Modeling Private Car Ownership in China

Modeling Private Car Ownership in China Investigation of Urban Form Impact Across Megacities Jieping Li, Joan L. Walker, Sumeeta Srinivasan, and William P. Anderson grown even faster. The rising prevalence of private cars in China is occurring mainly in megacities. Smaller cities and towns have seen modest growth in private car ownership (1). If the current car ownership growth trend continues, it will exacerbate numerous issues in megacities: land use scarcity, energy crisis, environmental pollution, health problems, urban smog, global warming, and traffic congestion. It is critical for policy makers to understand car ownership behavior even as they seek to restrain car ownership through nonauto-oriented urban planning. The paper seeks to address the following questions: How does urban form, such as population density and urban scale, affect auto ownership across megacities in China? What other socioeconomic and demographic variables drive car ownership across cities? The paper is organized as follows. First, it reviews previous related literature. The study methodology is proposed on the basis of the literature review. The empirical study which involved the use of both aggregate and disaggregate analyses, is then described. Finally, conclusions are drawn, limitations are discussed, and future work is recommended. The rising prevalence of private cars in the developing world is causing serious congestion and pollution. In China, private cars started to emerge as an important travel mode in the past decade. Prospective research on the relationship between urban form and car ownership is relatively uncommon in the developing world, and China offers a unique study opportunity, given the tremendous increases in private cars and fast-paced urbanization over the past decade. This study investigates the influence of urban form on car ownership as well as the impact of other socioeconomic and demographic factors on private car ownership across megacities in China. Analysis was conducted through the use of data from 36 megacities and two household survey data sets collected in Beijing and the city of Chengdu, China. Ordinary least squares regression and discrete choice models were employed to execute the aggregate and disaggregate analysis of the urban form impact on private car ownership across cities. The statistical model results demonstrate that urban affluence, urban scale, and road infrastructure supply factors have significant positive effects on the city level of private car ownership across cities. Population density calculated at the subdistrict level, however, had a significant negative effect on private car ownership across cities. Households with private cars were found to prefer to live close to urban centers where amenities were readily available. The results provide evidence for urban planners and policy makers. LITERATURE REVIEW Literature exists on various types of car-ownership modeling. In 2004, De Jong et al. classified into a number of model types the car ownership models for public-sector planning found in the academic literature after 1995 (2). According to data type, they can be categorized into aggregate or disaggregate models. By forecast horizon, they include short-term or long-term models. In terms of data span period, they are static or dynamic models. The research reported on here focuses on earlier studies that addressed, through the use of modeling approaches, how urban form, as well as other socioeconomic and demographic factors, affected car ownership. Aggregate car ownership models to examine the relationship between zonal–regional car ownership and the relevant explanatory variables were presented by Khan and Willumsen in 1986 (3), by De Jong in 1990 (4), by Ingram and Liu in 1997 (5), by Dargay and Gately in 1999 (6), by Romilly et al. in 2001 (7), by Lam and Tam in 2002 (8), by Holtzclaw et al. in 2002 (9), and by Clark in 2007 (10). Across this literature, the major determinant factor of economic development on car ownership was frequently tested by including the variables of gross national product (GNP), gross domestic product (GDP), and disposal income. Khan and Willumsen (3) and Ingram and Liu (5) used GNP per capita in their studies and found it had a positive effect on car ownership. Dargay and Gately (6), Romilly et al. (7), and Lam and Tam (8) concluded that income or disposal income was the major determinant of car ownership at the regional level. Population density was another critical factor with an impact on zonal–regional car ownership. Lam and Tam (8), Holtzclaw et al. (9), and Clark (10) have all demonstrated that population density had a negative Car ownership is an important travel behavior, for it interacts both directly and indirectly with trip frequency, trip length, mode choice, and vehicle miles driven. To understand car ownership behavior is especially important in the context of developing countries where the rapid growth of private cars causes serious congestion and pollution. Prospective research on the relationship between urban form and car ownership is relatively uncommon in the developing world. China offers a unique study opportunity, given the tremendous increase in private cars and fast-paced urbanization in the past decade. Car ownership increased 22.48% annually at the national level from the early 1990s to 2005. In 1985, only 285,000 private cars existed in China. By the end of 2005, the number had reached 18.5 million. This rapid growth of private car ownership is still occurring, and its speed has J. Li, Central Transportation Planning, State Transportation Building, Ten Park Plaza, Boston, MA 02116. J. L. Walker, Center for Global Metropolitan Studies, Department of Civil and Environmental Engineering, Institute of Transportation Studies, University of California, Berkeley, 111 McLaughlin Hall, Berkeley, CA 94720. S. Srinivasan, School of Engineering and Applied Sciences and Center for Government and International Studies, Knafel Center, Room 407, Harvard University, 1737 Cambridge Street, Cambridge, MA 02138. W. P. Anderson, Department of Political Science, University of Windsor, 401 Sunset Avenue, Windsor, Ontario N9B 3P4, Canada. Corresponding author: J. Li, [email protected]. Transportation Research Record: Journal of the Transportation Research Board, No. 2193, Transportation Research Board of the National Academies, Washington, D.C., 