does subway proximity discourage automobility? …...1 does subway proximity discourage...

23
1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhang a , Siqi Zheng b, c , Cong Sun d , Rui Wang e a. School of Economics and Management, Beijing Forestry University, Beijing, 100083, China b. Hang Lung Center for Real Estate, Tsinghua University, Beijing, 100084, China c. Department of Urban Studies and Planning, Center for Real Estate, and the STL Real Estate Entrepreneurship Lab, Massachusetts Institute of Technology, Cambridge, MA 02139, United States d. School of Urban and Regional Science, Shanghai University of Finance and Economics, Shanghai, 200433, China e. Johns Hopkins University, School of Advanced International Studies & UCLA Luskin School of Public Affairs, 1619 Massachusetts Avenue, NW, Washington, DC 20036, United States Abstract Many cities around the world are investing in rail transit, but whether it can effectively reduce road congestion and air pollution from automobiles remains an open question. A major challenge to empirically answering this question is the fact that the choices of residential location and travel mode are jointly made by households. The unique context of urban housing in Beijing provides us a natural experiment to separate residential location and travel choices of households living in the resettlement and reformed housing units. We take advantage of the largely exogenous residential locations of those living in the resettlement and reformed housing in Beijing and use the Heckman two-step method to correct a potential bias in estimating vehicle fuel consumption. To identify the heterogeneous effects of different subway stations, we use the travel time to city center by subway to proxy a subway station’s value to users. We find robust evidence supporting that subway proximity reduces a household’s probability of owning a car and subsequent fuel consumption. More valuable subway stations discourage nearby households’ car ownership rate by a greater extent. Evidence does suggest the existence of residential self-selection. Keywords Subway proximity; car ownership; fuel consumption; resettlement housing; reformed housing; Beijing.

Upload: others

Post on 22-May-2020

18 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

1

Does Subway Proximity Discourage Automobility? Evidence from Beijing

Yingjie Zhanga, Siqi Zhengb, c, Cong Sund, Rui Wange

a. School of Economics and Management, Beijing Forestry University, Beijing, 100083, China

b. Hang Lung Center for Real Estate, Tsinghua University, Beijing, 100084, China

c. Department of Urban Studies and Planning, Center for Real Estate, and the STL Real Estate

Entrepreneurship Lab, Massachusetts Institute of Technology, Cambridge, MA 02139, United

States

d. School of Urban and Regional Science, Shanghai University of Finance and Economics,

Shanghai, 200433, China

e. Johns Hopkins University, School of Advanced International Studies & UCLA Luskin School

of Public Affairs, 1619 Massachusetts Avenue, NW, Washington, DC 20036, United States

Abstract

Many cities around the world are investing in rail transit, but whether it can effectively reduce road

congestion and air pollution from automobiles remains an open question. A major challenge to

empirically answering this question is the fact that the choices of residential location and travel

mode are jointly made by households. The unique context of urban housing in Beijing provides us

a natural experiment to separate residential location and travel choices of households living in the

resettlement and reformed housing units. We take advantage of the largely exogenous residential

locations of those living in the resettlement and reformed housing in Beijing and use the Heckman

two-step method to correct a potential bias in estimating vehicle fuel consumption. To identify the

heterogeneous effects of different subway stations, we use the travel time to city center by subway

to proxy a subway station’s value to users. We find robust evidence supporting that subway

proximity reduces a household’s probability of owning a car and subsequent fuel consumption.

More valuable subway stations discourage nearby households’ car ownership rate by a greater

extent. Evidence does suggest the existence of residential self-selection.

Keywords

Subway proximity; car ownership; fuel consumption; resettlement housing; reformed housing;

Beijing.

Page 2: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

2

1. Introduction

Many cities around the world are investing in urban rail, but whether it can effectively reduce road

congestion and mobile air pollution remains an open question. A major challenge to empirically

answering this question is the fact that the choices of residential location and travel mode are often

jointly made by households, known as the self-selection of residential location by travelers –

household residential location choice is affected by travel needs and preferences (see, e.g., Guo

and Chen, 2007; Mokhtarian and Cao, 2008; Brownson et al., 2009; TRB, 2009; Ewing and Cevero,

2010). Without exogenous variations in the residential locations of households, one could not

cleanly identify the effects of rail transit on nearby households’ car ownership and travel behavior.

Taking advantage of the unique urban housing policies in China and the rapid expansion of urban

rail transit in Beijing, this study uses an empirical strategy (i.e., identifying a subsample of urban

households with exogenous residential locations) different from earlier studies to provide a robust

estimation of urban rail transit’s effects on automobility.

Decades of rapid economic growth and urbanization have dramatically changed China’s urban

transportation, making urban residents travel longer distances and more frequently and rely more

on modes using fossil fuels (Wang, 2010). Rapid motorization has led to a series of problems

including serious road congestion, severe air pollution, and rapidly rising demand for oil and

emissions of greenhouse gases. Beijing, China’s capital and one of its most motorized cities,

experienced an average annual increase rate of 11.8% in the number of motor vehicles from 2000

to 2010.1 Despite the government’s new policy of setting an annual quota on new car license plates

since 2011, the total number of motor vehicles in Beijing exceeded five million by 2012.2 To

combat road congestion and air pollution brought by rapid motorization, Beijing has been heavily

investing in public transit systems. By 2020, Beijing will have 30 urban rail lines in operation,

with 1,050 km in total route length and 450 subway stations.3

It is not a surprise that the new urban rail lines will be filled with passengers, especially as the

majority of Beijing’s residents do not own cars currently. However, it is unclear how the

development of urban rail will affect automobility (i.e., car ownership and usage) of residents. Will

the car owners reduce driving? Will additional rail service slow down the rise in car ownership?

The reason these questions are difficult to answer is that, on one hand, rail transit provides a

competing alternative to driving, but on the other hand, the fact that rail transit may reduce surface

road congestion (e.g., fewer buses are needed on the same route) can induce more driving from

those who can afford to drive. Thus we need empirically test urban rail transit development’s effect

on automobility. But this is not a straightforward task. One may observe that residents who live

near a subway station have a lower car-ownership rate, but we can’t confirm whether it is due to

that those who prefer subway to driving self-select to live nearby a subway station, or that

1 Data obtained from the Beijing Traffic Management Bureau. See http://www.bjjtgl.gov.cn/jgj/ywsj/index.html. 2 Ibid. 3 Data obtained from the Beijing Municipal Government. See http://zhengwu.beijing.gov.cn/gzdt/zyhy/t1114930.htm.

Page 3: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

3

improved access to subway does change residents’ car ownership and use behavior.

This study uses a 2009 household survey in Beijing to examine the impact of subway station

proximity on urban residents’ car ownership and fuel consumption. To address the potential bias

due to residence self-selection, we take advantage of the unique urban housing situation in China

as an opportunity of natural experiment and focus on households living in the resettlement

(chaiqian) housing and the reformed (fanggai) housing with pre-determined locations for causal

inference. We also explore the heterogeneous effects of subway stations due to their different travel

times to the city center via the subway network. Moreover, we employ the Heckman two-step

model to test and correct the potential sample selection bias when estimating rail transit’s effect

on fuel consumption using data from the car owners. Our findings show that subway proximity

does reduce an urban household’s probability of owning a car as well as the mileage driven, even

after controlling for the residential self-selection bias. The effect of subway on car ownership is

stronger where subway provides a shorter time of travel to the city center. Overall, the development

of urban rail in Beijing likely reduces overall car use as some would-be car owners choose not to

own a car and car owners drive less, producing positive traffic and environmental impacts.

