You are on page 1of 18

Transportation (2012) 39:705–722

DOI 10.1007/s11116-011-9368-3

An analysis of Metro ridership at the station-to-station


level in Seoul

Jinkyung Choi • Yong Jae Lee • Taewan Kim • Keemin Sohn

Published online: 25 August 2011


Ó Springer Science+Business Media, LLC. 2011

Abstract While most aggregate studies of transit ridership are conducted at either the
stop or the route level, the present study focused on factors affecting Metro ridership in the
Seoul metropolitan area at the station-to-station level. The station-to-station analysis made
it possible to distinguish the effect of origin factors on Metro ridership from that of
destination factors and to cut down the errors caused by the aggregation of travel
impedance-related variables. After adopting two types of direct-demand patronage fore-
casting models, the multiplicative model and the Poisson regression model, the former was
found to be superior to the latter because it clearly identified the negative influences of
competing modes on Metro ridership. Such results are rarely found with aggregate level
analyses. Moreover, the importance of built environment in explaining Metro demand was
confirmed by separating built environment variables for origin and destination stations and
by differentiating ridership by the time of day. For morning peak hours, the population-
related variables of the origin stations played a key role in accounting for Metro ridership,
while employment-related variables prevailed in destination stations. In evening peak
hours, both employment- and population-related variables were significant in accounting
for the Metro ridership at the destination station. This showed that a significant number of
people in the Seoul metropolitan area appear to take various non-home-based trips after
work, which is consistent with the results from direct household travel surveys.

Keywords Station-to-station level  Metro ridership  Built environment 


Direct demand model  Transit oriented development (TOD)

J. Choi  Y. J. Lee  T. Kim  K. Sohn (&)


Department of Urban Engineering, Chung-Ang University, Seoul, Korea
e-mail: kmsohn@cau.ac.kr
J. Choi
e-mail: chjiky83@naver.com
Y. J. Lee
e-mail: yjlee@cau.ac.kr
T. Kim
e-mail: twkim@cau.ac.kr

123
706 Transportation (2012) 39:705–722

Introduction

The role of mass transit in assuring the mobility of socially segregated groups, and ulti-
mately accomplishing a sustainable environment by reducing car use, has been emphasized
for the past few decades. Transit-oriented development (TOD) has also been spotlighted in
the same context. The three principles of TOD, Density, Diversity and Design, which were
suggested by Cervero and Kockelman (1997), are known to considerably increase transit
ridership. Recently, the importance of connectivity of transit lines has also been empha-
sized in addition to the existing three TOD principles (Ewing et al. 2008; Hamin and
Gurran 2009). For policymakers, understanding how various factors affect transit ridership
is crucial to the development of robust policies that will match the goals of sustainable
transport. Most researchers use empirical models to conduct studies on the relationship
between various factors from the TOD principles and transit ridership, and regression-type
macro models have been widely adopted at the stop level to determine this relationship
(Sohn and Shim 2010; Estupiñán and Rodrı́guez 2008; Yao 2007; Cervero 2006; Boyle
2006; Kuby et al. 2004; Chu 2004; Peng et al. 1997; Parsons Brinckerhoff 1996).
In a recent study, Sohn and Shim (2010) conducted a station-level investigation into the
effect of the built environment, inter-modal connection, and external connectivity on daily
Metro ridership by using data from the Seoul metropolitan area. Based on the results of a
regression analysis, they found a close interrelationship between Metro boardings and both
density-related variables and inter-modal connection variables, but found no external
connectivity contribution to Metro ridership. The external connectivity accounts for the
extent to which transit or car routes are efficient in connecting urban centers. They
managed to verify an indirect linkage between Metro ridership and connectivity of both the
Metro itself and competing modes by employing a structural equation model (SEM). This
weak linkage is attributable to the possibility that the spatial variation of connectivity is so
large that centrality-like variables could not properly reflect the original effect of con-
nectivity. Moreover, the effect of the population-related variables of the station area on
Metro ridership was not found in the station-level regression analysis, while the contri-
bution of an employment variable to Metro ridership was decisive. This imbalance can be
fixed by separating the impact of the built environment into origin and destination stations
and also by differentiating the daily average ridership by the time of day.
In another aggregate-level approach, Peng et al. (1997) dealt with route-level transit
ridership as a target variable. They confirmed the impact of the service travel time of the
subject transit line along with those of competing transit lines based on a simultaneous-
equations model. However, this approach has the limitation of not accounting for the
impact of the service level of competing modes. In addition, they recommended the need
for stop-level analysis in future research, since they had difficulty delineating the catch-
ment area of transit lines.
The most disaggregate approach would employ a discrete choice model. Many
researchers have attempted to examine factors affecting transit use at the level of the
individual traveler (Sohn and Yun 2009; Cao et al. 2009; Kim et al. 2007; Khattak and
Rodriguez 2005). The disaggregate approach has great flexibility in the choice of influ-
ential variables. A wide range of variables such as built environment, travel impedance,
individual characteristics, and even psychometric propensities of travelers has been
mobilized in the analysis dealing with various travel choices such as destination, modes,
routes, times, etc. Of course, the individual-level approach, understandably, ensures that all
the difficulties encountered when spatially aggregating independent variables in the
aggregate-level analysis can be sorted out. However, in order to apply the outcomes of the