2010, pp. 76–84. DOI: 10.3141/2193-10 76 Li, Walker, Srinivasan, and Anderson effect on car ownership. Clark discovered that the influence of population density on car ownership varied across different areas. Romilly et al. examined, in addition to economic development and density factors, the variables of income, traffic congestion, road length, motoring cost, bus fares, interest rate, and unemployment rate on car ownership in Great Britain over time (7 ). Lam and Tam concluded that the first registration tax had a significant negative impact on car ownership (8). Romilly et al. found that income, motoring costs, bus fares, and road traffic congestion had significant effects on car ownership (7). In the aggregate car ownership literature, two ways often have been taken to model the relationship between car ownership and influential factors. One way is to use linear regression and the other is to use a logit model. In both ways the zonal–regional car ownership level regresses as a function of the relevant explanatory variables. Another line of car ownership modeling literature used disaggregate household survey data to probe into the urban form, as well as other socioeconomic and demographic effects, on the individualhousehold, car-ownership decision. Four types of urban form measurements were often investigated in the reviewed disaggregate car ownership modeling literature, including land use measurement, location effect, transit accessibility, and urban design according to the urban form attributes involved. Land use measurements are characterized by gross population density, residential population density, dwelling unit density, road density, employment density, job–housing balance, mixed use, and diversity index. Cervero in 1996 (11), Schimek in 1996 (12), Ryan and Han in 1999 (13), and Hess and Ong in 2002 (14) examined the impact of population or household density on private car ownership. Their studies found that population and residential density had a negative effect on car ownership. Schimek found the effect of density on household automobile ownership was modest after use of a 1990 nationwide U.S. personal transportation survey (12). In addition to Cervero (11), Chu in 2002 (15), Potoglou and Kanaroglou in 2006 (16), and Zegras in 2006 (17) concluded that car ownership decreased when the land-use mixture increased. Location effects are commonly associated with the variables of urban location, rural location, distance to the central business district (CBD), centrality, and ring road location. In addition to Schimek (12) and Zegras (17), Bento et al. in 2005 (18), and Prevedouros and Schofer in 1992 (19), demonstrated that households had fewer cars when their locations were close to the center of the city, represented by the distance to the CBD or the centrality index. Transit accessibility generally is indicated by the nearest bus stop, distance to transit, proximity to transit stations, and transit supply. In addition to Ryan and Han (13), Potoglou and Kanaroglou (16), Bento et al. (18), and Prevedouros and Schofer (19), Kim and Kim in 2004 provided evidence that transit access and transit service had a significant negative effect on the number of automobiles owned (20). In 2001, however, Kitamura et al. concluded that “Accessibility no longer affects automobile ownership in the metropolises of industrialized countries where motorization has matured” (21). Urban design is normally measured by using the variables of the pleasantness of traveling on foot, pedestrian environment, street design, and city shape. Schimek (12) and Hess and Ong (14) illustrated that traditional neighborhoods with friendly walking and biking environments tended to reduce car ownership, whereas Zegras, who used 2001 data from Santiago, Chile, found that street patterns, block morphology, and intersection densities had no effect on car ownership (17). The disaggregate car ownership modeling literature usually takes the form of discrete choice models in which urban form effects are captured after controlling for other socioeconomic and demographic effects. The following sets of socioeconomic and demographic vari- 77 ables are often contained in addition to urban form measurements, including household characteristics (e.g., income, children, size, age-structure, and dwelling type); household head characteristics (e.g., age, gender, and education); car cost and fee variables (e.g., purchase cost, operation cost, registration fee, annual usage tax, and fuel or gas fee); and other relevant variables (e.g., parking availability and company car availability). The common feature of this literature is that researchers examined the urban form measurements as well as the impact of other socioeconomic factors on car ownership by using either aggregate analysis or disaggregate analysis. The present study sought to examine the impact of urban form on automobile ownership across megacities by using both aggregate and disaggregate analyses. The aggregate analysis makes it is easy to observe the overall urban form impact on the level of private cars in a city or region or across a city or region. The effects of transport network, city location, urban characteristics, and transport policy on private car ownership could be observed in the aggregate analysis. The aggregate analysis cannot, however, reveal the behavior mechanisms by which households make the decision on car ownership. Disaggregate analysis compensates for this drawback since it relies on microlevel information and uncovers the individual behavior explicitly. In addition, from the perspective of modeling specification and identification, some variables can only be identified in the aggregate models and vice versa. Another advantage of using both aggregate and disaggregate analyses is that all the available data can be used to examine many of the possible factors in addition to urban form measures. The earlier literature illustrated that both city and household locations play critical roles in the determination of household automobile ownership. The combination of aggregate and disaggregate analyses enable observation of the impact of both city and household characteristics on private car ownership. Research on the impact of urban form on auto ownership in the developing world is inadequate at the present time. As Khan and Willumsen pointed out, the limitations of data, the cost of conventional planning