Section 2 briefly reviews the literature. Section 3 describes the background of resettlement and

reformed housing in Beijing and our household survey data, as well as the measurement of the

heterogeneous locations of subway stations. Sections 4 discusses the analytical method and

hypotheses. Sections 5 and 6 present empirical findings and related robustness check results,

followed by conclusions in Section 7.

2. Literature

Multiple factors, especially income, fuel price, and road infrastructure, influence private passenger

motor vehicle ownership and travel behavior in cities. Income has been considered as a major

determinant of motorization. Many studies, mostly from the industrialized world, have estimated

the income elasticities of motor vehicle ownership and use. They indicate that motorization

increases rapidly with income, although the elasticities vary. Ingram and Liu (1999) summarize

studies since the mid-1960s and find that long-run income elasticities of car ownership (typically

based on cross-sectional data, e.g., Silberston, 1970; Wheaton, 1982; Kain, 1983) are greater than

1.0, while short-run elasticities (typically based on time series or panel data, e.g., Pindyck, 1979;

Button et al., 1993; Johansson and Schipper, 1997) are less than 1.0; income elasticities from

urban-level data (e.g., Beesley and Kain, 1964; Chin and Smith, 1997) are similar to or smaller

than those from country-level data largely due to the existence of competing modes of

transportation; income elasticities of motor vehicle use (e.g., Pindyck, 1979; Wheaton, 1982;

Mannering and Winston, 1985; Train, 1986; Hensher et al., 1990; Button et al., 1993; Johansson

and Schipper, 1997) are less than unity, indicating that motor vehicle use increases less rapidly

than ownership.

Page 4: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

4

Many have also studied the effects of fuel price on motorization (e.g., Pindyck, 1979; Wheaton,

1982; Train, 1986; Hensher et al., 1990; Johansson and Schipper, 1997). Compared to the

somewhat weak evidence on fuel price’s effect on vehicle ownership, studies generally confirm

that increase in fuel price negatively affect vehicle usage and positively affect the average fuel

efficiency of the vehicle stock, although evidence suggests that income elasticities are greater than

price elasticities in magnitude for both motor vehicle ownership and use (Ingram and Liu, 1999).

Road infrastructure at the national and city levels, usually provided publicly, is widely recognized

as closely related to motorization. However, due to the endogenous relationship between

infrastructure investment decisions made by governments and the growth in regional travel

demand, there has been limited robust evidence on how motorization is influenced by road

provision. Using a source of more plausible exogenous variations (the 1947 interstate highway

plan, 1898 rail routes and the major exploration routes during 1835-1850), Duranton and Turner

(2011) analyze the impact of interstate highway provision on city-level traffic in the continental

US between 1983 and 2003. They suggest the elasticity of metropolitan area interstate highway

vehicle-kilometer travelled with respect to lane kilometers to be 1.03 – a near proportional increase

in metropolitan traffic to the extension in interstate highways.

Some scholars, especially spatial planners, highlight land use and built environment’s effects on

motorization, primarily due to the interest in better planning built environment to reduce car

dependence, traffic congestion, and related environmental and health impacts (e.g., climate change,

energy shortage, air pollution, and the lack of physical activity). There have been several reviews

of this literature, such as Crane (2000), Ewing and Cervero (2001), Stead and Marshall (2001),

Handy (2005), Guo and Chen (2007), Mokhtarian and Cao (2008), Ewing and Cervero (2010), and

Cao (2015). Most studies have shown that features of the built environment, such as the “three Ds,”

namely “density,” “diversity” or land use mix, and “design” features related to pedestrian

friendliness of streets and street networks (Cervero and Kolkelman, 1997), are often associated

with travel behaviors including trip frequency, trip distance, mode choice, etc.

However, more and better empirical evidence is needed in order to advance our understanding of

the effects of land use on travel behavior (and/or associated health outcomes) for at least two

reasons. First, while a good number of studies have been conducted on urban land use, passenger

travel and health/environmental effects, the vast majority of existing evidence is based on cross-

sectional data and only confirms the correlations between land use patterns and travel, leaving

causality unexplained or falsely claimed, as in most studies reviewed in the meta-analyses by

Brownson et al. (2009) and Ewing and Cevero (2010). A small number of studies utilize a range

of often sophisticated statistical strategies to address the bias caused by the self-selection of

residential location, including (quasi-)longitudinal design (e.g., Boarnet et al., 2005; Handy et al.,

2005), structural equations (e.g., Cao et al., 2007), joint choice modeling (e.g., Bhat and Guo,

2007), propensity score matching (e.g., Boer et al., 2007; Cao and Fan, 2012), the explicit control

of residential/travel attitudes (e.g., Schwanen and Mokhtarian, 2005; De Vos and Witlox, 2016),

Page 5: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

5

among others. However, most of the results of the studies attempting to correct the self-selection

bias remain suggestive (Guo and Chen, 2007; Mokhtarian and Cao, 2008) and do not seem to be

very consistent with each other (TRB, 2009; Guo, 2009), although they generally confirm that the

built environment affects travel behavior even after controlling for residential self-selection.

Second, almost all major empirical studies are from industrialized countries, where travel behavior,

health background and the speed of land use change are completely different from those of

developing countries, where air pollution and carbon emissions grow as rapidly as urbanization

and motorization. Data and analyses are very much needed to enrich our knowledge in the

developing country setting for at least two reasons. On one hand, walking, cycling and transit use

are often much more important in the developing countries compared to the highly motorized

countries. On the other hand, significant and rapid socio-economic changes of developing-world

cities provide researchers with significant local built environment variations. Existing studies in

the developing world confirm the associations between land use density and travel mode choice

(Zhang, 2004), between road facility design (street density, connectivity, and location) and physical

activity (Cervero et al., 2009), between built environment characteristics and car ownership and

vehicle-kilometers traveled (Zegras, 2010), and between transit access and car ownership (Huang

et al., 2016). Unfortunately, these studies do not prove causality between the built environment

and travel behavior due to the potential self-selection bias in residential location.

Overall, while the effects of transit access on motorization (e.g., household car ownership and use)

may be studied from either the infrastructure provision perspective or the built environment

perspective, careful causal identification of such effects remain rare in the literature.

3. Research Context and Data

3.1 Housing in Beijing

The plan-to-market transition and the rapid growth of the Chinese economy have left cities with a

complicated housing stock. Most of the owner-occupied housing units belong to one of three types:

commodity housing traded in the free market, reformed (fanggai) housing resulted from the

privatization of work-unit housing under the previous socialist welfare housing regime, and

resettlement (chaiqian) housing typically due to urban renewal and redevelopment. See Chen and

Han (2014) for a detailed survey of the urban housing market in post-reform China.

