A study on possibility of commuting trip using private motorized modes in cities around the world: Application of multilevel model

Transportation Research Part D 41 (2015) 228–243

Contents lists available at ScienceDirect

Transportation Research Part D journal homepage: www.elsevier.com/locate/trd

A study on possibility of commuting trip using private motorized modes in cities around the world: Application of multilevel model Hyunsu Choi a,⇑, Yoongho Ahn b a b

Global Marketing Division, Korean Railroad Research Institute, 176 Cheoldo Bangmulgwan-ro, Uiwang-si, Gyeonggi-do, Republic of Korea Department of Civil Engineering, College of Science and Engineering, Ritsumeikan University, 1-1-1 Nojihigashi, Kusatsu-Shi, Shiga-Prefecture, Japan

a r t i c l e

i n f o

Article history: Available online 6 November 2015 Keywords: Commuting trip possibility Private motorized modes Disaggregated traffic data in the world Hierarchical logistic regression model Principle component analysis

a b s t r a c t The focus of the current research was to evaluate how the individual’s social characteristics and urban infrastructure impacts the usage of Private Motorized Modes (PMM). Based on individual and urban characteristics a multilevel analysis was conducted on the possibility of commuting trip by private motorized modes on the rush time of 78 cities around the world. Also the selected cities were classified through a principal component analysis, and based on the classification the impact of and urban variables on the possibility of commuting trips made by private motorized modes (PCTP) was verified. Results showed a diverse range of variables related to the usage of PMM, as well as the urban structure and railway lengths being an important variable in travel behavior. Ó 2015 Elsevier Ltd. All rights reserved.

Introduction In recent years, the range of individual travel behavior is expanding with the advancement in motorization, alongside the economic development of cities around the world. The world is shifting toward faster and energy intensive modes of transportation (Schafer and Victor, 1999). In order to combat this issue, new city planning methods and management strategies for technical development that shift popularity from cars to public transit and reduce the dependence on private motorized modes are required. Many planning techniques and research projects since 1970 have focused on developing the urban structure based on the concept of sustainable development. In addition, it is recognized that the importance of developing sufficiently high quality alternatives that induce drivers away from their cars is increasing. In particular, the heavy traffic demand during the commute hours is one of the factors that increase social costs, such as traffic congestion and delays. To cope with the increased social costs, congestion fees have been implemented in various cities to control traffic demand during peak commute hours (Suryo et al., 2007). Recently, several studies were carried out to derive policy measures to reduce the overuse of Private Motorized Modes (PMM) from an urban structure and transport environment perspective. A study by Newman that examined 46 cities around the world provided strong evidence that high population density contributed to controlling PMM traffic demands (Newman and Kenworthy, 1989). Studies on the correlation between population density and transport characteristics have been carried out in other countries, and Asian counties with high population densities, including South Korea and Japan, have shown high interest in the subject (Morimoto and Koike, 1995). However, there are differences in vehicle usage among cities or ⇑ Corresponding author. E-mail addresses: [email protected] (H. Choi), [email protected] (Y. Ahn). http://dx.doi.org/10.1016/j.trd.2015.10.008 1361-9209/Ó 2015 Elsevier Ltd. All rights reserved.

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

229

zones with similar population densities, which indicates that while population density is strongly correlated with the usage of PMM, there are individual factors or regional factors causing traffic demands created by PMM. As such, it is necessary to analyze the details of the relationships among the factors at different levels. The usage characteristics of PMM or factors expected to affect the usage characteristics can be examined based on a hierarchical structure. It could be considered that personal or household factors are in a lower hierarchy and zone, while factors of cities and countries are in a higher hierarchy. In general, social study data for a city or a country are strongly correlated with the features of the group in which the samples are included, and the individual samples are affected by the group’s characteristics. Therefore, to explain the individual features of PMM usage in the city, it is necessary to consider the structural hierarchy of individual samples and the groups from which the samples were drawn. However research regarding both urban characteristics and the individuals social attributes, such as traffic patterns has been rarely conducted. Also, there has not been any research on the reliance of personal nodes on the basis of individual and urban factors. In terms of the conditions, this study focuses on how the structural and economical characters of urban districts, as well as the social characteristics of individuals, affect the dependence of personal transportation. Based on this background, this research adopts a multi-level analysis method that enables us to understand the effects of hierarchical data on individual and urban level variables. Through the analysis on commuting trips of PMM, this research confirms the existence of significant variances in the possibility of commuting trips made by PMM (PCTP) among the cities, and the manner in which the variations and the relationship between dependent variables and independent variables at different levels are clarified. To achieve this, for the current study a database of cities was built that considered individual factors and travel behavior of PMM reflecting travel behaviors described in the National Household Travel Survey (NHTS), which describes some of the factors associated with demographic information on the individual, concerning an individual’s travel behavior in 78 cities in 14 countries. In addition, this study checked the validity of hypotheses for factors at different levels of the hierarchy that affect the possibility of commuting trips made by PMM.

Literature review Travel behaviors by urban structure, transport infrastructure and public transport Previous studies on the relationship between urban environment factors and transport behaviors have generally utilized data collected at the city or sub-municipal level (namely ward), or at the zone level due to the constraints in defining spatial ranges for a certain study purpose in the urban environment. Crane and Crepeau (1998) studied the relationship between shapes of road network and transport behavior cities in the United States. Suzuki and Muromachi (2009) analyzed the relationships between accessibility to railway stations and usages of PMM. Also, Ewing and Cervero (2001) reviewed factors that affect usages of PMM by categorizing urban space according to zones. On the other hand, an international survey study performed by Mackett and Sutcliffe (2003) reported several reasons for developing urban public transit systems, improving public transport, reducing traffic congestion, serving the city center better, improving the environment, and stimulating development. In addition, reducing transportation energy consumption by mitigating traffic congestion is widely considered as one of the common reasons for building new transit systems. Rail has been well received as a transit mode that promotes transitoriented development (TOD), which in the United States often translates into compact, mixed-use, and pedestrian-friendly development around transit stations, as an alternative to sprawl. In this context, Winston and Langer (2006) indicated that congestion costs of PMM decrease in a city as rail transit mileage expands. In several US cities, traffic congestion growth rates declined after a Tram service was established. Baum-Snow and Kahn (2005) found significantly lower average commute travel times in areas near rail transit than in otherwise comparable locations that lacked rail service, due to rail’s higher travel speeds compared with PMM or buses under the same conditions. In addition, Litman (2007) shows that per capita congestion delay is significantly lower in cities with high quality rail transit systems than in otherwise comparable cities with little or no rail service. Rail system expansion generally occurs in large and growing urban areas in response to increasing congestion. As a result, simplistic analysis often shows a negative correlation between rail transit and road congestion. Kenworthy et al. (1999), as well as Kenworthy and Laube (1999), have also shown that the more intensive the land use, the shorter the travel distance, the greater the viability of transit, the higher the car occupancy of vehicles, the less the need for a car, and these patterns suggest that the urban density is fundamental to shaping travel behavior. Giuliano and Dargay (2006) conducted an international comparative analysis of the relationships between car ownership, daily travel and urban structure. Using travel diary data for the US and Great Britain, they estimated models of car ownership and daily travel distance. Choi et al. (2011) also stated clearly that population density and number of vehicles have a negative correlation. In terms of the effect of the transport infrastructure on travel behavior, it is confirmed that an increase in road density and extension of the railway infrastructure have positively contributed to road transport demand and transit oriented urban development.

