Evaluating the Electric Vehicle Subsidy Program in China

Transcription

1 Evaluating the Electric Vehicle Subsidy Program in China Jing Qian October, 2018 Abstract China has become the world s largest market for electric vehicles (EVs) since 2015 and the government promotes the technology aggressively by providing large subsidies for EV buyers. The amount of subsidy is based on the driving range instead of the battery capacity as in the U.S. This paper evaluates the impacts of the subsidy program using detailed vehicle registration data in China from 2010 to 2015 and a household survey of vehicle ownership. I develop and estimate a market equilibrium model for China s automobile market in which the demand side consists of a random coefficient discrete choice model and the supply side characterizes automakers pricing decisions under the government subsidy program. The estimation suggests that while the subsidy program in 2015 contributed to 94 percent of EV sales in large cities, the program favored small and low-quality EV models that consumers do not value and led to a $2.88 billion loss in social welfare. The hypothetical subsidy program based on the battery capacity would have led to a $0.62 billion increase in consumer surplus and a $0.2 billion increase in social welfare compared with the subsidy program. Jing Qian is a Ph.D. candidate in the Dyson School of Applied Economics and Management at Cornell University. jq58@cornell.edu; Address: 410 Warren Hall, 137 Reservoir Ave, Ithaca NY I thank Shanjun Li, Panle Jia Barwick, Jura Liaukonyte, and Sumudu Watugala for their guidance and support of this project. I also thank the useful comments from Eric Zou, the 20th CU Environmental and Resource Economics Workshop participants, Camp Resources XXV participants, and the Dyson school AEP seminar participants. Data for this research is generously supported by China National Information Center, Shanjun Li, and Panle Jia Barwick. Lastly, I wish to thank Congyan Han and Binglin Wang for their help in collecting data.

2 1 Introduction Since 2015 China has become the world s largest market for electric vehicles (EVs), overtaking the United States which has been the forerunner in electrification. Sales of EVs in China have grown rapidly from 8,159 in 2011 to nearly 580,000 in 2017, accounting for half of the global EV sales. As the largest emissions producer 1 and oil importer, electrifying transportation is essential to addressing the environmental problems and reducing exposure to oil price validity and security risks since electricity is domestically sourced. In addition, accelerating the development of the EV industry is also important to the economy as the ban on internal combustion engine (ICE) vehicles is becoming more popular around the world. The trend of electrifying transportation provides an opportunity for China s domestic automakers in developing EV technologies to embrace the future of the automotive industry. 2 Growth of EVs is largely driven by government incentives. However, this tends to vary greatly across countries. All zero-emission cars in Norway, the country with the world s largest EV market share, are exempted from value-added tax (VAT) and registration tax. In the United States, the second largest EV market, the federal government provides tax credits capped at $7,500 for EV buyers which is based on the battery capacity. EV buyers also enjoy state incentives. For example, Clean Vehicle Rebate Project (CVRP) provides EV buyers with up to $2,500 in California. In China, this is quite different, and the subsidy is based on the driving range of battery electric vehicles (BEVs), which increases as the driving range rises. The central subsidy for BEVs ranged from 31,500 Yuan ($4,854) to 54,000 Yuan ($8,322). In the case of Plug-in Hybrid Electric Vehicles (PHEVs), the subsidy was 31,500 Yuan ($4,854) in In addition to the central subsidy, EVs also received subsidies from local governments. The majority of the local governments provided a subsidy proportional to the central subsidy in a fixed ratio for each city. Due to these aggressive incentives, China has witnessed a boom in EV sales. However, the program causes distortion in consumer choices. Figure 1 shows that bunching is found just above 150km according to the driving range of BEVs which accounted for nearly 70% of EV sales in BEVs with a driving range just above 150km saw a subsidy growth of about 43%. In fact, the most popular EVs in China are the small and low-quality vehicle models produced by young, domestic, and private firms (Ou et al., 2017). 3 One reason for the domination of small vehicles is that firms intend to obtain the subsidy using the lowest production cost in a short time. For instance, 1 In 2014, the transportation sector corresponded to 23% of global carbon dioxide emissions (International Energy Agency 2016) and 30% of PM2.5 (World Health Organization). In 2009, China became the world s largest market in the automobile sector. The transportation sector in China was estimated to be responsible for 7-8% of national carbon dioxide emissions (China s Energy Efficiency and Conservation). 2 Foreign automakers are leading in the ICE vehicle industry. In China, private automakers and SOEs only account for about 28% of gasoline car production. 3 In a Chinese article published by the Ministry of Finance in China, about 66% of the EVs are micro EVs regarding to the ten most popular EVs. Less than 20% of the EVs apply high technology. 1

3 the popular BEV model Kangdi K11 is based on its gasoline powered version dubbed the Panda, so that firm does not need to completely redesign the vehicle. The other main reason is that the willingness to pay (WTP) for EVs is low in China. According to a report by UBS Evidence Lab, Chinese consumers would consider purchasing an EV only if the price of the EV is less than that of an equivalent ICE vehicle. Small and low-quality EVs usually have a low price after subsidies so that they are more attractive to the consumers. To understand the impact of subsidies on consumer surplus and social welfare, I estimate a market equilibrium model in the framework of Berry, Levinsohn, and Pakes (1995) (henceforth BLP) and Petrin (2002). The demand side is a random coefficient discrete choice model and the supply side characterizes automakers pricing decisions under the subsidy program. I supplement the city-level sales data (macro-moments) with the household survey of vehicle ownership (micromoments) which relates household demographic characteristics to household choices. Based on the model and parameter estimation, I construct a counterfactual in which there were no subsidies on EVs. While 94% of EV sales in 2015 were driven by the subsidy program, it led to an billion Yuan ($2.88 billion) loss in social welfare. Then, I compare the consumer surplus under the subsidy program based on the driving range to a counterfactual where the subsidies for EVs were based on the battery capacity. The results show that there would be an increase in the range, size, weight, and horsepower of BEVs. The consumer surplus would increase by 4.03 billion Yuan ($0.62 billion), and the social welfare would also increase by 1.27 billion Yuan ($0.2 billion) in In addition, I confirm the opinion that the environmental benefits of EVs are highly related to locations. Since coal-generated electricity is still the primary energy source in most cities, growth in EV sales has little or even a negative impact on reducing air pollution. This paper makes the following four contributions to the existing literature. Firstly, this study examines the efficiency of the subsidy program, especially the impacts of the subsidy on consumer surplus and social welfare, which contributes to the literature on EV market from another aspect. Previous studies have examined the design and effects of financial incentives on EV adoption (Sierzchula et al. (2014), Borenstein and Davis (2016), Clinton and Steinberg (2016), and DeShazo et al. (2017)). Sierzchula et al. (2014) use the data from 30 countries. Borenstein and Davis (2016), Clinton and Steinberg (2016), and DeShazo et al. (2017) focus on the U.S. market. All papers find a significantly positive impact of subsidies on EV sales. Borenstein and Davis (2015) and DeShazo et al. (2017) also take subsidy distribution into consideration. Their studies show that high-income EV buyers receive most of the subsidies. Progressive rebates or aggressive rebates with price caps are superior to a single rebate for EVs regardless of family income from the cost-effectiveness analysis. However, they do not take network externalities into consideration. Springel (2016) and Li et al. (2017) find that charging station investment and EV adoption respond positively to each other in Norway and the U.S. market. Besides investment in charging stations, 2

4 Li (2016) concludes that manufacturer investment in charging standards to make them compatible would have a positive impact on EV adoption as well. Another stream of literature explores efficiency of subsidies from an aspect of pollution (Zivin et al. (2014) and Holland et al. (2016)). They find that the environmental benefits of EVs vary across locations depending on the electricity generation mix. This paper investigates the limitation of technology-based subsidies through focusing on the consumer purchasing decisions. In particular, I evaluate the distortion in consumer choices resulting from the unique subsidy program which subsidizes EVs in terms of their driving range. Secondly, the study adds to the existing literature on evaluating the consumption responses to energy efficient programs. Boomhower and Davis (2014), Houde and Aldy (2014), and Chen et al. (2017) point out the presence of inframarginal consumers in evaluating effectiveness of energy efficient programs. Chen et al. (2017) find that 53% of the subsidies for fuel efficient cars in China were ineffective and distributional. However, under the EV subsidy program, most of the EV consumers are marginal consumers rather than inframarginal consumers, and the program inefficiency mainly comes from the deviation from their best choices. Thirdly, to my knowledge, this paper is the first one to evaluate the EV subsidy program based on a structural method incorporating both demand and supply side in China. Helveston et al. (2015) design a survey to estimate consumer preferences for conventional vehicles, hybrid vehicles (HEVs), and EVs in China and the U.S. Ma et al. (2017) use city-level aggregate data and regressions to study the impact of various policies including subsidies on EV adoption. This paper differs from prior works on the EV market in China by utilizing registration data at the city-model-quarter level as well as an equilibrium model of the Chinese automobile market. The approach is closely related to Beresteanu and Li (2011) which examine the impacts of gasoline prices and income tax incentives on HEV adoption and Barwick et al. (2017) which investigates the local protection in China. This approach would allow me to do different counterfactuals to see the effects on the whole automobile market and social welfare. Finally, this paper provides policy implications to China, other developing countries, as well as developed countries. Recently the Chinese government issued a new energy vehicle (NEV) credit mandate which is supposed to go into effect in However, credits for BEVs are still dependent on the electric range. This paper addresses the welfare consequences of subsidizing EVs in terms of their driving range, which would provide a timely reference for the Chinese government. Since China is quite different from other developed countries and early EV adopters, other developing countries can learn from China s experience to understand what works and what does not when making policy decisions. More importantly, since we are all under the same dome (Chai Jing), air pollution produced in China can affect populations all over the world. The technological advancement in China also has a spillover impact on other countries. The rest of the paper is organized as follows. Section 2 describes the industry and policy 3