Under the socialist planning system, urban land was allocated to work units. A work unit typically

used part of its land to construct housing units and allocated them to its employees for a low rent

based on a worker’s office rank, occupational status, working experience, etc. (Fu et al., 2000). As

a result, most of urban workers did not choose their residential locations. Launched in the early

1990s, the housing reform established a housing market on one hand and privatized state-owned

housing units to sitting tenants at very low subsidized prices on the other hand. Such privatized

work-unit housing is the so called reformed (fanggai) housing, resale or lease of which to other

people outside the original work units have in general been prohibited.4 In other words, unlike the

4 The circulation ban of reformed housing has not been relaxed until recently, especially for the units owned by governmental

Page 6: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

6

commodity housing residents, households living in reformed housing do not have a free choice of

where to live.

Since the establishment of urban land and property markets, many Chinese cities have experienced

urban renewal and redevelopment in their old built-up area to maximize land value and utilization

rate. This process of demolition of properties and relocation of property users is called chaiqian,

which triggers the relocation of existing land users including enterprises and households. In many

cases the government will provide resettlement housing to the relocated households to enable the

process of urban redevelopment or renewal. While the size and condition of the resettlement

housing units are often better than the demolished properties, the location of the resettlement

properties is usually determined through a complicated and ad hoc planning process. Sometimes

resettlement housing is close to the original sites, while sometimes far away. In many cases,

resettlement housing units are mixed with free-market housing development because the

government requires or subsidizes commodity housing development to set aside some units

dedicated to resettlement households. Obviously, residents of the resettlement housing do not have

much control on their residential locations, i.e., they are unable to freely choose where to live.

In summary, due to the special institutional arrangement in Chinese cities’ housing stock, only

residents living in commodity housing had a free choice of residential location through housing

market, while those living in the reformed or resettlement housing did not. This unique context in

urban housing provides us a natural experiment in residential location to test the self-selection in

residential location and to estimate unbiased effects of neighborhood characteristics (e.g., subway

proximity) on household travel behavior.

3.2 Data

Beijing is a largely monocentric city (Zheng and Kahn, 2008), with a contiguous urban core

(central city) dominating most of the key public and private sector activities.5 Tian’anmen square

with its east and west extensions to the Second Ring Road are conventionally regarded as the city

center. Five ring roads (the Second to the Sixth Ring Roads) have been built from the center

outward, as shown in Figure 1. Beijing’ urbanized area is primarily within the Sixth Ring Road

(the largest ring in Figure 1).

*** Insert Figure 1 about here ***

This research uses data from the “Housing, Transportation and Energy Consumption Survey of

Beijing Households” conducted by the Institute of Real Estate Studies at Tsinghua University in

September 2009. This dataset covers all key demographic, income, residential location and travel

information of 826 urban households (in 38 residential complexes) and their household heads.6

employees, which account for more than 30% of the total stock of reformed housing in Beijing. According to the National Bureau

of Statistics of China, reformed housing remains the largest category in owner-occupied housing, indicating a slow conversion from

reformed housing to commodity housing. The cumulative transaction volume of reformed housing is less than 1.5% of its total

stock by the end of 2002 (http://bj.house.sina.com.cn/n/s/2003-09-27/29188.html). 5 Thanks to an anonymous referee’s comments that multiple sub-centers exist within the urban core, we also use multiple job

sub-centers in a robustness check in Section 5.2. 6 Quota sampling method is employed in this survey. The feature matrix of this quota sampling scheme is based on both housing

property type and zone-level population size. There are three types of housing in Beijing, namely reformed housing, commodity

housing and affordable housing, while the resettlement housing units are scattered among them. The respective shares of these

Page 7: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

7

Fig.1 marks the spatial distribution of the 38 complexes and the subway network in Beijing in

2009.

Overall, our sample contains 481 households living in commodity housing units, 204 households

living in reformed housing, and 107 households in resettlement housing, accounting for 58%, 25%

and 13% of the sample,7 respectively. As discussed previously, the resettlement housing units in

our sample are mixed with commodity housing units. Therefore, there are altogether three types

of housing units in two types neighborhoods (complexes) in our sample: commodity housing

complexes (with some resettlement housing units in them) and reformed housing complexes. Our

statistics show that, the average housing age (as of 2009), green space coverage ratio and floor-

area ratio of commodity (reformed) housing complexes are 9 (19) years, 33% (24%) and 3.03

(2.00), respectively. That is, commodity housing complexes are generally newer, taller and have

more green landscape than reformed housing complexes.

The car ownership rate in our sample is 47% by household, with resettlement households having

the lowest car ownership rate (30%), compared to those of households living in reformed housing

(50%) and in commodity housing (52%). 32% of the correspondents drive to work, with an average

one-way commuting time of 33 minutes. 12.6% of the sample residents commute by foot, with an

average one-way commuting time of 11 minutes. 14% commute by bicycle, on average spending

20 minutes one way. 20.6% of the correspondents commute by bus and their average one-way

commuting time is 49 minutes. Only 11.8% take subway to their jobs, and they have the longest

one-way door-to-door commuting time of 57 minutes.

We geo-code the household data and subway lines and stops using the map of Beijing. The straight-

line distance from a household’s residence to its closest subway stop is defined as this household’s

distance to subway. Overall, households living closer to a subway station are less likely to own

private cars. Car ownership rates among households living between 1-1.5 km, 0.5-1 km, and within

0.5 km from the nearest subway station are 56%, 46%, and 38%, respectively.

To explore the heterogeneity in the potential effects of different subway stations, we differentiate

subway stations by measuring the travel time to city center via the subway network. We obtain

typical travel time between each subway station and Tian’anmen Square during morning peak

hours (SUBTIME) using historical travel time data from GAODE Map, China’s version of Google

Map. We then use this SUBTIME variable to proxy the attractiveness of different subway stations.8

All else being equal (e.g., straight-line distance to city center), a larger value of SUBTIME of a

subway station indicates that nearby residents overall benefit less from using subway.

4. Method

three types of housing in Beijing’s then housing stock are 4:5:1. This survey divides Beijing’s urban area into three circles from

inside out (the first circle includes Xuanwu District, Dongcheng District, Xicheng District, and Chongwen District; the second circle

consists of Chaoyang District, Haidian District, Fengtai District, and Shijingshan District; districts in the third circle are Tongzhou

District, Daxing District, Shunyi District, and Changping District). The proportion of surveyed communities in each circle is based

on the total number of residents in the circle. Details of the sampling method are discussed in Zheng and Huo (2010). 7 The remaining 4% respondents reported their housing type as “other” in the survey. 8 Following an anonymous referee comment, we will also check the robustness of results using an alternative SUBTIME measure

defined as the subway travel time between a station and its nearest job sub-center (one of Beijing’s three sub-centers Jinrongjie,

Zhongguancun, and Yayuncun).

Page 8: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

8

Table 1 presents the variable definitions and summary statistics. CAROWN indicates whether a

household owns a private car. FUEL measures a household’s car usage by the monthly fuel

expenditure reported by car owners.

