未来能源研究所：收入对汽车需求的影响：来自中国新车市场的证据（英）

2021年07月15日
50 金币

The Effect of Income on Vehicle Demand: Evidence from China’s New Vehicle MarketJoshua Linn and Chang ShenWorking Paper 21-17 June 2021About the AuthorsJoshua Linn is an associate professor in the Department of Agricultural and Resource Economics at the University of Maryland and a senior fellow at RFF. His research centers on the effects of environmental policies and economic incentives for new technologies in the transportation, electricity, and industrial sectors. His transportation research assesses passenger vehicle taxation and fuel economy standards in the United States and Europe. He has examined the effects of Beijing’s vehicle ownership restrictions on travel behavior, labor supply, and fertility.Prior to joining RFF in 2010, Linn was an assistant professor in the economics department at the University of Illinois at Chicago and a research scientist at MIT. He was a senior economist at the Council of Economic Advisers from 2014–2015. He is serving on a National Academies of Sciences, Engineering, and Medicine committee on light-duty fuel economy. He is a co-editor for the Journal of Environmental Economics and Management.Chang Shen is a transportation economist with five years of research experience investigating environmental issues in the transportation sector. Her research focuses foremost on estimating consumers’ vehicle and travel demand, as well as evaluating policies and market incentives in promoting greener transportation and reducing GHG emissions. She holds a Ph.D. degree in Agricultural and Resource Economics (AREC) from University of Maryland.About RFFResources for the Future (RFF) is an independent, nonprofit research institution in Washington, DC. Its mission is to improve environmental, energy, and natural resource decisions through impartial economic research and policy engagement. RFF is committed to being the most widely trusted source of research insights and policy solutions leading to a healthy environment and a thriving economy.Working papers are research materials circulated by their authors for purposes of information and discussion. They have not necessarily undergone formal peer review. The views expressed here are those of the individual authors and may differ from those of other RFF experts, its officers, or its directors.Sharing Our WorkOur work is available for sharing and adaptation under an AttributionNonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) license. You can copy and redistribute our material in any medium or format; you must give appropriate credit, provide a link to the license, and indicate if changes were made, and you may not apply additional restrictions. You may do so in any reasonable manner, but not in any way that suggests the licensor endorses you or your use. You may not use the material for commercial purposes. If you remix, transform, or build upon the material, you may not distribute the modified material. For more information, visit https://creativecommons.org/licenses/by-nc-nd/4.0/.The Effect of Income on Vehicle Demand: Evidence from China’s New Vehicle MarketJOSHUA LINN CHANG SHEN*June 2021AbstractGrowth of private vehicle ownership in low-income and emerging countries is a dominant factor in forecasts of global oil demand and greenhouse gas emissions. Countries such as China are expected to experience rapid income growth over the next few decades, but little causal evidence exists on its effect on car ownership in these countries. Using city-level data on new car sales and income from 2005 to 2017, and using export-led growth to isolate plausibly exogenous income variation, we estimate an elasticity of new car sales to income of about 2.5. This estimate indicates that recent projections of vehicle sales in China have understated actual sales by 36 percent and carbon dioxide emissions by 18 million metric tons in 2017. The results suggest that, to meet its climate objectives, China’s climate policies will need to be substantially more aggressive than previous forecasts indicate.*Linn: University of Maryland and Resources for the Future (linn@umd.edu). Shen: University of Maryland. The authors thank Anna Alberini, Jim Archsmith, Erich Battistin, Jing Cai, and Cinzia Cirilo for helpful comments.11 IntroductionGlobal oil consumption and greenhouse gas (GHG) emissions from transportation are expected to increase over the next few decades, with lower-income countries causing most of the growth. The US Energy Information Agency predicts that oil consumption and transportation energy consumption will grow roughly 15 percent between 2020 and 2040. OECD and non-OECD countries are expected to follow diverging paths: consumption is expected to decline 3 percent for the former and increase nearly 30 percent for the latter (EIA, 2019). Projections from the International Energy Agency and other major organizations are broadly similar.China is a major driver of these diverging trends. China’s oil consumption is expected to grow 20 percent between 2020 and 2040 (EIA, 2017). Rising vehicle ownership explains much of this growth—in both China and other non-OECD countries. China’s vehicle stock is expected to grow by 200 million units between 2020 and 2040, accounting for nearly all of the global growth in the vehicle stock (BloombergNEF, 2020).1For China, and many other countries, climate policy depends crucially on these emissions forecasts. For example, under the United Nations Paris Agreement, China has pledged to peak its emissions by 2030 and then substantially reduce emissions. Total transportation sector emissions, which account for 9 percent of China’s GHG emissions (IEA, 2017), equal the emissions rate of vehicles multiplied by the number of vehicles and miles traveled per vehicle. China’s transportation policies focus mostly on reducing the new vehicle emissions rates and not the levels of emissions. Therefore, to achieve the emissions target under the Paris Agreement the greater the future vehicle ownership and use, the more China has to reduce vehicle emissions rates (Pan et al., 2018); if vehicle ownership turns out to be 10 percent higher than expected, policies would have to reduce emissions rates by an additional 10 percent. Globally, countries rely almost exclusively on GHG emissions rate standards for passenger vehicles to reduce transportation sector emissions, and similar policy uncertainty applies to these countries.Unfortunately, assumptions behind these forecasts rest on little empirical support. In the computational models that generate the forecasts, oil consumption and GHG emissions from passenger vehicles are closely linked to household vehicle ownership. An extensive literature correlates income with vehicle ownership in the United States, Europe, and other OECD countries (Dargay, 2001; Dargay, Gately, and Sommer, 2007; Nolan, 2010;1The situation is comparable to that for anticipated growth in electricity consumption, where growth is concentrated among non-OECD countries and driven largely by uptake of energy-consuming durable goods, such as refrigerators and air conditioners (Auffhammer and Wolfram, 2014; Davis, Fuchs, and Gertler, 2014).2Blumenberg and Pierce, 2012; Oakil, Manting, and Nijland, 2018). Most projections of future vehicle ownership in China and other non-OECD countries from the past two decades rely on the assumption that ownership and use will follow patterns observed in other countries. For example, Huo et al. (2007) forecast vehicle ownership in China using data from in Europe, Japan, and other countries, assuming that the effect of GDP per capita will be the same in China. He et al. (2005) and Yan and Crookes (2009) use similar methods, as do forecasts from organizations such as the International Energy Agency, which receive a lot of attention from policymakers. However, Wang, Teter, and Sperling (2011) argue that periods of early motorization in the United States and Europe may be more relevant to future motorization in China; in that case, basing projections on recent OECD data could yield overly conservative estimates of China’s future oil consumption and GHG emissions.Recently, income and vehicle ownership have exploded in China, with average income per capita growing 11.9 times and new vehicle sales growing 7.6 times between 2000 and 2017 (CEIC Data). This situation presents an opportunity to evaluate the assumptions that underlie the projections of future vehicle ownership and GHG emissions in China. That is, have recent projections of vehicle ownership in China proven to be accurate?We estimate the recent relationship between income and new car ownership in China and compare the results with recent forecasts of vehicle ownership. The main data include total new vehicle sales, income, and other socioeconomic variables by city and year for 2005–2017. This period includes 9.6 percent annual growth in income and 20 percent annual growth in new vehicle sales. During these years, sales grew from 5.8 to 29 million units, as China became the world’s largest new car market, roughly twice as large as markets in the United States or Europe.The objective is to estimate the causal effect of income on car ownership, and a major challenge is that income is endogenous to car ownership due to reverse causality and omitted variables. For example, if vehicle ownership reduces travel costs and allows people to ﬁnd better jobs, reverse causality could exist from ownership to income. Omitted factors that may be correlated with income and also affect car ownership, such as cultural trends related to car ownership, would cause omitted variables bias. Besides being potentially endogenous, city-level income may be measured with error.We adopt an instrumental variables strategy to address the endogeneity and measurement error. We employ a Bartik-style instrumental variable (IV) that is the interaction of a city’s education employment in 2004 with China’s annual high-technology exports. The relevance of the instrument is supported by a) high-technology exports having driven much of China’s economic growth over the past two decades; and b) cities with high initial education employment having large skilled worker populations who can produce3high-technology exports. The exclusion restriction is that 