5 background of the Chinese EV market, presents the data, and discusses descriptive evidence of positive correlation between subsidies and EV adoption. Section 3 describes the market equilibrium model. Section 4 reports results from the structural estimation. Section 5 discusses the quantitative impact of the subsidy program using simulations. Section 6 quantifies the welfare impact of the subsidy program. Section 7 concludes. 2 Background and Data In this section, I first present an introduction of the government support especially subsidies on the development of the Chinese EV industry and the outcome of the current policies for promoting EVs. I then discuss the data and descriptive evidence of the impact of subsidies on EV sales. 2.1 Policy Background On September 17, 2013, China s Ministry of Industry and Information Technology (MIIT), together with the Ministry of Finance (MoF), the Ministry of Science and Technology (MoST), and the National Development and Reform Commission (NDRC) issued a policy Regarding the Continuous Promotion and Application of New-Energy Vehicles which decided to provide a one-time subsidy to eligible 4 EVs solely in terms of their driving range. Import EVs, such as Tesla, are not included on the subsidy list. According to the Notice, BEVs with a driving range between 80km and 150km, 150km and 250km, and above 250km were granted a one-time subsidy of 35,000 Yuan ($5,394), 50,000 Yuan ($7,705), and 60,000 Yuan ($9,246) Yuan, respectively in PHEVs with a driving range above 50km were given 35,000 Yuan ($5,394) in Additionally, the subsidy would decrease by 10% and 20% in 2014 and 2015, respectively. However, the reductions were revised later to 5% and 10%. The central subsidies across years are summarized in Table 1. Compared with the earlier policies issued in May , this policy extended the subsidy coverage from 5 to 25 pilot cities and changed the subsidy criterion from battery capacity to driving range. Unlike other financial incentives like tax credits or rebates, the subsidies were directly allocated to firms. The price that consumers paid was the market suggested retail price (MSRP) reduced by subsidies. 4 Vehicles on the Energy Conservation and New Energy Vehicle (NEV) List can get a subsidy from the central government. Except for GX2, Lifan 320, and Panda, all the other domestic or joint-venture EVs on the market are on the list. 5 MoF issued the subsidy document Financial subsidy interim measures for private purchase of new energy vehicle in pilot cities. The 5 pilot cities were Shanghai, Changchun, Shenzhen, Hanghzou, and Hefei. The subsidies for private purchasing depended on the battery capacity. The standard was 3000 Yuan/kwh (462 $/kwh). The maximum subsidy for PHEV and BEV was 50,000 Yuan ($7,705) and 60,000 Yuan ($9,246), respectively. 4

6 Some local governments also provided subsidies for EVs. Figure 2 shows the dates when the local governments started to provide subsidies for the private purchasing of EVs in the 19 first two tier cities 6 that are covered in this analysis. All the 19 cities excluding Jinan offered local subsidies. Before 2014, only four cities, Shenzhen, Hangzhou, Guangzhou, and Shanghai offered local subsidies. In Shanghai, the subsidy amount was fixed to 30,000 Yuan ($4,623) for PHEVs and 40,000 Yuan ($6,164) for BEVs from 2013 to In order to get the subsidy, the EVs must be included in the local list of Shanghai. Besides Shanghai, Hangzhou also offered a fixed amount of subsidy for EVs starting from December Previously, the subsidy amount in Hangzhou was based on the battery capacity from 2011 to From May 2014, consumers in Nanjing and Wuxi were eligible to get a 25,000 Yuan ($3,953) subsidy for purchasing a BEV and a 15,000 Yuan ($2,312) subsidy for purchasing a PHEV from their provincial government. Later, the provincial government announced to subsidize EVs in terms of their wheelbase in March In addition to the subsidy from provincial government, EV consumers in Nanjing also received a fixed amount of subsidies, 35,000 Yuan ($5,394) for BEVs and 20,000 Yuan ($3,082) for PHEVs from local governments. Except for Shanghai, Hangzhou, Nanjing, and Wuxi, the other 14 cities began to offer EV buyers with a subsidy proportional to the central subsidy in a fixed ratio for each city starting from 2014, 7 which means the local subsidies were also based on the driving range of BEVs. Table 2 demonstrates the maximum local subsidy an EV consumer received in the 18 cities in The subsidies for PHEVs were less than those for BEVs. Similar to Shanghai, Beijing had its own subsidy catalogue which excluded PHEVs to get the local subsidy. The amount of subsidy was quite large relative to the MSRP of EVs especially small BEVs. For example, the maximum total subsidy for the popular BEV model Geely Zhidou was 95,000 Yuan ($14,640) which was almost 60% of its MSRP. An upper bound was set by governments to avoid excessive subsidies in some cities 8. In addition to subsidies, EVs included in the catalogue issued by MIIT are exempted from the 10% sales tax starting from September Besides financial incentives, in Shanghai, Beijing, Guangzhou, Tianjin, Hanghzou, and Shenzhen where consumers are restricted from purchasing new vehicles, owners of eligible EVs are exempted from the purchase restriction. EV buyers are either assigned to a separate lottery pool for EV applicants only or granted a plate. Owners of eligible EVs also receive exemption from driving restrictions in Beijing 9, Chengdu, and Wuhan. 6 The ranking is based on China Business Weekly in Guangzhou provided a flat subsidy of 10,000 Yuan for EV purchasing before December 2014, and then it started to subsidize EVs according to the 2013 central subsidies. Similar to Guangzhou, Shenzhen offered 60,000 Yuan ($9,246) for BEVs and 30,000 Yuan ($4,623) for PHEVs from July 2010 to May 2013, and then it started to subsidize EVs according to the 2013 central subsidies from The local government in Hangzhou set the limit at 50% of the MSRP. The upper bound in Changsha, Qingdao, Guangzhou, Wuhan, Chongqing, Xi an, and Shenyang was 60%, and the upper bound in Xiamen, Nanjing, and Wuxi was 80%. 9 In Beijing, only EVs on the local list are unrestricted from the car purchase and driving restrictions 5

7 2.2 The Chinese EV Industry Chinese EV market has experienced an explosive growth since 2014 (Figure 3). In 2017, China accounted for more than half of the global EV sales and had a market share of 2.2% (Global EV Outlook 2018). The first two tier cities account for 74.26% of the national EV sales from 2010 to City-level PHEV and BEV sales are illustrated in Figure 4. We can observe that there are a lot of variations in sales across both cities and fuel types. In general, BEVs are significantly preferred than PHEVs in China. However, the popular BEV models are mainly low-end products with small size, light weight, and small horsepower. Unlike the ICE vehicle industry where joint ventures (JVs) are the largest producers, the dominate firms in the EV industry are private and domestic firms, such as BYD, Geely and Zotye. Figure 5 illustrates that mini-compact BEVs dominate in the Chinese EV market and more than 50% of the BEVs are mini-compact or subcompact cars. As a result, almost 90% of the BEVs have a weight less than the average weight of all vehicles on the market (Figure 6). According to Anderson and Auffhammer (2013), fatality probability increases by 47% when being hit by a vehicle which is 1,000 pounds heavier, indicating that BEV drivers are exposed to higher fatality probability in an accident. For PHEVs, their size and weight are larger since they use ICE as a backup besides electric motor. The above market outcome is quite different from the U.S. market where the EVs are mainly produced by primary firms. U.S. consumers have high income and view EVs as a symbol of high social status. The most popular EV model, Tesla Model S with a range between 398 to 504 km, responded to 22% of all EV sales in There are two major reasons that small and low-quality EV models are popular in the Chinese EV market. One is that the government subsidies allow the private and domestic firms to quickly gain profits in the low-end segment under the rapidly expanding market. The other reason is that inexpensive EV models provide consumers an opportunity to obtain a vehicle plate which is really difficult to get in some megacities especially Beijing and Shanghai. Although EV buyers can also get a plate through purchasing an import EV model. Import EV models, such as Tesla S, are still too expensive for Chinese consumers to afford. Despite the high manufacturing cost, Import models are also subject to 25% tariff. In addition, the import models are excluded from subsidies and sales tax exemption. As a result, the market share of import EVs is low. The sales of Tesla in 2015 were 3,692 which accounted for more than 80% of all import EVs, but constituted only about 2% of total EV sales in China. 10 In summary, the Chinese EV market is still dominated by low-end products while some new firms, such as LVCHI Auto and NIO, have started to focus on developing EVs with high performance using advanced technologies. NIO introduced a high performance SUV model, ES8, in ES8 has a 70kWh battery capacity, a 355 km driving range, a 240kW horsepower, 10 The sales data of Tesla is from autohome website and the percentage of Tesla among all import EVs is from China Automotive Technology & Research Center (CATARC). 6

8 and a fast acceleration (4.4s from 0 to 100km/h). 2.3 Data The analysis is based on five main data sets: (1) individual vehicle registration data obtained from the State Administration of Industry and Commerce, (2) model-level attributes from major automotive websites, (3) Government incentives for EVs collected from government and major automotive websites, (4) household-level car ownership survey data compiled by Ministry of Industry and Information Technology (MIIT), (5) city-level household income data from City Annual Statistical Yearbook. The vehicle registration data contain the universe of car purchases in China from 2010 to For each record, we observe the month and county of registration, the owner s type, and the firm, brand, and model name of the purchased vehicle. Model is defined by model-fuel type-vehicle type-transmission type. Engine size and model code are also included to enable the match with detailed vehicle attribute data. This study focuses on the private purchasing which accounts for 90% of all registration records and excludes import vehicles due to data limitation. In China, the imports only account for 3.1% of total sales from 2009 to 2011 (Barwick et al. 2017). In addition, the market share of foreign EV automakers is low (around 2% in 2015). I aggregate the data of the first two tier cities to the model-quarter-city level. The first two tier cities account for 74.26% of the national EV sales from 2010 to The number of EV models and sales across years is shown in Table 3. The number of EV models increases significantly, and the largest increase is in In 2015, EV models accounts for 13% of all fuel type models. The number of BEV models is larger than that of PHEV models. The penetration rate of EVs researches 3.6% in 2015 which is more than triple of the national average rate. These cities also have large variations in EV sales due to the time variations in the implementation of incentive policies. To translate the aggregated sales into market shares, I compute market share through dividing sales by market size which is defined as one fourth of the annual number of households in each city since the observation is at quarter level. For each observation in the vehicle attribute data, I include MSRP, horsepower, fuel efficiency, and size in the analysis. The summary statistics for the 97,765 observations are reported in Table 4. According to Barwick et al. (2017), MSRP which includes value-added tax and consumption tax is a reasonable approximation of transaction price. Sales tax paid by consumers is on top of the MSRP, and usually set at 10% but reduced to 7.5% for vehicles with engine displacement no more than 1.6 liter in The price consumers pay for gasoline and hybrid vehicles is MSRP plus sales tax. Approved EVs receive sales tax exemption beginning from September 1,