*** Insert Table 1 about here ***

To show how the car ownership (CAROWN) and use (FUEL) of households relate to their subway

proximity (NEAR_SUB, a dummy variable indicating whether a household is less than 1,000 m

from the nearest subway station), we start with a probit regression of car ownership (Equation 1)

and an ordinary linear squares (OLS) regression of car-owning households’ fuel consumption

(Equation 2).

CAROWN = probit (Household attributes, NEAR_SUB, D_CBD) (1)

FUEL = OLS (Household attributes, REIM, NEAR_SUB, D_CBD) (2)

The coefficients of interest here are those of subway proximity (NEAR_SUB), controlling for

household attributes (INCOME, HHSIZE, and AGE of the head of household), distance to city

center (D_CBD), and whether fuel consumption is reimbursable from employer (REIM). Given

that driving and riding subway are substitutes (even if citywide drivers may benefit from the

expansion of subway service), we expect negative correlations between CAROWN and

NEAR_SUB in Equation (1) and between FUEL and NEAR_SUB in Equation (2) in subway

neighborhoods.

However, a sample self-selection problem arises in the regressions of both Equations (1) and (2)

if those who prefer subway to driving choose to live near a subway station and not own a car or

rarely use it, as previously discussed. There are at least two possible explanations for the expected

relationship – residents who live near subway stops are more likely to travel by subway instead of

car. That is, subway proximity can encourage residents to switch to rail transit; or alternatively,

residents who prefer subway travel will choose to live close to subway stops. This residential self-

selection happens when people have the freedom to choose not just travel mode, but also residential

location. Our solution to this problem is focusing on a group of households with exogenous

residential locations, i.e., those living in resettlement and reformed housing units. A comparison

of results using the full sample, the subsample with exogenous residential locations, and the

commodity housing subsample with endogenous residential locations, will provide us useful

information about the existence of residential self-selection and more robustly examine the causal

effects of subway proximity on car ownership and usage. If the estimated coefficient of

NEAR_SUB in the resettlement and reformed subsample regressions remains negative, we will be

more confident that subway proximity does reduce automobility.

Causal inference can be further strengthened by exploiting the heterogeneous effects of different

Page 9: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

9

subway stations on car ownership and use of urban households. The below probit regression

(Equation 3) and OLS regression (Equation 4) add an additional interaction term between

SUBTIME and NEAR_SUB to Equations (1) and (2) to explore such heterogeneous effects. When

a subway station is a shorter ride from city center (i.e., smaller SUBTIME), it is more appealing to

potential subway riders and more effective in discouraging nearby households’ car ownership and

use. We expect a positive sign of NEAR_SUB* SUBTIME in both Equations (2) and (4).

CAROWN = probit (Household attributes, NEAR_SUB, NEAR_SUB* SUBTIME, D_CBD) (3)

FUEL = OLS (Household attributes, REIM, NEAR_SUB, NEAR_SUB*SUBTIME, D_CBD) (4)

A separate challenge to the fuel consumption regressions (Equations 2 and 4) is a different sample

selection problem also due to the spatial sorting of households. If proximity to subway reduces the

need of nearby residents to drive, one observes the fuel consumption of residents nearby subway

only when they need to drive long distances (and consume a lot of fuel). The result is that a simple

OLS regression of fuel consumption on subway proximity will lead to downward-biased estimates,

because the sample (car owners) is unrepresentative of the population one is interested in (all

households in Beijing). Heckman (1976, 1979) has proposed a simple solution for such situations

by treating such sample selection as an omitted variable problem. This easy-to-implement method

is known as the two-step model, which involves the estimation of a probit model of the selection

process (e.g., Equations 1 and 3 in this study), followed by the insertion of a correction factor—

the inverse Mills ratio calculated from the probit model—into the second OLS model of interest.

The significance of the inverse Mills ratio (lambda) informs us about the severity of sample

selection bias. For all fuel consumption regressions, we check the Heckman second step results to

determine the most unbiased estimate as our primary findings.

5. Results

Table 2 reports the regression results of car ownership (Equation 1) and fuel consumption

(Equation 2) using firstly the full sample (columns 1-2), then the resettlement and reformed

housing subsample (columns 4-5), and lastly the commodity housing subsample (columns 7-8).

The Heckman second step regressions add the inverse Mills ratios from the estimation of Equation

(1) to the regressors of Equation (2), shown in columns 3, 6 and 9 for the three samples,

respectively.

*** Insert Table 2 about here ***

For the car ownership probit regressions (columns 1, 4 and 7), the estimated coefficients of

NEAR_SUB in all three samples are negative as expected. The estimated coefficient of NEAR_SUB

in the full sample is -0.210 (statistically significant at the 5% level), which means that all else

being equal, the closer a household lives to the nearest subway station, the less likely the household

Page 10: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

10

owns a car. As for the commodity housing subsample, where households’ residential locations are

most likely endogenous, the estimated coefficient of NEAR_SUB is -0.247, also significant at the

5% level. In the resettlement and reformed housing subsample, whose residential location are

likely exogenous, the coefficient of NEAR_SUB is -0.233 (statistically significant at the 10% level).

Compared to the commodity housing and the full samples, the resettlement and reformed housing

subsample has a smaller size and a larger p-value for the estimated NEAR_SUB coefficient. The

estimated coefficients of the control variables generally behave as expected. Household income,

size, and distance from CBD increase the likelihood of car ownership. The effect of the age of

household head is nonlinear – positive up to a certain age before turning into negative, although

the AGE coefficients are statistically insignificant in the smaller resettlement and reformed housing

subsample.

Due to sample differences, a direct comparison of the estimated coefficients in columns 1, 4 and 7

would be misleading. We calculate the marginal effect of NEAR_SUB in each regression. The

estimated marginal effect of NEAR_SUB on commodity housing households’ probability of

owning a private car is -7.7%, while for households living in the resettlement and reformed housing

the marginal effect is -6.6%. These estimates confirm that in a household sample with more

exogenous residential locations (the resettlement and reformed housing subsample), being close

to a subway station reduces household car ownership about 14.3% less than in a sample with

endogenous residential locations (the commodity housing subsample). Nevertheless, subway

proximity does decrease the likelihood of household car ownership even after accounting for the

potential self-selection in residential location.

In the fuel consumption regressions (columns 2, 3, 5, 6, 8 and 9), the lambda statistics suggest that

only the subsample of commodity housing suffers from a significant sample selection bias. This is

intuitive as we expect the commodity housing subsample to have the strongest residential sorting

based on travel demand. So we focus on findings in columns 2, 5 and 9 as our primary findings.

The estimated coefficients of NEAR_SUB are -0.293, -0.160 and -0.373 for the full, the

resettlement and reformed, and the commodity housing samples, respectively. All three estimates

are statistically significant, suggesting that being close to subway reduces household fuel

consumption even after controlling for residential self-selection. For the control variables,

INCOME, D_CBD and REIM all have positive and significant coefficients in the commodity

housing subsample (column 9) but not in the resettlement and reformed housing subsample

(column 5), suggesting important differences in travel behavior related to households’ ability to

make residential location choices.