230

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

McCann (2000) recommended that there be government support for the construction of more road infrastructures to restrain the sprawl that makes people dependent on private vehicles. Litman (2007) insisted that rail transit is a solution for traffic congestion and re-urbanization as an alternative mode of private vehicles from the following aspects. First, high quality transit service can reduce travel time costs for people who shift modes. Second, grade-separated transit reduces delays on parallel roadways. Urban transportation congestion tends to maintain equilibrium: congestion deters growth in peak-period trips. Finally, rail transit can stimulate Transit Oriented Development (TOD) – compact, mixed-use, walkable urban villages in which the residents tend to own fewer cars and drive less than if they lived in more automobiledependent neighborhoods. Also, Choi et al. (2013b) stated that PMM efficiency would be higher with a high railway infrastructure density in terms of urban energy efficiency from the macro-view point. From this result, it can be interpreted that PMM differs depending on the existence of the railway infrastructure. Effect of individual characteristics on travel behaviors Other studies have analyzed the effectiveness of individual household factors, including housing character, family size and household income level, or personal factors on travel behaviors. Two types of methodologies were adopted for these studies. One is to utilize personal factors directly to form a formula, and the other is to utilize data collected by categorizing personal factors according to certain spatial units. Cervero (1996), Sun et al. (1998) and Schimek (1996) analyzed the relationship between travel behaviors and individual factors by utilizing personal or household characteristics rather than by collecting those data by zone. Meanwhile, Choi et al. (2013b) utilized data collected by categorizing individual factors according to urban levels (Gu or Dong) and provided evidence that there is a specific pattern of travel behaviors depending on the level of urban density, and a robust correlation between the density and the patterns. Also, Dargay and Gately (1999) studied correlation between average income and average vehicle ownership at a national level. In this study, age and gender have been constructed to reflect the differentiation in experiences between ages, gender and travel behaviors, in order to elicit the unique factors affecting the probability of a commuting trip. The first set of variables is the capability constraint factors, which included gender and age. According to Elder (1985) and George (1996), the age categories represent major episodes in the life course. Rosenbloom (1995) and Chu (1994) both found that age is negatively related to trip distance. Meanwhile, car availability was found to be significant in the study of Vance and Iovanna (2007), which included the finding of a gender effect of a reduced probability of women to use cars and travel longer distances. They explained this as revealing the ‘‘patriarchal constraints” or traditional gender roles that limit women’s access to the car in cases in which a choice between drivers must be made and that ‘‘the general pattern is for husbands to have first choice of car use” (Pickup, 1985). According to Stradling et al. (2005), the effect of household size is hypothesized to have a negative effect on distance traveled via the effect of increasing need for social interactions at home in relation to an increased number of children at home. In addition, they found consistent positive effects of license ownership and number of cars owned by the household as indicators of car availability. Also, various theories regarding relations between trip behaviors and urban factors including spatial structure, economic status, and infrastructure for roads and railways are commonly accepted. In particular, correlations between urban density and trip behavior or modal share have been the subject of various studies, including one by Newman et al. (1989). Economic level of city and motorization On the other hand, it is revealed that the level of urban economic growth also has an effect on the patterns of traffic behaviors. Regarding the urban economic level, road transportation has gradually become the dominant part of the transportation system in China, where economic growth is rapidly increasing, and thus the amount of fuel consumed on the road is increasing with time. Furthermore, the energy situation related to road transportation in developed countries is the same as well (He et al., 2005). Also, Pucher et al. (2005) indicated that urban policies are prioritized in India to mitigate the urban transport crisis that accompanied rapid economic growth. They point out that there are two main obstacles to implementing the policies needed to deal with India’s urban transport crisis: financial and political problems according to the economic development. On the other hand, Choi et al. (2013a) clarified that there is a negative correlation between population density and motorized mode dependency, and the degree of correlation becomes stronger as the economic level of a city increases. From this, it can be interpreted that traffic congestion in the CBD area could negatively affect the PMM despite the growth of personal income level and accompanying urban motorization. Passenger car occupancy Finally, it is known that the increase in car occupancy due to car sharing programs could be a factor affecting the use of passenger cars. Car sharing programs were implemented as a measure to mitigate traffic congestion in high density urban areas, and other cities also implemented transport policies to reduce passenger vehicle use by utilizing HOV (High-

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

231

Occupancy-Vehicle) and car sharing programs (Wang, 2011). However, average car occupancy during peak commute hours is lower than at any other time of day, and passenger car use for commuting trips is relatively affected by the time factor, which is highly related to transport demand (KOTI report, 2008). Methodology and data The idea behind this research is to examine the factors in individual-urban level (travel behavior at the individual level and socio-economic-infra character at the urban level) influence possibility by PMM. In general, surveys on travel behaviors have been carried out for multi-level cities and zones. In particular, if a survey ignores the features of cities and zones and analyzes based on individual sample units, independent characteristics of the individual samples could not be guaranteed due to the assumption that the cities are homogeneous. On the other hand, analysis carried out only with urban level data could not consider individual sample errors. To this end, the current research utilizes a multilevel analysis model that enables the consideration of the data in a hierarchical structure at the same time; two levels in an individual-urban scale; by using NHTS and aggregated data of the cities in the world. Data in two levels are composed by individual characteristics related to demographic and travel duration by PMM at the individual trip level, and statistical data on urban characteristics is aggregated at the urban level for each area in Korea, Japan, the United States, and developing countries, which was originally collected by research institutes around the world. However due to the diversity of the samples, collected around the world, it was difficult to gather the data through a single institute. Although the method in which the data was gathered and processed was conducted in an identical manner, some modifications were made do the original data due to the criteria in which it was measured. As an example, conditions used to extract the trip data such as private motorized modes, trip purpose, and OD were the data that was measured equally by all institutes. Model specification Multilevel analysis model is appropriate when there is correlation among clusters of subjects. For example, data obtained from surveys of individuals within individuals across different urban areas may constitute a two level hierarchy-individual trip (level 1), urban area (level 2). It is the presence of within-cluster correlation that justifies the use of a multilevel model. Multilevel modeling is commonly used in social contexts and individual behaviors. The hierarchical structure of data is often seen in the field of urban and transport planning (Suzuki and Muromachi, 2011). In aggregation analysis on zone clusters, zones include individual travel behaviors and thus it is possible to consider that zones are macro level (level 2) and individual travel behaviors are micro level (level 1). Suppose we have collected data on i subjects (level 1) nested within j organization (level 2). With two-level structure data, three different equations can be formulated: individual-level model (level 1 model), organization-level model (level 2 model), and substitutional model. Assuming normally distributed errors, for subject ij we have a level 1 model as

b ij ; r2 Þ; rij
ð0; r2 Þ; Y ij
Nð Y ij b ij ¼ b ^0j þ b ^1j X 1ij þ b ^2j X 2ij þ b ^3j X 3ij . . . þ b ^Qj X Qij ; Y Y ij ¼ b0j þ

Q X

bqj X qij þ r ij

ð1Þ

ðLevel 1 modelÞ

q¼1

where b0j is the intercept, b1j the regression coefficient associated with the predictor Xij, and rij is the residual accounting for level 1 random effects. Although this formulation is similar to a linear regression model, there is an important difference in that both intercept and regression coefficients have subscript j, indicating that the intercept b0j and the slope coefficient b1j are permitted to vary across organizations (level 2). At the organization level, the units are organizations and the regression coefficients in the level 1 model for each organization are conceived as outcome variables depending on organization-level characteristics. Generally, there are three sub-models in multilevel models, depending on whether or not the intercept b0j and the slope coefficient b1j are assumed to vary across organizations. In this application, the intercept b0j is assumed to vary across organizations as a function of a grand mean, a single explanatory variable, and an error term; in addition, the slope coefficient b1j is assumed to vary across organizations (based on the assumption that variations could exist in the relationship between dependent and explanatory variables depending on their effects at city level (or a variable at a higher position in the hierarchy)). Then, the intercept b0j and the slope coefficient b1j are formulated as follows:

b0j ¼ r 00 þ

Q X r0n W qj þ u0j n¼1

Sq X bqj ¼ r q0 þ rqn W qj þ uqj n¼1

ð2Þ ðLevel 2 modelÞ

232

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

In model (2), note that the gammas (regression coefficients) have subscript j because they are assumed to vary across organizations (based on the assumption that variations could exist in the relationship between dependent and explanatory variables depending on their effects at the city level (or a variable in a higher position in the hierarchy)). This model corresponds with a random-intercept-slope model (Yannis et al., 2008; Kreft and Leeuw, 1998). Yij is the logit prediction for the ith subject at level 1 and the jth unit at level 2, c00 is the intercept meaning the grand mean, Wj the organization-level characteristic, Xij the individual-level characteristic, cq0 is the regression coefficient associated with organization-level characteristic and individual-level characteristic – in other words, fixed effects determined by limited factors – and cqs represents the regression coefficients associated with variance of slopes for each variable at the individual level which are explained by variable at the city level, respectively, uqj a random effect accounting for the random variation at level 2, where uj
(0, s00) and rij is the individual-level random effect, where rij
N(0, r2). For a standard multilevel model described above, a dependent variable is continuously distributed. If the observed outcomes Yij are binary, a binomial logit model is appropriate. Under hierarchical structure data, this translates to a multilevel binomial logistic model. A multilevel binomial logistic model is conceptually equivalent to a standard multilevel model with the exception of the dependent variable (Guo and Zhao, 2000). In this study, dependent variable Pij means the possibility of commuting trip by PMM and Pij = (0, 1). To consider the binominal variables as continuous variables, a logit transformation is required. Recall that the logit is the natural log of the odds for the commuting trip possibility. To get from logits to odds to predicted possibility of commuting trip by private modes, considering a binomial Yij = (0, 1) outcome, and Pij = (exp(Yij)/(1 + exp(Yij))), substitute this with formula (1) and (2) and it is written as

LogitðhÞ ¼ logðPðY ij Þ=1  PðY ij ÞÞ ¼ b0j þ

Q X bqj X qij þ r ij q¼1

¼ r 00 þ

Q X

Q X

n¼1

q¼1

r 0n W nj þ

Sq Q X Q X X rq0 X qij þ r qs W sj X qij þ u0j þ uqj X qij q¼1 s¼1

ð3Þ ðMultilevel logit modelÞ

q¼1

Evaluation and procedure of multilevel models To determine the proportion of the total variability that is accounted for by differences among cities, ICC has to be checked. The proportion of the variance in the outcome between the level 2 units is examined based on the ‘‘Intra-class Correlation Coefficient (ICC).” To determine the goodness of fit of two models; e.g. regression model versus random effect model, it is necessary to check the deviance, and the likelihood test is employed. This test requires that the difference between the log-likelihoods of the two models being compared be obtained. Deviance is represented as 2 of log-likelihood (i.e. 2logL1(2logL2)), and it is commonly interpreted that the smaller the deviance, the better the goodness of fit for the model concerned (Bryk and Raudenbush, 1992). In particular, if the reduction range of the deviance is greater than 2 with additional variable, it is interpreted that the goodness of fit for the model has improved. Data Person Trip data The significant drawback in using real life data for comparison analysis lies in the unreliability of the data collection process. To address this drawback, this research utilized disaggregate data (collected by NHTS, which is known for nonaggregated transport related data and similar ones in the world). This study defined it as Person Trip (PT) data, which is opened to the public and is mainly used for understanding an attribute of society and travel behavior on an individual level. The data are collected in 78 countries including the USA, South Korea, Japan, and other developing countries, and was originally collected by official statistical institutes of the countries (Table 1). The PT data is disaggregated data on individual characteristics. Economic and social activities in city are carried out by ‘‘persons.” A ‘‘person” who lives in an urban or rural area generates travel characteristics. PT data is used primarily to gain

Table 1 Person Trip data around the world. Nation

Year

Name used for ‘‘Person Trip survey”

Source

Japan Korea USA Developing countries

2005 2005 2001 1996–2005

The Nationwide Person Trip Survey Household Travel Survey National Household Travel Survey: (NHTS) Person Trip Survey

Ministry of Land, Infrastructure, Transport and Tourism (MLIT) Korean Transport Database (KTDB) Federal Highway Administration (FHWA) Japan International Cooperation Agency (JICA) 15 cities in Developing countries

233

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

a better understanding of individual factors and travel behavior. PT data is a tool in the urban transportation planning process; it provides data on personal travel behavior, trends in travel over time, trip generation rates, national data to use as a benchmark in reviewing local data, and data for various other planning and modeling applications. The important purpose of this research is to determine how the major factors generate the diversity of the relationship between commuting trip possibility by PMM and the features of variables at the city level. Since PT data offers individual information on demography and travel behavior in detail, it is appropriate to utilize for urban studies with hierarchical structures, especially for the cities in the world that have diverse characteristics. Therefore, the results of this study show which main variables at the city level and the individual trip level have an effect on possibility of commuting trip being made by PMM (PCTP), and provide evidence on how city level factors interact with individual trip factors, and how the effects vary. However, based on the fact that the data had to be collected from a number of countries around the world, and each country has a different method on how they collect their data, we have found it to be impossible to find data based on a unified method. So to resolve this matter, we used data that was based on identical survey methods as well as using the same data that was to be modified with the expansion rate. Also, terms defied in this research for trip collection, such as the private motorized mode, trip purpose, and OD is found to be defined in an identical way in all countries. Definition of trip in this research The current research extracted data for trips made by PMM. Hence, freight traffic, which is mainly through-traffic, making it difficult to determine the spatial zone, was excluded from this research. In addition, the trip mode used for the longest time in a complete trip was treated as the representative mode for the trip. Furthermore, extracted trips below 4 km/h on the representative mode were excluded from target trip as walking. In this research, trips that follow the above limitations were extracted from the total trips made within the target area, and used for estimation of the PCTP. Target major city areas The geographic focus of the current research is focused on 78 major city areas in 14 countries, from which PT data was drawn (Table 2). Target areas in this study were defined as a major city area with a minimum population of 800,000. The cities were also selected to represent a diverse economic status. The distribution of the target areas is as followed: 21 cities in Asia (7 cities in Korea and 14 cities in Japan), 45 cities in the United States, and 12 cities in developing countries.

Table 2 Target major city areas. No.

Country

Target major city area

No.

Country

Target major city area

Japan

Sapporo Sendai Saitama Chiba Tokyo

1331 1230 884 1007 284

27 28 29 30 31

USA

Yokohama Kawasaki Nagoya Kyoto Osaka

613 316 1391 675 198

32 33 34 35 36

Chicago Cincinnati Cleveland Columbus Dallas-Fort Worth Denver Detroit Hartford Houston Indianapolis

6 7 8 9 10 11 12

Kobe Hiroshima

674 1290

37 38

13 14 15 16

Kitakyushu Fukuoka Seoul Inchon

1552 882 28,132 73,240

39 40 41 42

17 18

Pusan Daegu

64,570 55,162

43 44

19 20 21 22

Kwangju Daejon Ulsan Atlanta

35,117 35,411 34,842 2707

45 46 47 48

2022 4284 4193 981

1 2 3 4 5

23 24 25 26

Korea

USA

Austin Boston BuffaloNiagara Falls Charlotte

Sample no. of trip

Sample no. of trip

No.