9 so the price consumers pay for these qualified vehicles is just MSRP 11 minus subsidy from then on. The mean of the real price, deflated to the 2015 level, is 167,566 Yuan ($25,823), and the mean of the real price of EVs is much lower (about $21,583) due to high subsidies. Fuel cost of each gasoline or hybrid model is based on the fuel consumption per 100km. I multiply it by province-year gasoline price to get the average fuel cost (yuan per km). For BEV models, I first obtain electricity consumption per km through dividing battery capacity by range. Then I multiply it by nation-year electricity price 12. For PHEV models, I assume 45 percent of miles are driven on gasoline and 55 percent of miles are driven on electricity based on the formula provided by the Office of Energy Efficiency and Renewable Energy (EERE). Besides attributes, incentives play an important role in EV purchase decisions in China. I collect incentives including central and local subsidies, free license plate, and exemption from driving restrictions at model-city-month level from government documents and major automotive websites, and then aggregate the data to model-city-quarter level. The cross-city variations in subsidies are shown in Table 2. For instance, subsidies for a BEV with driving range no less than 250km vary from 25,000 Yuan ($3,853) to 60,000 Yuan ($9,246). The above data sets demonstrate consumer purchase decisions and consumer choices. Since income is an important factor when making purchasing decisions, I obtain empirical distributions of household income at the city level from City Annual Statistical Yearbook. The data were collected from anuual city-level household surveys. In the survey, a number of urban households were randomly drawn and equally divided into 5 groups to get the mean of disposable income per capita in each group. For each city in each year, I assume a log normal distribution of disposable income per capita and obtain predicted group means and average mean of whole sample. Then I get the simulated mean and standard deviation for each city in each year by minimizing the differences between predicted and observed group means, and the difference between predicted and observed average sample mean as well. In the demand estimation, the income of pseudo individuals in each city in each quarter follow the simulated log normal mean and standard deviation in corresponding city and quarter. To convert disposable income per capita into disposable household income, I multiple the prior number by which is the average nationwide average family size in To link the household income and their purchasing decisions, I use a unique household-level car ownership survey data. The survey was conducted by China National Information Center from 2009 to It contains household level data on vehicle stocks, vehicle purchasing year, vehicle attributes, and household demographics, such as household income, family size, and education. I extract the 19 cities in my study from the survey and only keep new and domestic cars which were purchased between 2010 and The total number of households in these 19 cities is 6,097 from 11 The MSRP for some EV models are subtracted by firm subsidies which only apply to EV models. 12 The gasoline and electric price are collected from CEIC. 13 The average family size is obtained from the National Bureau of Statistics. 8

10 2010 to There are 1,617, 1,255, 1,290, 1,072, 625, and 238 households from year 2010 to The key variable is household after-tax income in purchasing year. Table 5 shows the income distribution among all vehicle buyers. About half of the households have annual income no less than 150,000 Yuan ($23,116). It is quite intuitive that households with high income are more likely to purchase cars. In addition, we see that the income distribution of household purchasing cars shifts to the right across years. Table 6 exhibits the fraction of buyers from each income group in each vehicle segment in It can be concluded that high income households are more likely to purchase medium/large cars and SUVs while low income households are more likely to purchase small cars. In the demand estimation, I match the simulated fraction of buyers from each income group in each year with the observed fraction in corresponding income group and year. Similarly, I match the simulated fraction of buyers from each income group in each vehicle segment in each year with the observed fraction in corresponding income group, vehicle segment, and year Descriptive Evidence To investigate the impact of subsidies on EV adoption, Figure 7 shows 12-month rolling average of EV sales in four representative cities. An increase in the sales of EVs is found in all four cities just after the implementation of local subsidies which are on top of the central subsidy. It seems that the central subsidy itself introduced in late September, is not large enough to stimulate EV sales. The sum of central and local subsidies contributes to the growth of EV sales. The significant increase in EV sales just after the subsidy is also observed in other cities with EV sales no less than 1,000 from 2010 to 2015 (Appendix Figure 1). The time variation indicates a significantly positive correlation between subsidies and EV sales. There are two other variations that help me identify the impact of subsidies on EV sales. We can see from the top right figure that the sales of BEVs are much higher than those of PHEVs in Beijing since Beijing doesn t provide any subsidies for PHEVs, which provides a cross variation. Another variation comes from the change in the subsidy criterion. Jiangsu provincial government first provided 25,000 Yuan ($3,853) for purchasing a BEV and 15,000 Yuan ($2,312) for purchasing a PHEV. In March 2015, the provincial government started to subsidize EVs based on their wheelbase. As a result, we observe a decrease in the sales of BEVs as the subsidies for small BEVs 16 decreases in the bottom right figure. To control for some confounding factors affecting EV sales, I examine the data by regressing 14 Since the number of total buyers in each segment in 2014 and 2015 is not large enough, I only use the information from 2011 to 2013 for the vehicle segment match. 15 Shanghai started to get central subsidies from June In Nanjing, the percentage of mini-compact or subcompact BEVs was one quarter among all EVs in

11 the logarithm of new vehicle registration on the set of EV incentives by controlling for a set of fixed effects given in regression equation (1). ln(sales ct j ) = βsubsidy ct j + γt E ct j + V ct je ω e + ν t + ν c j + ν c ν yr + ε ct j, e (1) where the observation is defined by vehicle model j in month t and city c, and all vehicle models are included. Subsidy and tax exemption (10,000 Yuan) are two financial incentives. Subsidy is the sum of available central and local subsidies only for EV buyers. Sales tax exemption (T E) is the amount of sales tax that is exempted which does not only apply to EVs. The rate of sales tax is usually set at 10% in China, while gasoline and hybrid models with engine displacement smaller than 1.6 liter were only subject to 7.5% in 2010 and 5% in November and December in Buyers of eligible EVs don t need to pay any sales tax starting from September Besides financial incentives, the other EV incentives (V ) include free plate and driving restriction exemption are controlled. I also include month fixed effects (ν t ) to control for seasonality and city-by-model fixed effects (ν c j ) to control for city specific vehicle model preferences, time-invariant city demand shocks, and time-invariant vehicle model attributes. The OLS regression results are demonstrated in Table 7. Compared with Column (1), Column (2) further controls for city-by-year fixed effects (ν c ν yr ) which provide city specific macroeconomic controls such as GDP, household income, and charging station expansion. The results of the two specifications are quite robust. Column (2) shows that a 10,000 Yuan ($1,541) increase in the subsidy amount is associated with a 9% increase in EV sales on average, holding the other control variables constant. Sales tax exemption also has a positive impact on stimulating EV adoption. The medium value of sales tax exemption for EVs is about 6,600 Yuan ($1,071). If the sales tax exemption incentive is removed, the sales of the EV with medium exemption amount would decrease by 8.6%. In addition to the financial incentives, we observe that both free plate and driving restriction exemption have significantly positive impact on EV adoption. These non-monetary incentives will be controlled in the demand estimation. 3 Empirical Model In this section, I discuss the structural model and estimation strategy. The model is closely related to recent empirical literature (e.g., BLP (1995), Petrin (2002), and Barwick et al. (2017)). Firstly, I model household s vehicle purchasing decision through a random coefficient discrete choice model. Then, I model the supply side taking the impact of subsidies into consideration and assuming Bertrand competition. Under optimal pricing, I can recover the marginal costs of vehicle models. 10

12 3.1 Demand A market is defined as a city. In a given quarter t, household i chooses from J mt vehicle models and an outside good to maximize his utility. The indirect utility ν of household i choosing product j in city m and quarter t is defined as u mti j = ν(p mt j,x t j,ξ mt j,y mti,d mti ) + ε mti j, (2) where p mt j is the vehicle price households pay which takes taxes and subsidies for electric vehicles into consideration. The price varies across markets due to the difference in local subsidy amount. X t j is a vector of of observed product attributes. ξ mt j is the unobserved product attribute. y mti and D mti represent for income and other household attributes. Due to the data limitation on the distribution of observed household attributes other than income, I only include the determinant factor income in this analysis. The specification of the indirect utility ν is assumed as ν mti j = (eᾱ e σ pν mti ) ln(p mt j) y mti + x t jk β mtik + ξ mt j, k (3) where eᾱ represents the base level of price sensitivity. e σ pν mti captures the consumer heterogeneity in disutility of price. In this term, ν mti has a standard normal distribution, and σ p is the standard deviation of the normal distribution. x t jk is the kth product attribute of product j. β mtik stands for consumer heterogeneous taste for attribute k which is defined as β mtik = β k + σ k ν mtik (4) where β k is the mean preference for product attribute k which is constant across all markets and quarters. ν mtik follows a normal distribution and σ k is its standard deviation which stands for household i s preference for attribute k. The utility function can be fully written as u mti j = x t jkβk + k e (eᾱ e σ pν mti ) ln(p mt j) y mti V mt je ω e + F j + η t + S j + τ m η yr + ξ mt j + x t jk σ k ν mtik + ε mti j, k (5) In the above equation, the first x t jk includes constant, a dummy for electric vehicles, horsepower, fuel cost, size, a dummy for automatic transmission. V mt j represents EV incentives including free plate and exemption from driving restriction. F j, η t, S j, and τ m η yr are firm dummies, quarter dummies, vehicle segment dummies, and year by city interaction dummies, respectively. Firm dummies, month dummies, and vehicle segment dummies control for firm loyalty, seasonality, 11