*** Insert Table 3 about here ***

Table 3 extends Table 2 with heterogeneous subway stations. Results using the full sample, the

resettlement and reformed housing subsample, and the commodity housing subsample are

presented in columns 1-3, 3-6, and 7-9, respectively. Consistent with the results in Table 2, only

Page 11: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

11

the commodity housing subsample has a significant Heckman sample selection issue when

estimating fuel consumption conditioning on household car ownership, so we focus on the findings

in columns 1, 2, 3, 4, 7 and 9. The estimated coefficients of NEAR_SUB in all CAROWN and FUEL

regressions remain negative and statistically significant as expected. The cross-station

heterogeneity in the effects of NEAR_SUB, shown by the expected positive coefficients of

NEAR_SUB*SUBTIME, is statistically significant in the car ownership regressions of all three

samples (columns 1, 4 and 7). The estimated coefficients are 0.017, 0.032 and 0.006 for the full

sample, the resettlement and reformed housing subsample, and the commodity housing subsample,

respectively. These results suggest that more convenient thus more valuable subway stations

discourage nearby households’ car ownership more. Car use (FUEL) also responds to station

heterogeneity as expected, but only statistically significant in the resettlement and reformed

subsample.

Overall, results in Tables 2 and 3 suggest that subway proximity does reduce an urban household’s

probability of owning a car and the subsequent fuel consumption, i.e., the use of car. Such a

conclusion holds in the resettlement and reformed housing subsample that is less subject to the

self-selection bias in residential location. As expected, the magnitude of subway station

proximity’s car ownership effect is smaller in the more exogenous resettlement and reformed

subsample, indicating the existence of residential self-selection bias when households have a

choice of where to live. Moreover, we find stronger “de-motorization” effects of more convenient

subway stations, further confirming the causal nature of subway proximity’s effects on household

automobility. Finally, correcting the Heckman sample selection problem when estimating

household fuel consumption using data from car owners seems necessary at least for the

commodity housing subsample.

6. Robustness Checks

We present two checks of the robustness of Section 5’s findings. First, Beijing’s dominant urban

core is not evenly developed and contains multiple job sub-centers. Using three job sub-centers

(Jinrongjie, Zhongguancun and Yayuncun) instead of the single geographic center of Tian’anmen

Square, we recalculate SUBTIME for each subway station by measuring subway travel time to the

nearest sub-center. Table 4 presents parallel results to those in Table 3’s columns 1, 4 and 7 using

the above alternative definition of SUBTIME.

*** Insert Table 4 about here ***

All the estimated coefficients of NEAR_SUB*SUBTIME in the CAROWN regressions are

statistically significant and consistent with those in Table 3. In fact, using this alternative

SUBTIME measure results in coefficients of NEAR_SUB slightly larger than those estimated in

Table 3, suggesting a stronger effect of a subway station on nearby household’s car ownership if

the uneven distribution of jobs within the urban core is taken into consideration.

Page 12: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

12

Using the dummy variable NEAR_SUB based on a cutoff of 1,000 m to indicate neighborhood

proximity to the nearest subway station follows our assumption that within 1,000 m people can

conveniently get to subway by walking or bicycling, both important access modes to subway in

Chinese cities. In fact, earlier studies on rail transit benefits such as Gibbons and Machin (2005)

and Debrezion et al. (2007) have also used the same cutoff of 1,000 m. While using the 1,000 m

cutoff to construct the NEAR_SUB dummy avoids the potential non-linearity in distance decay

from subway and simplifies the station heterogeneous effect (NEAR_SUB*SUBTIME), the 1,000

m cutoff is still somewhat arbitrary. Tables 5 and 6 serve as robustness checks of our key findings

with two different distance cutoffs (800 m and 1,200 m) for NEAR_SUB.

*** Insert Tables 5 & 6 about here ***

Table 5 reports the results of the car ownership models in all three samples, with columns 1-6 using

the 800 m cutoff and columns 7-12 using the 1,200 m cutoff. All the estimated coefficients of both

NEAR_SUB and NEAR_SUB*SUBTIME are statistically significant and consistent with findings

in Tables 2 and 3. Similarly, Table 6 presents the results of the fuel consumption models in all three

samples using the two alternative distance cutoffs. Again, the estimated coefficients of NEAR_SUB

are all negative and statistically significant, consistent with the results of Table 2.

7. Conclusion

Using the unique urban housing situation in China as a natural experiment, this study provides

robust evidence about urban rail transit’s effects on car ownership and use, accounting for the

potential endogeneity in household residential location. It is one of the earliest such studies in

Chinese cities, where urbanization, motorization and mass transit service have been expanding at

unprecedented speeds, along with intimidating challenges such as traffic congestion and emissions.

To deal with the longstanding concern of residential self-selection, we take advantage of China’s

unique context of urban housing reform and rapid urban redevelopment. We differentiate our

empirical strategy from previous studies by selecting a subsample of urban households in Beijing

who did not choose their residential locations. Using a survey of 826 urban households in Beijing

in 2009, we estimate how subway proximity, controlling for covariates and sample selection bias,

influence household car ownership choice and the fuel consumption of car owners. The key finding

is that proximity to a subway station does reduce a household’s likelihood of owning a car and

fuel consumption, which roughly equals mileage driven. This finding holds for all the household

samples used (full sample, the resettlement housing and reformed housing subsample that is less

subject to endogenous location choice, as well as the more endogenous commodity housing

subsample). Evidence does suggest the existence of residential self-selection bias in the

commodity housing and the full samples. Furthermore, we find heterogeneous effects for subway

stations at different locations – the proximity to a more attractive subway station (with a shorter

Page 13: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

13

travel time to city center or the nearest sub-center) reduces to a greater extent households’ car

ownership likelihood and fuel consumption.

Many large cities in China and other developing countries have been investing in urban rail transit.

Given that subway discourages automobility of nearby urban residents, policy makers interested

in mitigating the congestion and environmental problems of rapid motorization should match

investment in rail transit with significant housing development densities nearby subway stations

to maximize the effectiveness of public spending in transit. Urban rail transit, land use, and

redevelopment/renewal planning should be closely integrated. The important concept of “transit-

oriented development” is again supported by this study.

Page 14: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

14

References

Beesley, M. E., & Kain, J. F. (1964). Urban form, car ownership and public policy: an appraisal

of traffic in towns. Urban Studies, 1(2):174-203.

Bhat, C. R., & Guo, J. Y. (2007). A comprehensive analysis of built environment characteristics on

household residential choice and auto ownership levels. Transportation Research Part B:

Methodological, 41(5), 506-526.

Boarnet, M. G., Anderson, C. L., Day, K., McMillan, T., & Alfonzo, M. (2005). Evaluation of the

California Safe Routes to School legislation: urban form changes and children’s active

transportation to school. American Journal of Preventive Medicine, 28(2), 134-140.

Boer, R., Zheng, Y., Overton, A., Ridgeway, G. K., & Cohen, D. A. (2007). Neighborhood design

and walking trips in ten US metropolitan areas. American Journal of Preventive Medicine,

32(4), 298-304.

Brownson, R. C., Hoehner, C. M., Day, K., Forsyth, A., & Sallis, J. F. (2009). Measuring the built

environment for physical activity: state of the science. American Journal of Preventive

Medicine, 36(4), S99-S123.

