 2010 Shenzhen 322.11 104,266.37 94,296 22,807 2011 Shenzhen 332.30 114,990.88 110,421 24,080 2012 Shenzhen 328.58 130,781.43 123,247 26,728 2009 Tianjin 356.92 61,637.79 62,574 14,801 2010 Tianjin 366.20 69,476.84 72,994 16,562 2011 Tianjin 383.35 76,455.24 85,213 18,424 2012 Tianjin 399.92 84,435.27 93,173 20,024 Note: The data are from various issues of yearbook by cities and years. All money is nominal.

125 In case with policy, the household chooses product m, so

w w w w Uimt(yi, vi, pmt, xmt, δmt, εimt) = max uijt(yi, vi, pjt, xjt, δjt, εijt) j=1,..,Jt

The compensating variation (CV) is implicitly defined by

wo wo wo wo w w w w Uilt (yi, vi, plt , xlt , δlt , εilt) = Uimt(yi − CVit, vi, pmt, xmt, δmt, εimt)

Using equation (0.7), we rearrange the above equation and obtain

wo w w w Uilt = Uimt − η ln yi ln pmt + η ln(yi − CVit) ln pmt

Rearranging the equation gives the compensating variation

 wo w  Uilt − Uimt CVit = yi − exp ln yi + w η ln pmt

126 BIBLIOGRAPHY

Agency, Netherlands Environmental Assessment. 2014. “Trends in global CO2 emissions:

2014 Report.”Technical report.

Aguirre, Iaki, Simon Cowan, and John Vickers. 2010. “Monopoly Price Discrimination and

Demand Curvature.” American Economic Review, 100(4): 1601–1615.

Allcott, Hunt, and Nathan Wozny. 2014. “Gasoline Prices, Fuel Economy, and the Energy

Paradox.” Review of Economics and Statistics, 96(5): 779–795.

Alpert, Pinhas, Olga Shvainshtein, and Pavel Kishcha. 2012. “AOD trends over megaci-

ties based on space monitoring using MODIS and MISR.” American Journal of Climate

Change, 1(03): , p. 117.

Anderson, Simon P., Andre de Palma, and Jacques-Francois Thisse. 1989. “Spatial Price

Policies Reconsidered.” The Journal of Industrial Economics, 38(1): 1–18.

Andree, Kai. 2013. “Spatial Discrimination, Nations Size and Transportation Costs.” Inter-

national Economic Journal, 27(3): 385–397.

Atkinson, Scott E., and Robert Halvorsen. 1984. “A New Hedonic Technique for Estimating

Attribute Demand: An Application to the Demand for Automobile Fuel Efficiency.” The

Review of Economics and Statistics, 66(3): 417–426.

Austin, David, and Terry Dinan. 2005. “Clearing the air: The costs and consequences of high-

er CAFE standards and increased gasoline taxes.” Journal of Environmental Economics

and Management, 50(3): 562–582.

127 Baldwin, Richard, and James Harrigan. 2011. “Zeros, Quality, and Space: Trade Theory and

Trade Evidence.” American Economic Journal: Microeconomics, 3(2): 60–88.

Bank, World. 2007. World Development Indicators 2007. Washington, D.C.: World Bank.

Beckmann, Martin J. 1976. “Spatial Price Policies Revisited.” The Bell Journal of Eco-

nomics, 7(2): 619–630.

Beckmann, Martin J., and Jacques-Francois Thisse. 1987. The location of production activi-

ties. 21–95, New York: North-Holland.

Bento, Antonio M., Lawrence H. Goulder, Mark R. Jacobsen, and Roger H. von Haefen.

2009. “Distributional and Efficiency Impacts of Increased US Gasoline Taxes.” American

Economic Review, 99(3): 667–699.

Beresteanu, Arie, and Shanjun Li. 2011. “Gasoline Prices, Government Support, and the

Demand for Hybrid Vehicles in the United States.” International Economic Review, 52(1):

161–182.

Berliant, Marcus, and Hideo Konishi. 2000. “The endogenous formation of a city: popula-

tion agglomeration and marketplaces in a location-specific production economy.” Regional

Science and Urban Economics, 30(3): 289–324.

Berry, Steven, James Levinsohn, and Ariel Pakes. 1995. “Automobile Prices in Market E-

quilibrium.” Econometrica, 63(4): 841–890.

Braid, Ralph M. 2008. “Spatial price discrimination and the locations of firms with different

product selections or product varieties.” Economics Letters, 98(3): 342–347.

128 Center, Beijing Transport Research. 2004-2011. “Beijing Annual Reports of Urban Road

Transport Development (Various Issues) (in Chinese).”Technical report.