13 and preference for large cars, respectively. The interaction dummies of year by city absorb timevariant city attributes such as changes in the city public transportation and macro shocks such as gasoline price fluctuation. In addition, I include random coefficients for outside good (constant term), vehicle size, and price which are important attributes when making purchasing decisions. I denote the parameters to be estimated as θ which equals to (θ 1,θ 2 ). θ 1 and θ 2 represent linear and nonlinear parameters, respectively. The utility function can be decomposed into a common utility δ(θ 1 ) and a heterogeneous utility µ(θ 2 ). u mti j = δ mt j + µ mti j + ε mti j, (6) δ mt j = x t jkβk + V mt je ω e + F j + η t + S j + τ m η yr + ξ mt j, k e (7) µ mti j = (eᾱ e σ pν mti ) ln(p mt j) y mti + x t jk σ k ν mtik. k Household i chooses the vehicle with the highest utility. Let κ i be the vector of unobserved individual attributes. The market share of product j is given by exp(v mti j ) s mt j (p,x,ξ,y i ;θ 2 ) = 1 + J mt l=1 exp(v mtil) df(κ). 3.2 Supply A supply side is needed to back out marginal costs to do counterfactual analysis. I assume a Bertrand competition and national pricing. 17 A firm sets a national optimal price (MSRP) denoted by p 0 j for each vehicle model to maximize its annual national profit. In China, the price consumers pay consists of pre-tax price which includes consumption tax and value-added tax which is based on the pre-tax price for domestic vehicle models, the unit revenue firm gets is much smaller for vehicles with a large engine size. For example, the consumption tax for vehicles with a engine size between 2 to 2.5 liter, 2.5 to 3 liter, and 3 to 4 liter is 9%, 12%, and 25%, respectively. The unit revenue firm gets (p f m j ) is (I suppress subscript t for simplicity in this section. m not only represents cities, but also includes 4 quarters for each city in a given year. Thus, the definition of market is city by quarter instead of city): (8) (9) p f m j = p0 j b m j 1 + t va (1 tm c j) + b m j, j (10) where p 0 j is product j s MSRP which is the same across cities and within a year. b m j is the subsidy for EV model j which is directly allocated to the firm and not subject to any taxes. For models other 17 National pricing in this paper refers to optimize automakers profits in terms of the total profits in the 19 cities. 12

14 than EVs, b m j equals to 0. The price consumers pay for EV models and other models is p 0 j b m j and p 0 j, respectively. tva j is value-added tax which is equal to 17%. t c j is consumption tax which depends on the engine size and is constant within a year for a given model. For a given year, the annual profit for firm f is π f = = M m=1 M m=1 (p f m j mc j)m m s m j j F [τ j p 0 j + (1 τ j )b m j mc j ]M m s m j, j F (11) where τ j equals to 1 tc j 1+t va which represents the percentage of amount firm f gets among the price j consumers pay. Subsidy b m j is directly allocated to the firm and exempted from tax, which is adjusted by term (1 τ j )b m j. Each firm chooses {p 0 j, j F } to maximize its total profits. Given this assumption, p0 j satisfies the following first-order condition (FOC): M τ j m =1 M m s m j + r F (1 τ r ) M m =1 M m b mr s mr p 0 j + (τ r p 0 r mc r ) r F M m =1 M m s mr p 0 j = 0, j (12) To compute the term s mr, I use the price consumers pay (p p 0 m j ) to connect the demand and j supply side. The price consumers pay taking subsidy and sales tax into account is p m j = (p 0 j b m j ) + p0 j b m j 1 + t va t s j, j (13) s mr p 0 j = s mr p m j p m j p 0 j = s mr p m j t j, (14) where t j = 1+tva j +t s j 1+t va and t s j is sales tax. j Define the national sales M m s m j as S j, and as J by J matrix, whose ( j,r) term is S r if r p 0 j and j belong to the same firm, and 0 otherwise. B j represents the second term in equation 12. The FOC can be organized as: τp 0 1 B = mc + 1 (τs) (15) 13

15 3.3 Estimation The key variable to identify the effectiveness of subsidies on the total sales of EVs is price. In addition to subsidies, EVs are also exempted from purchase and driving restrictions. I control for these two factors in mean utility δ mt j. In this study, I do not distinguish the effects of exemption from tax and subsidies. Both of them are reflected by price. Another issue is that price is correlated with unobserved product attributes which are represented by ξ mt j. Besides excluded variables, I use the number of products produced by other firms in the same vehicle segment and with the same fuel type, the number of products produced by the same firms in the same vehicle segment and with the same fuel type, and the central subsidy amount EVs received as the exogenous variables. The first two variables capture the market power and competition affecting firms pricing decision. Central subsidies are for EVs only and based on the driving range. The subsidy for BEVs ranged from 31,500 Yuan ($4,854) to 54,000 Yuan ($8,322) in It increases as driving range rises and decreases across years. EVs with a high driving range (above 250km for BEVs) need more technology and manufacturing investment and usually have high quality. The assumption that the instruments are uncorrelated with unobservables provides the first moment: E[ξ mt j (θ 1,θ 2 ) Z mt j ] = 0, (16) where the unobserved individual attributes have been integrated over in equation (9) and Z mt j includes all excluded and exogenous instruments. The second set of moment conditions follow (Barwick et al. (2017)) but have different data source. Table 5 and Table 6 contain the observed fractions of buyers in each income group, and the observed fractions of buyers in each income group by vehicle segments from the householdlevel car ownership survey data. The estimation requires model predicted fractions to match with the observed fractions. For example, to compute the predicted fractions of buyers in each income group in 2015, I multiply the predicted fraction of buyers in each income group in each city-quarter by the observed total sales in that city-quarter in Then I aggregate the sales of each group to the 2015 national level. The fractions of buyers in each income group in 2015 is obtained through dividing the sales of each group by the 2015 national sales in the certain year. There are 90 micromoments in total. With an initial value of θ 2, I use contraction mapping to recover δ mt j. The objective function is formed by stacking the above two sets of moment conditions. The estimation is carried out using the identity matrix as the weighting matrix to obtain the optimal weighting matrix. I then estimate the model using the optimal weighting matrix. 14

16 4 Estimation Results In this section, I first discuss evidence from the reduced-form regressions on the sales impact of the subsidy policy in the top two tier cities using two basic sets of results. These are a simple logit specification and an instrumental variables logit specification. Then I present the parameter estimates from the random coefficient discrete choice model. 4.1 Reduced-form Regressions To provide evidence on the sales impact of subsidy policy, I estimate the following logit specification for the utility function. lns mt j lns mt0 = x t jkβk + V mt je ω e + F j + η t + S j + τ m η yr + ξ mt j, (17) k e where the dependent variable is the log market share of product j minus log market share of outside good in city m and quarter t. x t jk are vehicle attributes including constant, log vehicle price, EV dummy, horsepower, fuel cost, size, transmission type. V mt j represents EV incentives including free plate and exemption from driving restriction. Firm dummies, month dummies, vehicle segment dummies, and city-by-year interactions are also controlled as I discuss in Section 3. Column (1) in Table 8 reports the results of OLS applied to the logit specification. Due to the endogeneity of price, the coefficient of log price is biased towards zero, implying inelastic demands of the products. To control for unobserved quality which is correlated with price and product market share, I use instruments for price in Column (2). The coefficient of log price increases a lot in absolute value. The first F-test of the instruments is 12.35, indicating that the instruments are valid. The own price elasticity is -4.52, which is plausible. Although we obtain a plausible own elasticity, the IV logit can not provide plausible cross elasticities. In addition, it does not incorporate income impact which is an important factor in making purchasing decisions and household heterogeneous preference for vehicle attributes. Using the own elasticity, I can compute the EV sales without subsidies. I find that the sales of PHEVs would reduce from 54,770 to 999, and all BEVs would exit the market in 2015 due to a extremely high proportion of subsidy to the BEV price. If subsidy is reduced to the half, the sales of PHEVs would decrease to 9,471, and the sales of BEVs would decline to 271. I will come back to the estimates for comparison after obtaining parameter estimates from the random coefficient discrete choice model. 4.2 Parameter Estimates from the Random Coefficient Model The consumer demand for vehicles derived from the utility function in equation (5) is contained in Table 9. The first specification is the benchmark specification. I will compare the specification 15

17 with the other three specifications in the next section. The linear parameters are those in the mean utility function in equation (7) which reveal households references for vehicle attributes. The price coefficient and random coefficients are in the household-specific utility function in equation (8). The first specification indicates that households prefer vehicle models with large horsepower, more fuel-efficiency, large size, and automatic transmission. Exemption from purchase and driving restrictions are the two non-monetary incentives affecting households purchasing decisions. The results confirm that these two factors play a significant role in stimulating EV sales. The channel is that the price of vehicle plates is significantly larger than that of EV models in megacities such as Shanghai. In Shanghai, the average bid for a license was more than 92,000 Yuan ($14,178) in March 2013 (Li, 2014), much higher than the after subsidy price of many low-end EVs. I control for these two incentives to exclude their positive impact on the EV adoption. The random coefficients stand for the standard deviation of household preferences from the mean for vehicle attributes. The random coefficient of constant captures households heterogeneous preference for purchasing a car. The preference parameter on it has a standard normal distribution with mean 0.2 and standard deviation -6.03, indicating that 95% of the households have a parameter on size in the range of [-11.62,12.02]. The key variables capturing the effects of subsidies on the sales of EVs are base level price sensitivity (eᾱ) and price random coefficient σ p. The term eᾱ/y i represents price sensitivity of consumers with different income. It is intuitive that consumers with higher income are less price sensitive. σ p represents dispersion in price disutility indicating that consumers with the same income have different price sensitivity. For example, given two consumers (with price random draw equals to 1 and -1, respectively) purchasing the same vehicle model in the same market at the same quarter, the demand elasticity of the former consumer is nearly 2.3 times of the latter one. The price random coefficient helps to reflect the phenomenon that Chinese consumers have lower income but are willing to purchase expensive cars compared with the U.S. consumers. To understand the magnitude of the estimation results in the specification (1), I plot the 2015 own-price elasticities and Lerner index against price for the 305 models in Figure 8 and 9. The own-price elasticity ranges from -21 to For vehicle models with the same price, the elasticity of EV model is smaller than that of the gasoline or hybrid model, indicating households are less price sensitive to EV models. The sales-weighted elasticity in 2015 is -6.42, while the magnitude of sales-weighted elasticity of EVs is a bit larger which is This is because low-end EVs dominate the Chinese EV market, and buyers of these models tend to be more price sensitive compared with buyers of high-end EV models such as Geely S60L and BMW-Brilliance 530Le which have the lowest elasticities. Figure 8 shows that vehicle models with a higher price tend to 18 The second largest elasticity is -13. The model with the largest elasticity is a low-end EV, Chery QQ3 which has the lowest price in our data. 16