Button, K., Ngoe, N., & Hine, J. (1993). Modelling vehicle ownership and use in low income

countries. Journal of Transport Economics and Policy, 27(1), 51-67.

Cao, X., Mokhtarian, P. L., & Handy, S. L. (2007). Do changes in neighborhood characteristics

lead to changes in travel behavior? A structural equations modeling approach.

Transportation, 34(5), 535-556.

Cao, X., & Fan, Y. (2012). Exploring the influences of density on travel behavior using propensity

score matching. Environment and Planning B: Planning and Design, 39(3), 459-470.

Cao, J. (2015). Examining the relationship between neighborhood built environment and travel

behavior: a review from the US perspective. Urban Planning International, 30(4), 46-52.

(in Chinese)

Cervero, R., Sarmiento, O. L., Jacoby, E., Gomez, L. F., & Neiman, A. (2009). Influences of built

environments on walking and cycling: lessons from Bogotá. International Journal of

Sustainable Transportation, 3(4), 203-226.

Cervero, R., & Kockelman, K. (1997). Travel demand and the 3Ds: density, diversity, and design.

Transportation Research Part D: Transport and Environment, 2(3), 199-219.

Chen, J., & Han, X. (2014). The Evolution of Housing Market and Its Socio-economic Impacts in

Post-Reform China: A Survey of the Literature. Journal of Economic Survey, 28(40), 652-

670.

Chin, A., & Smith, P. (1997). Automobile ownership and government policy: The economics of

Singapore's vehicle quota scheme. Transportation Research Part A: Policy and Practice,

31(2), 129-140.

Crane, R. (2000). The influence of urban form on travel: an interpretive review. Journal of

Planning Literature, 15(1), 3-23.

De Vos, J. & Witlox, F. (2016). Do people live in urban neighbourhoods because they do not like

to travel? Analysing an alternative residential self-selection hypothesis. Travel Behaviour

and Society, 4, 29-39.

Page 15: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

15

Debrezion, G., Pels, E., & Rietveld, P. (2007). The effects of railway investments in a polycentric

city: a comparison of competitive and segmented land markets. Environment and Planning

A, 39(9), 2048-2067.

Duranton, G., & Turner, M. A. (2011). The fundamental law of road congestion: Evidence from

US cities. The American Economic Review, 101(6), 2616-2652.

Ewing, R., & Cervero, R. (2001). Travel and the built environment: a synthesis. Transportation

Research Record: Journal of the Transportation Research Board, (1780), 87-114.

Ewing, R., & Cervero, R. (2010). Travel and the built environment: a meta-analysis. Journal of

the American Planning Association, 76(3), 265-294.

Fu, Y., David, K. T., & Zhou, N. (2000). Housing choice behavior of urban workers in China's

transition to a housing market. Journal of Urban Economics, 47(1), 61-87.

Gibbons, S., & Machin, S. (2005). Valuing rail access using transport innovations. Journal of

urban Economics, 57(1), 148-169.

Guo, J. Y., & Chen, C. (2007). The built environment and travel behavior: making the connection.

Transportation, 34(5), 529-533.

Guo, Z. (2009). Does the pedestrian environment affect the utility of walking? A case of path

choice in downtown Boston. Transportation Research Part D: Transport and Environment,

14(5), 343-352.

Handy, S. (2005). Smart growth and the transportation-land use connection: What does the

research tell us? International Regional Science Review, 28(2), 146-167.

Handy, S., Cao, X., & Mokhtarian, P. (2005). Correlation or causality between the built

environment and travel behavior? Evidence from Northern California. Transportation

Research Part D: Transport and Environment, 10(6), 427-444.

Heckman, J. J. (1976). The common structure of statistical models of truncation: Sample selection

and limited dependent variables and a simple estimator for such models. Annals of

Economic Social Measurement, 5(4), 475-492.

Heckman, J. J. (1979). Sample Selection Bias as a Specification Error. Econometrica, 47(1), 53-

161.

Hensher, D. A., Milthorpe, F. W., & Smith, N. C. (1990). The demand for vehicle use in the

urban household sector: theory and empirical evidence. Journal of Transport Economics

and Policy, 24(2), 119-137.

Huang, X., X. Cao, & J. Cao. (2016). The association between transit access and auto ownership:

evidence from Guangzhou, China. Transportation Planning and Technology, 39(3): 269-

283.

Ingram, G. K. & Z. Liu (1999). Determinants of motorization and road provision. In Gomez-Ibanez,

J., W. B. Tye, and C. Winston (eds.), Essays on Transportation Economics and Policy – A

Handbook in Honor of John R. Meyer. Brookings.

Johansson, O., & Schipper, L. (1997). Measuring the long-run fuel demand of cars: Separate

estimations of vehicle stock, mean fuel intensity, and mean annual driving distance.

Journal of Transport Economics and policy, 31(3), 277-292.

Page 16: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

16

Kain, J. F. (1983). Impacts of higher petroleum prices on transportation patterns and urban

development. Research in Transportation Economics, 1, edited by Theodore E. Keeler.

Stamford, Conn.: JAI Press Inc.

Mannering, F., & Winston, C. (1985). A dynamic empirical analysis of household vehicle

ownership and utilization. The RAND Journal of Economics, 16(2), 215-236.

Mokhtarian, P. L. & X. Cao (2008). Examining the impacts of residential self-selection on travel

behavior: A focus on methodologies. Transportation Research Part B: Methodological,

43(3), 204-228.

Pindyck, R. S. (1979). Structure of World Energy Demand. MIT Press.

Silberston, A. (1970). Automobile use and the standard of living in east and west. Journal of

Transport Economics and Policy, 4(1), 3-14.

Schwanen, T. & Mokhtarian P. L. (2005). What affects commute mode choice: neighborhood

physical structure or preferences toward neighborhoods? Journal of Transport Geography,

13(1), 83-99

Stead, D., & Marshall, S. (2001). The relationships between urban form and travel patterns. An

international review and evaluation. European Journal of Transport and Infrastructure

Research, 1(2), 113-141.

Train, K. (1986). Qualitative choice analysis: Theory, econometrics, and an application to

automobile demand. MIT press.

TRB (Transportation Research Board) (2009). Driving and the built environment: The effects of

compact development on motorized travel, energy use, and CO2 emissions. Special Report

298.

Wang, R. (2010). Shaping urban transport policies in China: Will copying foreign policies work?

Transport Policy, 17(3), 147-152.

Wheaton, W. C. (1982). The long-run structure of transportation and fuel demand. The Bell

Journal of Economics, 13(2), 439-454.

Zegras, C. (2010). The built environment and motor vehicle ownership and use: Evidence from

Santiago de Chile. Urban Studies, 47(8), 1793-1817.

Zhang, M. (2004). The role of land use in travel mode choice: Evidence from Boston and Hong

Kong. Journal of the American planning association, 70(3), 344-360.

Zheng, S. & Y. Huo (2010). Low-carbon urban spatial structure: An analysis on private car

traveling. World Economic Papers, (6), 50-65. (In Chinese)

Zheng, S. & M. E. Kahn (2008). Land and residential property markets in a booming economy:

New evidence from Beijing. Journal of Urban Economics, 63(2), 743-757.