2004 education employment is uncorrelated with subsequent unobserved factors that affect vehicle ownership via channels other than income. We provide evidence supporting this assumption, including a lack of correlation between 2004 education employment and subsequent shocks to other drivers of vehicle ownership, such as the quality of public transportation.We ﬁnd that a 1 percent increase in income causes total new vehicle sales to increase by 2.5 percent. Income does not affect the sales-weighted average price of new vehicles sold, meaning that as income has grown, sales of low- and high-price vehicles have grown by the same proportion. Likewise, the elasticity of new vehicle sales to income does not appear to be correlated with a city’s initial income, again suggesting proportional growth. The estimate is robust to alternative functional forms and controlling for other socioeconomic variables.Comparing our results with the literature, we conclude that recent projections of future new vehicle sales in China may be vastly understated. Our estimates mean that in 2017, rising income increased sales by about 36 percent and increased carbon dioxide emissions by about 18 million metric tons more than predicted by recent forecasts. In the long run, annual sales are proportional to the new vehicle stock, suggesting that recent forecasts of new vehicle sales underpredict the effect of rising income on emissions by roughly 26 percent. Our results indicate that recent forecasts may substantially underestimate China’s future oil consumption and GHG emissions. This implies that China will need to adopt more aggressive climate policies to meet its GHG targets.We contribute to several literature strands. First, several studies project China’s future vehicle stock. A typical method is to assume that vehicle ownership is an S-shaped function of per-capita gross domestic product (GDP). The rationale for this is that in OECD countries, vehicle ownership increased slowly at low levels of GDP, rose steeply, and then leveled off (Lu et al., 2018). For instance, Huo et al. (2007) assume the vehicle ownership rate follows an S-shaped Gompertz function of per-capita GDP and conclude that the Chinese highway vehicle stock will reach 389–495 million units by 2040. Huo and Wang (2012) compare several S-shaped functional forms and account for income inequality and vehicle prices. Lu et al. (2018) and Gan et al. (2020) use similar methods and more recent data, projecting that China’s vehicle stock will reach 400–600 million units by 2050.These forecasts assume parameters for the function linking vehicle stock to income, including a saturation rate. Most previous studies assume saturation rates of 200–800 cars per thousand people (He et al., 2005; Dargay, Gately, and Sommer, 2007; Huo et al., 2007; Huo and Wang, 2012; Lu et al., 2018), which is based on observations from other countries. Gan et al. (2020) comment that transferring parameters from other countries to China is4arbitrary, and they use household survey data instead to calibrate their model. In contrast to this literature, rather than calibrating a curve to data from other countries,we use historical data from China to estimate the effect of income on total new vehicle sales, accounting for the potential endogeneity of income. To our knowledge, ours is the ﬁrst study to investigate the effect of income at the city level. Nearly all prior research estimates vehicle growth using national data. However, as we illustrate, Chinese cities have had imbalanced development and it was a national strategy to prioritize the development of certain regions (Shen, Teng, and Song, 2018). Each city also has its own preferences for public transportation and road systems. Different cities might exhibit very different vehicle growth patterns. Our balanced panel of city-level data allows us to exploit crosssectional and time-series variation of income and new car sales and consider whether the income–sales relationship varies across cities.We also contribute to the broader literature on income and energy-consuming durables and future GHG and oil demand. As noted, little research exists on the effect of income on new vehicle demand in non-OECD countries, although some research has been undertaken on household appliances and residential energy-efﬁcient and renewable energy products, such as solar panels. McNeil and Letschert (2010) ﬁnd that ownership of refrigerators, washing machines, televisions, and air conditioners increases with household income, urbanization, and electriﬁcation rates. They also document an S-shaped relationship between income and appliance ownership. Auffhammer and Wolfram (2014) and Li et al. (2019) report somewhat conﬂicting evidence on the relationship between income and appliance ownership. Auffhammer and Wolfram (2014) show that the proportion of households above the poverty line affects the uptake of energy-using durable goods in rural China. However, Li et al. (2019) show that the income threshold for ownership is correlated with the cost of the appliance. In contrast to Auffhammer and Wolfram (2014), they ﬁnd that changes in the income distribution have negligible effects on penetration rate of household appliances.2Closely connected to the literature on income and energy-consuming durables is the extensive literature on the Environmental Kuznets Curve (EKC; for example, Shaﬁk and Bandyopadhyay, 1992; Dasgupta et al., 2002; Wagner, 2008; Yao, Zhang, and Zhang, 2019). This literature examines the relationship between income (or GDP) and pollution, and most studies use aggregate national or regional data. Many studies document an inverted U-shaped relationship between income and emission: at low levels of income, emissions rise with income, but at high levels of income, emissions decline with income. However,2A vast literature exists on income and appliance ownership in OECD countries, such as Zhao et al. (2012) and Mundaca and Samahita (2020).5studies disagree on the factors explaining this relationship (Kaika and Zervas, 2013); many argue that in the early stages of economic development, economic expansion is fuelled by heavy and polluting industry. Subsequently, the economy shifts from heavy industry to services and environmental regulation strengthens, causing pollution to decline. Yin, Zheng, and Chen (2015) conﬁrm an inverted U-shaped relationship between per-capita income and carbon dioxide emissions in China.Our analysis contributes to the EKC literature in two main ways. First, ours is among the small number of papers that addresses the endogeneity of income to other factors that affect pollution, such as road networks and public transportation. Second, we demonstrate that income can affect pollution through microlevel household behavior, which contrasts with the predominant use of macrolevel data in the EKC literature.Finally, the literature on long-run climate policy using integrated assessment models (IAMs) and other aggregate models calibrates the relationship between GDP and emissions absent policy intervention (Nordhaus and Yang, 1996; Cantore, 2011; Krey et al., 2012; Vliet et al., 2012; Ruijven et al., 2012; Calvin et al., 2013; Steckel et al., 2013; Luderer et al., 2015; Cherp et al., 2016; Calderón et al., 2016; Zwaan et al., 2018; Nieto et al., 2020). These models can be used to estimate the efﬁcient carbon price or the costs of achieving long-run policy objectives, such as maintaining expected temperature changes below a certain threshold. Often, IAMs are calibrated to forecasts of future GDP and emissions. We ﬁnd that at least for China, those forecasts may vastly understate future transportation emissions. Given China’s contribution to global emissions, that would cause the forecasts to understate global transportation emissions by a nontrivial amount—and by even more if our results pertain to other non-OECD countries. Therefore, more accurate predictions of future vehicle ownership in China and perhaps other non-OECD countries could have implications for climate policy analysis in IAMs.2 Data and Summary StatisticsWe use data on new vehicle registrations in all of China from 2005 to 2017. The Chinese Department of National Security collects the data, which contain information on the total number of new vehicles registered by city, month, model, and usage purpose (personal or business). The data include vehicle attributes, such as engine size and manufacturer suggested retail price. Because we are interested in the effect of personal income on vehicle ownership, we exclude business purchases. Imported cars are also excluded due to a lack of price information (the results are similar if we include these, which account for a small share of total registrations).6We use counts of new vehicle registrations as proxies for vehicle sales. Tan, Xiao, and Zhou (2019) compare the new vehicle registration data with statistics on vehicle sales from the China Automotive Industry Yearbook and ﬁnd that the registration data account for 70 percent of total vehicle sales. However, further investigation suggests that the sales data, rather than the registration data, are misleading. Many agencies and organizations in China compile their own sales statistics, such as the China Association of Automobile Manufacturers, the State Information Center, and the Chinese Passenger Cars Association. Each organization uses its own data collection methodology. For instance, the China Association of Automobile Manufacturers includes unrealized orders from manufacturers. In addition, most of these statistics rely on self-reported data from manufacturers. It is not uncommon for vehicle manufactures to fake sales, and many automobile industry analysts are turning to registration data. In the remainder of the paper, we use the terms sales and registrations interchangeably.One potential concern about using registration data is that consumers may not immediately register their vehicles after purchase. In that case, registrations would lag sales. However, the month of registration is a good proxy for the month of sales. Most Chinese car buyers choose to pay an additional service fee and apply for registration through the car dealers immediately after they complete their purchases. The application process usually takes 