18 have a lower elasticity. The magnitudes of the own-price elasticities are similar to those estimated from the U.S. automotive market in Petrin (2002) and Beresteanu and Li (2011). This study focuses on the first two tier cities where the new car buyers have relatively high income with the mean at 188,076 Yuan ($28,984) and cars have become a necessity for households, though average income of U.S households is much higher than that of Chinese households. Besides, the sales-weighted own-price elasticity is between those estimated from the Chinese automotive market in Barwick et al. (2017) and Li (2014). In addition to own-price elasticities, I compute the marginal cost for each vehicle model from firms FOC in (12). Based on the marginal cost, the price-cost margins can be computed as p f j mc j, where p f j is the unit revenue firm gets. The price-cost margins are reported in Figure 9. The salesweighted margin in 2015 is 17.46% which is similar to Petrin (2002) (16.7%) and Beresteanu and Li (2011) (17.72%). For EV models, I observe that some of them with small own-price elasticities have low margins. The reason is that the marginal costs for EVs are still high especially in early years due to the high battery cost. According to the world s largest battery production company Contemporary Amperex Technology Co. Limited (CATL), a Chinese company, the average selling price of EV batteries for automakers produced by itself was 2.89 and 2.28 Yuan/wh (0.45 and 0.35 $/wh) in 2014 and 2015, respectively. 19 For a BEV with a 20 kwh battery capacity, the average battery cost was roughly 45,600 to 57,800 Yuan ($7,027 to $8,907) in 2015 given that the battery supply was from CATL. The battery cost is even larger than the price of low-end gasoline vehicle models. Since battery accounts for up to half of an EV s cost of production 20, the estimated lower bound of the cost of an EV with 20 kwh battery capacity could research 91,200 to 115,600 Yuan ($14,055 to $17,815). p f j 4.3 Robustness Checks The model fit of specification 1 is shown in Table 10 and 11. From Table 10, the model fit is decent with the largest prediction error at 3% and the average prediction error at about 1%. For the second set of Micro-moments, the prediction also fits well, which is demonstrated in Table 11 with the average prediction error at 2% in The fit in year 2010 to 2012 is also decent. Table 9 also shows three robustness checks for the benchmark specification. The robustness checks show that the benchmark estimation is robust to: (1) the definition of market size; (2) the distribution of draws for random coefficients; (3) the income group cutoffs to generate micromoments. In the bench specification, I assume that all households are potential new car buyers 19 The data is described in an article published by Bloomberg Opinion. The detailed information on CATL can also be found in the Chinese article. 20 In a article published by China Briefing. 17

19 in the market as is often used in the study on U.S. automotive market since this study focuses on the first two tier cities where the income of the households is relatively high compared with other cities, and the demand for cars grows rapidly in China. Specification 2 assumes that half of the households are potential new car buyers, implying that the market size of a quarter is the number of annual households divided by 8. The coefficients of the specification 2 are similar to those of the benchmark specification except that the random coefficient of size is smaller. This implies that the coefficient of size in the mean utility of the specification 2 is larger than that of the benchmark specification. Specification 3 removes the impact of extreme draws for the random coefficients. Instead of using unbounded random draws in the benchmark specification, this specification drops the bottom 2.5% and top 97.5% draws. The estimates are again close to those in the benchmark specification and the larger random coefficient of size implies a smaller coefficient of size in the mean utility. In the specification 1, I group households with annual income less than 60,000 Yuan as the first group, households with annual income no less than 60,000 Yuan and less than 100,000 as the second income group, households with annual income no less than 100,000 and less than 150,000 as the third income group, and households with annual income no less than 150,000 as the fourth income group. The cutoffs are close to the 25th, 50th, and 75th percentile income according to the city-level income distribution from the demographics data. Another reason is that about 16% and 12% of the surveyed new car buyers have nominal household income at 100,000 and 150,000, respectively. setting the cutoffs at 100,000 and 150,000 makes sure that the peaks are included in the same group across years. In the specification 4, I set the cutoffs at 60,000, 115,000, and 165,000. The estimated coefficients are close to those in the benchmark specification as well. I will show the simulation results and welfare analysis in the below sections which are similar to the corresponding results of the benchmark specification. 5 Counterfactuals The previous sections estimate a random coefficient model and recover the marginal cost for each vehicle model. In this section, I first conduct a counterfactual scenario to evaluate the subsidy program based on the driving range (baseline program), especially the effects of the program on consumer choices. Then I conduct a counterfactual scenario to compare the baseline program with an alternative program which subsidizes EVs based on their battery capacity in terms of their effects on consumer choices and social welfare. I use the following methodology. Firstly, I change the FOC according to different scenarios. Then, I solve the new equilibrium prices and sales using the demand parameter estimates and product marginal costs. In this study, the product on the market would stay the same under the different scenarios. To the extent that the baseline program favors 18

20 small and low-quality EVs and therefore induce automakers to offer more such kinds of EVs, the counterfactual analysis would underestimate the true effects of the program on consumer choices. The other thing is that Nanjing and Wuxi are excluded in the following counterfactual analysis since the provincial subsidies for those two cities were based on the wheelbase in Sales without subsidy To evaluate the impact of the baseline program on prices, I solve for new equilibrium prices and sales without any subsidies using equation (12) and assuming the second term to be zero. Table 12 shows the changes in sales by fuel type after removing subsidies in It is not surprising that 94% of the EV sales were induced by the baseline program. The reduction in the sales of PHEVs is less than that of BEVs. One reason is that subsidies for BEVs are much higher than those for PHEVs. Compared with the results using logit IV, we see that the sales of EVs removing subsidies would be larger using the random coefficient model. One of the reason is that the random coefficient model incorporates consumer heterogeneity. For example, the random coefficient of constant is large in Table 9, indicating that some consumers have strong preference to purchase a car even if the subsidy is removed. The random coefficient model also captures the substitutions among cars with different fuel types. We can find that about 63% of EV buyers would purchase gasoline or hybrid models instead. Regarding to the specific vehicle models, high-end models, Geely S60L and BMW-Brilliance 530Le, would have the least percentage decrease (27% and 38%, respectively) in their sales since these models have low own-price elasticities. The sales of most of the EVs would decrease by more than 90%, and some low-end models would even exit the market. Figure 10 shows the distribution of BEV range without subsidies. It is observed that the percentage of BEVs with driving range between 150km and 160km would decline significantly, from 55% to only 6%. BEVs with driving range no less than 360km would be the most popular models in the market. Correspondingly, Figure 11 and 12 illustrate that large and heavy BEVs would dominate the EV market without subsidies. More specific, the sales of mini-compact BEVs would decline the most from 50,392 to only 102, followed by subcompact BEVs, while the sales of BEVs belonging to MPV would decline the least from 3,939 to 437. For PHEVs, the sales of compact cars would decrease the most from 39,253 to 3,313, while the sales of medium or large cars would reduce the least from 668 to 416. On the other hand, we would see a growth in the sales of gasoline vehicles. The largest increase would be found in compact and small SUV segments. In addition to performance improvement in driving range, size, and weight, BEVs with higher horsepower are preferred shown in Figure 13. All results imply that Chinese households have an underlying preference for BEVs with high performance. However, the subsidy policy based on the driving range results in distortions in their choices. 19

21 The sales without subsidies under the robustness checks (specification 2 to 4) are displayed in Table 13 Panel A. The sales without subsidies by fuel type are consistent across different specifications and the sales of EVs would decrease by 93% to 95% without subsidies. 5.2 Subsidies Based on the Battery Capacity The first counterfactual analysis provides evidence that the baseline program does favor small and low-quality BEVs, which deviates government s policy goal aimed at the development of EV technology. Since battery is the most crucial part for an EV, one alternative subsidy method is to subsidize EVs based on their battery capacity instead of driving range. The alternative subsidy program is similar to the U.S. subsidy program. To achieve the same environmental goal, the sales-weighted average fuel efficiency of all vehicle models under the alternative subsidy program should be the same as that under the baseline program. To have the same miles per gallon (MPG), the alternative subsidy program is designed as follows. The program would provide 6,500 Yuan (about $1,000) as the base subsidy for any EVs. Every additional kwh battery capacity greater than 8kwh would receive 4,900 Yuan (about $755), and the largest subsidy amount would be 130,000 Yuan (about $20,000). The new equilibrium prices and sales are contained in Table 14. Since BEVs usually have a larger battery, the alternative subsidy program favors BEVs than PHEVs. We can see that the sales of BEVs would increase by 38%, and the sales of PHEVs would decline by 55%. The total sales of EVs would rise by 726. Under the alternative program, given the same driving range, BEVs with a large battery would receive a larger subsidy, which reduces the sales of BEVs with a long driving range but a small battery. It implies that the alternative subsidy program prevents automakers to make BEVs smaller instead of applying a larger battery to achieve a longer driving range. Figure 14 compares the range distribution under the baseline program and the alternative subsidy program, respectively. Under the alternative subsidy program, the bunching at the 150km would decline significantly. The percentage of BEVs concentrated between 150km and 160km would be 14%. It implies that the large subsidies under the baseline program induced households to choose the BEVs with driving range just above 150km but with a small battery capacity. Under the alternative subsidy program, 30% of the BEVs have a driving range between 200km to 240km. We can also observe from Figure 15 that the percentage of mini-compact BEVs would decrease from more than 60% to less than 15% under the alternative subsidy program. Most of the BEVs would be subcompact cars, and the percentage of BEVs belonging to compact and MPV segments would grow from about 10% to 30%. As a result, BEVs on the road would be heavier (Figure 16), leading to a better safety for BEV drivers. The distribution of horsepower would also shift to the right (Figure 17), leading to a higher horsepower. More specifically, the average sales- 20

22 weighted range, size, weight, and horsepower would rise from 167km to 203km, 5.95 m 2 to 7.15 m 2, 32.18kw to 45.20kw, and kg to kg, respectively. The results indicate that under the alternative subsidy program, it would be more affordable for households to purchase BEVs with better performance compared with the baseline program. The results also imply that the WTP for small and low-quality BEVs is rather small (I simulate the WTP of some small and low-quality BEV models in the Appendix). The subsidies for these kinds of BEVs under the alternative subsidy program would not be large enough to stimulate their sales. However, the performance of the BEVs under the alternative scenario is still worse than that without subsidies. It implies that there still exist consumer distortions under the alternative subsidy policies though this policy could correct some consumer distortions caused by the baseline program. The sales under the alternative subsidy policy in the robustness checks (specification 2 to 4) are demonstrated in Table 13 Panel B. The sales under the alternative policy by fuel type are again consistent across different specifications. The sales of PHEVs would decrease by 54% to 58% and the sales of BEVs would grow by 36% to 40%. 6 Welfare Analysis We have observed distortions in households choices due to the baseline program in the above two counterfactual analysis. This section examines the welfare consequences of this subsidy program separately for consumer surplus, CO 2 emissions, firm profits, and government expenditures. 6.1 Theoretical Framework for Welfare Estimation The purpose of this section is to illustrate the welfare impact of the subsidy program. To simplify the analysis, I look at the marginal (e.g., the increase in the sales of EVs itself) social welfare impacts of the program and assume linear supply and demand curves in the market for EVs. According to Holland et al. (2016), the second-best subsidy for an EV should be the difference in lifetime damages between an EV and a gasoline vehicle. In China, since the generation mix highly relies on coal, the environmental benefits from driving EVs are different across cities. Besides operation, EV production and battery disposal add extra environmental costs. Qiao et al. (2017) find that the GHG emissions from a BEV production is about 50% higher than that from an ICE vehicle production. Battery waste is another environmental issue in China. The amount of lithium battery waste is estimated to be up to 170,000 metric tons by 2020, 21 but there are no explicit regulations on lithium waste. 21 The data is according to a news by Reuters. 21