Page 17: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

17

Figures

Figure 1. The spatial distribution of the 38 residential complexes in Beijing in 2009

Page 18: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

18

Tables

Table 1. Variable definitions and summary statistics

Variable Definition

Full sample

(N=826)

Resettlement &

reformed

subsample

(N=311)

Commodity

subsample

(N=481)

Mean S.D. Mean S.D. Mean S.D.

CAROWN Whether household owns a car (1=YES; 0=NO) 0.47 0.50 0.43 0.49 0.52 0.50

FUEL* ln (household monthly fuel consumption in RMB) 6.59 0.77 6.53 0.68 6.62 0.81

INCOME ln (household annual income in RMB) 11.39 0.77 11.15 0.63 11.54 0.81

HHSIZE Household size 3.13 1.10 3.01 1.02 3.20 1.45

AGE Age of household head 46.50 13.80 48.75 13.57 44.98 13.11

REIM Whether fuel expenditure can be reimbursed (1=YES;

0=NO) 0.17 0.38 0.19 0.39 0.16 0.36

D_CBD ln (distance from residence to Tian’anmen Square in m) 9.30 0.66 9.45 0.60 9.20 0.68

NEAR_SUB** Whether distance from a household to the closest subway

station is less than 1,000 m (1=YES; 0=NO) 0.32 0.47 0.44 0.50 0.24 0.43

SUBTIME (to center) Travel time from a station to Tian’anmen Square by subway,

in minutes 30.89 11.40 29.94 11.28 31.49 11.44

SUBTIME (to subcenter)

Travel time from a station to the nearest of the sub-centers

(Jinrongjie, Zhonguancun, and Yayuncun) by subway, in

minutes

24.02 10.49 23.21 9.11 24.52 11.25

*Only includes households owning automobile(s).

**We also use alternative cutoffs of 800m and 1,200m to define this dummy in robustness checks.

Page 19: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

19

Table 2. Regression results by housing group

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Full sample Resettlement & reformed subsample Commodity housing subsample

DEP. VAR.

MODEL

CAROWN

Probit

FUEL

OLS

FUEL

Heckman

CAROWN

Probit

FUEL

OLS

FUEL

Heckman

CAROWN

Probit

FUEL

OLS

FUEL

Heckman

INCOME 0.866*** 0.206** 0.277 1.064*** 0.076 -0.470 0.929*** 0.210* 0.565**

(11.333) (2.060) (1.218) (7.860) (0.725) (-1.259) (10.291) (1.863) (2.092)

HHSIZE 0.154*** -0.034 -0.020 0.212** -0.027 -0.128 0.121** -0.056 -0.004

(3.015) (-0.852) (-0.377) (2.067) (-0.447) (-1.564) (2.003) (-1.127) (-0.075)

AGE 0.081** -0.032 -0.025 0.017 -0.075** -0.085*** 0.102*** -0.040 0.004

(2.487) (-1.539) (-0.827) (0.398) (-2.734) (-3.300) (3.025) (-1.269) (0.097)

AGE2 -0.001*** 0.001 0.001 -0.001 0.001** 0.001*** -0.001*** 0.001 -0.001

(-2.740) (1.394) (0.609) (-0.602) (2.801) (3.418) (-3.182) (1.113) (-0.277)

D_CBD 0.260** 0.229*** 0.242** 0.429*** 0.003 -0.231 0.392*** 0.196* 0.354*

(2.177) (2.725) (2.180) (3.735) (0.027) (-1.144) (4.834) (1.836) (1.842)

NEAR_SUB -0.210** -0.293** -0.301** -0.233* -0.160* -0.016 -0.247** -0.281* -0.373*

(-2.045) (-2.237) (-2.028) (-1.779) (-1.950) (-0.099) (-2.324) (-1.839) (-2.032)

REIM 0.279** 0.274** 0.077 0.056 0.322* 0.300*

(2.326) (2.258) (0.336) (0.251) (1.949) (1.816)

Constant -14.349*** 3.021* 1.768 -17.057*** 7.653*** 17.312** -16.814*** 3.600 -3.690

(-9.487) (1.714) (0.412) (-7.488) (4.480) (2.617) (-9.331) (1.593) (-0.660)

Pseudo R2 0.208 0.245 0.209

R2 0.115 0.116 0.091 0.099 0.119 0.129

Lambda 0.162 -0.903 0.814*

(robust t-statistic) (0.356) (-1.592) (1.719)

N 826 385 385 311 134 134 481 251 251

Note: robust z-statistics (CAROWN), and t-statistics (FUEL) in parentheses; standard errors adjusted for 38 clusters.

*** p<0.01, ** p<0.05, * p<0.1.

Page 20: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

20

Table 3. Regression results with station heterogeneity by housing group

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Full sample Resettlement & reformed subsample Commodity housing subsample

DEP. VAR.

MODEL

CAROWN

Probit

FUEL

OLS

FUEL

Heckman

CAROWN

Probit

FUEL

OLS

FUEL

Heckman

CAROWN

Probit

FUEL

OLS

FUEL

Heckman

INCOME 0.881*** 0.210** 0.322 1.127*** 0.091 -0.250 0.935*** 0.215 0.612**

(11.488) (2.073) (1.290) (8.702) (0.904) (-0.695) (10.241) (1.668) (2.106)

HHSIZE 0.145*** -0.038 -0.018 0.181* -0.043 -0.091 0.118* -0.061 -0.007

(2.864) (-0.932) (-0.324) (1.803) (-0.686) (-1.095) (1.953) (-1.227) (-0.149)

AGE 0.080** -0.033 -0.022 0.009 -0.079** -0.081*** 0.102*** -0.040 0.009

(2.421) (-1.582) (-0.704) (0.204) (-2.856) (-3.068) (3.009) (-1.278) (0.198)

AGE2 -0.001*** 0.001 0.001 -0.001 0.001*** 0.001*** -0.001*** 0.001 -0.001

(-2.684) (1.433) (0.471) (-0.406) (2.919) (3.159) (-3.167) (1.117) (-0.376)

D_CBD 0.244** 0.222** 0.245** 0.337*** -0.040 -0.155 0.380*** 0.170 0.333

(2.072) (2.546) (2.224) (3.156) (-0.366) (-0.897) (4.592) (1.120) (1.692)

NEAR_SUB -0.648*** -0.486* -0.586* -1.068*** -0.556** -0.218 -0.416*** -0.579* -0.851**

(-3.319) (-1.994) (-1.965) (-5.424) (-2.659) (-0.493) (-2.650) (-1.874) (-2.357)

NEAR_SUB*SUBTIME 0.017** 0.007 0.010 0.032*** 0.014** 0.005 0.006* 0.011 0.017

(2.292) (0.866) (1.100) (4.887) (2.597) (0.442) (1.740) (0.945) (1.553)

REIM 0.275** 0.269** 0.072 0.048 0.308* 0.282*

(2.312) (2.233) (0.319) (0.213) (1.889) (1.731)

Constant -14.336*** 3.076* 1.086 -16.620*** 8.026*** 13.599** -16.763*** 3.803 -4.187

(-9.538) (1.735) (0.238) (-7.503) (4.719) (2.241) (-9.285) (1.661) (-0.720)

Pseudo R2 0.211 0.258 0.210

R2 0.118 0.118 0.104 0.108 0.124 0.136

Lambda 0.255 -0.547 0.900*

(robust t-statistic) (0.519) (-1.045) (1.782)

N 826 385 385 311 134 134 481 251 251

Note: z-statistics (CAROWN) and t-statistics (FUEL) in parentheses, standard errors adjusted for 38 clusters.