methods, and the scarcity of technical resources make transport modeling difficult in the developing world (3). Another reason that car ownership modeling has rarely been studied in developing countries, such as China, is that the prevalence of private car ownership only recently emerged in the 1990s, along with economic expansion and improvement in quality of life. Either aggregate or disaggregate car ownership models from the urban planning perspective are nonexistent in China. This study seeks to fill the gap and to explore many of the possible factors, in particular the impact of urban form, on private car ownership by using city-level data from 36 megacities and two household survey data sets collected in the cities of Beijing and Chengdu, China. EMPIRICAL STUDY The study employed linear regression methods in the aggregate analysis and binary logit models in the disaggregate analysis to examine the association between car ownership and urban form across cities and between households. The empirical study was conducted in two stages. In the first stage, aggregate data from 36 cities and variables that indicated urban form were collected to examine the association with car ownership. Factor analysis was employed to reduce the correlation among the correlated variables and extract the factors that indicated urban form. A linear regression model was then constructed to examine the correlation between the resulting urban form factors 78 Transportation Research Record 2193 and the number of private cars per hundred households at city level. The second stage of the study used two sets of household survey data collected in Beijing in 2006 (2006 Yearbook of Beijing) and Chengdu in 2005 (2006 Yearbook of Chengdu) and employed binary logit models to explore the impact of neighborhood urban form as well as other sociodemographic factors on the household car ownership decision. Aggregate Analysis The aggregate analysis employed an ordinary least squares (OLS) regression model to examine the urban form influence on car ownership across 36 megacities. Factor analysis was first employed to identify the urban form factors after the capture of the correlation among the correlated nondummy variables. OLS linear regression was then applied to examine the relationship between the dependent variables— city-level, private car ownership per hundred households—and the independent variables—resultant urban form factors and other explanatory variables. FIGURE 1 Spatial distribution of private cars in 36 megacities in China. Aggregate Data and Variable Description Aggregate city-level data were collected from 36 megacities. These cities were undergoing fast urbanization and private motorization. They included four, direct-controlled municipalities, 27 provincial capitals, and five subprovincial cities. Municipalities are the highest level cities in China and are controlled directly by central government. They include Beijing, Shanghai, Tianjin, and Chongqing. Provincial cities are the capitals of the provinces. Subprovincial cities are prefecture-level cities. They are ruled by provinces but administered independently with regard to economics and law. Available variables that potentially affected city-level car ownership were collected from statistical yearbooks (i.e., 2006 Yearbook of Beijing, 2006 Yearbook of Chengdu, 2006 Yearbook of China’s Cities, 2006 China Statistical Yearbook) and official websites. The number of households that owned private cars per 100 households was used to represent the car ownership at city level. The number was used as the dependent variable in the OLS regression. The distribution of automobile ownership was quite uneven among the 36 cities (Figure 1). In 2006, Guangzhou, Beijing, Chengdu, and Shenzhen, Li, Walker, Srinivasan, and Anderson 79 • Annual average household disposable income (yuan), • Per capita GDP (yuan), • Nonagricultural population (10,000), • Built-up area (square kilometers), • Road density (kilometers per square kilometer), • Road area per capita (square meters), • Number of buses (per 10,000 persons), • Number of taxis (per 10,000 persons), • Metro dummy (six cities have metro), • Bus rapid transit (BRT) dummy (15 cities have BRT), • Auto industry dummy (14 cities have auto industry), • Coast dummy (13 cities are coastal cities), and • License plate fee policy dummy (Shanghai was the only city that had a license plate fee in 2006). Factor Analysis and OLS Regression Results Three urban form factors were extracted from the following six original nondummy explanatory variables: (a) household disposal income, (b) GDP per capita, (c) total population, (d) built-up area, (e) road density, and ( f ) road area per person. The factors are defined on the basis of which original variables had the highest correlations with the principal component factors. Disposable income and GDP per capita had the highest component loadings–correlation coefficients in Component 1. Thus Component 1 was designated city affluence factor. Similarly, on the basis of the value of correlation coefficients of Components 2 and 3, an urban scale factor resulted from the combination of built-up area and nonagricultural population; a road supply factor was obtained from road density and road area per capita. Figure 2 plots the number of households that owned private cars per hundred households and the three urban form factors that resulted across the 36 cities. The plot was sorted according to the descending trend of the affluence factor. Of the two y-axes in the figure, the one on the left displays the number of private cars per hundred households, and the one on the right shows the relative value of the four 50.0 35 40.0 30 30.0 25 20.0 20 10.0 15 0.0 10 -10.0 5 -20.0 0 -30.0 zh B o C eiji u he ng Sh n en gdu zh e H Da n an li a Sh gzh n ou a n W g ul ha u Ku mu i Sh nm qi en in ya g n Ji g H nan a Ta iko iy u H ua ae n r Zh Na bin en nji gz ng La ho nz u N hou in gb L o Xi asa a W me n G uha ui n Q yan in g gd a N X o an ia ch n a C T ng Sh ha ianj n ijia g in zh sha u Fu ang z N ho Yi ann u C nch ing h u C ong an ha q ng ing ch u H n e H Xi fei uh ni eh ng ao te 40 G ua ng Number of Private Cars per 100 HHs Across 36 Cities Correlation between the collected nondummy explanatory variables was examined. Some variables with high coefficient values might correlate with each other, such as household disposal income and GDP per capita with a high