2–7 days, and driving without a registration would add penalty points to the driver’s license. Therefore, the registration month and purchase month are the same for the majority of vehicles, and the two may differ by at most a month. This situation likely introduces little measurement error because we aggregate the monthly data to the annual level.We combine the registration data with a set of socioeconomic variables from the China City Statistical Yearbook, which is published annually by the National Bureau of Statistics of China (NBSC). Each year, NBSC distributes questionnaires to municipal statistics departments. Province-level statistics departments and the NBSC check the validity of the responses. For each city, the yearbook includes average income per capita, which is the key independent variable in the econometric analysis; the built area (constructed areas for residential, commercial, or industrial use); area of paved roads (road length multiplied by width); population; number of buses and taxis in the public transportation system; total retail revenue; and the share of education sector employment in total employment. We also gather information on national-level high-technology exports from the World Bank, which deﬁnes these as products with high R&D intensity, such as aerospace, computers, pharmaceuticals, scientiﬁc instruments, and electrical machinery.Table 1 summarizes the car registration and socioeconomic variables. The dataset contains 3,627 unique city-year observations (a balanced panel of 279 cities over 13 years).7New car expenditures and the number of cars sold are aggregated to the city-year level, and average car price is weighted by the number of cars sold. Car sales and socioeconomic variables vary substantially across cities.Table 1: Summary statisticsMain variables New cars sold (thousand units) New car expenditure (billion RMB) Car price (thousand RMB) Income per capita (thousand RMB) Built-up area (square km) Area of paved roads (square km) Population (thousand people) Buses (thousand units) Taxis (thousand units) Total retail expenditure (billion RMB) Employment percentage in education (2004) National high-tech export (billion RMB)mean 37.9 5.9 153.0 41.4 121.0 15.9 4,354.2 1.3 3.1 73.0 4.8 3,786.0std.dev 62.0 10.5 27.0 16.3 165.2 21.73,098.4 2.8 6.0 111.0 4.5 590.2coef. var. 1.6 1.8 0.2 0.4 1.4 1.4 0.7 2.1 2.0 1.5 0.9 0.2min 0.4 0.1 94.1 8.8 6.0 0.4 172.2 0.0 0.1 1.6 0.1 2,461.6max 727.0 140.1 259.3 135.0 1446.0 214.9 34,332.3 35.8 68.5 1183.0 41.6 4,509.9median 17.0 2.4 148.9 40.4 70.0 8.43,700.0 0.5 1.5 38.6 3.94,001.6Notes: The data contain 3,627 observations. "Built area" is the area of land constructed for residential, commercial, or industrial use. "Bus" and "taxi" are the numbers of buses and taxis operating in the city. Allmonetary variables are adjusted for inﬂation and measured in 2017 RMB. "Coef. var." is the coefﬁcient of variation.Figure 1 illustrates income growth for ﬁve groups of cities between 2005 and 2017. We compute quantiles of the distribution of city-level average income per capita using 2005 income data, and we assign each city to a quintile group. The ﬁgure shows the average income of cities in each group, with average income normalized to 1 in 2005 to facilitate comparison of income growth across cities. All ﬁve quintiles saw tremendous growth, as well as a steady pattern of converging income levels across cities. Between 2005 and 2017, the average income for cities in the lowest quantile grew by a factor of 3.64, indicating a remarkable 28 percent average annual growth rate. In contrast, the average income for cities in the highest quantile grew by a factor of 2.51. This convergence eliminated 67 percent of the difference between the average income of the ﬁfth and ﬁrst quantiles: in 2005, the average income of the highest quantile is 214 percent of the average of the lowest quantile, whereas in 2017, this number decreased to 148 percent. The income growth and variation across cities helps identify the effect of income on car ownership and facilitates an analysis of whether the effect varies across cities.8Figure 1: Income Growth by City Income QuintileNotes: All cities are divided into ﬁve quintiles according to their income in the initial period (2005). For each year, we calculate the average income of cities within each of the ﬁve quintiles. We normalize the quintile-level averages by their 2005 values.Table 2 provides further insight into the income dynamics during the study period. Each column indicates a city’s 2005 income quintile, and each row indicates a city’s 2017 income quintile (based on the 2017 rather than the 2005 income distribution). Quintile 1 is the lowest, and Quintile 5 is the highest. Each cell reports the percentage of cities that were in the indicated 2005 income quintile and that belong to the 2017 income quintile. For example, 50 percent of cities in the lowest 2005 income quintile belong to the lowest 2017 income quintile, whereas 36 percent of cities in the lowest 2005 income quintile belong to the second 2017 quintile. The table shows that many initially low-income cities catch up to and pass many initially higher-income cities; for example, 15 percent of cities in the lowest income quintile in 2005 belong to the top three quintiles in 2017.The next two ﬁgures present summary statistics about new car registrations and attributes. All monetary values are converted to 2017 RMB using the annual consumer price index. Figure 2 reports the growth rate of vehicle sales (panel a) and revenue (panel b) by income quintile, with quintiles deﬁned as in Figure 1 and 2005 levels normalized to 1 for comparability across quintiles. Total sales and expenditure have a similar pattern to income growth from Figure 1, with sales and revenue increasing more quickly in low-income cities than in high-income cities. The similarity of the patterns across the two ﬁgures previews our main ﬁnding of a strong connection between income and new car demand; in fact, panel (d) shows that registrations outpaced income growth for each group of cities.9Table 2: Quintile Switching2017 income level1 (lowest) 2 3 4 5 (highest)1 50 % 36 % 9% 4% 2%2005 income level23430 % 13 % 5 %23 % 27 % 14 %29 % 27 % 25 %16 % 32 % 30 %2 % 2 % 25 %5 2% 0% 11 % 18 % 69 %Notes: Each column indicates a city’s 2005 income quintile, and each row indicates a city’s 2017 income quintile (based on the 2017 rather than the 2005 income distribution). Each cell reports the percentage of cities that were in the indicated 2005 income quintile and that belong to the 2017 income quintile.Panel (c) shows that the average real new car price has been declining during this period, indicating that although the aggregate demand for new cars increased, households are not systematically buying more expensive cars at the end of the period relative to what they were buying at the beginning. This is consistent with domestic car manufacturing evolving during the sample, with most domestic brands targeting low- and middle-price vehicles, putting downward pressure on average prices.10Figure 2: Growth of Sales, Expenditure, and Price by Income Quintile(a) Sales(b) Expenditure(c) Average price(d) Expenditure share in incomeNotes: All monetary values are converted to 2017 RMB using the annual CPI.The average new car price declined between 2008 and 2016. During this period, income growth slowed and vehicle policies changed. From January through December 2009 and October 2015 through December 2016, China reduced purchase taxes from 10 percent to 5 percent for cars with small engines. The change likely increased demand for such cars, which also tend to be less expensive than cars with large engines. Moreover, in 2008, China introduced a fuel economy standard that required a 10 percent reduction in fuel consumption rates (the fuel consumption rate, measured in liters per kilometer, is the inverse of fuel economy, in miles per gallon). Manufacturers attempted to meet this standard by incentivizing consumers to purchase cars with low fuel consumption rates,11which also tend to have low purchase prices. Panel (b) shows that fuel consumption rates increased from 2005 through 2008 but declined after 2008, as the fuel economy standards tightened.Panels (c) and (d) show that average horsepower and weight increased over the sample period at similar rates across quintiles. The overall upward trends are interrupted by temporary decreases in 2008 and 2016, which coincide with the engine tax policy changes.Overall, the data show dramatic growth of new vehicle sales and income. The rates varied considerably across cities, with income and sales converging over time. Average prices decreased between 2005 and 2017, and much of the decrease coincided with tax policy changes and fuel economy regulation. The empirical strategy controls for the policies.Figure 3: Average Vehicle Attributes by Income Quintile(a) Engine size(b) Fuel consumption rate(c) Horsepower(d) Gross weight123 Empirical StrategyThe ﬁrst subsection provides a theoretical framework that yields the estimating equation; the second subsection discusses the IV estimation that accounts for the endogeneity of income.3.1. Economic Framework and Estimating EquationWe motivate the estimating equation by examining a framework that links income to new car ownership. For an individual household, purchasing a new car would increase the household’s utility because of the comfort and convenience of travel. For example, suppose a household member commutes by public transportation and purchasing a car would reduce commuting time. The car may also allow household members to take new trips. Besides comfort and convenience, owning a car may confer status to the household.When deciding whether to purchase the car, the household compares the beneﬁt of ownership with the costs. The costs include the purchase price (the forgone consumption of other goods), fuel costs, and maintenance. If car ownership is a normal good, it increases with income. If we aggregate across households, total new car sales increase with income.3The objective is to estimate the causal effect of household income on new vehicle purchases and expenditure, conditional on other factors that could affect new car demand. Because population determines