*** p<0.01, ** p<0.05, * p<0.1.

Page 21: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

21

Table 4. Robustness check of CAROWN regressions using multiple job sub-centers for station heterogeneity

(1) (2) (3)

Full sample Resettlement & reformed

subsample

Commodity housing

subsample

INCOME 0.871*** 1.074*** 0.927***

(11.796) (8.056) (10.662)

HHSIZE 0.151*** 0.185* 0.120**

(3.181) (1.830) (2.091)

AGE 0.073** 0.011 0.096***

(2.326) (0.261) (2.985)

AGE2 -0.001*** -0.001 -0.001***

(-2.605) (-0.484) (-3.125)

D_CBD 0.234** 0.379*** 0.344***

(1.997) (3.302) (3.671)

NEAR_SUB -0.688*** -0.758*** -0.559***

(-4.980) (-3.173) (-4.849)

NEAR_SUB*SUBTIME 0.024*** 0.026*** 0.016**

(4.126) (3.138) (2.193)

Constant -13.964*** -16.457*** -16.212***

(-9.762) (-8.181) (-9.176)

Pseudo R2 0.214 0.253 0.212

N 826 311 481

Note: robust z-statistics in parentheses, standard errors adjusted for 38 clusters.

*** p<0.01, ** p<0.05, * p<0.1.

Page 22: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

22

Table 5. Robustness check of CAROWN regressions with different cutoffs for NEAR_SUB

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12)

Cutoff 800m 1200m

Sample full sample resettlement and

reformed subsample commodity subsample full sample

resettlement and

reformed subsample commodity subsample

INCOME 0.869*** 0.883*** 1.072*** 1.133*** 0.929*** 0.932*** 0.865*** 0.880*** 1.064*** 1.126*** 0.929*** 0.935***

(11.231) (11.272) (7.783) (8.382) (10.328) (10.258) (11.310) (11.435) (7.870) (8.693) (10.291) (10.241)

HHSIZE 0.153*** 0.149*** 0.201** 0.186* 0.116* 0.116* 0.152*** 0.144*** 0.212** 0.181* 0.121** 0.118*

(3.002) (2.904) (1.964) (1.773) (1.941) (1.923) (2.960) (2.819) (2.067) (1.803) (2.003) (1.953)

AGE 0.079** 0.079** 0.015 0.008 0.102*** 0.102*** 0.079** 0.078** 0.017 0.008 0.102*** 0.102***

(2.437) (2.396) (0.353) (0.180) (2.974) (2.973) (2.397) (2.335) (0.395) (0.198) (3.025) (3.009)

AGE2 -0.001*** -0.001*** -0.001 -0.001 -0.001*** -0.001*** -0.001*** -0.001*** -0.001 -0.001 -0.001*** -0.001***

(-2.689) (-2.651) (-0.542) (-0.372) (-3.136) (-3.136) (-2.655) (-2.602) (-0.599) (-0.400) (-3.182) (-3.167)

D_CBD 0.256** 0.259** 0.445*** 0.472*** 0.368*** 0.365*** 0.257** 0.245** 0.431*** 0.342*** 0.392*** 0.380***

(2.168) (2.193) (3.950) (4.230) (5.270) (5.118) (2.247) (2.181) (3.758) (3.223) (4.834) (4.592)

NEAR_SUB -0.224** -0.583*** -0.324** -1.297*** -0.270*** -0.348*** -0.235** -0.668*** -0.245* -1.078*** -0.247** -0.416***

(-2.135) (-2.728) (-1.966) (-5.089) (-2.663) (-2.994) (-2.410) (-3.333) (-1.869) (-5.425) (-2.324) (-2.650)

NEAR_SUB

*SUBTIME 0.014* 0.036*** 0.003* 0.017** 0.032*** 0.006*

(1.773) (4.393) (1.737) (2.269) (4.893) (1.840)

Constant -14.319*** -14.493*** -17.235*** -17.944*** -16.554*** -16.562*** -14.251*** -14.263*** -17.067*** -16.640*** -16.814*** -16.763***

(-9.455) (-9.386) (-8.179) (-7.706) (-9.699) (-9.658) (-9.478) (-9.605) (-7.491) (-7.508) (-9.331) (-9.285)

Pseudo R2 0.208 0.210 0.248 0.260 0.210 0.210 0.209 0.212 0.246 0.258 0.209 0.210

N 826 826 311 311 481 481 826 826 311 311 481 481

Note: robust z-statistics in parentheses; standard errors adjusted for 38 clusters.

*** p<0.01, ** p<0.05, * p<0.1.

Page 23: Does Subway Proximity Discourage Automobility? …...1 Does Subway Proximity Discourage Automobility? Evidence from Beijing Yingjie Zhanga, Siqi Zhengb,c,Cong Sund, Rui Wange a. School

23

Table 6. Robustness check of FUEL regressions with different cutoffs for NEAR_SUB

(1) (2) (3) (4) (5) (6)

Cutoff 800m 1200m

Sample full sample

resettlement and

reformed

subsample

commodity subsample

full sample

resettlement and

reformed

subsample

commodity subsample

Model OLS OLS Heckman OLS OLS Heckman

INCOME 0.208** 0.074 0.539* 0.208** 0.076 0.565**

(2.060) (0.671) (2.024) (2.066) (0.725) (2.092)

HHSIZE -0.040 -0.040 -0.019 -0.036 -0.027 -0.004

(-0.942) (-0.715) (-0.364) (-0.872) (-0.447) (-0.075)

AGE -0.032 -0.074** 0.001 -0.033 -0.075** 0.004

(-1.518) (-2.659) (0.014) (-1.597) (-2.734) (0.097)

AGE2 0.001 0.001** -0.001 0.001 0.001** -0.001

(1.371) (2.738) (-0.200) (1.444) (2.801) (-0.277)

D_CBD 0.209** -0.006 0.287 0.217** 0.003 0.354*

(2.552) (-0.068) (1.567) (2.603) (0.027) (1.842)

NEAR_SUB -0.305** -0.276** -0.368* -0.269** -0.160* -0.373*

(-2.291) (-2.146) (-1.959) (-2.109) (-1.650) (-2.032)

REIM 0.282** 0.125 0.297* 0.268** 0.077 0.300*

(2.348) (0.529) (1.788) (2.206) (0.336) (1.816)

Constant 3.182* 7.761*** -2.592 3.139* 7.653*** -3.690

(1.801) (4.390) (-0.480) (1.797) (4.480) (-0.660)

R2 0.116 0.110 0.124 0.113 0.091 0.129

N 385 134 251 385 134 251

Note: t-statistics in parentheses; standard errors adjusted for 38 clusters.

*** p<0.01, ** p<0.05, * p<0.1.