value of correlation coefficient. A high degree of correlation may lead to a multicollinearity problem and cause incorrect magnitude or wrong signs or both in the regression coefficients. Lam and Tam omitted one or more highly correlated variables to avoid multicollinearity when they examined the correlation among the independent variables (8). In the present study, it was inappropriate to exclude the correlated variables because the entangling individual variables all possibly contributed to the car ownership decision. Factor analysis was employed to reduce the correlation among the explanatory variables. Principal components analysis (PCA), which is the most common form of factor analysis, was used to identify latent variables that contributed to the common variance of the set of the original explanatory variables. PCA generated uncorrelated factors through removal of the maximum variance of the linear combination of the original explanatory variables; the data dimensions were reduced. # Cars FIGURE 2 Affluence Scale Plot of private car level and resulting urban form factors. Road Supply Urban Form Factor Scores China, had the highest ratio of private cars—36, 34, 29, and 23 per hundred households, respectively. This ratio was under 10 in most other cities. What factors drove the spatial distribution of private cars so differently across cities? On the basis of the literature review and on perceptions, the following available explanatory variables that potentially affected city-level private car ownership were collected: 80 Transportation Research Record 2193 resulting urban form factors. It is difficult to observe the relationship between the city level of private cars per hundred households and the resultant urban form factors purely by looking at the figure, although all the plots appear to have the descending trend line. An OLS linear regression was employed to examine this. The dependent variable was the number of private car owners per hundred households. Treated as independent variables in the OLS regression were the following: the three resultant urban form factors of city affluence, urban scale, and road supply and the variables of the number of buses per 10,000 persons, the number of taxis per 10,000 persons; the population density calculated at city level, and five dummy variables: (a) supply of metro service, (b) supply of BRT service, (c) auto industry dummy, (d) license fee, and (e) coastal city. The standardized parameter estimates of the OLS regression are listed in Table 1. The overall model displays a goodness-of-fit measured by the adjusted R squared with a value of 0.549. The statistical tests demonstrated that the three resulting urban factors were significant. The influence of the relevant urban form factors can be compared by looking at the standardized estimates. The city scale factor had the strongest, significant, positive impact to boost car ownership across cities. It indicated that the larger scale cities had a higher level of private car ownership than did relatively small-scale cities. Also, city affluence and road supply factors had significant, positive effects on auto number. This means that car ownership will rise in richer cities with sufficient road supply. The estimates of the number of buses and of taxis per 10,000 capita were insignificant. This implies that the supply of bus and taxi service did not affect city-level car ownership. Five dummy variables were estimated in the OLS regression. The parameter of BRT supply was positive and significant. This indicates that households tended to own more cars in the cities with BRT service than in the cities without it. This estimate seems implausible at first glance. Intuitively, BRT service should reduce car ownership rather than increase it, for it provides a substitute for private cars. The estimate was interpretable once car ownership in cities with BRT was compared with cities without BRT. The average number of private cars per hundred households was 13.9 for cities with BRT, whereas the average number of private cars per hundred households for cities without BRT was only 3.1. This indicates a positive relationship between BRT and car ownership. The TABLE 1 OLS Linear Regression Results Explanatory Variable City affluence factor Urban scale factor Road supply factor BRT (dummy) License fee (dummy) Population density Number buses per 10,000 capita Number taxis per 10,000 capita Metro (dummy) Auto industry (dummy) Coast (dummy) Standardized Estimates T-Statistic 0.609 0.769 0.360 0.305 −0.343 −0.116 −0.032 0.021 −0.187 −0.204 −0.030 2.5* 3.1* 2.5* 1.8** −1.9** −0.5 −0.2 0.2 −0.8 −1.2 −0.2 NOTE: Dependent variable = number of private cars per 100 households. Adjusted R-squared = .549. *Significant at 5%. **Significant at 10%. presence of BRT may induce more travel demand and thus increase car ownership. The license plate fee parameter was the only variable that appeared to be negative and significant. This indicates that cities that carried out the license plate fee policy would reduce private car ownership. Beijing had lower factor scores for city affluence and road supply than did Shanghai, and these factors had a significant, positive effect on city-level private car ownership. The level of private car ownership in Beijing was much higher than it was in Shanghai. Similarly, Chengdu should have had fewer private cars than Shanghai, on the basis of the analysis of the estimates for city affluence and urban scale. Its private car ownership level was higher than Shanghai’s, however. Shanghai’s use of license plate fees mostly explained why it had much less private car ownership, even though it scored relatively high for affluence, scale, and road supply urban form factors. The average private car license plate fee in Shanghai was about 50,000 renminbi (about $7,000 U.S. dollars in 2007), whereas the plates were free in Beijing and Chengdu. The dummy estimates of metro service supply, auto industry, and coastal area were insignificant, which implies that these factors had no significant impact on city-level private car ownership. The above aggregate analysis explained the impact of resultant urban form factors and other urban form variables on the level of private car ownership across cities. The results indicate that larger and more affluent cities with more road supply tend to encourage city-level private car ownership and that car license plate fee policy