the size of the potential market for new vehicles, it is natural to begin by assuming that car purchases and expenditure are proportional to income conditional on population, giving rise to the following regression:lnYjt = αN lnNjt + αP lnPjt + Xjtδ + γj + τt + jt(1)The dependent variable is the log of either new vehicle purchases or expenditure in cityj and year t. The variable Njt is income, Pjt is population, Xjt is a vector of controls, γj includes city ﬁxed effects, τt includes year ﬁxed effects, and jt is an error term. Note that instead of controlling for the log of population, we could normalize the dependent variableand income by population. Normalizing by income would amount to setting αP equal to -1. Because equation (1) allows the coefﬁcient to differ from -1, this speciﬁcation is moreﬂexible.3This statement could be formalized by considering a model in which a household derives utility from a car and a composite good. The utility from car ownership depends on its attributes (such as interior space or performance) and an idiosyncratic preference shock. If the utility function exhibits decreasing marginal utility for the composite good, an increase in income raises the probability that the household purchases the car. Aggregating across households, we conclude that total new car sales increase with average income.13The vector Xjt includes factors that may affect new car demand independently of income, such as the built-up area in the city (constructed area for residential, commercial or industrial use), area of paved roads, population, number of buses and taxis in the city’s public transportation system, and total retail revenue (that is, a proxy for retail expenditures). The city ﬁxed effects control for time-invariant attributes, such as geographic proximity to other cities (which could affect travel demand), and the year ﬁxed effects control for aggregate shocks that affect car sales proportionately, such as tax incentives.The main coefﬁcient of interest in equation (1) is αN. Because the dependent variable and income enter the equation in logs, the coefﬁcient is interpreted as an elasticity; a coefﬁcient of 1 means that a 1 percent increase in income is associated with a 1 percent increase in car purchases or expenditure. We expect αN to be positive because an increase in income raises new vehicle demand.We consider the log-log relationship between average city income and sales to be an approximation of a potentially more complex relationship. For example, sales could be a function of the household income distribution if a threshold level of income exists below which households do not purchase new vehicles. As the household income distribution shifts to the right over time, sales increase but the relationship between average city income and total sales may not be iso-elastic. We show that we use the log-log approximation because it appears to ﬁt the data reasonably well.Note that we do not model explicitly the effects of vehicle attributes on household purchase decisions, as we might if we were to estimate a discrete choice model rather than the reduced-form equation (1). We choose the reduced-form approach because a discrete choice model is not necessary given the scope of our paper; implementing it introduces unnecessary structure and the need to instrument for endogenous vehicle attributes, such as vehicle price. In equation (1), the log income coefﬁcient (αN) includes the mediating effects of vehicle attribute changes caused by income changes. The next subsection explains our approach to controlling for attribute changes that are not caused by income changes, such as the vehicle regulations that affect fuel economy and engine size discussed in the previous section.3.2. IV Estimation and InterpretationThe theoretical framework at the beginning of the previous subsection indicates three reasons why estimating equation (1) by ordinary least squares (OLS) would yield inconsistent estimates of αN: reverse causality, omitted variables bias, and measurement error. Reverse causality could arise if owning a car reduces commuting costs, expanding an14individual’s job opportunities and income from employment. Omitted variables bias could occur if variables other than income affect the costs andbeneﬁts of owning a car. We mentioned the quality of transportation as one example, and many others exist, such as vehicle operating and maintenance costs. Although we attempt to control for variables that affect new car demand independently of income, such as the number of buses and taxis operating in a city, many such variables are unobservable or difﬁcult to measure, such as the quality of public transportation.Finally, income may be measured with error. We use the average income of a city’s population, but the relevant measure may be the income of households considering buying new vehicles. Note that a city’s average income and the average income of potential new car buyers are likely to be highly correlated with one another, but using average citywide income likely introduces some measurement error.Given these concerns, we use a Bartik-style instrument based on high-technology exportdriven income growth. A classic Bartik instrument is formed by interacting local industry shares and national industry growth rates. It is commonly used across many ﬁelds in economics, including labor, public, development, macroeconomics, international trade, and ﬁnance (Goldsmith-Pinkham, Sorkin, and Swift, 2020; Beaudry, Green, and Sand, 2012; Nunn and Qian, 2014; Baum-Snow and Ferreira, 2015; Jaeger, Ruist, and Stuhler, 2018).The literature on export-driven income growth motivates the IV, which is the interaction of national high-technology exports with the city’s 2004 education sector employment. We use high-technology exports deﬁned by the World Bank, which include products with high R&D intensity in aerospace, computers, pharmaceuticals, scientiﬁc instruments, and electrical machinery. Numerous studies ﬁnd that high-technology exports have substantially improved economic growth. For instance, Hausmann, Hwang, and Rodrik (2007) show that export quality is positively correlated with growth, and Falk (2009) shows that high-technology exports have a positive effect on economic growth in OECD countries. Jarreau and Poncet (2012) conﬁrm that high-technology exports promote economic growth in China. They exploit variation in export sophistication at the province and prefecture levels and ﬁnd that regions specializing in more sophisticated goods subsequently grow faster.Moreover, human capital growth has contributed to high-technology exports (Stokey, 1991; Levin and Raut, 1997; Mehrara, Seijani, and Karsalari, 2017; Mulliqi, Adnett, and Hisarciklilar, 2019). Thus, the literature documents a strong connection from human capital growth to high-technology export growth to income growth. Given these ﬁndings, we specify the ﬁrst stage as15lnNjt = βX ln(Et) · Hj + βP lnPjt + Xjtη + γj + τt + µjt(2)where Et is national high-technology exports and Hj is the city’s education employment in 2005. The second stage islnYjt = αN lnNjt + αP lnPjt + Xjtδ + γj + τt + jt(3)The IV speciﬁcation is similar to equation (1), except that we use a Bartik-style instrument for income. The aforementioned literature on export-driven economic growth establishes the relevance of the instrument. We show that the instrument is a strong predictor of income, reducing potential concern about weak instruments bias. Moreover, using presample education employment and aggregate exports addresses potential concerns about reverse causality and omitted variables bias. Speciﬁcally, it eliminates reverse causality because city-level new vehicle purchases cannot plausibly affect a city’s 2004 education employment. Moreover, the IV reduces the likelihood that changes in a city’s predicted (second-stage) income are correlated with other factors affecting demand for cars, such as public transportation.The IV reduces measurement error because export-driven high-income growth likely affected workers with high human capital, who are more likely than other workers to purchase new cars. That is, if we were using an instrument based on income to low-skilled or agricultural workers, who purchased relatively few cars, we might be exacerbating rather than reducing measurement error.The exclusion restriction is that a city’s 2004 educational employment is uncorrelated with factors that subsequently affect new car sales independently of income and are not included in the IV estimation. Omitted variables correlated with the instrument are likely the most important remaining concern about the IV strategy. Although unobserved factors may be correlated with initial employment, we show that the city’s 2004 educational employment is uncorrelated with 2004 levels and 2004–2017 growth of variables that may affect car ownership independently of income, such as road space, built area, and the number of buses and taxis. That observed factors are uncorrelated with initial employment supports the exclusion restriction, but that restriction cannot be tested directly.Care must be taken when interpreting the IV coefﬁcient in equation (1). It identiﬁes the effect of income driven by expanding exports and includes effects of income mediated through other factors that are not included in the estimation. For example, if rising income makes owning a new car more fashionable, the IV coefﬁcient includes that effect.As another example, consider trafﬁc congestion. If rising income increases driving16and raises congestion, the coefﬁcient includes that (presumably negative) effect of trafﬁc congestion. If congestion increases for reasons besides income, the IV estimate would be consistent as long as the instrument is uncorrelated with the initial congestion level.A similar argument pertains to other factors affecting car demand, such as public transportation quality and vehicle attribute changes. If rising income causes cities to invest more in public transportation, reducing demand for cars, the IV estimate would capture the effect of income on car sales, net of the opposing effect of public transportation quality. Likewise, the IV estimate includes vehicle attribute changes caused by increasing demand for new cars. At the same time, the