23 Figure 18 Panel (a) presents a scenario in which social benefits of an EV are smaller than private benefits. Due to the subsidy, the equilibrium quantity of BEVs rises from Q 0 to Q 1. Although there is a gain in the consumer and producer surplus which is represented by ACGE, government needs to pay AFGE to stimulate the sales. In addition, there is a loss (CGHD) induced by the externalities since the marginal social benefits are smaller than the marginal private benefits. The social welfare loss in the scenario is shown by CFHD. For an EV which brings positive externalities, the efficient way is to provide a subsidy which is equivalent to the marginal benefit. Figure 18 Panel (b) illustrates the scenario that the social benefits of an EV are larger than the private benefits, but the subsidy is higher than the optimal subsidy. As a result, the producer provides more products from Q to Q 1 under the excessive subsidy. Compared with the optimal subsidy scenario, there is a gain in consumer and producer surplus which is EDHI and an environmental benefit represented by CGHD. However, the government has to pay the sum of EDHI, CGDH, and CFG to achieve the new equilibrium quantity. The social welfare loss is CFG. Besides marginal social welfare impacts, cross-sectional substitution between vehicle types and within vehicle types is also an important dimension in our setting. As we have observed in Section 5, the baseline subsidy program which subsidizes EVs in terms of their driving range favors small and low-quality EVs induces households to deviate from their top choices to a product they would never choose without the subsidy program. This deviation results in a loss in consumer surplus. For example, without subsidies household i in market m would choose the top product A which gives a consumer surplus of $5,000. Due to the large subsidies preferring small and low-quality products, the household would choose product B which provides only $2,000 consumer surplus. Suppose that the government provides the household with $4,000 to make the deviation happen. In this case, there would be a $3,000 loss in consumer surplus. 6.2 Consumer Surplus This section examines consumer surplus change in details that comes from the deviation from the best choices under the baseline program. The consumer surplus can be computed using the log sum method (I suppress subscript t for simplicity. The definition of market m is city by quarter instead of city): EU mi = E[max(u mi j )] = max j=0,...,j {δ m j + µ mi j } = ln( exp(δ m j + µ mi j )) + C, j (18) where C is Euler constant, and the term ln( j exp(δ m j + µ mi j )) effectively represents the expected consumer i s welfare in utility. The monetary value of consumer surplus can be defined as follows: 22

24 CS mi = EU mi / EU mi p m j, (19) where EU mi p m j is the marginal utility from a dollar. Since I am using a logarithm form of price in the utility function, I assume the marginal utility from a dollar as: EU mi p m j = α mi E( 1 p m j ), where α mi = (eᾱ e σ pν mi) y mi and E( 1 p m j ) is the expected value of the inverse price of each vehicle model. After getting the individual consumer surplus in dollars, I average the monetary consumer surplus across consumers for each market and multiply it with the market size to obtain the total consumer surplus in dollars. The results are contained in Table 15. We can observe that the baseline subsidy program resulted in a 7.83 billion Yuan ($1.21 billion) consumer surplus loss compared with the scenario that subsidies were removed in Compared with the alternative subsidy program, the baseline program led to a 4.03 billion Yuan ($0.62 billion) consumer surplus loss in The results confirm the existence of consumer distortions due to the subsidy program based on the driving range. It also implies that households WTP for EVs is still low in China, subsidies are necessary to stimulate EV sales. (20) 6.3 CO 2 Emissions One possible benefit of the baseline program is that it favors small EV models which are usually considered to be more environmental friendly compared with gasoline models and relatively large EVs. However, the production and exploitation of EVs in China actually results in more greenhouse gas (GHG) emissions by Mark Buchanan, 22 and the environmental consequences are highly dependent on locations. In this analysis, I focus on the fuel cycle, which follows the study of Peng et al. (2017) and Ou et al. (2012). According to their study, the fuel cycle includes resource exploitation and transportation, fuel production, transmission, distribution, storage and filling, and vehicle use. I exclude damages from other pollutants: NO x, VOCs, PM 2.5, and SO 2 which are included in Holland et al. (2016). The intertemporal substitution of vehicle purchases is excluded either. The GHG emissions are calculated based on the lifetime of the vehicle. The GHG emissions intensities of generating electricity and gasoline are contained in Table 16. Electricity generated by fossil fuels produces much larger carbon emissions compared with electricity generated by non-fossil fuels. The unit GHG emissions produced by gasoline is one third of that produced by fossil fuels, which indicates that for provinces where electricity generation is highly dependent on coal, such as Tianjin and Shanghai (Table 17), driving EVs with low 22 This is from an article in Bloomberg. It points out that only if the electricity is produced in a relatively clean way, the EVs are environmental benign. 23

25 fuel efficiency even worses the environment. However, for provinces relying on non-fossil fuels, such as Sichuan and Hubei (Table 17), driving EVs does help reduce carbon emissions. Table 17 demonstrates the variations in GHG emissions intensity across provinces. Taking fuel efficiency into consideration, for a gasoline vehicle which has a MPG at the medium value of all gasoline vehicles (33 MPG), it emits 20.78kg CO 2 to travel 100km. For a BEV with a MPGe at 105, the CO 2 emissions vary from 3.13kg to 20.81kg for the same distance, which is highly correlated with the electricity generation locations. To compute CO 2 emissions of the whole fleet, I assume vehicles of all types travel 10,500 miles per year and last 12 years, 23 which means the lifetime travel distance of a vehicle is 201,600 km, and the social cost of carbon is $6.61 in 2010 dollars per ton in 2015 (Nordhaus, 2016) which is equal to in 2015 Yuan. Based on the total miles travelled and fuel efficiency, I can derive the lifetime CO 2 emissions of each EV model using Table 17 and the lifetime CO 2 emissions of gasoline and hybrid models using the GHG emissions intensity of gasoline in Table 16. Then I aggregate the emissions over model and time to obtain the total CO 2 emissions by new vehicles in that year. The underlying assumptions are that the generation mix is quite the same across years and the method is based on the average electricity generation emissions instead of electricity on the margin. However, the CO 2 emissions of EVs in this analysis are lower bounds which only account for fuel cycle, the actual emissions of an EV is much higher when taking vehicle cycle into consideration. According to the study of Qiao (2017), the GHG emissions of a BEV prodcution range from 15 to 15.2t CO 2 eq which is about twice than those of an ICE vehicle. Table 15 demonstrates the CO 2 emissions under different scenarios. The CO 2 emissions under the baseline program are a bit larger than those without subsidies, which implies that the adoption of EVs did not bring environmental benefits in 2015 since most of the EVs are widely adopted in cities with relatively dirty electricity, such as Shanghai and Beijing. It is true that the electricity generation mix is becoming cleaner, but coal still accounts for 70.31% of electricity generation in China in 2017 from the World Bank. In addition, the CO 2 emissions only include the fuel circle in this study, the pollution would be worse if incorporating EV production. When more renewable energy is used, the alternative subsidy program can achieve similar environmental benefits as the baseline program since the fleets under the two programs have the same average MPG. 6.4 Firm Profits, Government Spending, and Social Welfare The baseline program resulted in the least firm profit, while removing subsidies would lead to the largest firm profit, though the total vehicle sales under the baseline program is larger than those under the scenario without subsidies. If subsidies were removed, the sales of EVs would decrease 23 The data is from an articlein Bloomberg 24

26 by 128,683, and the sales of gasoline and hybrid vehicles would increase by 80,582 (Table 12). Since the price-cost margins of gasoline and hybrid vehicles are much larger compared with those of EVs (Figure 9), the increased profits from the gasoline vehicle sector would not only offset the lost from the EV sector, but also lead to a higher total profit. More specifically, the EV automakers would lose 2.01 billion Yuan ($0.31 billion), while automakers producing ICE vehicles would gain 2.37 billion Yuan ($0.37 billion) without subsidies. Thus, the total profits without subsidies would have increased by $0.36 billion Yuan ($0.06 billion) compared with the baseline subsidy. Under the alternative subsidy program, the sales of PHEVs would decrease by 31,075 and the sales of BEVs would increase by 30,349 compared with the baseline program (Table 14), implying that the subsidy program based on the battery capacity has a more emphasis on BEV adoption than the baseline program. As a result, BEV producers would gain additional 0.78 billion Yuan ($0.12 billion) while the PHEV producers would lose 0.59 billion Yuan ($0.09 billion). The total profits under the alternative program would just increase by 0.08 billion Yuan ($0.012 billion), which means the gasoline and hybrid vehicle firms would experience a 0.11 billion Yuan ($0.017 billion) loss in profits. Take the welfare change in consumer surplus, firm profit, CO 2 emissions, and government spending into consideration, compared with the baseline program, removing subsidies and subsidizing EVs based on the battery capacity would have increased the social welfare by billion Yuan ($2.88 billion) and 1.27 billion Yuan ($0.2 billion) in 2015 (Table 15), respectively. Although the government expenditure would be larger under the alternative subsidy program, the consumer surplus would have a significant increase. Compared with the social welfare of the baseline subsidy program, the social welfare change under the scenario that subsidies were removed ranges from to billion Yuan and the social welfare change under the alternative program ranges from 0.93 to 2.18 billion Yuan using the results of specification 2 to 4 in Table13, which shows the consistency of the robustness checks with the preferred specification in this study. 7 Conclusion This paper evaluates the EV subsidy program in China. The program provides a natural experiment to look at the unique subsidy program which is solely based on the driving range of EVs. As the largest EV market in the world, the adoption of EVs in China is important not only for Chinese automotive industry development, energy security, and pollution reduction but also for the global EV industry and oil market. The EV market itself is also a crucial industry since electrification has become a trend in automotive industry. This paper estimates a random coefficient discrete choice model to reveal consumer preference, and the demand estimators are used to recover the marginal costs under the subsidy program based 25