appears to reduce car ownership. The aggregate analysis sheds some light on how urban form measured at the city level influences car ownership across cities, but it cannot reveal how the urban form affects individual household private car ownership behavior across cities. Further efforts are needed to understand the impact. The disaggregate analysis provided additional insights. Disaggregate Analysis A discrete choice model was employed to estimate the urban form impact on the decision whether or not to purchase a private car at the individual household level. The car ownership choice models were random utility models. Utility equations were defined for car-owning households and no-car households. Discrete choice analysis was employed to develop a disaggregate car ownership choice model to examine the responses to various factors on household car ownership (22). Since in the survey data most families owned only one car, the interest here was to examine how urban form, as well as other socioeconomic and demographic factors, affected the individual household’s car ownership decision. The binary logit model equation captured the household tradeoff between the choice of owning a private car or not owning a private car. The form of the binary logit choice model was as follows: Pi = 1 1 + e − ui i = 1, 2 where P1 = probability that a household owns one or more cars, P2 = probability that a household does not own any car, and ui = utility of mode i. In the disaggregate analysis, two household survey data sets were used, which had been collected in Beijing and Chengdu. The China Social Science Institute and Spatial Structures in the Social Sciences group at Brown University, Providence, Rhode Island, collected the Li, Walker, Srinivasan, and Anderson Beijing survey data in 2006. Households surveyed numbered 1,200 in eight urban districts. The China Project at Harvard University, Cambridge, Massachusetts, collected the Chengdu survey data in 2005 in collaboration with the Research Center for Contemporary China at Beijing University. Data were collected on the sociodemographic, travel behavior, and location characteristics of 1,001 households. The purpose of the estimation of two-car ownership models for Beijing and Chengdu was not to compare the car ownership behavior in the two cities but to understand how the urban form and other socioeconomic and demographic factors affected car ownership decisions across cities at the individual household level. Context of Beijing and Chengdu Among the 36 megacities surveyed, Beijing and Chengdu had the highest level of private car ownership. Both cities have thousands of years of history, and their road networks are circular systems. Each city is composed of a series of ring roads. Beijing’s five rings stretch out for about 16 km and Chengdu’s three rings extend out 7 km. Local governments in the two cities have not carried out regulation policy to actively suppress car ownership. The private car license plate bidding policy, which Shanghai adopted in 1994, has suppressed the growth of private cars in that city at a moderate rate. In 2006, Beijing had 34 and Chengdu had 29 private cars per hundred households, whereas Shanghai had only 11, although its household disposable income and per capita GDP were higher than those of the other two cities. Disaggregate Data and Variable Description 81 – OwnHouse. Dummy. It was one if the family owned their house; otherwise zero. – Child. Dummy. It was one if the family had children; otherwise zero. – OwnBike. Dummy. It was one if the family owned at least one bicycle; otherwise zero. – OwnEbike. Dummy. It was one if the family owned at least one electric bicycle; otherwise zero. – OwnMotor. Dummy. It was one if the family owned at least one motorcycle; otherwise zero. – CloseWorkUnit. Dummy. It was one if the family lived close to the husband or wife’s workplace; otherwise zero. – HHsize. Ordinal. Number of persons in household. 3. Household head characteristics – EduHi and EduMed. Categorical. Education level of household heads was divided into three groups—high, medium, and low. Low education was treated as the base. – Age. Ordinal. Age of the household head. – Married. Dummy. It was one if household head was in a marriage; otherwise zero. – StateOwnedJob. Dummy. It was one if household head worked in a state-owned company; otherwise zero. – ProfessionalJob and StaffJob. Categorical. Household head jobs were categorized into professional, staff, and other. Other jobs were treated as the base. – TransConvenience. It was one if the household head answered that the transport network was convenient to their neighborhood versus inconvenient; otherwise zero. – BusPass. Dummy. It was one if the household head had a monthly bus pass; otherwise zero. – TakeBus. Dummy. It was one if the household head took the bus regularly; otherwise zero. The explanatory variables that entered the discrete choice models were selected on the basis of the available collected survey information and review of previous empirical work on car ownership introduced in the literature review. Although the focus of the research was to investigate the role of urban form on household private car ownership, it was necessary to include other important socioeconomic and demographic variables, such as household income, household mobility information, and so forth. Three categories of explanatory variables were included in the discrete choice models: urban form measurements, household characteristics, and household head characteristics. The three urban form measurements were generated by integrating the available spatial information and by applying geographic information system tools. Descriptions of the explanatory variables were listed according to their name, type, and definition. Because the surveys were designed by different institutes, which collected different data items, an attempt was made to enter into the Beijing and Chengdu car ownership models only the socioeconomic and demographic variables in the survey data sets that might influence car ownership in one or the other city. The variables estimated in the two models therefore were not exactly the same. The parameters Child, StateOwnedJob, and CloseWorkUnit appeared only in the Beijing model, whereas ProfessionalJob, StaffJob, BusPass, and TakeBus appeared only in the Chengdu model. 