123
Transportation (2012) 39:705–722 707

discrete choice model to a practical context, considerable effort is required to aggregate the
estimated results. Furthermore, the more individual-based the approach, the more time-
consuming and labor-intensive it becomes.
Considering the strengths and drawbacks of the three levels of transit ridership analysis
(i.e. stop, route, and individual levels), an intermediate level of spatial aggregation of
Metro ridership, the station-to-station level, was adopted for the present study. We
expected that the complications that Sohn and Shim (2010) confronted while examining
the factors at the station level would be resolved by adopting a station-to-station-level
model. That is, it is possible to directly investigate the effect of supply variables of
competing modes, and at the same time to allow a separate accounting for built-envi-
ronments of origin and destination stations. On the other hand, the station-to-station level
model can be understood as a generalized spatial interaction model that has a structural
similarity with the multiplicative model that will be used in the present study. In this
context, Metro travel time definitely is acting as the gravity model denominator, while
some of the other independent variables definitely capture the gravity model numerators of
opportunity (Golledge and Stimson 1997; Hannon et al. 1991; Ivy 1995).
Many researchers have dealt with the station-to-station demand with respect to inter-
urban rail transport (Wardman 1997; Jones and Nichols 1983), but few studies have
focused on the intra-urban transit demand at the station-to-station level. A reason for this
disparity might be because data regarding Metro ridership have been very difficult to
collect at the station-to-station level prior to the introduction of automatic fare collection
systems in urban transit. In one similar case, Thompson (1997) utilized a direct demand
model to address factors associated with transit demand at the zone-to-zone level and
identified the separate effect of production and attraction variables. His study actually was
the inspiration for the present investigation, and his findings provided many insights in
setting up the present analysis models.
The present study is a sequel to the previous investigation (Sohn and Shim 2010), and
thus the results from the two studies will be compared throughout the subsequent sections.
Factors that were expected to affect Metro ridership will be presented in the next section.
Then, a brief introduction of models to be used will be followed by a detailed explanation
of data collection for both dependent and independent variables. The model estimation
results and the corresponding discussions will then be addressed. Finally, conclusions will
be drawn and recommendations to increase Metro ridership will be offered.

Expected factors influencing station-to-station ridership

The previous work by Sohn and Shim (2010), on which the present study was based, dealt
with average weekday boardings at each Metro station as the dependent variable. Daily
boardings have an intrinsic shortcoming in that they show no indicator for the trip purpose.
Actually, there is no way of recognizing the trip purpose of data from the automatic fare
collection system. The best way to approximately estimate the trip purpose is to separate
daily ridership by the time of day. Peng et al. (1997) analyzed transit ridership for five time
periods (morning peak, midday, afternoon peak, evening and night). However, the present
study encompassed only the former three time periods, based on results from the household
travel survey that home-based work trips compose most ridership in the morning peak
hours, most trips returning home are included in the ridership in the evening peak hours,

123
708 Transportation (2012) 39:705–722

and various home- and non-home-based mid-day trips are mixed (Seoul Metropolitan
Government 2008). The morning peak hours were set as 07:00–09:00, the midday as
12:00–14:00, and the evening peak hours as 18:00–20:00.
Factors that were expected to affect the station-to-station ridership were chosen based
on results from previous studies (Estupiñán and Rodrı́guez 2008; Yao 2007; Cervero 2006;
Boyle 2006; Kuby et al. 2004; Chu 2004; Parsons Brinckerhoff 1996). The chosen factors
were categorized into four areas: built-environment variables, travel impedance variables,
intermodal connection variables, and other variables.

Built-environment variables

The population and employment within walking distance of each station were first chosen
as variables that influence Metro ridership. The net floor area of buildings within walking
distance was then hypothesized to be a complementary variable. Floor area was catego-
rized into four types: residential, office, commercial, and other uses. It was crucial to
determine the walking distance to Metro stations. A considerable number of studies
(Canepa 2007; Murray et al. 1998; O’Sullivan and Morral 1996) have evaluated the rel-
evant walking distance of transit riders and results ranged from 400 to 800 m, in either
Euclidean or network distance. Various walking distances were also investigated in the
Seoul Metropolitan area. However, for convenience a radius of 500 m has been widely
adopted in transportation studies (Seoul Metropolitan Government 2008). The area within
this walking distance from a station will hereafter be referred to as the pedestrian catch-
ment area (PCA). All of the built-environment-related variables within a PCA were ana-
lyzed with the assistance of a geographic information system (GIS). In addition to the
conventional built-environment variables above, the impact of a specific trip generator was
taken into account. There are 35 universities in the Seoul metropolitan area that have a
Metro station adjacent, or close, to them. A dummy variable (University dummy) to
accommodate the impact of universities was employed under the hypothesis that such
stations are related to additional boardings and alightings from students and employees
working in the vicinity of universities. However, ignoring the distance between a uni-
versity and a Metro station might have eroded the significance of the variable.
Many researchers have dealt with variables associated with land-use mix, connectivity
of walk and bike paths, street network density, etc. (Forsyth et al. 2009; Bhat and Guo
2007; Kim et al. 2007; Khattak and Rodriguez 2005). Walkability and land-use diversity
were adopted in the present study. The walkability indicator was calculated based on the
methodology suggested by Schlossberg and Brown (2004). Among six indicators they
proposed, only three were employed due to the availability of data, i.e., the quantity of
impedance paths (Impedance road), the pedestrian-friendly intersection density, and the
density of dead ends (Dead ends), respectively. In order to obtain the first indicator, an
assumption was made that pedestrian use is interrupted by heavily automobile-dominated
roadways (e.g., major arterials and freeways). The total length of such impeding streets
was recorded for each station’s PCA. In calculating the second indicator, the number of
intersections within each station’s PCA that are crossed by only pedestrian-friendly streets
was adopted as an independent variable. A pedestrian-friendly street was defined as a street
with a sidewalk on each side, having fewer than two car lanes in both directions with no
signal control. Of course, streets exclusively reserved for pedestrians belong to this cate-
gory, while heavily automobile-dominated streets do not. When a pedestrian-friendly street
reached a major arterial and did not extend further, the street was designated as a dead end.