Chay, Kenneth Y., and Michael Greenstone. 2003. “The Impact of Air Pollution on Infant

Mortality: Evidence from Geographic Variation in Pollution Shocks Induced by a Reces-

sion.” Quarterly Journal of Economics, 118(3): 1121–1167.

Chen, Chin-Sheng, and Hong Hwang. 2014. “Spatial Price Discrimination in Input Markets

with an Endogenous Market Boundary.” Review of Industrial Organization, 45(2): 139–

152.

Chen, Fei, and Dajian Zhu. 2013. “Theoretical research on low-carbon city and empirical

study of Shanghai.” Habitat International, 37 33–42.

Chen, Yuyu, Avraham Ebenstein, Michael Greenstone, and Hongbin Li. 2013. “Evidence on

the impact of sustained exposure to air pollution on life expectancy from Chinas Huai

River policy.” Proceedings of the National Academy of Sciences, 110(32): 12936–12941.

Chetty, Raj. 2015. “Behavioral Economics and Public Policy: A Pragmatic Perspective.”

American Economic Review, 105(5): 1–33.

Cheung, Francis K., and Xinghe Wang. 1995. “Spatial price discrimination and location

choice with non-uniform demands.” Regional Science and Urban Economics, 25(1): 59–

73.

Claycombe, Richard J. 1996. “Mill pricing and spatial price discrimination: Monopoly.”

Journal of Regional Science, 36(1): , p. 111.

129 Creutzig, Felix, and Dongquan He. 2009. “Climate change mitigation and co-benefits of

feasible transport demand policies in Beijing.” Transportation Research Part D: Transport

and Environment, 14(2): 120–131.

Cropper, Maureen. 2010. What are the Health Effects of Air Pollution in China?. 10–46:

Springer.

Currie, Janet, and Matthew Neidell. 2005. “Air Pollution and Infant Health: What Can We

Learn from California’s Recent Experience?” Quarterly Journal of Economics, 120(3):

1003–1030.

Davis, Lucas W. 2008. “The Effect of Driving Restrictions on Air Quality in Mexico City.”

Journal of Political Economy, 116(1): 38–81.

Deng, Haiyan, and Alyson C. Ma. 2010. “Market Structure and Pricing Strategy of China’s

Automobile Industry.” The Journal of Industrial Economics, 58(4): 818–845.

Dube, Jean-Pierre, Jeremy T. Fox, and Che-Lin Su. 2012. “Improving the Numerical Per-

formance of Static and Dynamic Aggregate Discrete Choice Random Coefficients Demand

Estimation.” Econometrica, 80(5): 2231–2267.

Edlund, Lena, Cecilia Machado, and Maria Micaela Sviatschi. 2015. “Bright minds, big

rent: gentrification and the rising returns to skill.” National Bureau of Economic Research

Working Paper Series, w21729.

130 Eliasson, Jonas, Lars Hultkrantz, Lena Nerhagen, and Lena Smidfelt Rosqvist. 2009. “The

Stockholm congestion charging trial 2006: Overview of effects.” Transportation Research

Part A: Policy and Practice, 43(3): 240–250. of Environmental Protection, Ministry. 2011. “China Vehicle Emission Control Annual Re-

port.”Technical report.

EPA. 2004. “Air Quality Criteria for Particulate Matter.” EPA/600/P-99/002aF (October).

Fullerton, Don, and Li Gan. 2005. “Cost-Effective Policies to Reduce Vehicle Emissions.”

American Economic Review, 95(2): 300–304.

Gallagher, Kelly Sims, and Erich Muehlegger. 2011. “Giving green to get green? Incen-

tives and consumer adoption of hybrid vehicle technology.” Journal of Environmental

Economics and Management, 61(1): 1–15.

Gallego, Francisco, Juan-Pablo Montero, and Christian Salas. 2013. “The effect of transport

policies on car use: Evidence from Latin American cities.” Journal of Public Economics,

107 47–62.

Gibson, Matthew, and Maria Carnovale. 2015. “The effects of road pricing on driver behavior

and air pollution.” Journal of Urban Economics, 89 62–73. de Grange, Louis, and Rodrigo Troncoso. 2011. “Impacts of vehicle restrictions on urban

transport flows: The case of Santiago, Chile.” Transport Policy, 18(6): 862–869.

Greenhut, M. L., and H. Ohta. 1972. “Monopoly Output Under Alternative Spatial Pricing

Techniques.” American Economic Review, 62(4): 705–713.

131 Greenstone, Michael, and Rema Hanna. 2014. “Environmental Regulations, Air and Water

Pollution, and Infant Mortality in India.” American Economic Review, 104(10): 3038–72.

Grg, Holger, Lszl Halpern, and Balazs Murakozy. 2010. “Why do within firm-product export

prices differ across markets?” Working paper.