1. Urban form measurements – PopDensity. Continuous. Population density was calculated at the subdistrict level within the city. The Beijing survey covered 48 subdistricts, whereas the Chengdu survey covered 32. – DisttoCBD. Continuous. Distance was computed from the household residential location to the CBD. – NearestBusStop. Continuous. The distance was computed from the residential location to the nearest bus stop in Chengdu. – LiveWithinR4. Dummy. It was one if the family lived within R4; otherwise zero. Residential location was within Beijing’s fourth ring road, which is the city’s urban fringe. 2. Household characteristics – IncomeHi and IncomeMed. Categorical. Household income was categorized into three levels—high, medium, and low. Low income low was treated as the base. Table 2 presents the binary logit model results obtained through the use of the Beijing and Chengdu survey data sets. The probability that a household owned an automobile was modeled. Therefore, a positive coefficient sign indicated that the explanatory variable had a positive effect on the household’s likelihood to own a car; a negative coefficient sign indicated that the explanatory variable had a negative effect on the household’s likelihood to own a car. Urban form measurements estimated in either the Beijing or Chengdu car ownership models included neighborhood population density calculated at the subdistrict level, distance from household residential location to the CBD, distance from household residential location to the nearest bus stop, and household spatial location relative to the major ring roads. Population and household density have been the most researched urban form characteristics. In the Beijing car ownership model, neighborhood population density was calculated for the Discrete Choice Model Estimation Results 82 Transportation Research Record 2193 TABLE 2 Discrete Choice Model Estimates Beijing Car Ownership Model Factor Parameter Urban form PopDensity DisttoCBD LiveWithinR4 IncomeHi IncomeMed HHsize OwnHouse OwnBike OwnMotor Child CloseWorkUnit Age Married EduHi EduMed TransConvenience StateOwnedJob Household characteristics Household head characteristics Constant Car HHs Non-car HHs Chengdu Car Ownership Model Estimate T-Stat. Parameter Estimate T-Stat. −0.286 −0.045 0.794 1.982 0.899 0.139 0.673 −1.199 −0.089 0.141 0.018 −0.010 0.091 1.889 1.387 −0.237 −0.415 −2.6* −2.5* 2.7* 4.8* 2.3* 1.3 2.5* −4.8* −0.2 0.7 0.1 −1.2 0.3 3.5* 2.7* −1.0 −2.1* PopDensity DisttoCBD NearestBusStop IncomeHi IncomeMed HHsize OwnHouse OwnBike OwnEbike −0.048 −0.248 −0.001 4.497 2.363 0.223 1.126 −0.652 −0.941 −2.7* −1.9** −1.0 4.3* 2.3* 2.3* 3.2* −2.4* −2.9* Age Married EduHi EduMed TransConvenience ProfessionalJob StaffJob BusPass TakeBus Car HHs Non-car HHs −0.011 0.436 −0.309 0.234 0.617 0.435 −1.422 −1.199 −1.162 Base 3.665 −0.9 1.2 −0.7 0.7 2.3* 1.2 −2.9* −1.4 −3.5* 0 2.8* Base 3.365 0 4.5* NOTE: Beijing car ownership model: number of parameters = 18, number of observations = 1,200, final log likelihood = −407.87, and adjusted rho-squared = 0.488. Chengdu car ownership model: number of parameters = 19, number of observations = 1,001, final log likelihood = −217.15, and adjusted rho-squared = 0.660. *Significant at 5%. **Significant at 10%. surveyed 48 subdistricts; in the Chengdu car ownership model, neighborhood population density was calculated for the surveyed 32 subdistricts. In both cities, household car ownership was lower in neighborhoods with higher population densities. This result is consistent with what has been found in the OLS model results, namely that city-level density has a significant negative effect on private car ownership, and it lends some support to Schimek’s finding that high population density measured by the U.S. standard can help reduce car dependency (12). The estimate of residential distance to the CBD had a significant negative effect on car ownership in both the Beijing and Chengdu models. This implies that households tend to have fewer private cars when they live further away from the urban center. This result is contrary to what has been found in the developed world. Prevedouros and Schofer demonstrated that residency relocation into outer-ring, low-density suburbs affected car ownership positively in a study done on the basis of a survey of Chicago, Illinois (19). Beijing and Chengdu have developed and expanded along circular ring road systems. The fourth ring road in Beijing is the border between well-developed urban areas and urban fringe areas. The model results indicated that car ownership tended to be higher in households within the fourth ring road in Beijing. The CBD distance and the fourth ring road location parameters are quite different from the situation in North America, where people tend to have more cars when they live further away from the local CBD, but this is not true in China currently. In North American cities, car ownership is essential for families that dwell outside of CBD areas because of the lack of frequent public transportation. One possible reason that Chinese households that own cars prefer to live in urban centers is because that is where good amenities are to be found. The urban form measurement of the distance to the nearest bus stop in residential locations was included in the Chengdu car ownership model. Its estimate was negative but insignificant. Both household characteristics and household head characteristics also played critical roles in the determination of car ownership, which the literature has demonstrated. In the Beijing and Chengdu car ownership models, the urban form effects on private car ownership were examined after controlling for household characteristics and household head characteristics. Household income exerted a strong influence on car ownership in both the Beijing and Chengdu models. High-income households showed a stronger preference to own private cars compared with medium- and low-income families. Household size affected car ownership in the Chengdu model but was not significant in the Beijing model. Families with housing tenure tended to be more likely to own private cars than families that rented. In both models, families tended to be less likely to own private cars when they owned bicycles. The parameter estimate of the electric bicycle had a positive effect on noncar households in Chengdu but was not significant in the Beijing model. Household head level of education affected car ownership significantly in the Beijing model but was