123
Transportation (2012) 39:705–722 709

The number of dead ends was also counted for each station’s PCA. According to
Schlossberg and Brown (2004), the second indicator was hypothesized to positively affect
boardings, while the other two indicators were assumed to negatively influence boardings.
The walkability indices were not all found to be statistically significant in the previous
regression analysis at the station level (Sohn and Shim 2010). However, at the more precise
aggregation level (e.g. station-to-station level), they were expected to have statistical
significance.
The land-use mix diversity (Land-use mix) was quantified using a formula suggested by
Bhat and Guo (2007). They introduced Eq. 1 as a land-use composition measure to
compute a land-use diversity indicator (LI).
 r 1 o 1  c 1 ot 1
 þ L4 þ L4 þ L 4
LI ¼ 1  L 4 3=2
ð1Þ

where L = r ? o ? c ? ot, and r is the floor area in residential buildings, o is the floor
area in office buildings, c is the floor area in commercial buildings, and ot is the floor area
in other-use buildings. This indicator was expected to have a positive impact on Metro
ridership when daily trips were dealt with as a dependent variable. However, the index is
unlikely to be significant for the specific time period of day, because a trip purpose
governing the specific land-use in the vicinity of origin and destination stations might
prevail for each time period.

Travel impedance variables

Sohn and Shim (2010) adopted centrality indices to investigate the impact of travel
impedance on Metro boardings. This study also considered the impedance of competing
modes in the same form, but found no strong evidence for that. In the present study, the
average travel time from an origin station to a destination station was directly used as a
travel impedance variable. Car and bus travel times were used to examine how Metro
station ridership was affected by the travel impedance of competing modes. An efficient
connection of alternative modes might be negatively associated with Metro ridership.
Metro travel times were expected to have a negative effect and travel times of competing
modes were considered positive. Thompson (1997) adopted zone-to-zone travel times of
transit, but did not reflect the travel time of competing modes.
An additional variable was added to supplement the travel impedance. A well-known
reason for car-dependence is the avoidance of frequent, inconvenient transfers (Seoul
Metropolitan Government 2007). To judge the legitimacy of this reason, the present study
adopted the number of transfers across an itinerary with minimum travel time among all
feasible itineraries connecting an origin station and a destination station.

Intermodal connection variable

It is common for feeder buses to supply service between a Metro station and a surrounding
area (Chu 2004). In order to consider the impact of intermodal connections on Metro
ridership, many researchers adopted variables regarding the level of bus access (Kim et al.
2007; Kuby et al. 2004). The number of feeder bus lines stopping at Metro stations was
counted in the present study. The number of feeder lines was hypothesized to be positively
connected to station boardings and alightings, because discounted transfers between a bus

123
710 Transportation (2012) 39:705–722

and the Metro are allowed in the Seoul metropolitan area, which served as the sample area
for the present study.

Other variables

There are two city centers in the Seoul metropolitan area; the historic CBD is located in the
city’s geographic center, and the growing Kangnam district in the south-east (Sohn and
Kim 2010). Two dummy variables were created to investigate the effect of whether either
end station of a station-to-station trip would fall on one of these two city centers (CBD
dummy and Kangnam dummy).
A variable was included to identify the influence of net population density on Metro
boardings, which was synthesized for each station’s PCA by dividing the population by the
total residential floor area. This variable was assumed to more intensively affect the
ridership than the population variable based on the results found by Spillar and Rutherford
(1998). Another meaning of the variable is that it can be a proxy variable for indicating
socio-economic status in a station’s PCA. Wealthy people tend to consume more housing
and thus the variable can substitute for the average income level. That is, if more people
live in residential areas with smaller floor space, this means that the income level of the
corresponding area is lower.

Models for analysis

A direct demand model was adopted to address the association between the station-to-
station Metro ridership and the chosen variables. The direct demand model was widely used
in early transport demand studies (Kraft 1968; Lave 1972), and is known to overcome
drawbacks of the sequential approach (Ortúzar and Willumsen 2001). For example, the
direct model can sort out the problem of errors accumulating step-by-step in the sequential
approach. In the present study, two types of direct demand models were tested to account for
the station-to-station Metro ridership. The first model takes a multiplicative form. The
earlier form of the model was slightly modified, so that influences of origin and destination
can be differentiated. The model also can separately consider the impact of travel imped-
ances of different modes. These characteristics matched the main purpose of the present
study, which was to separate the impact of variables by origin and destination and by
different modes. Equation 2 represents a multiplicative model adopted in the present study.
Y
P
a
Y
P
b Y
M
hm
Tij ¼ / Xipp Xjpp tijm ð2Þ
p¼1 p¼1 m¼1

where Tij is the Metro travel from station i to station j, Xip the pth independent variable
with respect to the origin station i, Xjp the pth independent variable with respect to the
destination station j, tijm the travel impedance of mode m from station i to station j, / the
scale parameter, ap, bp, hm the model parameters to be estimated, P the number of inde-
pendent variables, and M the number of modes under consideration.
Equation 2 can be easily transformed into the linear form by taking the logarithm for
both sides. Another advantage of employing a multiplicative model is that each estimated
coefficient represents the constant elasticity of Metro ridership with respect to the change
of the corresponding variable, and, thus, the magnitude of the coefficients is directly
comparable.

123
Transportation (2012) 39:705–722 711

!
X
P X
P
Tij ¼ exp u þ ap Xip þ bp Xjp þ hm tijm ð3Þ
p¼1 p¼1

The other direct model adopted in the present study is the Poisson regression model, as
shown in Eq. 3, which will be referred to as Poisson model. The Poisson model is
appropriate for the case where a count variable with a minimum value of zero is chosen as
the dependent variable (Flowerdew 1991; Gurmu and Trivedo 1994). When analyzing
transit demand, Thompson (1997) employed the model to account for factors associated
with the level of zone-to-zone transit patronage. The Poisson model is also easily tractable
by taking the logarithm of both sides. The right-hand side converts to a simple linear
combination of independent variables. The Poisson model can be distinguished from the
multiplicative model in that it reflects the varying elasticity of the dependent variable
according to the level of independent variables.