Gronberg, Timothy J., and Jack Meyer. 1982. “Spatial Pricing, Spatial Rents, and Spatial

Welfare.” The Quarterly Journal of Economics, 97(4): 633–644.

Guo, Wen-Chung, and Fu-Chuan Lai. 2014. “Spatial price discrimination and location choice

with labor markets.” The Annals of Regional Science, 52(1): 103–119.

Harrington, Winston. 1997. “Fuel Economy and Motor Vehicle Emissions.” Journal of En-

vironmental Economics and Management, 33(3): 240–252.

Haughwout, Andrew, James Orr, and David Bedoll. 2008. “The price of land in the New

York metropolitan area.” Current Issues in Economics and Finance, 14(3): .

Herriges, Joseph A., and Catherine L. Kling. 1999. “Nonlinear Income Effects in Random

Utility Models.” The Review of Economics and Statistics, 81(1): 62–72.

Hinloopen, Jeroen, and Stephen Martin. 2016. “Costly location in Hotelling duopoly.” Re-

search in Economics, forthcoming.

Hobbs, Benjamin F. 1986. “Mill Pricing Versus Spatial Price Discrimination Under Bertrand

and Cournot Spatial Competition.” The Journal of Industrial Economics, 35(2): 173–191.

132 Holahan, William L. 1975. “The welfare effects of spatial price discrimination.” The American

Economic Review, 65(3): 498–503.

Hwang, Hong, and Chao-Cheng Mai. 1990. “Effects of spatial price discrimination on output,

Welfare, and Location.” American Economic Review, 80(3): 567–575.

Lee, Jinhyuk, and Kyoungwon Seo. 2015. “A computationally fast estimator for random

coefficients logit demand models using aggregate data.” The RAND Journal of Economics,

46(1): 86–102.

Li, Peiheng, and Steven Jones. 2015. “Vehicle restrictions and CO2 emissions in Beijing A

simple projection using available data.” Transportation Research Part D: Transport and

Environment, 41 467–476.

Li, Shanjun. 2015. “Better Lucky Than Rich? Welfare Analysis of Automobile License

Allocations in Beijing and Shanghai.” Working paper.

Li, Shanjun, Junji Xiao, and Yimin Liu. 2015. “The Price Evolution in China’s Automobile

Market.” Journal of Economics and Management Strategy, 24(4): 786–810.

Liu, Yingqi, and Ari Kokko. 2013. “Who does what in Chinas new energy vehicle industry?”

Energy Policy, 57 21–29.

Martin, Stephen. 2008. Industrial organization in context. Oxford: Oxford University Press.

McMillen, Daniel P. 2003. “The return of centralization to Chicago: using repeat sales to i-

dentify changes in house price distance gradients.” Regional Science and Urban Economics,

33(3): 287–304.

133 Miller, Nathan H., and Matthew Osborne. 2014. “Spatial differentiation and price discrimi-

nation in the cement industry: evidence from a structural model.” The RAND Journal of

Economics, 45(2): 221–247.

Neidell, Matthew J. 2004. “Air pollution, health, and socio-economic status: the effect of

outdoor air quality on childhood asthma.” Journal of health economics, 23(6): 1209–1236.

Nevo, Aviv. 2000. “Mergers with differentiated products: the case of the ready-to-eat cereal

industry.” RAND Journal of Economics, 31(3): 395–421.

Nevo, Aviv. 2001. “Measuring Market Power in the Ready-to-Eat Cereal Industry.” Econo-

metrica, 69(2): 307–342.

Nevo, Aviv. 2003. “New Products, Quality Changes, and Welfare Measures Computed from

Estimated Demand Systems.” Review of Economics and Statistics, 85(2): 266–275.

Ohta, H., and T. Wako. 1988. “The Output Effects of Spatial Price Discrimination Revisit-

ed.” Journal of Regional Science, 28(1): 83–87.

Parry, Ian W. H., and Kenneth A. Small. 2005. “Does Britain or the United States Have the

Right Gasoline Tax?” American Economic Review, 95(4): 1276–1289.

Parry, Ian W. H., Margaret Walls, and Harrington Winston. 2007. “Automobile Externalities

and Policies.” Journal of Economic Literature, 45(2): 373–399.

Parry, Ian WH, and Govinda R Timilsina. 2009. “Pricing externalities from passenger trans-

portation in Mexico city.” World Bank Policy Research Working Paper Series.

134 Petrin, Amil. 2002. “Quantifying the Benefits of New Products: The Case of the Minivan.”

Journal of Political Economy, 110(4): 705–729.

Small, Kenneth A., and Harvey S. Rosen. 1981. “Applied Welfare Economics with Discrete

Choice Models.” Econometrica, 49(1): 105–130.