27 on the driving range denoted as the baseline program. Through counterfactual analysis, I find that 94% of the EV sales in 2015 were induced by the program in the 19 cities under the study. However, this caused inefficiency in that some households had deviated their choices to small and low-quality EVs. This deviation from their top choices resulted in consumer surplus loss and a $2.88 billion social welfare loss in The hypothetical subsidy program which subsidizes EVs based on the battery capacity would be a more efficient program which would achieve the same environmental goal, correct consumer distortions to some extent, and increase social welfare by $0.2 billion compared with the baseline program, though it would need more subsidies from the government. At the early stages, EV adoption still depends on financial incentives from governments. In addition to stimulating EV sales, increasing electricity generated from clean energy source is another important issue for Chinese government. EVs are still dirty in megacities such as Shanghai and Beijing since coal-generated electricity is still the primary energy source in these cities. While this study is the first one to measure the efficiency in terms of welfare of the subsidy program which subsidizes EVs based on the driving range, further research could explore the impact of the program on firm responses apart from pricing decisions, such as introductions of new products that are more attractive under the subsidy framework and a merge or partner between EV and ICE vehicle automakers. References [1] Anderson, M. L., & Auffhammer, M. (2013). Pounds that kill: The external costs of vehicle weight. Review of Economic Studies, 81(2), [2] Barwick, P. J., Cao, S., & Li, S. (2017). Local protectionism, market structure, and social welfare: China s automobile market (No. w23678). National Bureau of Economic Research. [3] Beresteanu, A., & Li, S. (2011). Gasoline prices, government support, and the demand for hybrid vehicles in the United States. International Economic Review, 52(1), [4] Berry, S., Levinsohn, J., & Pakes, A. (1995). Automobile prices in market equilibrium. Econometrica: Journal of the Econometric Society, [5] Boomhower, J., & Davis, L. W. (2014). A credible approach for measuring inframarginal participation in energy efficiency programs. Journal of Public Economics, 113, [6] Borenstein, S., & Davis, L. W. (2016). The distributional effects of US clean energy tax credits. Tax Policy and the Economy, 30(1),

28 [7] Chen, C. W., Hu, W. M., & Knittel, C. R. (2017). Subsidizing Fuel Efficient Cars: Evidence from China s Automobile Industry (No. w23045). National Bureau of Economic Research. [8] Clinton, B. & D. Steinberg (2016). Providing the Spark: Impact of Financial Incentives on Battery Electric Vehicle Adoption. Working paper. [9] DeShazo, J. R., Sheldon, T. L., & Carson, R. T. (2017). Designing policy incentives for cleaner technologies: Lessons from California s plug-in electric vehicle rebate program. Journal of Environmental Economics and Management, 84, [10] Houde, S., & Aldy, J. E. (2014). Belt and suspenders and more: the incremental impact of energy efficiency subsidies in the presence of existing policy instruments (No. w20541). National Bureau of Economic Research. [11] Helveston, J. P., Liu, Y., Feit, E. M., Fuchs, E., Klampfl, E., & Michalek, J. J. (2015). Will subsidies drive electric vehicle adoption? Measuring consumer preferences in the US and China. Transportation Research Part A: Policy and Practice, 73, [12] Holland, S. P., Mansur, E. T., Muller, N. Z., & Yates, A. J. (2016). Are there environmental benefits from driving electric vehicles? The importance of local factors. American Economic Review, 106(12), [13] Li, J. (2016). Compatibility and Investment in the US Electric Vehicle Market. Working paper. Cambridge, MA: Harvard University, Department of Economics. [14] Li, S. (2014). Better lucky than rich? Welfare analysis of automobile license allocations in Beijing and Shanghai. Welfare Analysis of Automobile License Allocations in Beijing and Shanghai (March 1, 2014). [15] Li, S., Tong, L., Xing, J., & Zhou, Y. (2017). The market for electric vehicles: indirect network effects and policy design. Journal of the Association of Environmental and Resource Economists, 4(1), [16] Ma, S. C., Fan, Y., & Feng, L. (2017). An evaluation of government incentives for new energy vehicles in China focusing on vehicle purchasing restrictions. Energy Policy, 110, [17] Nordhaus, W. D. (2017). Revisiting the social cost of carbon. Proceedings of the National Academy of Sciences, [18] Ou, S., Lin, Z., Wu, Z., Zheng, J., Lyu, R., Przesmitzki, S., & He, X. (2017). A Study of China s Explosive Growth in the Plug-in Electric Vehicle Market (No. ORNL/TM-2016/750). 27

29 Oak Ridge National Lab.(ORNL), Oak Ridge, TN (United States). National Transportation Research Center (NTRC). [19] Ou, X., Yan, X., Zhang, X., & Liu, Z. (2012). Life-cycle analysis on energy consumption and GHG emission intensities of alternative vehicle fuels in China. Applied Energy, 90(1), [20] Peng, T., Zhou, S., Yuan, Z., & Ou, X. (2017). Life cycle greenhouse gas analysis of multiple vehicle fuel pathways in China. Sustainability, 9(12), [21] Petrin, A. (2002). Quantifying the benefits of new products: The case of the minivan. Journal of political Economy, 110(4), [22] Qiao, Q., Zhao, F., Liu, Z., Jiang, S., & Hao, H. (2017). Cradle-to-gate greenhouse gas emissions of battery electric and internal combustion engine vehicles in China. Applied Energy, 204, [23] Sierzchula, W., Bakker, S., Maat, K., & Van Wee, B. (2014). The influence of financial incentives and other socio-economic factors on electric vehicle adoption. Energy Policy, 68, [24] Springel, K. (2016). Network Externality and Subsidy Structure in Two-Sided Markets: Evidence from Electric Vehicle Incentives. Berkeley: University of California. [25] Zivin, J. S. G., Kotchen, M. J., & Mansur, E. T. (2014). Spatial and temporal heterogeneity of marginal emissions: Implications for electric cars and other electricity-shifting policies. Journal of Economic Behavior Organization, 107,

30 Figure 1: Distribution of BEV Range in China Notes: The figure shows the distribution of sales-weighted BEV range in the top two tier cities between 2010 and The subsidies for BEVs with driving range no less than 80km and less than 150km are the lowest; The subsidies for BEVs with driving range no less than 150km and less than 250km are the second highest; The subsidies for BEVs with driving range no less than 250km are the highest. Figure 2: The Timeline of Local Subsidies Notes: The red line represents the beginning of national subsidies which subsidizing EVs based on their driving range. Except for Shanghai, Hangzhou, Nanjing, and Wuxi, the other 14 cities began to offer EV buyers with a subsidy proportional to the central subsidy in a fixed ratio for each city starting from Shanghai and Hangzhou provided a fixed amount of subsidy in 2014 and 2015, while the provincial government of Nanjing and Wuxi started to subsidize EVs based on their wheelbase in March

31 Figure 3: EV Sales in China from 2010 to 2015 Notes: The figure shows EV sales between 2010 and 2015 in China. Figure 4: EV Sales across cities Notes: The figure shows aggregate city sales from 2010 to

32 Figure 5: Distribution of EV Size in China Notes: The figure shows the distribution of sales-weighted EV segments in the top two tier cities between 2010 and Figure 6: Distribution of EV Weight in China Notes: The figure shows the distribution of sales-weighted EV weights in the top two tier cities between 2010 and Red vertical line represents the average weight of all vehicles including gasoline and hybrid vehicle models. 31

33 Figure 7: EV Sales in Representative Cities, Notes: The figure shows the 12-month rolling average of EV sales. The vertical solid red line denotes the starting date of local subsidy policy. In Shanghai, only EVs included in the local subsidy list can receive local subsidies. BEVs entered in the list first, and then PHEVs entered in the list. The yellow short dash line represents the change in subsidy method from a fate rate to subsidizing EVs based on their wheelbase. 32

34 Figure 8: Price Elasticities in 2015 Notes: The blue dots represent own-price elasticities of gasoline and hybrid models. The red diamonds stand for own-price elasticities of EVs. Chery QQ3 is excluded from the graph whose elasticity is about -21. Figure 9: Price-Cost Margins in 2015 Notes: The blue dots represent price-cost margins of gasoline and hybrid models. The red diamonds stand for price-cost margins of EVs. 33

35 Figure 10: Distribution of BEV Range without Subsidies in 2015 Notes: The total number of BEVs under the baseline program and the absence of subsidies are 82,153 and 1,297, respectively. The range is weighted by sales. Figure 11: Distribution of BEV Size without Subsidies in

36 Figure 12: Distribution of BEV Weight without Subsidies in 2015 Figure 13: Distribution of BEV Horsepower without Subsidies in

37 Figure 14: Distribution of BEV Range under Different Subsidy Criteria in 2015 Figure 15: Distribution of BEV Size under Different Subsidy Criteria in

38 Figure 16: Distribution of BEV Weight under Different Subsidy Criteria in 2015 Figure 17: Distribution of BEV Horsepower under Different Subsidy Criteria in

39 Figure 18: Welfare Impact of the Current Subsidy Program Notes: Panel (a) is the case that EVs introduce negative externalities. Panel (b) is the case that the subsidy for EVs is too large, though the EVs introduce positive externalities. 38

40 Table 1: Central Subsidies for EVs from 2013 to 2015 Table 2: Maximum city-level subsidies for EVs in China in 2015 City BEVs (yuan) PHEVs (yuan) City BEVs (yuan) PHEVs (yuan) Shenzhen 60,000 35,000 Dalian 43,200 25,200 Hanghzou 30,000 20,000 Qingdao 60,000 35,000 Guangzhou 60,000 35,000 Wuhan 54,000 31,500 Shanghai 40,000 30,000 Changsha 45,000 31,500 Beijing 54,000 0 Chongqing 54,000 31,500 Nanjing 60,000 35,000 Xiamen 54,000 31,500 Wuxi 25,000 14,000 Chengdu 27,000 18,900 Tianjin 54,000 31,500 Fuzhou 45,000 31,500 Xi an 54,000 31,500 Shenyang Not Available 28,350 Notes: According to the local policies, the maximum subsidies for BEVs in Changsha, Chengdu, and Fuzhou with a driving range above 250 kilometers were 54,000, 32,400, and 54,000 in 2015, respectively. However, the largest driving range of the BEVs in Changsha, Chengdu, and Fuzhou was 170 kilometers so the actually maximum subsidy the buyers received was 45,000, 27,000, and 45,000. There were no sales of PHEVs in Wuxi in January and February in 2015, so the maximum subsidy the PHEVs received was based on the policy issued in March to subsidize PHEVs based on their wheelbase. There were no BEVs in the market of Shenyang in