Data collection

The data used in the present study were collected from various sources. Data for the
station-to-station ridership were obtained from a commercial agency that is committed to
collecting transit fares by the Seoul metropolitan government (Korea Smart Card 2010). In
the Seoul metropolitan area, transit fares are charged when travelers put their electronic
cards to the collection equipment as they begin, transfer, and end their trip. Thus, station-
to-station boardings can be counted exactly by the time of day. The weekday counts were
averaged separately for three time periods (morning peak, midday, and evening peak),
which were then set as dependent variables in the two models.
As shown in Fig. 1, there are 270 stations in the area (where each transfer station that has
multiple platforms is counted as a single station). However, because of data availability, we
omitted 19 stations that belong to line 9, which was opened recently, and thus the total number
of origin–destination pairs is 251 9 251 = 63,001. Although data for each station are
available, sampling was needed because it was very hard to collect car and bus travel times
from station to station. The data were obtained manually through a number of time-
consuming queries to two websites that provide real-time data for car and bus travel times.
The car travel time was collected from a commercial agency (ROADI 2010), and the bus
travel time from a public agency in charge of a bus information system (BIS) in the Seoul
metropolitan area (ALGOGA 2010). Using the actual travel time instead of estimating it
through network calculation enhanced the modeling accuracy. A sample of 2,000 (repre-
senting about 3.2% of total 63,001) was extracted randomly from the total of origin–
destination pairs. The descriptive statistics for the sample are summarized in Table 1.
The built-environment data were obtained from the Seoul Development Institute (SDI).
The floor area of buildings in a station’s catchment area was calculated using GIS tech-
nology, because the raw data are available on an individual building basis. After com-
pleting the previous study (Sohn and Shim 2010), the SDI discovered and corrected errors
in collecting the raw data for floor area. At the same time, the raw data for floor area were
partly updated by newly conducted surveys. This is why results of the previous study
erroneously showed that floor-area-related variables had little significant effect in the
regression analysis. Thus, the present study with the newly updated floor data was expected
to contribute to a more accurate analysis. The population and employment within a sta-
tion’s PCA were approximated, because the spatial unit of traffic analysis zone (TAZ) was

123
712 Transportation (2012) 39:705–722

Fig. 1 Map of Seoul Metro

not consistent with that of the station’s PCA. A simple rule of data fusion was employed
under the assumption that population and employment are uniformly distributed within
each urbanized block in a TAZ. The walkability-related data were also retrieved from the
SDI roadway database using GIS technology. Data regarding the variables related to
intermodal connectivity were obtained from the BIS of the Seoul metropolitan area
(ALGOGA 2010).

Empirical findings

The final specifications for both multiplicative and Poisson Models were derived after
taking empirical steps that were adopted by Sohn and Shim (2010). Starting from the initial
model with the full set of variables, the variables for the second model were drawn from
the initial model by selecting those with a p-value of less than 40%. For the third model,
variables from the second model that had p-values of less than 5% and variance inflation
factors (VIFs) of less than 6.5 were selected. Regarding multi-collinearity, a cut-off value
of 6.5 measured by VIF is relatively conservative since many textbooks propose 10, with a
rule of thumb, as a criterion for concern about multi-collinearity among independent
variables (Hair et al. 1987; Kline 1998). After testing as many plausible specifications as
possible by adding or deleting variables from the third model according to the insights and

123
Transportation (2012) 39:705–722 713

Table 1 Summary of sample data


Variable Mean Minimum Maximum

Dependent variable (station-to-station ridership)


Morning peak 67 0 1970
Midday 26 0 575
Evening peak 65 0 1766
Built-environment variables
Origin
Population 54,162 5707 129,131
Employment 16,401 566 140,245
Residential area 480,551 0 1,774,748
Office area 180,086 0 1,767,568
Commercial area 298,864 0 1,381,481
University dummy 0.126 0 1
Impedance road 2665 0 7759
Pedestrian friendly intersections 186 0 608
Dead ends 12 0.000 64
Land-use mix 0.494 0.010 0.930
Destination
Population 54,296 5707 129,131
Employment 17,255 566 140245
Residential area 477,415 0 1,774,748
Office area 193,193 0 1,767,568
Commercial area 303,318 0 1,381,481
University dummy 0.138 0 1
Impedance road 2755 0 7759
Pedestrian friendly intersections 183 0 608
Dead ends 12 0.000 64
Land-use mix 0.498 0.010 0.930
Intermodal connection variables
Origin
Feeder bus lines 7.543 0 32
Destination
Feeder bus lines 7.564 0 32
Other variables
Origin
CBD dummy 0.087 0 1
Kangnam dummy 0.117 0 1
Net population density 10.6 0 56
Destination
CBD dummy 0.091 0 1
Kangnam dummy 0.133 0 1
Net population density 10.5 0 56

123
714 Transportation (2012) 39:705–722

Table 1 continued

Variable Mean Minimum Maximum

Travel impedance variable (average across morning peak, midday


and evening peak hours)
Transfers 0.937 0 4
Metro travel time 32.823 2 85
Car travel time 32.634 3 101
Bus travel time 43.303 12 132
Sample size 2000