IV strategy controls for the effects of other factors on vehicle attributes, such as fuel economy regulation, to the extent that these factors are uncorrelated with the IV.Before turning to the estimation results, we comment on dynamics. We have assumed a contemporaneous relationship between income and new car sales. However, new car sales could respond to lagged income or a moving average of recent income if household-level income shocks are transitory. We allow for such possibilities in the robustness analysis.4 ResultsThis section reports the main results and robustness analysis and compares our estimates with recent forecasts of new vehicle sales in China.4.1. Main ResultsTable 3 reports estimates of equation (1) (OLS) and equation (3) (IV). Column 1 shows the OLS estimate of the key coefﬁcient, αN, from equation (1). The speciﬁcation includes city ﬁxed effects, year ﬁxed effects, and province by year interactions. Standard errors are reported in parentheses, clustered by city. The coefﬁcient on log income is 0.73 and statistically signiﬁcant at the 1 percent level. The estimate means that a 1 percent increase in income is associated with a 0.73 percent increase in new car sales.As the previous section mentioned, the OLS estimate of αN is likely to be inconsistent because of reverse causality, omitted variables bias, and measurement error. Column 2 of Table 3 reports the IV coefﬁcient, using the interaction of the city’s 2004 education employment with aggregate high-technology exports in the corresponding year. The IV estimate is 2.53, which is signiﬁcant at the 1 percent level.The IV coefﬁcient is about three times greater than the OLS coefﬁcient in column 1, which could be explained by reverse causality, omitted variables that are negatively17Table 3: Effect of Income on Vehicle Registrations, Expenditure, and Average PriceDependent var: Estimated by: Log incomeLog income * post 2010Log income instrumentObservations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val(1) Log new registrationsOLS 0.73 (0.11)3,627(2) Log new registrationsIV 2.53 (0.39)3,627 83.33 44.95 0.00(3) Log income OLS2.29 (0.25) 3,627 83.33 44.95 0.00(4) Log new registrationsIV 2.07 (0.52) -0.50 (0.32)3,627 22.88 40.36 0.00(5) Log new expenditureIV 2.64 (0.40)3,627 83.33 44.95 0.00(6) Log averageprice IV 0.10(0.08)3,627 83.33 44.95 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. Column headings state the dependent variable and estimation method. All regressions include city ﬁxed effects, year ﬁxed effects, and province–year interactions.correlated with income, or measurement error that causes attenuation bias. We return to the economic interpretation of this estimate at the end of this section. Column 3 shows the ﬁrst-stage coefﬁcient on the instrument, which is precisely estimated. Column 2 shows that the ﬁrst-stage effective F-statistic is 83, reducing concerns about weak instruments bias.4Column 4 allows for different effects of income on registrations for two periods: 2005–2010 (particularly high income growth) and 2011–2017 (slower income growth). The negative coefﬁcient on the interaction term indicates that income had a smaller effect in the latter period, although the estimate is not statistically signiﬁcant, and it indicates just a 25 percent decline in the elasticity across the two periods.Having shown that rising income causes total new registrations to increase, we consider how income affects the total expenditure on new vehicles and average prices. Columns 5 and 6 are the same as column 2, except that the dependent variable is the log of new car expenditures (column 5) or the log of the sales-weighted average price (column 6). The income coefﬁcient in column 5 is similar to that in column 2, indicating that an increase in income causes new car registrations and expenditures to increase by the same proportion. Consistent with that result is that the coefﬁcient in column 6 is small and not statistically4We use effective F statistics developed by Olea and Pﬂueger (2013) for detecting weak instruments. The approach is robust to heteroskedasticity, time series autocorrelation, and clustering, which are likely to occur in our data. However, it is only available for settings with one endogenous regressor; we are not aware of a heteroskedasticity-consistent weak instrument test for multiple regressors. Since we ﬁnd that in speciﬁcations with a single endogenous regressor, the effective F statistic is within 1 percent of regular F statistics, we report regular F statistics for speciﬁcations with multiple endogenous regressors as an approximation.18signiﬁcant; the data reject the hypothesis that the coefﬁcient equals 1 at the 1 percent level. These estimates mean that rising income causes total new car sales to increase but does not affect the average price of those cars. In other words, as incomes increased during the sample period, consumers purchased more cars but did not substitute systematically toward more expensive cars.The ﬁnding that income has not affected average prices is perhaps surprising, given that one might expect rising income to increase demand for relatively expensive vehicles. We consider two possible explanations for this result. First, the fuel economy and taxation policy discussed in section 2 could encourage sales of small and relatively inexpensive cars. This effect could counteract the effect of rising income on demand for new cars. However, Table A2 shows that rising income tends to increase the average engine size, fuel consumption, horsepower, and weight. Therefore, the regulation and policy do not appear to explain the ﬁnding in columns 5 and 6.A second possibility is that middle- rather than high-income consumers may have been driving the growth in new car sales. That is, a change in the composition of new car buyers over time could counteract the effect of within-household income growth. For example, middle-income households may have higher demand for domestic brands, which tend to be relatively inexpensive. Unfortunately, household-level data on income and vehicle purchases across multiple years are not available, preventing us from testing this hypothesis.4.2. RobustnessThis subsection presents additional estimation results. We consider omitted variables bias, dynamics, and functional form assumptions.As discussed, the IV strategy rests on the assumption that the initial level of education employment is uncorrelated with subsequent unobserved shocks to new vehicle sales. Although we cannot test this assumption directly, we can provide some supporting evidence. Speciﬁcally, if one assumes that unobserved variables in the IV regression are correlated with observed variables, we can check whether the results are sensitive to adding or dropping control variables. This assumption seems reasonable, as omitted variables, such as trafﬁc congestion (which would negatively affect new car demand), are likely to be correlated with observable variables, such as the size of the public transportation system.For convenience, column 1 in Table 4 repeats the main IV regression from column 2 of Table 3, which we refer to as the baseline. Column 2 of Table 4 shows that omitting the province by year interactions causes the income coefﬁcient to increase by about one-third.19Table 4: Effects of Adding Controls on IV EstimatesDependent var: Log new registrations(1)(2)(3)(4)(5)Log incomeLog income * license cap dummyProvince by year FE Socioeconomic controls 2005 income quintile * trend Observations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val2.53 (0.39)YES NO NO 3,627 83.33 44.95 0.003.48 (0.53)NO NO NO 3,627 59.73 46.85 0.002.40 (0.39)YES YES NO 3,627 78.85 45.30 0.002.38 (0.38) -0.05 (0.01) YES NO NO 3,627 41.16 44.80 0.005.58 (2.15)YES NO YES 3,627 8.09 7.93 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. The dependent variable is log registrations. All regressions include city and year ﬁxed effects. Columns 1, 3, 4, and 5 include province–year interactions. Column 3 includes built area, area of paved roads, population, number of buses and taxis in the public transportation system, and total retail revenue. Column 4 includes the interaction of log income and a dummy variable indicating if the city caps new license plates at the time. Each city is assigned to a quintile based on its 2005 income. Column 5 includes the interaction of a linear time trend and ﬁxed effects for the city’s quintile.In column 3, we add several socioeconomic controls that are likely to be correlated withnew car demand, independently of income: built area of the city, area of paved roads,population, number of buses and taxis in the public transportation system, and total retailsales. That including these variables causes the income coefﬁcient to decrease only slightlysupports the identiﬁcation strategy.Moreover, Tables A3 and A4 show that 2004 education employment, which is used toconstruct the instrument, is uncorrelated with the other socioeconomic variables in column3. Speciﬁcally, the appendix tables include interactions of each of the socioeconomicvariables with a linear time trend or year ﬁxed effects (the latter is more ﬂexible). Ifeducation employment were correlated with these variables, adding these controls wouldaffect the IV coefﬁcient. However, the income coefﬁcient is reasonably stable across these speciﬁcations, further supporting the empirical strategy.55China’s Poverty Reduction Program is another potential source of endogeneity. It was ﬁrst introduced in 1984 and aimed to promote economic development in impoverished areas. Between 300 and 500 counties were designated as a "county of extreme poverty" and could receive special aid from the central government for building infrastructure and funding local entrepreneurs (in China, a city is a larger geographic area than a county). The list of counties was updated in 2001 and 2012. To assess whether this program biases our results, we compute the percentage of a city’s population living in a county of extreme poverty. We ﬁnd that the variable is weakly correlated with the instrumented income in equation (1); the results are not affected by including the variable as an independent variable or as an instrument.20As noted, Table A2 indicates that emissions policies are not driving our