Country

Target major city area

Sample no. of trip

5658 1345 2128 1068 4476

53 54 55 56 57

USA

Pittsburgh Portland Providence Rochester Sacramento

1782 1656 783 6663 1321

2273 3840 848 3955 1135

58 59 60 61 62

2054 964 1846 1652 3974

Jacksonville Kansas City

881 1525

63 64

8081 769 715 1493

65 66 67 68

Lebanon Syria

24,537 11,855 3794 21,507

7954 2898

69 70

Philippines China

Manila Changdu

13,691 8096

970 773 33,481 1078

71 72 73 74

Nicaragua Romania Egypt Malaysia

Managua Bucharest Cairo Kuala Lumpur

9670 18,588 12,636 9392

49 50 51

Los Angeles Louisville Memphis Miami-Fort Lauderdale Milwaukee MinneapolisSt. Paul Nashville New Orleans New York NorfolkVirginia Oklahoma City Orlando Philadelphia

St. Louis Salt Lake City San Antonio San Diego San FranciscoOakland Seattle Tampa-St. Petersburg Washington Honolulu Tripoli Damascus

668 932 3516

75 76 77

Vietnam Vietnam Kenya

Ho-Chi-Minh Hanoi Nairobi

13,295 6615 4567

52

Phoenix

2187

78

Peru

Lima

19,508

2955 1689

234

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

For data acquisition, the research used established data from major cities in the United States, Japan, Korea and other developing countries. Common variables were selected for a consistent data set, due to the diversity of each countries data configuration. Also, the term ‘‘major city” was defined based on the administrative definition of each country. For example the cities in the United States were selected based on the definition of the Federal Highway Administration’s ‘‘major city area”. And for Korea, Japan and developing countries, the term major city area was based on the individual countries administrative definition. All data acquired in this research was based on this criterion. Classification of the global cities A multilevel analysis was conducted to verify the influence of individual and urban variables on PCTP. The multimodal method was selected to analyze the deviation between the samples and within the samples as well. However this method is limited in a manner where it only explains the variation between or within the selected cities, making it difficult to characterize the underlining cause behind the variation. In the current study the selected cities were classified through a principal component analysis, and based on the classification the impact of and urban variables on PCTP was verified. This study use PCA to derive new variables for private passenger vehicle dependence and railway transport dependence by integrating common factors in the used variables in city level, and selects varimax which is an orthogonal rotation for simplifying factor loading. The result is described below. The upper part of Table 3 illustrates Kaiser–Meyer–Olkin (KMO) test value of 0.831 carried out to identify how the unobserved latent variables describe the correlation between variables used in the analysis. Kaiser (1974) classifies the measure of KMO greater than 0.90 as ‘‘marvelous”, 0.80–0.89 as ‘‘meritorious”, 0.70–0.79 as ‘‘middling”, 0.60–0.69 as ‘‘mediocre”, 0.50–0.59 as ‘‘miserable” and 0.5 or smaller as ‘‘unacceptable”. Therefore, the Sampling Adequacy (0.732) in this study is regarded as satisfactory. Because significance probability is .000, application of the sample in this study to factorial analysis is regarded as high significance in terms of statistics. Table 4 shows the weight for each principal component of Principle Component 1 (PC1), Principle Component 2 (PC2) derived with transport behaviors and transport infrastructure in the 78 cities used in PCA. The Rotated Component Matrix of Table 4 illustrates total road length, the number of registered passenger vehicles, PMM share (modal share), and GRDP which is an indicator showing the economic level of a city to describe the common factors in PC1. In particular, note that the factor loading of each variable is 0.6 or greater, and is in a highly positive relation with PC1.

Table 3 KMO and Bartlett’s test & communalities.a Kaiser–Meyer–Olkin measure of sampling adequacy Bartlett’s test of sphericity

Approx. Chi-Square Df Sig.

.831 315.460 36 .000

Initial

Extraction

1.000 1.000 1.000 1.000 1.000 1.000 1.000 1.000

.642 .844 .693 .776 .787 .833 .465 .710

Communalities Urban density Share of PMM Share of PUB Registered passenger vehicle GRDP Road length Rail length Passenger car occupancy a

Extraction method: Principal component analysis.

Table 4 Rotated component matrix.a Variable

Road length Registered passenger vehicle Share of PMM Share of PUB GRDP Urban density Passenger car occupancy Rail length a

Component 1

2

.910 .881 .851 .767 .684 .639 .114 .153

.069 .029 .346 .324 .565 .483 .835 .492

Rotation method: Varimax with Kaiser normalization.

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

235

Such a tendency of PC1 implies that the common factors of PC1 is a component showing dependence on private passenger vehicles in that the factor loading by the hardware variable for using private passenger vehicles, for example, total road length and the number of private passenger vehicles has a great weight, and positive correlation is observed between the service level variable related to using road transportation like modal share of PMM and PC1. Moreover, this is based on that modal share of PUB and the factor loading of daily trip number by PUB is negatively correlated with PC1, and no correlation of total rail length with PC1. In this context, PC1 can be analyzed with the ‘‘Passenger car dependence” with correlation between variables of PC1 and the factor loading values described by each variable. With respect to factor loading showing correlation between PC2 and each variable, PC2 is described with total length of railway, GRDP, modal share of PUB. On the contrary, while the factor loading of variables related to using private passenger vehicles and road infrastructure, modal share of PMM, and total road length shows it is in a negative relation with PC2. The tendency of common factors of PC2 is totally opposite to PC1, and indicates high dependence on public transportation because total length of railway, the share of public transportation, urban density, GRDP are high. It is regarded that the greatest point of factor loading related to total rail length implies that PC2 has a feature of the ‘‘index of railway dependence”. Prior studies on correlation between railway infrastructure and characteristics of travel behaviors include Winston and Langer (2006) indicating that congestion costs of PMM decrease in a city as rail transit mileage expands. Litman (2007) insisted that rail transit is a solution for traffic congestion and re-urbanization as an alternative mode of private vehicles from the aspects that high quality rail transit service can reduce travel time costs for people who shift modes. Therefore, this study defines PC1 as ‘‘Passenger car dependence” and PC2 as ‘‘Railway dependence” in consideration of the relation between PC1 and PC2. Model variables and hypothesis Dependent variable-possibility of commuting trip by PMM In this study, dependent variable Pij means the possibility of commuting trip being made by PMM (PCTP) and Pij = (0, 1). The data sets were drawn from PT data for individual trips made for commuting within the spatial ranges of the subject cities. Most of the trips carried out by an individual in the study area were done using a single mode. Explanation variables The selection of the variables employed in the current research basically has been informed by previous research on transportation energy consumption in the study area (Choi et al., 2011), which strongly reflects the extent of the dependence on private passenger vehicles in a city. Input variables at the city level for this study include urban density, GRDP, number of private vehicles, and status of the transport infrastructure, which are considered as primary factors affecting transport behaviors within urban areas. Independent variables consist of level 1 (individual trip level) and level 2 (urban level). The level 1 database was constructed with the data sets drawn from the Person Trip Survey done by each country (n = 652,725). Variables at level 2 were drawn from data sets based on the urban level, which are published by the official statistic institutes of each country (n = 78). The spatial range of the data sets for levels 1 and 2 were managed to coincide to ensure the consistency. Table 5 below shows the variables in the hierarchical structure; individual trips (level 1) and urban (level 2) for this research. Hypothesis In this research, the city level and individual attribute index which affect the city’s PCTP were selected based upon the views of existing studies related to traffic characteristics. Also, a hypothesis on the influence of the selected cityindividual identity was formed, and verification on the hypothesis was performed. The index setting was made using the correlation between city-individual attributes and traffic characteristics, which was basically obtained through the literature study as the background. Based on the variables, which mentioned in the second chapter, at a city level including urban density, GRDP, road density, metro length, tram length, car occupancy, and number of registered private vehicles were analyzed as important factors affecting trip behaviors in the urban area. These variables have also been the focus in recent urban planning studies on reducing traffic congestion and energy consumption in the transport sector. However, the results of these previous studies (Newman and Kenworthy, 1989; Kenworthy et al., 1999; He et al., 2005, etc.) in the second chapter had some limitations, such as spatial constraints due to their focus on local areas, or their failure to reflect the innate characteristics of data that have a hierarchical structure. Therefore, it is unclear whether the provided results on urban-social patterns could be able to predict homogeneously for similar cluster samples across the global contexts. It might be generally agreed that individual social features have effects on dependent variables. However, it is unclear through review whether the effects vary depending on city size, and which factors at the urban level cause the differences.