results of previous studies, a final specification with the maximum F-value was chosen.
Of course, the p-value and the VIF of each coefficient in the final specification were all
confirmed at the above acceptable level (i.e., 5% and 6.5, respectively). The estimation
results for the final specifications of both multiplicative and Poisson models are shown in
Table 2.
As a whole, the multiplicative model was found to fit better than the Poisson model in
terms of the F-statistic, the R2 value, and the number of statistically significant variables.
The best model indicated an R2 value of 0.793, and an F-statistic value of 504.24. These
results can be compared with outcomes of previously published work, in which regression
analysis was conducted based on variables at the station level. The final model of Sohn and
Shim (2010) presented an R2 value of 0.60, and an F-statistic value of 50.938 with seven
statistically significant variables. Kuby et al. (2004) proposed the final model with 12
significant variables at the 0.054 level with an R2 value of 0.727 after the correction for
heteroscedasticity, while Estupiñán and Rodrı́guez (2008) had 5 variables with an R2 value
of 0.45, one of which even had a significance level of 0.44. Cervero and Murakami (2008)
also proposed a series of regression models to predict Metro ridership in Hong Kong as a
function of several TOD-related variables. Their results showed that the number of
independent variables ranged from 7 to 10 and the value of R2 ranged from 0.61 to 0.746.
Although the relatively high R2 of the present study might stem from the intrinsic property
of Gravity-type models, intuitive comparison confirmed that the performance of the present
models is not inferior to the previous models. Hereafter, the discussion will be presented
based on the multiplicative model because the multiplicative model is superior to the
Poisson model in performance and the coefficient estimation results from the Poisson
model are almost identical to those from multiplicative model (see Table 2).

Results from the morning peak hours

Regarding Metro ridership in the morning peak hours, the impact of built environment
variables was consistent with conventional results. That is, the population of the origin
station’s PCA and the employment of the destination station’s PCA were found to be
statistically significant in accounting for the morning-peak ridership. In the same context,
the ridership was associated with the residential floor area of buildings within an origin
station’s PCA, and also was related to the office floor area of buildings within a destination
station’s PCA. This reflects the natural fact that Metro travel in the morning peak hours
consists of many home-based work trips. An advance in the present study is that a direct
statistical significance was verified for population-related variables (Population and Res-
idential area), whereas the previous study was based on station boardings and found none

123
Transportation (2012) 39:705–722 715

Table 2 The estimation results of the final model specification


Independent variables Multiplicative model Poisson model

B p-Value VIF B p-Value E VIF

(a) Morning peak hours


Origin
Pedestrian friendly intersections 0.063 0.000 1.069
Impedance roads -0.000079 0.000 -0.211 1.138
Residential area 0.049 0.041 1.184 0.000001 0.000 0.481 1.256
Office area
Commercial area
Land-use mix
Net population density
Feeder bus lines 0.401 0.000 1.266 0.057 0.000 0.430 1.098
University dummy
Population 0.329 0.000 1.247 0.000011 0.000 0.596 1.268
Employment
Dead ends
CBD dummy
Kangnam dummy
Destination
Pedestrian friendly intersections 0.085 0.000 1.434
Impedance roads
Residential area
Office area 0.064 0.002 2.836 0.000001 0.000 0.193 2.665
Commercial area 0.000002 0.000 0.607 2.827
Land-use mix
Net population density
Feeder bus lines 0.326 0.000 1.389 0.027 0.000 0.204 1.084
University dummy 0.346 0.000 1.074 0.306 0.000 0.042 1.042
Population
Employment 0.450 0.000 3.278
Dead ends
CBD dummy 0.429 0.000 1.625
Kangnam dummy 0.411 0.000 1.732 0.547 0.000 0.072 1.165
Travel impedance
Transfers -0.447 0.000 1.556 -0.824 0.000 -0.772 1.901
Metro travel time -2.206 0.000 5.907 -0.061 0.000 -2.002 4.171
Bus travel time 0.430 0.000 4.005 0.019 0.000 0.823 2.874
Car travel time 1.155 0.000 6.324
No. of observation = 2000 No. of observation = 2000
F-statistic = 438.742 F-statistic = 305.657
R2 = 0.769 R2 = 0.740
Adj. R2 = 0.767 Adj. R2 = 0.738
(b) Evening peak hours
Origin

123
716 Transportation (2012) 39:705–722

Table 2 continued

Independent variables Multiplicative model Poisson model

B p-Value VIF B p-Value E VIF

Pedestrian friendly intersections


Impedance roads -0.000061 0.035 -0.163 1.505
Residential area
Office area 0.098 0.000 2.060
Commercial area 0.000002 0.000 0.597 2.019
Land-use mix
Net population density
Feeder bus lines 0.366 0.000 1.351 0.020 0.002 0.154 1.323
University dummy 0.268 0.000 1.062
Population
Employment 0.374 0.000 2.809
Dead ends
CBD dummy 0.492 0.000 1.636
Kangnam dummy 0.540 0.000 1.652
Destination
Pedestrian friendly intersections 0.094 0.000 1.082
Impedance roads -0.000042 0.121 -0.114 1.265
Residential area
Office area
Commercial area 0.000001 0.001 0.165 1.723
Land-use mix
Net population density
Feeder bus lines 0.475 0.000 1.251
University dummy 0.282 0.000 1.074
Population 0.121 0.003 1.232
Employment 0.134 0.000 1.306
Dead ends
CBD dummy
Kangnam dummy
Travel impedance
Transfers -0.520 0.000 1.585 -0.669 0.000 -0.626 3.126
Metro travel time -1.958 0.000 5.819 -0.023 0.000 -0.752 6.050
Bus travel time 0.509 0.000 3.861 0.003 0.390 0.121 3.345
Car travel time 0.695 0.000 6.133
No. of observation = 2000 No. of observation=2000
F-statistic = 504.24 F-statistic = 242.718
R2 = 0.793 R2 = 0.595
2
Adj. R = 0.791 Adj. R2 = 0.592
(c) Midday
Origin
Pedestrian friendly intersections
Impedance roads -0.000066 0.001 -0.177 1.478