41 Table 3: History of Electric Vehicles No. of No. of No. of New Vehicle Percent of Year BEV Models BEV Sales PHEV Models PHEV Sales All Models Sales (Million) EVs (%) , , , , Table 4: Summary Statistics of Key Variables Variable Mean Std. Dev. Min Max Sales Real Price (10,000 yuan) EV Power (100 kw) Fuel cost (yuan/km) Size (10 m 2 ) Auto transmission Subsidy (10,000 yuan) Free Plate Driving Restriction Exemption Notes: The number of observations is 97,765. Sales are quarterly sales by model and city. Table 5: Fraction of Households among Vehicle Buyers by Annual Income (1,000 yuan) Year < 60k 60k-100k 100k-150k 150k

42 Table 6: Fraction of Buyers by Annual Income and Vehicle Segment in 2013 Segment < 60k 60k-100k 100k-150k 150k Mini/small car Compact car Medium/large car SUV Table 7: Impacts of EV Incentives on EV Sales log(no. of Registered Cars) (1) (2) Subsidy 0.11*** 0.09** (0.04) (0.04) Sales Tax Exemption 0.13*** 0.13*** (0.04) (0.04) Free Plate 0.89*** 1.72*** (0.27) (0.29) Driving Restriction Exemption 1.07*** 1.06*** (0.29) (0.30) Year Yes Absorbed City Year No Yes N R Notes: The dependent variable is the logarithm of new vehicle sales of all vehicle models. The unit of subsidy and sales tax exemption is ten thousand. All regressions control for month and city-by-model fixed effects. Standard errors are reported in parentheses and two-way clustered at the city and the model level. *p<0.1, **p<0.05, ***p<

43 Table 8: Results from the Reduced-form Regressions OLS Logit Demand IV Logit Demand log(price) *** (0.40) (1.26) EV dummy -2.22*** 0.46 (0.51) (0.98) Power *** (0.49) (1.17) Fuel cost -1.65** -1.64** (0.79) (0.73) Size 3.44*** 10.19*** (0.90) (2.35) Auto Transmission 0.46*** 1.09*** (0.11) (0.22) Free Plate 1.89*** 2.17*** (0.29) (0.44) Driving Restriction Exemption 1.02*** 1.17*** (0.26) (0.40) Cons -9.75*** -8.89*** (0.57) (0.80) N 97,765 97,765 R-squared Notes: Free plate and driving restriction exemption are EV incentives. The regressions control for firm, vehicle segments, quarter fixed effects, and city by year interactions. Standard errors are reported in parentheses and clustered at the model level. *p<0.1, **p<0.05, ***p<

44 Table 9: Results from the Random Coefficient Model (1) (2) (3) (4) Est. S.E. Est. S.E. Est. S.E. Est. S.E. Linear parameters Cons * EV dummy * Power 4.55*** *** *** *** 0.44 Fuel cost -1.89* * * ** 0.97 Size 17.71*** *** *** *** 1.23 Auto Transmission 1.54*** *** *** *** 0.15 Free Plate 2.99*** *** *** *** 0.73 Driving Restriction Exemption 1.47** ** * ** 0.65 Price coefficient eᾱ *** *** *** *** Random coefficients Cons -6.03*** *** *** *** 0.35 Size -3.51*** *** log(price) 0.41*** *** *** *** 0.03 Notes: Free plate and driving restriction exemption are EV incentives. All columns control for firm, vehicle segments, month fixed effects, and city by year interactions in the mean utility. Specification 1 is the benchmark and preferred model. Specification 2 assumes half of the households as the potential buyers in a year so that market size in each quarter is the total number of households divided by 8. Specification 3 drops extreme random draws below 2.5 and above 97.5 percentiles for unobserved household heterogeneous preference for vehicle attributes. Specification 4 use different income group cutoffs. *p<0.1, **p<0.05, ***p<

45 Table 10: Model Fit in the First Set of Micro-moments year Observed share Predicted share <60k k k <60k k k Table 11: Model Fit in the Second Set of Micro-moments in 2013 Segment < 60k 60k-100k 100k-150k 150k Panel A: Observed Share Mini/small car Compact car Medium/large car SUV Panel B: Predicted Share Mini/small car Compact car Medium/large car SUV

46 Table 12: Impacts of Removing Subsidies on Sales in 2015 Fuel type Baseline Subsidy W/O Subsidy Change %Change Based on Range Gasoline 3,347,263 3,427,732 80, % Hybrid 3,912 4, % PHEV 54,770 6,943-47, % BEV 82,153 1,297-80, % Table 13: Robustness Checks: Impacts of Subsidies on Sales in 2015 Fuel type Observed Specification Specification Specification Specification Sales Panel A: Counterfactural Sales without Subsidies Gasoline 3,347,263 3,427,732 3,423,744 3,433,702 3,429,161 Hybrid 3,912 4,025 4,023 4,034 4,021 PHEV 54,770 6,943 7,534 5,687 7,718 BEV 82,153 1,297 1,410 1,087 1,489 Panel B: Counterfactural Sales in Terms of Battery Capactiy Gasoline 3,347,263 3,346,793 3,346,274 3,343,253 3,348,007 Hybrid 3,912 3,866 3,869 3,863 3,868 PHEV 54,770 24,421 25,140 22,768 25,386 BEV 82, , , , ,118 Notes: Observed sales refer to the sales under the baseline program which subsidizes EVs based on their driving range. The specifications are corresponding to the specifications in Table 9. Specification 1 is the benchmark specification; specification 2 uses a different definition of the market size; specification 3 removes extreme draws for random coefficients; specification 4 uses different cutoffs for income group. 45

47 Table 14: Impacts of Subsidies Based on the Battery Capacity on Sales in 2015 Fuel type Baseline Subsidy Subsidy Change %Change Based on Range Battery Capacity Gasoline 3,347,263 3,346, % Hybrid 3,912 3, % PHEV 54,770 24,421-30, % BEV 82, ,228 31, % Notes: The MPG of the two subsidy programs are the same at 37 MPG. Table 15: Welfare under Different Scenarios in 2015 Welfare (billion Yuan) Baseline Subsidy W/O Subsidy Subsidy Base on Range Battery Capacity Consumer surplus Firm profit CO 2 emissions Government spending Total social welfare change ($2.88) 1.27 ($0.2) Notes: The welfare changes in the table is based on the baseline scenario, subsidizing EVs in terms of their driving range. Table 16: Fuel Cycle GHG Emissions Intensity of Electricity and Gasoline Energy Type GHG Emissions (gco 2,e /kwh) Data Source Panel A: Electricity Thermal (Coal) (2015) Peng et al. (2017) Nuclear (2008) Ou et al. (2012) Hydro, Wind, Solar, and Other 18 (2008) Ou et al. (2012) Panel B: Gasoline (2015) Peng et al. (2017) Notes: According to Peng et al. (2017), vehicle cycle (including production and transportation of raw materials, vehicle manufacturing, and vehicle decommissioning and recycling) is excluded in the estimation. The GHG emissions of gasoline can be converted to gco 2,e /L considering the energy density of gasoline at 32 MJ/liter. 46

48 Table 17: Electricity Supply Structure and GHG Emissions Intensity of Power Grid in 2015 Hydro Thermal Nuclear Wind Solar Other GHG Emissions (%) (%) (%) (%) (%) (%) (g CO 2,e /kwh) Beijing Hunan Sichuan Chongqing Liaoning Fujian Guangdong Zhejiang Shandong Shanghai Tianjin Hubei Shaanxi Nation Source: China Electricity Power Yearbook,

49 Appendices APPENDIX: ONLINE MATERIALS As I mentioned in the paper that the WTP for small and low-quality EVs is quite low. We can obtain the WTP for any EVs through the following steps. Case 1: For a given EV, the WTP of a consumer who purchases it is the increased price which lets him switch to his second best option. The utility of the consumer who purchases this EV is as follows: u mti j = δ mt j α mti ln(p mt j ) + x t j kσ k ν mtik + ε mti j. k The utility of purchasing the second best choice (product g) is: (21) u mtig = max{0,δ mtg α mti ln(p mtg ) + x tgk σ k ν mtik + ε mtig }. k (22) To prevent consumer switching from product j to g, the price of j can be raised to γ mt j p mt j (WTP) at the most. γ mt j is calculated as: u mti j u mtig = α mti ln(γ mt j ), (23) γ mt j = exp( u mti j u mtig α mti ). (24) Case 2: For consumers who purchase outside good or other vehicle models, to let the consumer purchase the EV, the price of this EV should be reduced to γ mtg p mtg (WTP) at least, in which case γ mtg is defined as: u mti j u mtig = α mti ln(γ mtg ), (25) γ mtg = exp( u mti j u mtig α mti ) = exp( u mtig u mti j α mti ). (26) where u mti j is the utility of optimal choice (outside good or other vehicle models), and u mtig is the utility of purchasing the EV. Using the steps, I can compute γ and consumers WTP. In the simulation, I draw 1,000 pseudo people for each city in the last quarter of Then I plot the 200 highest WTP out of 1,000. The WTP for two popular small BEVs is shown in Figure 2. We can see that the WTP is far away from the marginal cost. Without subsidies, these two popular BEVs would exit the market. 1

50 Figure 1: EV Sales across Cities, Notes: The figure shows the 12-month rolling average of EV sales. The vertical red line denotes the starting date of local subsidy policy. For Shenzhen and Guangzhou, the red line represents the starting date of the subsidy that is proportional to the central subsidy. For Hangzhou, the local subsidy in 2011 and 2012 was based on the battery capacity, and then the local subsidy stopped. 2

51 Figure 2: WTP for Zhidou and K11 Notes: The red line is the marginal costs for these two EV models, and the blue dot represents one s WTP. There are 200 people in total. 3