236

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243 Table 5 Description of variables used in analysis. Variable

Description and unit

Number of samples

Dependent variable  Probability of commuting trip by private motorized modes

652,725

City level characteristics (Level 2)  Urban densityb  GRDPb  Road densityb  Metro lengthb  Tram lengthb  Passenger carb  Car occupancya

78 78 78 78 78 78 78

      

Continuous Continuous Continuous Continuous Continuous Continuous Continuous

variable variable variable variable variable variable variable

(persons/ha) ($/person) (m/vehicle) (m/1000 persons) (m/1000 persons) (vehicle/1000 persons) (persons/vehicle)

Individual level characteristics (Level 1)  Gendera  1 if male, 0 female  Agea  Continuous variable (age) a  Continuous variable (person)  Family size a  Passenger car  1 if owns passenger car, 0 otherwise  Trip durationa  Continuous variable (Min/trip)

652,725 652,725 652,725 652,725 652,725

a KTDB: Korean Transport Database, MLIT: Ministry of Land, Infrastructure, Transport and Tourism, JICA: Japan International Cooperation Agency, FHWA: Federal Highway Administration U.S. Department of Transportation. b Korea: The Official Statistics Report of each city (2005), Japan: The Official Statistics Report of each city (2005), U.S.A: U.S. Department of Transportation, Federal Highway Administration. Highway Statistics 2001 and statistics in Bureau of Economic Analysis of U.S Department of Commerce, Developing countries: The reports ‘‘The study on master plan for urban transport in the each metropolitan area-(Cairo. Tripoli (2001); Chengdu, Kuala Lumpur (2000); Damascus. Managua (1998); Manila (1997); Bucharest (1999); Lima. Hanoi (2005); Ho-Chi-Minh (2003); Nairobi (2004))”.

For example, previously utilized regression methods assume that the samples are linearly independent, and this has been tested by OLS (Ordinary Least Squares). In general, samples from the same group display show similar characters compared to different groups. On that account, if the influence of PCTP is only focused on urban scale data, Ecological fallacy as well as information loss of data may occur. Also it will not meet the assumptions made on residual independence (Hox and Kreft, 1994). And if the analysis relies only on individual data, the urban level influences are ignored, leading to atomistic fallacy. As so traffic patterns are influenced not only by an individual’s character but also by group that it is affiliated to. And, if the averaged values within a group and averaged values between groups are confused, the variance of the individual samples may misrepresent the characters of the data set it is included to. To overcome this type of error, the current study includes both personal and urban level variables into the multilevel model to verify their influence. Individual level characteristics First of all, individual variables (level 1) including gender, age, family size, driving license ownership, passenger mode ownership, and trip duration were employed as independent variables for the model. In terms of individual capacity constraints, age and gender were considered as the major factors causing individual differences in mode choices and travel behaviors, as stated in Literature Review. The second set of individual factors referred to were coupling and authority constraint factors, including household size, driving license ownership and passenger car ownership. This is meaningful, as it provides evidence that the family size and availability of private motorized mode represented by variables such as passenger car ownership and license ownership are related to individual mode choice and travel behavior. Next, this study utilized trip duration (min/trip) as the representative indicator for travel behavior. Due to the inconsistency in the information on trip distance among the chosen cities, this study regards trip duration as a substitutable variable, which was commonly drawn from PT data of all the subject cities and was adopted into the model. City level characteristics Meanwhile, the fact that the level of economic growth and rises in personal income are greatly affecting the relationship between the factors in the urban area, and several studies have confirmed that economic growth has played a crucial role in driving a shift to personal vehicle-based modes of transport. Thus, this research adopted GRDP into the analysis model as an index representing the urban economic level. Regarding the influence of the traffic infrastructure on the urban traffic behavior characteristics, the increase of traffic density and establishment of railroad infrastructure seem to make a positive contribution to the demand reduction in road traffic and public transportation-centered downtown development. Therefore, the status of urban road extension and railroad infrastructure (e.g., Metro, Tram, etc.) establishment was adopted into the analysis model in this research. Finally, the increase of car occupancy by car sharing was adopted into the model by setting average urban car occupancy as the index of city level, considering factors affecting the use of passenger cars. In particular, in view of the fact that the

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

237

Table 6 Hypothesis in this research. Hypothesis

Description

 (H 1)

Effects of individual factors including gender, age, family size, passenger car ownership, diving license ownership, and trip duration on PCTP vary by city Effects of variables including gender, age, diving license ownership, and trip duration on PCTP vary by population density As population density increases, effects of license ownership on PCTP diminishes Effects of passenger vehicle ownership on PCTP vary by car occupancy and metro length at the city level As car occupancy of city increases, effects of passenger vehicle ownership on PCTP become higher As metro length of city becomes longer, effects of passenger vehicle ownership on PCTP become smaller Effects of license ownership and trip duration on PCTP vary by number of passenger vehicles in cities

 (H 2) – (H 2–1)  (H 3) – (H 3–1) – (H 3–2)  (H4)

average car occupancy in rush hour is lower than in other time zones, whether or not the urbanization will affect PCTP will be reviewed. Based on this background, this research employed a multilevel analysis method to observe the differences in the effects of individual variables on PCTP among the subject cities and to observe which urban factors caused the differences. Also, this research made the following hypothesis to check whether variances deviated depending on city features, and which features affect the deviance. The assumptions made for this research has been listed in Table 6. Analysis Unconditional model This research set up three models including the ‘‘Unconditional model,” the ‘‘random coefficient model (Case 1),” and the ‘‘random coefficients model (Case 2)” to analyze the effectiveness of variables at the individual and urban levels on PCTP. The Unconditional model has no predictors and has been included to get preliminary information on the extent to which variation in the outcome can be explained by each of the two levels of the hierarchy. The unconditional two-level model can be expressed as follows (Raudenbush and Bryk, 2002). Here, ICC (intra-class correlation) provides an assessment of the extent to which there is variability in responses at the group level. In general, the variance of the outcome in standard multilevel models consists of two components: the variance of uj (s00) and the variance of rij (r2). The r2 parameter captures variability within groups and s00 captures variability between groups. With these two variances, the intra-class correlation coefficient for standard multilevel models is calculated using Eq. (4) below to measure the proportion of the variance in the outcome between the level 2 units.

ICC ¼ s00 =ðs00 þ r2 Þ

ð4Þ

If ICC is sufficiently close to zero, then there is effectively no variation in the subjects between the level 2 units, suggesting that standard subject level models are adequate for these data. Here, when the logistic model is applied into multilevel analysis, the level 1 residuals are assumed to follow the standard logistic distribution, which has a mean of 0 and a variance (r2) of p2 =3. This variance represents the within-group variance for ICC calculations for dichotomous data (Raudenbush and Bryk, 2002). Table 7 shows the results of an estimation on Unconditional models. The purpose of the unconditional models is to identify ICC and the difference in impact between 2-level variables, as well as the first status of deviance. According to the results of unconditional model analysis, variance of intercept including u0j is 0.0328, which is statistically significant. It means that the variance of city level accounted for 3.28% of total variance and PCTP among cities varies at some level. Further, it indicates that the characteristics of groups have effects on dependent variables, and this result would have not been attained with the simple regression analysis. Random coefficient model – Case1 Random coefficient model (Case 1) is formulated by plugging variables at the individual level into the unconditional model to check the effectiveness of individual trip factors. Also, the model considered the random effects of explanatory