123
Transportation (2012) 39:705–722 717

Table 2 continued

Independent variables Multiplicative model Poisson model

B p-Value VIF B p-Value E VIF

Residential area
Office area
Commercial area 0.000002 0.000 0.467 1.520
Land-use mix
Net population density
Feeder bus lines 0.334 0.000 1.165 0.046 0.000 0.344 1.103
University dummy 0.168 0.006 1.044 0.148 0.023 0.019 1.085
Population 0.000005 0.000 0.295 1.221
Employment 0.256 0.000 1.376
Dead ends
CBD dummy
Kangnam dummy 0.134 0.059 1.335
Destination
Pedestrian friendly intersections
Impedance roads -0.000057 0.004 -0.158 1.432
Residential area
Office area
Commercial area 0.047 0.010 1.723 0.000001 0.000 0.428 1.960
Land-use mix
Net population density
Feeder bus lines 0.302 0.000 1.342 0.050 0.000 0.381 1.168
University dummy 0.240 0.000 1.058
Population
Employment 0.245 0.000 1.745
Dead ends
CBD dummy 0.238 0.002 1.313 0.260 0.003 0.024 1.682
Kangnam dummy 0.289 0.000 0.038 1.413
Travel impedance
Transfers -0.744 0.000 -0.697 1.973
Metro travel time -2.018 0.000 5.022 -0.062 0.000 -2.028 4.186
Bus travel time 0.408 0.000 3.848 0.007 0.018 0.288 2.813
Car travel time 0.698 0.000 5.813
No. of observation = 2000 No. of observation = 2000
F-statistic = 559.429 F-statistic = 336.064
R2 = 0.772 R2 = 0.780
2
Adj. R = 0.771 Adj. R2 = 0.777

B coefficient, p-value probability of significance, VIF variance inflation factor


The elasticity (E) computed for the mean value of the independent variable is presented for Poisson model,
for which the elasticity varied with the level of the variable

123
718 Transportation (2012) 39:705–722

(Sohn and Shim 2010). The two city centers in the Seoul metropolitan area had a sig-
nificant impact on station-to-station ridership especially when these were destination sta-
tions (CBD dummy and Kangnam dummy). These two city centers include many station
areas of large employment. So, a station within these centers would derive extra Metro
demand in excess of the local employment level. Such a neighborhood effect was also
found in (Sohn and Kim 2010). If a University was within a destination station’s PCA
(University dummy), Metro ridership increased considerably.
The number of feeder bus lines and the number of pedestrian-friendly intersections were
positively related to Metro ridership for both origin and destination stations. However, the
results on feeder bus lines should be interpreted with care, because the cyclic relationship
between this variable and the ridership was found (Sohn and Shim 2010) by using a SEM.
Unfortunately, the present models could not explicitly deal with the endogeneity. In any
case, this result was in accordance with the conventional results that conditions of access
and egress to transit stations are important in increasing transit ridership. In particular,
identifying the impact of the walking environment on Metro ridership is meaningful
because the previous study, which was based on station-level analysis, found no direct
statistical significance to such variables (Sohn and Shim 2010).
Travel impedance variables were found to have great influence on Metro ridership in the
morning peak hours. The present study clearly confirmed that travel times of competing
modes (car and bus) positively affect Metro ridership, while Metro travel time has a
negative influence on ridership. The former finding was very meaningful, because Taylor
and Fink (2007) pointed out the difficulty of quantifying the impact of competing modes as
variables in aggregate-level analysis, although a variation in transit ridership due to the
utility of competing modes has often been observed in the disaggregate analysis of mode
choice. Our previous study found no direct association between impedance-related vari-
ables of competing modes and Metro ridership (Sohn and Shim 2010), but the present
study of station-to-station ridership observed such a relationship. Moreover, we found that
frequent transfers have a negative effect on Metro ridership.
As mentioned earlier, the magnitude of the estimated coefficient of the multiplicative
model directly represented the elasticity of ridership with respect to the corresponding
variable. According to the results, only two variables, Metro travel time and car travel
time, are elastic. The effect of Metro travel time represents the basic distance-decay
relationship in trip length distribution, which is incorporated into the spatial interaction
model. On the other hand, the car travel time plays a more decisive role in generating
Metro ridership than any other variables such as built-environment variables and inter-
modal connection variables. The previous study based on station-level ridership found no
counter-effect of car travel time on Metro demand (Sohn and Shim 2010). This might be
due to the errors entailed by aggregating travel impedances into a station-level variable
such as a centrality index.

Results from the evening peak hours

Regarding Metro ridership in the evening peak hours, two employment-related variables of
an origin station’s PCA (Office area and Employment) were found to be statistically
significant. However, with respect to the impact of the destination station on Metro rid-
ership, the employment variable was also significant, as was the population variable. It can
be hypothesized that in the evening peak hours some Metro users would travel to an
employment-oriented area instead of returning home. This could mean that people are
traveling to stations with a lot of employment-intensive services like restaurants and bars.

123
Transportation (2012) 39:705–722 719

An additional interpretation is that people would be going to night jobs. The effect of the
two city centers on ridership turned out to be significant only for the origin station (CBD
dummy and Kangnam dummy). This is contrary to the results from the morning peak hours
when the two variables were significant only for the destination station. Metro ridership
increased when a University was located within either the origin or destination station’s
PCA (University dummy). This implies that students leave school after class and, at the
same time, other people gather in the vicinity of college towns in the evening. In the Seoul
metropolitan area, there is a dense concentration of commercial and recreational facilities
that attract young people in the vicinity of Universities.
As in the case of morning peak hours, in the evening peak hours the number of feeder
bus lines turned out to be positively related to Metro ridership for both origin and desti-
nation stations. However, the number of pedestrian-friendly intersections was significant
only within the destination station’s PCA. The result can be interpreted as a Metro user’s
preference of walking from a destination station to walking to an origin station. This means
commuters are more likely to take a bus to an origin station. Otherwise, this might be due
to an unknown error that should be investigated in further studies. With respect to travel
impedance-related variables, the results from evening peak hours were almost the same as
those from morning peak hours.