estimated relationship between income and new registrations. Another potentially relevant policy is the cap that some Chinese cities, such as Beijing and Shanghai, impose on new license plates (Li, 2018): a household wishing to add a new vehicle must ﬁrst obtain a license plate in a lottery or auction. Income should have a smaller effect on new registrations in such cities, although if a household already owns a car, it can avoid the lottery or auction by discarding that car and buying a new one. Column 4 shows that income has a smaller effect for cities with license plate caps than other cities, although the coefﬁcient is small.Next, we turn to dynamics. In principle, mean reversion in income and new car registrations could explain the large effect that we estimate. To illustrate this possibility, consider a hypothetical city that experiences simultaneous negative shocks to income and new vehicle demand prior to our sample. If both income and new vehicle demand are mean reverting processes, we would estimate a positive relationship between the two variables even if the relationship is only spurious. To allow for the possibility of such mean reversion, we compute quintiles of city income using the 2005 distribution. Column 5 of Table 4 adds to the baseline the interactions of quintile ﬁxed effects with a linear time trend. These time trends control for potential mean reversion, and adding these variables would decrease the income coefﬁcient if mean reversion were an important factor. However, as column 5 shows, adding them causes the coefﬁcient to increase. The estimate is signiﬁcant at the one percent level, but the standard error is also larger than in column 1. This reﬂects the correlations among the instrumented income and the quintile-trend interactions; the ﬁrst-stage F-statistic in column 5 is substantially smaller than in column 1. Thus, notwithstanding the large standard errors, we do not ﬁnd evidence that mean reversion causes a spurious estimate.Another issue related to dynamics is the possibility that income has a noncontemporaneous effect on vehicle demand. That is, the baseline IV speciﬁcation includes the implicit assumption that income affects new vehicle demand within a year. In practice, consumers may delay making a new car purchase after their incomes increase for a variety of reasons, such as waiting to determine whether the increase is permanent. Ideally, we would test for such dynamics by adding lags of income to the baseline speciﬁcation, but unfortunately, current income is highly correlated with lagged income. Therefore, in Table 5 in columns 2–4, we replace current income with the one-, two-, or three-year lag (column 1 repeats the baseline). If consumers respond to rising income with a lag, the income coefﬁcient on lagged income would be larger than on current income, but the table shows that this is not the case. Therefore, we do not ﬁnd evidence refuting the hypothesis that new vehicle purchases respond to income within a year; put differently, if purchases respond with a lag,21the lagged response is no larger than the estimated contemporaneous response.Table 5: Effects of Lagged Income on Vehicle RegistrationsDependent var: log new registrations Log incomeObservations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val(1) Current2.53 (0.39) 3,627 83.33 44.95 0.00(2) 1-year lag2.85 (0.42) 3,348 76.06 43.62 0.00(3) 2-year lag2.96 (0.42) 3,069 67.99 41.89 0.00(4) 3-year lag2.59 (0.45) 2,790 60.59 39.24 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. The dependent variable is the log of registrations. All regressions include city and year ﬁxed effects and province–year interactions. The ﬁrst column uses current period log income. Columns 2–4 replace current log income with one-, two-, or three-year lags of log income, instrumented by the corresponding lag of the instrument.Estimating equation 3 yields the sample average elasticity of new registrations to income. As noted, because the income distribution may affect registrations rather than average income, the elasticity could vary across cities or with income. In Table 6, we allow the income coefﬁcient to vary across cities according to the city’s 2005 income. Columns 1 and 2 assign each city to one of two groups, depending on whether its 2005 income is below or above the median 2005 income. Columns 3 and 4 include three equal-sized groups based on 2005 income. The table shows that the effect of income on new car registrations is larger for initially high- than low-income cities, but the difference across city groups is small; columns 1 and 2 show that the high-income city coefﬁcient is about 5–10 percent higher, and columns 3 and 4 show that the effect is 10–20 percent higher, depending on the speciﬁcation and group. However, note that the ﬁrst-stage F statistics are smaller than in the baseline, particularly when we consider three groups. This indicates that although the instrument has sufﬁcient variation to identify the baseline speciﬁcation, unfortunately, variation is insufﬁcient to consider much heterogeneity.22Table 6: Effect of Income by Initial Income LevelDependent var: log new registrations Log incomeLog income * above medianLog income * middle incomeLog income * high income2005 income quintile * trend Observations F statistics for IV Kleibergen-Paap stat Underidentiﬁcation p valTwo groups(1)(2)3.195.78(0.63) (2.33)0.310.27(0.13) (0.17)NO 3,627 17.53 24.48 0.00YES 3,627 3.49 6.88 0.01Three groups(3)(4)4.467.07(1.08) (3.18)0.45 (0.17) 1.01 (0.36) NO 3,6275.6 14.85 0.000.55 (0.29) 0.99 (0.48) YES 3,627 1.60 4.93 0.03Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. The dependent variable is the log of registrations. All regressions include city and year ﬁxed effects. Columns 2 and 4 include interactions of a time trend with 2005 income quintiles. We compute the median income across cities in 2005. Columns 1 and 2 include interactions of log income with a dummy variable equal to one if the city’s 2005 income is greater than the median. We construct three equal-sized groups of cities according to 2005 income, and columns 3 and 4 include the interaction of log income with income group ﬁxed effects. For each column, we form IVs by interacting the income instrument with the corresponding dummy variable or ﬁxed effects.234.3. Comparison with Recent ForecastsThis subsection compares our estimated elasticity of new registrations to income with recent forecasts of the elasticity of new vehicle sales to income. We begin by summarizing recent forecasts published in the literature.Table 7 shows the implied elasticity of new vehicle sales to income from recent forecasts. Some studies report forecasts of new vehicle sales, and others report forecasts of the entire on-road stock. For the latter, we impute sales following Hsieh, Kishimoto, and Green (2018) and Gan et al. (2020) and assuming a survival rate of Chinese vehicles estimated by Lu et al. (2018). A few of the studies report estimates for different scenarios, which we show in different rows in the table.Table 7: Elasticity of Vehicle Sales to GDP from Other StudiesHuo et al. (2007)—High Huo et al. (2007)—Mid Huo et al. (2007)—Low Wang, Teter, and Sperling (2011)—High Wang, Teter, and Sperling (2011)—Low Huo and Wang (2012)—High Huo and Wang (2012)—Low Hsieh, Kishimoto, and Green (2018) Gan et al. (2020) Average2005–2010 3.10 3.43 3.063.202010–2020 1.54 1.61 1.38 0.92 0.93 1.03 1.00 0.541.122020–2030 1.23 0.72 1.07 0.94 0.83 0.64 0.61 0.85 0.54 0.832030–2040 0.910.730.45 0.41 0.12 0.30 0.492040–2050 0.760.510.37 0.34 0.050.41Notes: The table shows the elasticity of vehicle sales to income from recently published forecasts of oil consumption and GHG emissions in China. For studies that project the vehicle stock rather than new vehicle sales, we use the method adopted from Hsieh, Kishimoto, and Green (2018) and Gan et al. (2020) to impute new vehicle sales. We follow the survival rate schedule for Chinese vehicles estimated by Lu et al. (2018).The table reports the elasticity of sales to income by decade. Overall, studies forecast a declining elasticity over time, from an average of 3.2 for 2005–2010 to 0.41 for 2040–2050.We use the observed income changes for 2005–2010 and 2010–2017 to compute the sales change predicted for each city. Speciﬁcally, we calculate total income across all cities in our data for each year. The predicted change in log total sales for 2005–2010 equals the 2005–2010 income elasticity multiplied by the change in total log income for the corresponding years. We repeat this calculation for 2010–2017. Table A7 shows calculations for each study in Table 7, and the ﬁrst row of Table 8 shows the results using the average elasticities in the literature. The literature predicts a change in log sales of 1.8 for 2005–2010 and predicts 0.6 for 2010–2017. Thus, the literature predicts a dramatic slowdown in sales growth from 2010 to 2017.24Because the literature uses national data and we use city-level data, to facilitate comparisons, we need to adjust our estimates. For each city, we predict the 2005–2010 change in log new registrations by multiplying αN by the 2005–2010 change in log income. Exponentiating this expression, summing across cities, and taking logs yields the predicted change in log national sales for 2005–2010.6 We repeat the calculation for 2010–2017. Note that we use the baseline speciﬁcation, which includes a single income coefﬁcient for 2005–2017 because we reject the hypothesis of a large change in the coefﬁcient between the two time periods.Table 8: Comparison of Implied Elasticity Between Our Model and Previous StudiesLiterature average Prediction with single coefﬁcient∆ln Sale 05−10 1.8 1.4∆ln Sale 10−17 0.6 1.3Total: ∆ln Sale 05−17 2.4 2.7Notes: See text for details on the calculations. For studies lacking projections for 2005–2010, we use the average income elasticity of other studies during this period.The second row of Table 8 shows the results using our estimates. Given observed income growth, our estimates predict a 2005–2010 increase in log sales of 1.4, which is 0.4 less than the 1.8 predicted by the literature. In contrast, for 2010–2017, our estimates predict about twice the increase in log sales as the literature—1.3 versus 0.6. Across the entire 2005–2017 period, our estimates predict greater growth by 0.3 log points, which translates to about 36 percent.7The additional vehicles sold during the 2005–2017 period caused China’s 2017 emissions to be higher than projected. Using assumptions on driving and scrappage of older vehicles from (Shen, 2021), we calculate that the additional vehicles sold between 2005 and 2017 caused emissions to be 18 million tons higher in 2017 than projected, which represents about 2 percent of China’s carbon dioxide emissions from the transportation sector. Note that Table 8 shows that the literature underpredicted log sales growth by half between 2010 and 2017; if this difference continued in the late 2010s and continues in the 2020s, the gap between predicted and actual emissions will grow over time.6That is, ∆ ln Sale 05−17 = ln ∑i exp (βIV · ∆ln Incomei + ln Salei2005) − ln (∑i ln Sale i2005). 