Small, Kenneth A., Clifford Winston, and Jia Yan. 2005. “Uncovering the Distribution of

Motorists’ Preferences for Travel Time and Reliability.” Econometrica, 73(4): 1367–1382.

StevenBerry, JamesLevinsohn, and ArielPakes. 2004. “Differentiated Products Demand Sys-

tems from a Combination of Micro and Macro Data: The New Car Market.” Journal of

Political Economy, 112(1): 68–105.

Su, CheLin, and Kenneth L Judd. 2012. “Constrained optimization approaches to estimation

of structural models.” Econometrica, 80(5): 2213–2230.

Tan, Lin-Ti. 2001. “Spatial Pricing Policies Reconsidered: Monopoly Performance and Lo-

cation.” Journal of Regional Science, 41(4): , p. 601.

Tang, Bao-jun, Xiao-feng Wu, and Xian Zhang. 2013. “Modeling the CO2 emissions and

energy saved from new energy vehicles based on the logistic-curve.” Energy Policy, 57

30–35.

Tang, Siman, and Hong K. Lo. 2008. “The impact of public transport policy on the viability

and sustainability of mass railway transit The Hong Kong experience.” Transportation

Research Part A: Policy and Practice, 42(4): 563–576.

135 Train, Kenneth E. 2009. Discrete choice methods with simulation.: Cambridge University

Press.

Unit, Economist Intelligence. 2010. “Latin American green city index: assessing the environ-

mental performance of Latin Americas major cities.” Study sponsored by Siemens. Siemens

AG, Germany.

Venkatesh, R., and Wagner Kamakura. 2003. “Optimal Bundling and Pricing under a

Monopoly: Contrasting Complements and Substitutes from Independently Valued Prod-

ucts.” Journal of Business, 76(2): 211–231.

Viard, V. Brian, and Shihe Fu. 2015. “The effect of Beijing’s driving restrictions on pollution

and economic activity.” Journal of Public Economics, 125 98–115.

Vogel, Jonathan. 2011. “Spatial Price Discrimination with Heterogeneous Firms.” The Jour-

nal of Industrial Economics, 59(4): 661–676.

Wang, Yiming, Suwei Feng, Zhongwei Deng, and Shuangyu Cheng. 2016. “Transit premium

and rent segmentation: A spatial quantile hedonic analysis of Shanghai Metro.” Transport

Policy, forthcoming.

Winston, Clifford, and Jia Yan. 2016. “Vehicle Size Choice and Automobile Externalities: A

Dynamic Analysis.” Working paper.

Wolff, Hendrik. 2014. “KEEP YOUR CLUNKER IN THE SUBURB: LOW-EMISSION

ZONES AND ADOPTION OF GREEN VEHICLES.” Economic Journal, 124(578): F481–

F512, Wolff, Hendrik 1; Affiliations: 1: University of Washington; Issue Info: Aug2014,

136 Vol. 124 Issue 578, pF481; Thesaurus Term: AIR pollution; Thesaurus Term: COST

effectiveness; Subject Term: AUTOMOBILE emissions – Environmental aspects; Subject

Term: EMISSION control; Subject Term: ENVIRONMENTAL aspects; Subject Term:

PUBLIC health – Germany; Number of Pages: 32p; Illustrations: 14 Charts, 8 Graphs;

Document Type: Article.

Wolff, Hendrik, and Lisa Perry. 2010. “Policy Monitor: Trends in Clean Air Legislation in

Europe: Particulate Matter and Low Emission Zones1.” Review of Environmental Eco-

nomics and Policy, 4(2): 293–308.

Xiao, Junji, and Heng Ju. 2014. “Market Equilibrium and the Environmental Effects of Tax

Adjustments in China’s Automobile Industry.” Review of Economics and Statistics, 96(2):

306–317.

Xiao, Junji, and Xiaolan Zhou. 2013. “An Empirical Assessment of the Impact of the Vehicle

Quota System on Environment: Evidence from China.” School of Management, Fudan

University.

Xiao, Junji, Xiaolan Zhou, and Wei-Min Hu. 2017. “Welfare Analysis of the Vehicle Quota

System in China.” International Economic Review, 58(2): 617–650.

Yang, Jun, Ying Liu, Ping Qin, and Antung A. Liu. 2014. “A review of Beijings vehicle

registration lottery: Short-term effects on vehicle growth and fuel consumption.” Energy

Policy, 75 157–166.

137 Yang, Ziying, Felix Munoz-Garcia, and Manping Tang. 2017. “Effective or Efficient? Effects

of Beijing’s Vehicle Lottery System on Fleet Composition and Environment.” Working

paper.

Zheng, Siqi, and Matthew E. Kahn. 2013. “Understanding China’s Urban Pollution Dynam-

ics.” Journal of Economic Literature, 51(3): 731–72.

138