Results from midday

Mid-day trips realistically contain various kinds of home-based trips such as shopping
trips, recreation trips, medical/service trips, and tourism, etc. However, regarding mid-day
Metro ridership, the employment variable for the origin station and the employment and
commercial area variables for the destination station were found to be statistically sig-
nificant. That is, only employment-related variables were significant for both origin and
destination stations. This result is not consistent with prior expectations that home-based
trips would prevail during midday. Non-home-based trips in mid-day are likely to be made
up of workers visiting their clients or business counterparts and workers going out to lunch
or shopping, as well as tourists. In the same context, the city centers were statistically
significant to both origin and destination stations (Kangnam dummy for origin and CBD
dummy for destination). Many Metro trips, of both origin and destination, predictably
occurred within the city centers during midday. Metro ridership was positively associated
with the existence of a University within both origin and destination stations (University
dummy), which reflects a generic land-use pattern that many attractive facilities are located
in the vicinity of Universities. As in the case of the other two time periods, the number of
pedestrian friendly intersections and the number of feeder bus lines also turned out to have
a positive relationship with Metro ridership for both origin and destination stations. With
respect to travel impedance-related variables, the results from midday were almost iden-
tical to those from morning or evening peak hours, except for the exclusion of the impact
of transfers between transit lines.

Conclusion and further research

By adopting a station-to-station analysis of Metro ridership, the present study found several
significant results. It was found that built-environment variables associated with generating
Metro ridership vary by both trip ends (origin/destination) and time of day. For example, in
the morning peak hours the population is a key variable for Metro boardings in an origin

123
720 Transportation (2012) 39:705–722

station, while employment becomes a key variable in the same origin station during
evening peak hours. This strongly implies that developing the vicinity of stations with
mixed land use will increase Metro ridership on a daily basis, which has been proclaimed
by many researchers who support new urbanism. In the same context, in a specific time
period, the variable directly representing the extent of land use mix did not account for an
increase in Metro ridership.
The importance of inter-modal connectivity at stations, which had been argued by Sohn
and Shim (2010), was also confirmed by the present study. Providing good feeder bus
services in both origin and destination stations at all time periods considerably enhanced
Metro ridership. At the same time, the walkability for pedestrians in a station’s PCA
proved to be important for increasing Metro ridership.
Regarding travel impedances, the results showing that Metro travel time has the most
influence on Metro ridership represent an intrinsic property of trip length distribution.
In order to separate the net effect of Metro travel time from the total basic distance-decay
effect, a more robust modeling technique would be needed in further studies. The car travel
time was the second most influential variable that affects Metro ridership. It can be con-
cluded that making Metro lines more efficient than highway alignment should take pre-
cedence over any other measures to increase Metro ridership. The idea of developing the
vicinity of Metro stations would not succeed in the Seoul metropolitan area without
considering the connectivity efficiency of the Metro line. In addition, the number of
transfers deteriorated the utility of Metro travel. This supports the current policy of the
Seoul Metropolitan Government to reduce inconvenient transfers by directly connecting
two independent Metro lines (Seoul Metropolitan Government 2007).
The proposed model will be extended to account for all origin–destination pairs in the
near future, which was not possible in the present study because the data collection for
travel impedance variables was very costly and time-consuming. Moreover, our future
plans include a survey of the details of design elements for pedestrians within each sta-
tion’s PCA. The new survey is expected to yield many interesting results regarding design
elements in the vicinity of Metro stations.

References

ALGOGA: http://www.algoga.org/ (2010)


Bhat, C.R., Guo, J.Y.: A comprehensive analysis of built environment characteristics on household
residential choice and auto ownership levels. Transp. Res. Part B 41, 506–526 (2007)
Boyle, D.K.: Fixed-route Transit Ridership Forecasting and Service Planning Methods. TCRP synthesis 66.
Transportation Research Board, Washington (2006)
Canepa, B.: Bursting the bubble: determining the transit-oriented development’s walkable limits. Transp.
Res. Rec. J. Transp. Res. Board, No. 1992, 28–34 (2007)
Cao, X., Mokhtarian, P.L., Handy, S.L.: The relationship between the built environment and non-work
travel: a case study of Northern California. Transp. Res. Part A 43, 548–559 (2009)
Cervero, R.: Alternative approaches to modeling the travel-demand impacts of smart growth. J. Am. Plan.
Assoc. 72(3), 285–295 (2006)
Cervero, R., Kockelman, K.: Travel demand and the 3Ds: density, diversity, and design. Transp. Res. Part D
2(3), 199–219 (1997)
Cervero, R., Murakami J.: Rail ? property development: a model of sustainable transit finance and
urbanism. Working paper. UC Berkeley Center for Future Urban Transport, UCB-ITS-VWP-2008-5,
pp. 82–88 (2008)
Chu, X.: Ridership models at the stop level. National Center for Transit Research, Center for Urban
Transportation Research, University of South Florida, Tampa (2004)