7Between 2005 and 2017, actual log sales increased by about 1.5. This number is smaller than either our estimates or the literature predicts, likely because both isolate the effect of income on sales. Other developments in China oppose the effect of rising income, such as expanding public transportation and trafﬁc congestion.255 ConclusionThis paper reports a strong connection between income growth and new vehicle sales in China. We assemble a unique data set of city-level sales, income, and socioeconomic characteristics. We use an instrumental variables strategy that isolates income growth driven by high-technology exports and addresses potential concerns about endogeneity and measurement error of income. The preferred speciﬁcation indicates an elasticity of city-level new car sales to income of 2.5.We show that recent forecasts of vehicle sales in China appear to have substantially underestimated the effect of income on sales between 2005 and 2017. Our estimates indicate that income growth has caused new car sales to grow by 36 percent more than the average growth anticipated in forecasts conducted in the 2000s or early 2010s.The results suggest that China’s future oil consumption and GHG emissions may be higher than recent studies have predicted. Back-of-the-envelope calculations show that past forecasts have underestimated China’s transportation sector GHG emissions by about 2 percent in 2017.Estimating the implications of our estimates for future GHG emissions would require a computational model of China’s economy, which lies outside the scope of our paper. Previous forecasts anticipate that the effect of income growth on sales growth will diminish over the coming decades, as vehicle ownership follows an S-shaped curve. However, given how dramatically these studies underpredicted sales growth in the 2010s, it seems unlikely that sales growth in the 2020s will slow to the low levels these studies anticipate. Nonetheless, because GHG emissions from passenger cars are roughly proportional to the size of the car stock in such models, our results imply that China’s GHG emissions from cars will be substantially higher than recent forecasts anticipate. In that case, future oil consumption and GHG emissions (in the absence of policy intervention) would be much higher than expected, and meeting China’s pledge under the UN Paris Agreement would require more aggressive policies than if the forecasts had proven to be accurate. Future work may assess how much policies would have to account for China’s unexpectedly large increase in car ownership.26ReferencesAuffhammer, Maximilian and Catherine D Wolfram (2014). “Powering up China: Income distributions and residential electricity consumption”. In: American Economic Review 104.5, pp. 575–80.Baum-Snow, Nathaniel and Fernando Ferreira (2015). “Causal inference in urban and regional economics”. In: Handbook of regional and urban economics. Vol. 5. Elsevier, pp. 3– 68.Beaudry, Paul, David A Green, and Benjamin Sand (2012). “Does industrial composition matter for wages? A test of search and bargaining theory”. In: Econometrica 80.3, pp. 1063–1104.BloombergNEF (2020). “Electric Vehicle Outlook”. In: URL: https://about.bnef.com/ electric-vehicle-outlook/.Blumenberg, Evelyn and Gregory Pierce (2012). “Automobile ownership and travel by the poor: Evidence from the 2009 National Household Travel Survey”. In: Transportation Research Record 2320.1, pp. 28–36.Calderón, Silvia et al. (2016). “Achieving CO2 reductions in Colombia: Effects of carbon taxes and abatement targets”. In: Energy economics 56, pp. 575–586.Calvin, Katherine et al. (2013). “Implications of simultaneously mitigating and adapting to climate change: initial experiments using GCAM”. In: Climatic Change 117.3, pp. 545– 560.Cantore, Nicola (2011). “Distributional aspects of emissions in climate change integrated assessment models”. In: Energy Policy 39.5, pp. 2919–2924.Cherp, Aleh et al. (2016). “Global energy security under different climate policies, GDP growth rates and fossil resource availabilities”. In: Climatic Change 136.1, pp. 83–94.Dargay, Joyce, Dermot Gately, and Martin Sommer (2007). “Vehicle ownership and income growth, worldwide: 1960-2030”. In: The energy journal 28.4.Dargay, Joyce M (2001). “The effect of income on car ownership: evidence of asymmetry”. In: Transportation Research Part A: Policy and Practice 35.9, pp. 807–821.Dasgupta, Susmita et al. (2002). “Confronting the environmental Kuznets curve”. In: Journal of economic perspectives 16.1, pp. 147–168.Davis, Lucas W, Alan Fuchs, and Paul Gertler (2014). “Cash for coolers: evaluating a large-scale appliance replacement program in Mexico”. In: American Economic Journal: Economic Policy 6.4, pp. 207–38.EIA (2017). International Energy Outlook. — (2019). International Energy Outlook.27Falk, Martin (2009). “High-tech exports and economic growth in industrialized countries”. In: Applied Economics Letters 16.10, pp. 1025–1028.Gan, Yu et al. (Mar. 2020). “Future private car stock in China: current growth pattern and effects of car sales restriction”. In: Mitigation and Adaptation Strategies for Global Change 25, pp. 289–306.Goldsmith-Pinkham, Paul, Isaac Sorkin, and Henry Swift (2020). “Bartik instruments: What, when, why, and how”. In: American Economic Review 110.8, pp. 2586–2624.Hausmann, Ricardo, Jason Hwang, and Dani Rodrik (2007). “What you export matters”. In: Journal of economic growth 12.1, pp. 1–25.He, Kebin et al. (2005). “Oil consumption and CO2 emissions in China’s road transport: current status, future trends, and policy implications”. In: Energy policy 33.12, pp. 1499– 1507.Hsieh, I-Yun Lisa, Paul Natsuo Kishimoto, and William H Green (2018). “Incorporating multiple uncertainties into projections of Chinese private car sales and stock”. In: Transportation Research Record 2672.47, pp. 182–193.Huo, Hong and Michael Wang (Apr. 2012). “Modeling future vehicle sales and stock in China”. In: Energy Policy 43, pp. 17–29.Huo, Hong et al. (Dec. 2007). “Projection of Chinese Motor Vehicle Growth, Oil Demand, and CO 2 Emissions Through 2050”. In: Transportation Research Record 2038, pp. 69–77.IEA (2017). World Energy balances and CO2 emissions from fuel combustion. Jaeger, David A, Joakim Ruist, and Jan Stuhler (2018). Shift-share instruments and the impactof immigration. Tech. rep. National Bureau of Economic Research. Jarreau, Joachim and Sandra Poncet (2012). “Export sophistication and economic growth:Evidence from China”. In: Journal of development Economics 97.2, pp. 281–292. Kaika, Dimitra and Efthimios Zervas (2013). “The environmental Kuznets curve (EKC)theory. Part B: Critical issues”. In: Energy Policy 62, pp. 1403–1411. Krey, Volker et al. (2012). “Urban and rural energy use and carbon dioxide emissions inAsia”. In: Energy Economics 34, S272–S283. Levin, Andrew and Lakshmi K Raut (1997). “Complementarities between exports andhuman capital in economic growth: evidence from the semi-industrialized countries”. In: Economic development and cultural change 46.1, pp. 155–174. Li, Shanjun (2018). “Better Lucky Than Rich? Welfare Analysis of Automobile License Allocations in Beijing and Shanghai”. In: Review of Economic Studies 85, pp. 2389–2428. Li, Yating et al. (Aug. 2019). “Household appliance ownership and income inequality: Evidence from micro data in China”. In: China Economic Review 56, p. 101309.28Lu, Zifeng et al. (2018). China Vehicle Fleet Model: Estimation of Vehicle Stocks, Usage, Emissions, and Energy Use-Model Description, Technical Documentation, and User Guide. Tech. rep. Argonne National Lab.(ANL), Argonne, IL (United States).Luderer, Gunnar et al. (2015). “Description of the REMIND model (Version 1.6)”. In: McNeil, Michael A and Virginie E Letschert (2010). “Modeling diffusion of electricalappliances in the residential sector”. In: Energy and Buildings 42.6, pp. 783–790. Mehrara, Mohsen, Samaneh Seijani, and Abbas Rezazadeh Karsalari (2017). “Determi-nants of high-tech export in developing countries based on Bayesian model averaging”. In: Zbornik radova Ekonomskog fakulteta u Rijeci: cˇasopis za ekonomsku teoriju i praksu/Proceedings of Rijeka Faculty of Economics: Journal of Economics and Business 35.1, pp. 199–215. Mulliqi, Arta, Nick Adnett, and Mehtap Hisarciklilar (2019). “Human capital and exports: A micro-level analysis of transition countries”. In: The Journal of International Trade & Economic Development 28.7, pp. 775–800. Mundaca, Luis and Margaret Samahita (2020). “What drives home solar PV uptake? Subsidies, peer effects and visibility in Sweden”. In: Energy Research & Social Science 60, p. 101319. Nieto, Jaime et al. (2020). “An ecological macroeconomics model: The energy transition in the EU”. In: Energy Policy 145, p. 111726. Nolan, Anne (2010). “A dynamic analysis of household car ownership”. In: Transportation research part A: policy and practice 44.6, pp. 446–455. Nordhaus, William D and Zili Yang (1996). “A regional dynamic general-equilibrium model of alternative climate-change strategies”. In: The American Economic Review, pp. 741–765. Nunn, Nathan and Nancy Qian (2014). “US food aid and civil conﬂict”. In: American Economic Review 104.6, pp. 1630–66. Oakil, Abu Toasin, Dorien Manting, and Hans Nijland (2018). “The role of individual characteristics in car ownership shortly after relationship dissolution”. In: Transportation 45.6, pp. 1871–1882. Olea, José