not significant in the Chengdu model. Household heads with high education preferred to own private cars in Beijing. The parameter estimates of household head age and marital status were insignificant in both the Beijing and Chengdu models. Families tended to be less likely to own cars if the household heads worked in a state-owned company in the Beijing model. In the Chengdu model, car ownership increased when the household head had a professional job, whereas it decreased when the household head had a staff job. The TransConvenience parameter carried a positive sign and was significant in the Chengdu model, which means Li, Walker, Srinivasan, and Anderson the households tended to be more likely to own cars when their heads thought that the local road network was good. This finding supports the Downs-Thomson paradox that expansion of a road system as a remedy for congestion is not only ineffective but also counterproductive. Unsurprisingly, the parameters BusPass and TakeBus were significant in the Chengdu model. If household heads owned a monthly bus pass and took the bus regularly, their families tended not to own private cars. CONCLUSIONS AND FUTURE WORK Conclusions This paper presents the results of a pilot empirical study of private car ownership across megacities in China. The study used both aggregate and disaggregate analyses and examined the impact of urban form on car ownership, measured at both the city and neighborhood level. From the output of aggregate OLS regressions and disaggregate discrete choice models, the following results are summarized. First, the estimates demonstrate that population density calculated at the neighborhood level has a significant, negative effect on car ownership across cities. This result is consistent with that of previous studies [Cervero (11); Schimek (12); Ryan and Han (13); and Holtzclaw et al. (9)]. Second, affluence, urban scale, and road infrastructure supply factors have significant, positive impacts on car ownership across cities. These results indicate that richer and larger cities with adequate infrastructure are inclined to have more private cars, whereas compact cities with high population densities tend to suppress car ownership. These findings reinforce the results found in previous car ownership studies in the developed world [Hess and Ong (14) and Potoglou and Kanaroglou (16)]. The model estimates, however, imply that households that own cars in megacities prefer to live in the urban center rather than further away from CBD. This situation is contrary to what is currently observed in the developed world, where households closer to the center of cities tend to own fewer automobiles than those located further away [Schimek (12) and Bento et al. (18)]. This finding provides evidence to help understand the differences in car ownership behavior between the developed and developing worlds. Other socioeconomic and demographic factors that might affect household car ownership also were analyzed. Education level influenced car ownership decisions significantly in the Beijing model but not in the Chengdu model. Household size had a significant effect on the car ownership decision in the Chengdu model but not in the Beijing model. These estimate differences, it is argued, point to the fact that the household car ownership model specification is not entirely transferable between cities. The bicycle ownership estimate had a significant, negative impact on car ownership in both the Beijing and Chengdu models. This finding provides evidence for policy makers to reduce automobile dependency by promoting bicycle ownership. In the past decade, however, bicycle purchases dropped dramatically, whereas private automobile purchases increased greatly. 83 They should also be aware that households that own cars tend to live in urban centers where good amenities are provided. Caution should be exercised for several reasons when the research results are applied. Car ownership is not solely determined by urban form measurements and the other variables included in the research. Many other factors contribute to car ownership, such as transport policy, traffic regulation, parking space and cost, local history and culture, and investments in public transit. People’s attitudes toward cars also play a critical role in the determination of car ownership (23). Although density is important to lower car ownership, density alone is not sufficient to control private car growth (12). Schimek concluded that each 1% increase in population density is associated with a 0.11 reduction in the average number of vehicles per household. The effect of population density change on private car ownership was simulated in both the Chengdu and Beijing car ownership models. If population density increased by 10% in the Beijing and Chengdu car ownership models, car ownership would decrease 0.486% in Chengdu and 0.491% in Beijing on the basis of the estimated parameters in Table 2. These results imply that population density has a small effect on the suppression of car ownership in Chengdu and Beijing. Furthermore, the population density in the urban core of Chengdu was already up to 20,000 persons per square kilometer. An increase in population density should not be a strategy to reduce automobile ownership in already densely populated areas. In terms of policy planning, car ownership is in some ways now used as a placeholder for the more direct measure of private car impact–miles driven. Although a person may have a car because he or she can afford it, that person may still use transit to commute if adequate and convenient transit service is available. The private automobile may not be the suitable, dominant mode of urban transportation in China, as Liu and Guan have pointed out, given the country’s large population and limited per capita land use, and especially not for commuting purposes in large metropolitan areas (24). Urban form should be promoted to support nonmotorized mode and mass transit use. Future Work Travel behavior in China is undergoing rapid change because of tremendous economic growth and fast motorization. The private automobile is changing from a luxury to a necessary commodity. The urban form is also experiencing rapid change. Many megacities are expanding and transportation infrastructure is growing. Longitudinal studies should be developed to understand the car ownership behavior across time (25). With the growth