123
Transportation (2012) 39:705–722 721

Estupiñán, N., Rodrı́guez, D.A.: The relationship between urban form and station boardings for Bogota’s
BRT. Transp. Res. Part A 42(2), 296–306 (2008)
Ewing, R., Bartholomew, K., Winkelman, S., Walters, J., Chen, D.: Growing Cooler: The Evidence on
Urban Development and Climate Change. Urban Land Institute, Washington, DC (2008)
Flowerdew, R.: Poisson regression modeling of migration. In: Stillwell, J., Congdon, P. (eds.) Migration
Models: Macro and Micro Approaches, pp. 92–112. Belhaven, London (1991)
Forsyth, A., Oakes, J.M., Schmitz, K.H.: The built environment, walking, and physical activity: is the
environment more important to some people than others? Transp. Res. Part D 14(1), 42–49 (2009)
Golledge, R.G., Stimson, R.J.: Spatial Behaviour. The Guilford Press, New York (1997)
Gurmu, S., Trivedo P.K.: Recent Developments in Models of Event Counts: A Survey. Discussion Paper
261. Department of Economics, University of Virginia, Charlottesville (1994)
Hair, J.F., Anderson, R.E., Tatham, R.L.: Multivariate Data Analysis. Macmillan, New York (1987)
Hamin, E.M., Gurran, N.: Urban form and climate change: balancing adaptation and mitigation in the
U.S. and Australia. Habitat Int. 33, 238–245 (2009)
Hannon, J.T., Moore, R.P., Ryan, L.W.: Burns Waterway boat capacity study. In: Proceeding on the First
International Conference on World Marina ‘91, Long Beach, CA. ASCE, New York (1991)
Ivy, R.L.: The restructuring of air travel linkages in the new Europe. Prof. Geogr. 47(3), 280–288 (1995)
Jones, I.S., Nichols, A.J.: Demand for inter-city rail travel in the United Kingdom: some evidence. J. Transp.
Econ. Policy 17, 133–154 (1983)
Khattak, A.J., Rodriguez, D.: Travel behavior in neo-traditional neighborhood developments: a case study in
USA. Transp. Res. Part A 39(6), 481–500 (2005)
Kim, S., Ulfarsson, G.F., Hennessy, J.T.: Analysis of light rail rider travel behavior: impacts of individual,
built environment, and crime characteristics on transit access. Transp. Res. Part A 41, 511–522 (2007)
Kline, R.B.: Principles and Practice of Structural Equation Modeling. The Guilford Press, New York (1998)
Korea Smart Card: http://www.koreasmartcard.com/ (2010)
Kraft, G.: Demand for Intercity Passenger Travel in the Washington-Boston Corridor. North-East Corridor.
Project Report. Systems Analysis and Research Corporation, Boston (1968)
Kuby, M., Barranda, A., Upchurch, C.: Factors influencing light-rail station boardings in the United States.
Transp. Res. Part A 38(3), 223–247 (2004)
Lave, L.B.: The demand for intercity passenger transportation. J. Reg. Sci. 12, 71–84 (1972)
Master plan of Seoul Metro (2007) Seoul Metropolitan Government
Murray, A.T., Davis, R., Stimson, R.J., Ferreira, L.: Public transportation access. Transp. Res. Part D 3(5),
319–328 (1998)
O’Sullivan, S., Morral, J.: Walking distances to and from light-rail transit stations. Transp. Res. Rec.
J. Transp. Res. Board, No. 1538, 19–26 (1996)
Ortúza, J.D., Willumsen, L.G.: Modeling Transport, 3rd edn. Wiley, New York (2001)
Parsons Brinckerhoff Quade & Douglas, Inc: Transit and Urban Form. National Academy Press, Wash-
ington, DC (1996)
Peng, Z., Dueker, K.J., Strathman, J., Hopper, J.: A simultaneous route-level transit patronage model:
demand, supply, and inter-route relationship. Transportation 24, 159–181 (1997)
ROADI: http://www.roadi.com/ (2010)
Schlossberg, M., Brown, N.: Comparing transit-oriented development sites by walkability indicators.
Transp. Res. Rec. J. Transp. Res. Board, No. 1887, 34–42 (2004)
Sohn, K., Kim, D.: Zonal centrality measures and the neighborhood effect. Transp. Res. Part A 44(9),
733–743 (2010)
Sohn, K., Shim, H.: Factors generating boardings at metro stations in the Seoul metropolitan area. Cities 27,
358–368 (2010)
Sohn, K., Yun, J.: Separation of car-dependent commuters from normal-choice riders in mode-choice
analysis. Transportation 36, 423–436 (2009)
Spillar, R.J., Rutherford, G.S.: The effects of population density and income on per capita transit ridership in
Western American cities. In: Institute of Transportation Engineers’ Compendium of Technical Papers:
60th Annual Meeting, 5–8 August, pp. 327–331 (1998)
Taylor, B.D., Fink, C.N.Y.: The factors influencing transit ridership: a review and analysis of the ridership
literature. Working paper of UCLA Institute of Transportation Studies (2007)
Thompson, G.L.: Achieving suburban transit potential. Transportation Research Record 1571, pp. 151–162
(1997)
Household Transportation Survey in Seoul Metropolitan Area (2008) Seoul Metropolitan Government
Wardman, M.: Inter-urban rail demand, elasticities and competition in Great Britain: evidence from direct
demand models. Transp. Res. Part E 33(1), 15–28 (1997)

123
722 Transportation (2012) 39:705–722

Yao, X.: Where are public transit needed—examining potential demand for public transit for commuting
trips. Comput. Environ. Urban Syst. 31, 535–550 (2007)

Author Biographies

Jinkyung Choi is a graduate student in the Department of Urban Engineering, Chung-Ang University,
Seoul, Korea. He is preparing a dissertation for a master’s degree based on the present study.

Yong Jae Lee is a Professor in the Department of Urban Engineering, Chung-Ang University, Seoul, Korea.
He is a recognized expert in the field of transportation engineering and has more than 33 years experience in
transportation planning, design, policy and management. He has served the President of Korea Society of
Transportation (KST) for 2 years.

Taewan Kim is an Associate Professor in the Department of Urban Engineering at Chung-Ang University
in Seoul, Korea. He obtained his Ph.D. degree from the University of California, Davis in 2003. His research
interests cover traffic flow theory, traffic operation, ITS, and urban transportation planning.

Keemin Sohn is an Assistant Professor in Department of Urban Engineering at Chung-Ang University in


Seoul, Korea. He obtained the degree of Doctor of Engineering from Seoul National University, Korea.
His research interests cover urban transportation planning and its interactions with policies directed at
controlling travel demand.

123

You might also like