Luis Montiel and Carolin Pﬂueger (2013). “A robust test for weak instruments”. In: Journal of Business & Economic Statistics 31.3, pp. 358–369. Pan, Xunzhang et al. (2018). “Decarbonization of China’s transportation sector: in light of national mitigation toward the Paris Agreement goals”. In: Energy 155, pp. 853–864. Ruijven, Bas J van et al. (2012). “Implications of greenhouse gas emission mitigation scenarios for the main Asian regions”. In: Energy economics 34, S459–S469. Shaﬁk, Nemat and Sushenjit Bandyopadhyay (1992). Economic growth and environmental quality: time-series and cross-country evidence. Vol. 904. World Bank Publications.29Shen, Chang (2021). The value of Fuel Economy: Evidence from the Chinese Passenger Vehicle Market. Tech. rep. University of Maryland.Shen, Hongyan, Fei Teng, and Jinping Song (2018). “Evaluation of spatial balance of China’s regional development”. In: Sustainability 10.9, p. 3314.Steckel, Jan Christoph et al. (2013). “Development without energy? Assessing future scenarios of energy consumption in developing countries”. In: Ecological Economics 90, pp. 53–67.Stokey, Nancy L (1991). “Human capital, product quality, and growth”. In: The Quarterly Journal of Economics 106.2, pp. 587–616.Tan, Jijun, Junji Xiao, and Xiaolan Zhou (2019). “Market equilibrium and welfare effects of a fuel tax in China: The impact of consumers’ response through driving patterns”. In: Journal of Environmental Economics and Management 93, pp. 20–43.Vliet, Oscar van et al. (2012). “Synergies in the Asian energy system: Climate change, energy security, energy access and air pollution”. In: Energy Economics 34, S470–S480.Wagner, Martin (2008). “The carbon Kuznets curve: a cloudy picture emitted by bad econometrics?” In: Resource and Energy Economics 30.3, pp. 388–408.Wang, Yunshi, Jacob Teter, and Daniel Sperling (June 2011). “China’s soaring vehicle population: Even greater than forecasted?” In: Energy Policy 39, pp. 3296–3306.Yan, Xiaoyu and Roy J Crookes (2009). “Reduction potentials of energy demand and GHG emissions in China’s road transport sector”. In: Energy Policy 37.2, pp. 658–668.Yao, Shujie, Shuai Zhang, and Xingmin Zhang (2019). “Renewable energy, carbon emission and economic growth: A revised environmental Kuznets Curve perspective”. In: Journal of Cleaner Production 235, pp. 1338–1352.Yin, Jianhua, Mingzheng Zheng, and Jian Chen (2015). “The effects of environmental regulation and technical progress on CO2 Kuznets curve: An evidence from China”. In: Energy Policy 77, pp. 97–108.Zhao, Tingting et al. (2012). “Consumer responses towards home energy ﬁnancial incentives: A survey-based study”. In: Energy Policy 47, pp. 291–297.Zwaan, Bob van der et al. (2018). “An integrated assessment of pathways for low-carbon development in Africa”. In: Energy Policy 117, pp. 387–395.30AppendixFigure A4: New registrations growth trend for cities with license cap policyNotes: On the y-axis, new registrations are expressed as an index relative to the city’s 2005 level. On the x-axis, we normalize time to zero in the year that the license cap policy starts. For instance, Beijing’s lottery started in 2011. In this chart, we put the 2011 registration data for Beijing at t = 0 and 2012 data at t = 1, and so on.31Table A1: Summary statistics by initial income quintilesQuantiles of 2005 income (1st is lowest quantile) New car sold New car expenditure Average car price Income Built-up area Area of paved roads Population Bus Taxi Total retail Employment ratio in education (2004) National high-tech export1st19.1 (20.6)2.5 (2.5) 142.6 (22.4) 33.0 (13.1) 63.7 (33.4) 8.1 (5.5) 4,821.6 (2,776.8) 0.4 (0.4) 1.5 (1.0) 39.0 (30.1) 4.5 (2.4) 3,786.0 (590.5)2nd21.6 (30.4)2.9 (3.9) 145.4 (23.3) 37.5 (13.8) 67.4 (57.3) 8.6 (9.2) 3,899.1 (2,203.5) 0.5 (0.7) 1.7 (2.3) 37.9 (37.6) 3.9 (2.4) 3,786.0 (590.5)3rd23.4 (28.4)3.2 (3.6) 149.6 (25.0) 39.0 (13.6) 71.8 (44.1) 9.6 (7.4) 4,229.7 (2,106.8) 0.6 (0.5) 1.6 (1.6) 45.2 (37.9) 4.6 (4.3) 3,786.0 (590.5)4th37.8 (61.5)5.7 (9.2) 155.6 (25.3) 43.2 (14.3) 135.5 (159.4) 16.4 (21.0) 4,306.4 (4,523.9) 1.4 (1.9) 3.3 (3.8) 74.6 (100.2) 4.9 (4.7) 3,786.0 (590.5)5th88.5 (100.5)15.2 (18.0) 172.2 (28.0) 54.4 (17.9) 269.5 (274.3) 37.1 (34.3) 4,517.2 (3,181.4)3.7 (5.2) 7.3 (11.6) 170.0 (189.5) 6.3 (6.7) 3,786.0 (590.5)Notes: The table shows the mean value for each variable and standard deviation in parentheses. The data contain 3,627 observations. "New car expenditures" equals the total expenditure on new cars. "CV" is the coefﬁcient of variation. All expenditure variables (new car expenditure, total retail, and national high-tech export) are reported in billion RMB. The weighted average of vehicle price and income are expressed in thousand RMB. "Built-up area" is the area of land constructed for residential, commercial, or industrial use. Both "built-up area" and "area of paved roads" are in square kilometers. "New car sold," "population," "bus" and "taxi" are expressed in thousand units. "Bus" and "taxi" are the numbers of buses and taxis, respectively, operating in the city. The employment ratio is expressed in percentage points. All monetary variables are adjusted for inﬂation and measured in 2017 RMB.32Table A2: Impact of income on vehicle attributesDependent variable isLog incomeObservations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val First stage IV(1) Log average engine size0.17 (0.05) 3,627 83.33 44.95 0.00 2.29 (0.25)(2) Log average fuel consumption0.13 (0.04) 3,627 83.33 44.95 0.00 2.29 (0.25)(3) Log average horsepower0.18 (0.05) 3,627 83.33 44.95 0.00 2.29 (0.25)(4) Log average gross weight0.07 (0.03) 3,627 83.33 44.95 0.00 2.29 (0.25)Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. Column headings state the dependent variable. All regressions include city and year ﬁxed effects and province–year interactions.Table A3: Adding initial level of socioeconomic variable * linear time trendDependent: Log new registrations Including t * 2005 level of Log incomeObservations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val(1) Built-uparea 2.39 (0.42) 3,627 64.40 38.33 0.00(2) Paved road2.55 (0.44) 3,627 60.54 35.85 0.00(3)population2.11 (0.42) 3,627 70.98 40.49 0.00(4) Numberof bus 2.41 (0.41) 3,627 67.03 39.40 0.00(5) Number of taxi2.48 (0.40) 3,627 75.81 42.00 0.00(6) Total retail 2.49 (0.41) 3,627 69.34 39.70 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. All columns use log new registration as dependent variable. All regressions include city ﬁxed effects and province by year ﬁxed effects. The regression is based on our main regression (Table 3 column 2), but for each column, we add in the interaction between one initial socioeconomic variable and a linear time trend. For instance, column 1 adds 2005 level built-up area interacted with a linear time trend t.33Table A4: Adding initial level of socioeconomic variable * year FEDependent: Log new registrations Including year FE * 2005 level of Log incomeObservations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val(1) Built-uparea 2.46 (0.42) 3,627 61.61 36.96 0.00(2) Paved road2.55 (0.44) 3,627 59.63 35.13 0.00(3)population2.11 (0.42) 3,627 74.70 41.18 0.00(4) Numberof bus 2.47 (0.41) 3,627 64.27 38.25 0.00(5) Number of taxi2.51 (0.41) 3,627 73.38 41.11 0.00(6) Total retail 2.53 (0.42) 3,627 66.10 38.31 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. All columns use log new registration as dependent variable. All regressions include city ﬁxed effects and province by year ﬁxed effects. The regression is based on our main regression (Table 3 column 2), but for each column we add in the interaction between one initial socioeconomic variable and year ﬁxed effects. For instance, column 1 adds 2005 level built-up area interacted with a linear time trend t.Table A5: First stage for Table 4Dependent: Log income(1)(2)(3)(4)Log income IVProvince by year FE Socioeconomic controls Avg quantile income in 2005 by year tend Observations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val2.29 (0.25) YES NO NO 3,627 83.33 44.95 0.002.01 (0.26) NO NO NO 3,627 59.73 46.85 0.002.21 (0.25) YES YES NO 3,627 78.85 45.30 0.000.69 (0.24) YES NO YES 3,627 8.09 7.93 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. All columns use log income as dependent variable. All regressions include city ﬁxed effects. The 1, 3, and 4 columns include province–year interactions, whereas column 2 only includes year ﬁxed effects. Column 3 adds a set of city-level socioeconomic controls, including built area, area of paved roads, population, number of buses and taxis in the public transportation system, and total retail revenue. Column 4 also controls for the interaction between the average initial income for each quantile and year trend. The quantiles are also deﬁned using initial income into ﬁve groups.34Table A6: First stage for Table 5Dependent: Log income Log income IVObservations Effective F stat for IV Kleibergen-Paap stat Underidentiﬁcation p val(1) Current2.29 (0.25) 3,627 83.33 44.95 0.00(2) 1 year lag2.16 (0.25) 3,348 76.06 43.62 0.00(3) 2 years lag1.98 (0.24) 3,069 67.99 41.89 0.00(4) 3 years lag1.76 (0.23) 2,790 60.59 39.24 0.00Notes: The table reports coefﬁcient estimates with standard errors in parentheses, clustered by city. Column headings state the dependent variable. All regressions include city and year ﬁxed effects and province–year interactions. The ﬁrst column uses current period log income, instrumented with current period IV. The second to the fourth columns replace current log income with one-, two-, or three-year lags, instrumented by the corresponding lag of instrument.Table A7: Vehicle sales projection from previous studiesHuo et al. (2007)—High Huo et al. (2007)—Mid Huo et al. (2007)—Low Wang, Teter, and Sperling (2011)—High Wang, Teter, and Sperling (2011)—Low Huo and Wang (2012)—High Huo and Wang (2012)—Low Hsieh, Kishimoto, and Green (2018)∆ln Sale 05−101.8 1.7 1.8 1.9 1.7 1.8 1.8 1.8∆ln Sale 10−200.8 0.9 0.8 0.5 0.5 0.6 0.5 0.3Total: ∆ln Sale 05−202.6 2.6 2.5 2.4 2.2 2.3 2.3 2.1Sale201732.7 32.3 30.1 26.6 21.8 24.9 24.4 19.1Notes: Predicted gross sale in millions of cars. For studies that do not have projections for 2005–2010, we use the average of implied income elasticity of vehicle sales of available studies during this period.35