of private car ownership, parking demand will increase greatly in the near future. As a result of the lack of data, the cost of parking is not now included in either aggregate or disaggregate analysis. The introduction of a parking cost variable into disaggregate analysis would enable examination of the individual response to parking cost change. In the future, it will be interesting to see how this cost influences car ownership. Discussion of Results ACKNOWLEDGMENTS The results of the research provide useful implications for China, where many megacities are transforming from compact areas with high densities to sprawling metropolises with lower densities in urban, peripheral areas. Policy makers should keep in mind that high density and compact urban form are conducive to constraints on private car ownership, if they want to plan for less car-dependent cities. Beijing University conducted the Chengdu survey in collaboration with the China Project of the School for Engineering and Applied Sciences, Harvard University, and Harvard University Center for the Environment. The authors thank the V. Kann Rasmussen Foundation, Harvard University Asia Center, and Volvo Educational and Research Foundation for their support. They also thank the Spatial 84 Structures in the Social Sciences initiative at Brown University for the use of its Beijing survey data. REFERENCES 1. Cherry, C. China’s Urban Transportation System: Issues and Policies Facing Cities. Working paper. Center for Future Urban Transport: A Volvo Center of Excellence, University of California, Berkeley, 2005. 2. De Jong, G. C., J. Fox, M. Pieters, L. Vonk, and A. Daly. Comparison of Car Ownership Models. Transport Reviews, Vol., 24, No. 4, 2004, pp. 379–408. 3. Khan, M. A., and L. G. Willumsen. Modeling Car Ownership and Use in Developing Countries. Traffic Engineering and Control, Vol. 27, 1986, pp. 554–560. 4. De Jong, G. C. An Indirect Utility Model of Car Ownership and Private Car Use. European Economic Review, Vol. 34, No. 5, 1990, pp. 971–985. 5. Ingram, G. K., and Z. Liu. Motorization and Road Provision in Countries and Cities. Policy Research Working Paper 1842. World Bank, Washington D.C., 1997. 6. Dargay, J., and D. Gately. Income’s Effect on Car and Vehicle Ownership, Worldwide: 1960–2015. Transportation Research Part A, Vol. 33, 1999, pp. 101–138. 7. Romilly, P., H. Song, and X. Liu. Car Ownership and Use in Britain: A Comparison of the Empirical Results of Alternative Cointegration Estimation Methods and Forecasts. Applied Economics, Vol. 33, 2001, pp. 1803–1818. 8. Lam, W., and M.-L. Tam. Reliability of Territory-wide Car Ownership Estimates in Hong Kong. Journal of Transport Geography, Vol. 10, No. 1, 2002, pp. 51–61. 9. Holtzclaw, J., R. Clear, H. Dittmar, D. Goldstein, and P. Haas. Location Efficiency: Neighborhood and Socioeconomic Characteristics Determine Auto Ownership and Use—Studies in Chicago, Los Angeles, and San Francisco, Transportation Planning and Technology, Vol. 25, No. 1, 2002, pp. 1–27. 10. Clark, S. D. Estimating Local Car Ownership Models. Journal of Transport Geography, Vol. 15, 2007, pp. 184–197. 11. Cervero, R. Mixed Land-uses and Commuting: Evidence from the American Housing Survey. Transportation Research Part A: Policy and Practice, Vol. 30, No. 5, 1996, pp. 361–377. 12. Schimek, P. Household Motor Vehicle Ownership and Use: How Much Does Residential Density Matter? In Transportation Research Record 1552, TRB, National Research Council, Washington, D.C., 1996, pp. 120–125. Transportation Research Record 2193 13. Ryan, J. M., and G. Han. Vehicle-Ownership Model Using Family Structure and Accessibility Application to Honolulu, Hawaii. In Transportation Research Record: Journal of the Transportation Research Board, No. 1676, TRB, National Research Council, Washington, D.C., 1999, pp. 1–10. 14. Hess, D. B., and P. M. Ong. Traditional Neighborhoods and Automobile Ownership. In Transportation Research Record: Journal of the Transportation Research Board, No. 1805, Transportation Research Board of the National Academies, Washington, D.C., 2002, pp. 35–44. 15. Chu, Y.-L. Automobile Ownership Analysis Using Ordered Probit Models. In Transportation Research Record: Journal of the Transportation Research Board, No. 1805, Transportation Research Board of the National Academies, Washington, D.C., 2002, pp. 60–67. 16. Potoglou, D., and P. S. Kanaroglou. Modelling Car Ownership in Urban Areas: A Case Study of Hamilton, Canada. Working Paper. Center for Spatial Analysis, McMaster University, Hamilton, Ontario, Canada, 2006. 17. Zegras, P. C. Sustainable Suburban Mobility: Exploring the Role of the Built Environment. PhD dissertation. Massachusetts Institute of Technology, Cambridge, 2006. 18. Bento, A., M. Cropper, A. Moborak, and K. Vinha. The Impact of Urban Spatial Structure on Travel Demand in the United States. Review of Economics and Statistics, Vol. 87, No. 3, 2005, pp. 466–478. 19. Prevedouros, P. D., and J. L. Schofer. Factors Affecting Automobile Ownership and Use. In Transportation Research Record 1364, TRB, National Research Council, Washington, D.C., 1992, pp. 152–160. 20. Kim, H. S., and E. Kim. Effects of Public Transit on Automobile Ownership and Use in Households of the USA. Review of Urban & Regional Development Studies, Vol. 16, No. 3, 2004, pp. 245–262. 21. Kitamura, R., T. Akiyama, T. Yamamot, and T. F. Golob. Accessibility in a Metropolis: Toward a Better Understanding of Land Use and Travel. In Transportation Research Record: Journal of the Transportation Research Board, No. 1780, TRB, National Research Council, Washington, D.C., 2001, pp. 64–75. 22. Ben-Akiva, M., and S. R. Lerman. Discrete Choice Analysis: Theory and Application to Travel Demand. MIT Press, Cambridge, Mass., 1985. 23. Cullinane, S. Hong Kong’s Low Car Dependence: Lessons and Prospects. Journal of Transport Geography, Vol., 11, No. 1, 2003, pp. 25–35. 24. Liu, R., and C. Guan. Mode Biases of Urban Transportation Polices in China and Their Implications. Journal of Urban Planning and Development, Vol. 131, No. 2, 2005, pp. 58–70. 25. Kitamura, R., and D. S. Bunch. Heterogeneity and State Dependence in Household Car Ownership: A Panel Analysis Using Ordered-response Probit Models with Error Components. Research Report. Institute of Transportation Studies, University of California, Davis, 1989. The Transportation in the Developing Countries Committee peer-reviewed this paper.