Table 7 ICC of unconditional model. Variance Level 1 (s00) Level 2 (r2) ICC

0.112 3.290 0.0328

238

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

variables at the individual level to check whether the variance of explanatory variables at individual level varied by city characteristics. The left-hand side of Table 8 presents the results of the multilevel model (Case 1), and all of the explanatory variables at the individual level were significant at p-values less than 1%. The results show that social characters at the individual level affected PCTP at significant levels for the subject cities. In other words, each city showed integrated variances, and therefore, the effectiveness of explanatory variables at the individual level on PCTP varies by city and satisfies ‘‘H 1.” For the explanatory variables at the individual level, coefficients of gender dummy, license ownership, and trip duration are positively related to PCTP, while coefficients of age, family size, and passenger car ownership are negatively related to PCTP. The odds ratio in Table 8 is an indicator to estimate the possibility of a commuting trip being made by one of the PMM, and is used to interpret the actual effects of estimated coefficients. The odds ratios of estimated coefficients indicates how the odds of an event are affected by the presence of an intersection characteristics. In this research, the odds ratio of gender is equal to 1.401, and male subjects commute by PMM more than female subjects by 1.4 times. In particular, the odds ratio of license ownership is equal to 8.2, which means that the PCTP of license owners is 8.2 times higher than those without a license, and the effectiveness of this variable is dominant compared to other variables at the individual level (c50 = 2.138). This result could be evidence that modal share of commuting trip varies by individual car availability and gender, which

Table 8 Estimation results. Fixed effect

Random coefficient model-Case 1

Random coefficient model-Case 2

Coefficient

Standard error

Odds Ratio

INTRCPT1, b0 INTRCPT2, c00 Density, c01 GRDP, c02 Road density, c03 Metro length, c04 Tram length, c05

2.09717 – – – – –

0.101129 – – – – –

– – – – –

GENDER slope, b1 INTRCPT2, c10 Density, c11

–

0.032729 –

–

AGE slope, b2 INTRCPT2, c20 Density, c21

0.01137 –

0.00201 –

–

Family person slope, b3 INTRCPT2, c30

0.337457

pvalue

Coefficient

Standard error

Odds Ratio

pvalue

0.123

<0.001 – – – – –

2.10044 0.01279 0.00001 0.01354 0.01181 0.00410

0.07432 0.00198 0.00000 0.00705 0.00399 0.00142

0.122 1.013 1.000 1.014 1.012 1.004

<0.001 <0.001 0.014 0.059 0.004 0.005

1.401

<0.001 –

0.34615 0.00269

0.03155 0.00132

1.414 1.003

<0.001 0.045

0.989

<0.001 –

0.01122 0.00016

0.00193 0.00004

0.989 1.000

<0.001 <0.001

0.943

<0.001

0.05416

0.00870

0.947

<0.001

0.513

<0.001 – –

0.22579 0.29158 0.00884

0.05320 0.10997 0.00371

0.798 1.339 1.009

<0.001 0.010 0.020

8.202

<0.001 – – –

2.13802 0.00479 0.00455 0.03370

0.09889 0.00140 0.00044 0.00713

8.483 0.995 1.005 1.034

<0.001 0.001 <0.001 <0.001

1.013

<0.001 – – –

0.01330 0.00007 0.00002 0.00009

0.00079 0.00002 0.00000 0.00001

1.013 1.000 1.000 1.000

<0.001 <0.001 <0.001 <0.001

Standard deviation

Variance component

v2

pvalue

0.05835

0.00848

Passenger car ownership slope, b4 INTRCPT2, c40 0.66721 Car occupancy, c41 – Metro length, c42 –

0.07023 – –

– –

License ownership slope, b5 INTRCPT2, c50 Density, c51 Registration car, c52 Road density, c53

– – –

0.16380 – – –

– – –

Trip duration slope, b6 INTRCPT2, c60 Density, c61 Registration car, c62 Tram length, c63

– – –

0.00120 – – –

– – –

Standard deviation

Variance component

v2

pvalue

0.892 0.248 0.017 0.058 0.545

0.7959 0.0617 0.0003 0.0033 0.2970

117210.3 1198.1 3087.9 526.4 2971.8

<0.001 <0.001 <0.001 <0.001 <0.001

0.917 0.184 0.050 0.186 0.330

0.8410 0.0340 0.0025 0.0344 0.1087

10334.3 148.7 8031.3 1860.1 267.7

<0.001 <0.001 <0.001 <0.001 <0.001

1.376 0.010

1.8931 0.0001

24392.9 7266.9

<0.001 <0.001

0.498 0.013

0.2483 0.0002

188.4 1466.7

<0.001 <0.001

Random effect INTRCPT1, u0 Gender slope, u1 AGE slope, u2 Family person slope, u3 Private car dummy slope, u4 License dummy slope, u5 Trip duration slope, u6 2⁄ log-likelihood at convergence

2.10439

0.01339

1,841,858

Deviance in unconditional model: 1,852,386.

1,839,496

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

239

is opposed to the result of Stradling et al. (2005). In contrast to the effectiveness of license ownership, the odds ratio of passenger car ownership is equal to 0.513, which is lower compared to other variables, and the coefficient was shown to be negative. This means that license ownership as an authority of vehicle use could be an inducement for commuting trips by vehicles, while the influence of passenger car ownership to mode choice in commuting trips may be small. Therefore, it could be necessary to separately analyze the effects of vehicle authority and passenger car ownership on travel behavior for commuting trips. In terms of the relationship between PCTP and age and family size as representative variables for capability and coupling constraint, subjects with older age and a larger family showed lower PCTP. Even though the effectiveness by age range spectrum was not available in this study, it reflects the finding of a previous study (Schmöcker et al., 2005) that trip distance made by vehicle and age have an inverse correlation. However, the effectiveness of age on PCTP is less significant compared to other variables (odds ratio = 0.989). However, family size showed negative effects on PCTP. On the other hand, effectiveness of family size showed relatively small effects (odds ratio = 0.943) similar to the age effect. It is also confirmed that effectiveness of trip duration on PCTP varies by city. The results basically showed that propensity of PCTP with respect to commute distance are positive. However, travel behaviors represented by trip length or trip duration are sensitively affected by urban structure, land use features, and status of the transport infrastructure (Holz-Rau, 1997; Giuliano and Dhiraj, 2003; Givoni and Rietveld, 2007). Therefore, an additional analysis on the subordinate relationship among the variables at the city level for checking the effects of trip duration on commuting trip by PMM is required. More detailed analyses on these aspects are presented in Case 2. Random coefficient model – Case2 To check the variation in the effectiveness of individual explanatory variables on PCTP by urban character at level 2, the case 2 model was constructed by substituting independent variables of the Case 1 model at level 1 with level 2 variables, which represent urban characters. The right-hand side of Table 8 presents that all fixed parts of the model in individual trip level including intercept c00 are statistically significant at p-values less than 1%. This means that independent variables at the individual level showed integrated variance for groups in the Case 2 model, which considered city effects. In other words, it showed various urban averages. Considering random effects, random coefficient effects of all the explanatory variables including spatial structure and status of transport infrastructure at city level were significant at p-values less than 1%. Also, for significance of the input variables at the city level, all fixed parts of the model in the city level, except road density (p = 0.059) were significant at p-values less than 5%. This means that there exist deviances for relevancy of each factor at the individual level with PCTP depending on the size of substituted variables at the city level, and it satisfies H2, H3, and H4. In other words, this result provides evidence that accepted claims of a relationship between urban social factors and travel behaviors could be rejected if city level characteristics are considered through a multilevel analysis. In more detail, it is not only confirmed that the averages of the effectiveness of gender, age, and license ownership on PCTP vary by plugging population density into the model, but also that the effects of individual factors on PCTP become larger by considering variables at the city level when compared to the results of Case 1, such as for gender dummy (odds ratio: 1.401 ? 1.414), Passenger car ownership (odds ratio: 0.513 ? 0.798), and license ownership (odds ratio: 8.202 ? 8.483). This means that the variance of PCTP, which would not have been explained with individual explanatory variables, could be explained by considering population density at the city level. The result satisfies the hypothesis that the character of young adults who are supposed to engage in more social activities has a positive impact on PCTP, and the pre-conception that the vehicle is a male-dominant mode. At the same time, it provides the interesting interpretation that these trends vary by urban population density. Also, the hypothesis that subjects with a driver’s license would have lower PCTP in large cities with high population density (H 2-1) was satisfied. With population density, the effectiveness of license ownership on PCTP was negative, which is the opposite result to cities with lower population density. According to Bussiere et al. (1996), the progress of motorization is positively associated with the distance from the city center, and car ownership increases as the distance from the city center increases. Also, public transport is more developed in the city center, and competes with the automobile. Conversely, where access to public transport is more difficult, particularly in rural areas, the share of households with at least one car is higher (Roux et al., 2010). Therefore, in terms of urban structure, travel behaviors show a high dependence on PMM in cities with low population density, and the dependence on motorized modes for people with a driver’s license varies by population density. This result is consistent with the results of previous research. Regarding the pattern of PCTP depending on car occupancy at the urban level, in cities with a higher car sharing rate, passenger car owners showed higher PCTP than ones without car ownership, by approximately 1.3 times (odds ratio = 1.399). This result satisfies H3-1 and shows the travel behavior of passenger car owners in cities with a higher car sharing level. According to a report by KOTI (2008), the rate of single occupancy vehicles is relatively high during the commuting peak hours while the rate of high occupancy vehicles is high during non-peak hours for shopping and leisure activities. This result shows that trip patterns vary by time of day. In general, the level of utilization of car sharing programs implemented to mitigate traffic congestion is related to the rate of parking space, the status of the public transport system, the infrastructure for green transport, and the population density

240

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

of the city (KOTI report, 2008). In relation to this aspect, this research categorized urban types based on status of railway development to check the effectiveness of the urban environment in 78 cities around the world utilizing an urban transport database, and reviewed the relationship between population density and average car occupancy for each urban type (Fig. 1). By reviewing Fig. 1, it is confirmed that the cities without a metro or tram infrastructure and with a relatively low population showed higher car occupancy. On the other hand, cities with a railway infrastructure in high-density urban areas showed relatively high rates of single car occupancy. This pattern explains the urban transport character for places where the necessity of vehicle ownership is relatively high considering the urban spatial structure and status of public transport system development. And also, this could be a basis for H-3. Regarding the effectiveness of Metro length and private car ownership on PCTP at the city level, the effectiveness varies depending on the city (p < 0.001), and the effectiveness of private car owners on PCTP increases with a longer metro length. This result contradicts previous research results (Litman, 2007; Choi et al., 2013b), which hold that the effectiveness of metro length on PCTP raised by private car owners would constrain railway infrastructure on private car dependency and also reject H 3-2. However, effectiveness of urban environment by metro length is relatively weak compared to factors at the individual level considering urban characters (c42 = 0.0088, odds ratio = 1.009). Regarding the effectiveness of license ownership and trip duration on PCTP depending on the number of registered vehicles and the road infrastructure density at the city level, averages of all the variables vary by city, with the significant levels at 1%. In particular, for cities with a longer road network, the relatively high effectiveness of license owners on PCTP is remarkable. From this, it can be interpreted that the possibility of private vehicle use might increase with respect to there being an advantageous transport environment for vehicle use. However, the effectiveness of city level by road length is smaller than the effectiveness of the individual level (odds ratio = 1.034). While it is confirmed that the effectiveness of variables at the individual level on PCTP vary by the size of the urban spatial structure and the number of passenger vehicles or tram length (c63, p < 0.001), effectiveness of the urban environment is very weak, or close to 0. Examination of the random effects suggests that there are significant variations in the PCTP occurring at a city level (i.e. all of the coefficients of random effects are significantly different from zero). It should be noted that the city level fixed effects for predicting both Case 1 and Case 2 capture a significant portion of the variation across cities, as reflected by random effect coefficients close to zero. Average of PCTP varies by city depending on the status of transport infrastructure development and the urban spatial structure, including population density, car occupancy, urban spatial structure, and length of railway development, among the variables related to the urban environment. Also, it is confirmed that the possibility of a relationship between individual characters during commuting hours and modal choice varies depending on city level characteristics. In more detail, the effectiveness of variables at the city level showed a weak relevance compared to the effectiveness of variables at the personal and household level. However, the results explicitly explain that similar variables at the personal level could affect travel behavior in different ways, depending on the variables in the urban environment. In particular, this result implies the appropriateness of adopting analytical models for predicting PCTP that consider fixed and random effects both at the individual and the urban levels. Multilevel analysis is based on the relationship between variables of different levels, indicating whether or not an individual’s character influences an urban variable. However, this method extends to only the existence of a relationship without

Fig. 1. Relationship between urban density and car occupancy by urban types.

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

241

Fig. 2. Passenger car and railway dependence of the target major city areas.

specifying the variables and how it impacts the variance. Therefore in this research 78 metropolitan areas were segmented through principle component analysis (PCA) to explain the importance and impact of individual and urban variables on PCTP. Urban scale variables used in the multilevel analysis, such as population density, modal share of PMM, train and road length, and car occupancy were some of the main variables used for the analysis. Fig. 2 shows the relationship between the factor scores deduced by PCA. Results show the characteristics of a given city based on the given variables. PC1, indicate high road length, modal share of PMM, car occupancy, and GRDP with low urban density, representing major city areas commonly found in the United States. On the other hand, PC2 represents cities commonly found in Japan where the modal share of PUB, urban density is high and total railway lengths are long. For developing countries show highly populated cites with short road length and no railways. For Korea two cases were shown with cities with and without long railways with a constant high levels of population densities. Based on the results found above, the relationship between urban density and total railway lengths may be deducted. In other words, PMM usage and PCTP may be characterized based on the population density and railway length. In other words PMM usage and PCTP values vary based on a countries urban characteristics and the policy behind it. This method may help reinforce the finding of multilevel analysis where it prevents the averaging out of the impact of individual variables. Also it can be deducted that individual variables can be used to categorize cities of various characters. Conclusions This research was performed by collecting datasets showing individual trip behaviors from 78 large cities around the world, and estimating factors of individual and urban characteristics that affect the PCTP through a multi-level analysis that enabled the consideration of hierarchical datasets simultaneously. Specifically, PCTP were quantitatively estimated by dividing variances into changes caused by individual decision makers and changes caused by city characteristics such as the urban space and transport infrastructure. If you look at the results you may see that the relationship between urban characteristics and PCTP have a statistical significance. This means that policies that rely on reducing PCTP and encouraging public transportation may differ in its results. In other words it is important to plan policies based on the urban characteristics. For example, in developing countries the PCTP is low for cities with high population density, however in these conditions by expanding the roads instead of building railway infrastructure will increase the PCTP. While the spatial scope of this research is for cities around the world, the transport data sets were drawn from representative cities of each country. Therefore, the expectation is that the samples have relatively similar characteristics, which might have lowered the intra-correlation between cities. For the aspect of urban characteristics such as various urban spatial structures, economic status, and transport infrastructure development status and context effects of data sets were considered, and various interpretations were possible for the factors that affect the individual trip behaviors. To review the urban effects, this research utilized the macro urban space defined for administrative purposes.

242

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

Therefore, care should be taken in attempting to draw homogenous conclusions regarding the effects of urban characteristics on PCTP among the cities. The reason is that the effects of individual characteristics on trip behaviors might vary by the characteristics of sub-spaces that are categorized into micro levels. In addition to the spatial scope, time factors should be considered for a more precise analysis on trip patterns depending on the purpose, since trip patterns vary by time. Acknowledgements Providing individual travel data from Korean Transport Database (KTDB), Ministry of Land, Infrastructure, Transport and Tourism (MLIT), Japan International Cooperation Agency (JICA), Federal Highway Administration U.S. Department of Transportation (FHWA). References Baum-Snow, N., Kahn, M.E., 2005. The Effects of Urban Rail Transit Expansions: Evidence from Sixteen Cities, 1970 to 2000. Brookings Papers on Urban Affairs. Brookings Institute. Bryk, A.S., Raudenbush, S.W., 1992. Hierarchical linear models. Sage, Beverly Hills, CA. Bussiere, Y., Armoogum, J., Madre, J.L., 1996. Vers la saturation? Une approche demographique de l’equipement des menages en automobile dans trois regions urbaines. Rev. Popul. l’INED 50 (4/5), 955–977. Choi, H., Nakagawa, D., Matsunaka, R., Oba, T., 2011. Building a database on transportation energy consumption in cities of the world with information related to travel behavior. J. Int. City Plan. 2011, 263–272. Choi, H., Nakagawa, D., Matsunaka, R., Oba, T., Yoon, J., 2013a. Research on the causal relationship between urban structure and transport energy consumption by economic level. Int. J. Urban Sci. 17 (3), 362–384. Choi, H., Nakagawa, D., Matsunaka, R., Oba, T., Yoon, J., 2013b. Estimation on the efficiency of transportation energy consumption from the viewpoint on urban transportation infrastructure and characteristics of travel behavior. Int. J. Railway 6 (2), 33–44. Chu, X., 1994. The Effects of Age on the Driving Habits of the Elderly: Evidence from 1990 National Personal Transportation Studies. Research and Special Programs Advisory Report No. DOT-T-95-12. US Department of Transportation. Crane, R., Crepeau, R., 1998. Does neighborhood design influence travel? A behavioral analysis of travel diary and GIS data. Transp. Res. Part D 3 (4), 225– 238. Dargay, J., Gately, D., 1999. Income’s effect on car and vehicle ownership worldwide: 1960–2015. Transp. Res. Part A 33 (2), 101–138. Elder, G.H., 1985. Perspectives on the life course. In: Elder, G.H., Jr. (Ed.), Life Course Dynamics. Cornell University Press, Inthaca, pp. 23–49. Ewing, R., Cervero, R., 2001. Travel and the built environment: a synthesis. Transp. Res. Rec. 1780, 87–144. George, L.K., 1996. Missing links: he case for a social psychology of the life course. Gerontologist 36 (2), 248–255. Giuliano, G., Dargay, J., 2006. Car ownership, travel and land use: a comparison of the US and Great Britain. Transp. Res. Part A 40 (2), 106–124. Giuliano, G., Dhiraj, N., 2003. Another look at travel patterns and urban form: the US and Great Britain. Urban Stud. 40 (11), 2295–2312. Givoni, M., Rietveld, P., 2007. The access journey to the railway station and its role in passenger’s satisfaction with rail travel. Transp. Policy 14 (5), 357– 365. Guo, G., Zhao, H., 2000. Multilevel modeling for binary data. Annu. Rev. Sociol. 26, 441–462. He, K., Huo, H., Zhang, Q., He, D., An, F., Wang, M., Walsh, M.P., 2005. Oil consumption and CO2 emissions in China’s road transport: current status, future trends, and policy implications. Energy Policy 33 (12), 1499–1507. Holz-Rau, C., 1997. ‘‘Siedlungsstrukturen und Verkehr”. ‘‘Materialien zur Raumentwicklung”, Heft 84, Bonn. Hox, J.J., Kreft, I.G.G., 1994. Multilevel analysis methods. Sociol. Methods Res. 22 (3), 283–299. Jang, W., Park, J., Kim, D., 2008. A Study on Introducing Car-Sharing Schemes. Annual Report 2008. The Korea Transport Institute (KOTI). Kaiser, H.F., 1974. An index of factorial simplicity. Psychometrika 39, 31–36. Kenworthy, J.R., Laube, F.B., 1999. Patterns of automobile dependence in cities: an international overview of key physical and economic dimensions with some implications for urban policy. Transp. Res. Part A 33 (7), 691–723. Kenworthy, J.R., Laube, F.B., Newman, P.W.G., Bater, P.A., Raad, T., Poboon, C., Guia, B., 1999. An International Sourcebook of Automobile Dependence in Cities 1960–1990. University Press of Colorado, Boulder, CO. Kreft, I., Leeuw, J., 1998. Introducing Multilevel Modeling. Sage Publications Ltd.. Litman, T., 2007. Evaluating rail transit benefits: a comment. Transp. Policy 14 (1), 94–97. Mackett, R., Sutcliffe, E.B., 2003. New urban rail systems: a policy-based technique to make them more successful. J. Transp. Geogr. 11 (2), 151–164. McCann, B., 2000. Driven to Spend: The Impact of Sprawl on Household Transportation Expenses. Surface Transportation Policy Project, Washington, D.C. Morimoto, A., Koike, H., 1995. A comparison of the urban structure impact upon transportation energy. J. City Plan. Inst. Jpn. 30, 685–690. Newman, P.W.G., Kenworthy, J.R., 1989. Gasoline consumption and cities. A comparison of US cities with a global survey. J. Am. Plan. Assoc. 55 (1), 24– 37. Pickup, L., 1985. Women’s gender-roll and its influence on travel behavior. Built Environ. 10 (1), 61–68. Pucher, J., Korattyswaropam, N., Mittal, N., Ittyerah, N., 2005. Urban transport crisis in India. Transp. Policy 12 (3), 185–198. Raudenbush, S.W., Bryk, A.S., 2002. Partial proportional odds models for ordinal response variables. Appl. Stat. 39 (2), 205–217. Rosenbloom, S., 1995. Travel by the elderly. In: 1990 Nationwide Personal Transportation Survey. Demographic Special Reports. U.S. Department of Transportation. Roux, S., Armoogum, J., Madre, J., 2010. Dynamic of car ownership and car use in France since the 1960s. In: 12th World Conference on Transport Research, Lisbonne, Portugal. Schafer, A., Victor, D., 1999. Global passenger travel: implications for carbon dioxide emissions. Energy 24 (8), 657–679. Schmöcker, J.D., Quddus, M., Noland, R.B., Bell, M.G.H., 2005. Estimating trip generation of elderly and disabled people: analysis of London data. Transp. Res. Rec.: J. Transp. Res. Board 1924, 9–18. Stradling, S.G., Carreno, M., Ferguson, N., Rye, T., Halden, D., Davidson, P., Anable, J., Hope, S., Alder, B., Ryley, T., Wigan, M., 2005. Scottish Household Survey Analytical Topic Report: Accessibility and Transport. Scottish Executive, Edinburgh. Sun, X., Wilmot, C.G., Kasturi, T., 1998. Household travel, household characteristics, and land use: an empirical study from the 1994 Portland activity-based travel survey. Transp. Res. Rec. 1617, 10–17. Suryo, R., Fan, C., Weiler, S., 2007. Commuting choices and congestion taxes in industrializing Indonesia. Soc. Sci. J. 44 (2), 253–273. Suzuki, T., Muromachi, Y., 2009. Empirical analysis on the possible impact of railroad development on population density and car use. Pap. City Plan. 44 (3), 73–78 (in Japanese). Suzuki, T., Muromachi, Y., 2011. A multilevel analysis on the effect of urban environmental and personal factors on car use. J. City Plan. Inst. Jpn. 1, 13–18 (in Japanese).

H. Choi, Y. Ahn / Transportation Research Part D 41 (2015) 228–243

243

Vance, C., Iovanna, R., 2007. Gender and the automobile: an analysis of non-work service trips. In: Paper Presented in the 86th Annual Meeting of the Transportation Research Board, January 2007, Washington, D.C. Wang, R., 2011. Shaping carpool policies under rapid motorization: the case of Chinese cities. Transp. Policy 18 (4), 631–635. Winston, C., Langer, A., 2006. Effect of government highway spending on road users’ congestion costs. J. Urban Econ. 60 (3), 463–483. Yannis, G., Papadimutriou, E., Antoniou, C., 2008. Impact of enforcement on traffic accidents and fatalities: a multivariate multilevel analysis. Saf. Sci. 46 (5), 738–750.