Journal of Transport Geography 30 (2013) 127–137

Contents lists available at SciVerse ScienceDirect

Journal of Transport Geography

journal homepage: www.elsevier.com/locate/jtrangeo

Factors influencing modal split of commuting journeys in medium-size

European cities
Georgina Santos a,b,⇑, Hanna Maoh c, Dimitris Potoglou a,d, Thomas von Brunn a,1
a
School of Planning and Geography, Cardiff University, Glamorgan Building, King Edward VII Avenue, Cardiff CF10 3WA, UK
b
Transport Studies Unit, School of Geography and the Environment, University of Oxford, South Parks Road, Oxford OX1 3QY, UK
c
Department of Civil and Environmental Engineering, University of Windsor, 401 Sunset Avenue, Windsor, Ontario, Canada N9B 3P4
d
RAND Europe, Westbrook Centre, Milton Road, Cambridge CB4 1YG, UK

a r t i c l e i n f o a b s t r a c t

Keywords: This paper attempts to identify factors that influence modal split for journeys to work in 112 medium-
Mode share size cities in Europe. Using a discrete choice modelling approach we find that: (a) car share increases with
Mode choice car ownership and GDP per capita; (b) motorcycle share decreases with petrol price and increases with
Sustainable transport motorcycle ownership; (c) bicycle share increases with the length of the bicycle network in the city; (d)
Commuting journeys
public transport share increases with resident population, GDP per capita and the number of buses (or
Journeys to work
Logit model
bus equivalents) operating per 1000 population, and decreases with public transport fares, number of
days of rain per year, proportion of people aged 65 and over living in the city and the proportion of house-
holds with children; (e) the number of students in universities and further education establishments per
1000 resident population is positively associated with the shares of public transport, motorcycle, bicycle
and walking. Policies aimed at increasing the bicycle network are likely to increase cycling share. Policies
aimed at increasing the number of buses (or bus equivalents) and reducing public transport fares are
likely to increase public transport share. Policies aimed at discouraging car ownership are likely to reduce
car share.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction This paper works towards that objective by analysing the fac-
tors that influence the modal split for journeys to work2 in 112
Sustainable transport has become an important topic on most medium-size cities in Europe. These are defined, for the purposes
national, regional and local governments’ agendas. Sustainability of our study, as cities with populations of 100,000–500,000. We con-
in transport, typically defined along economic, environmental tend that understanding travel behaviour by identifying common
and social (or equity) dimensions (European Commission, 2008, factors that have an impact on modal split in medium-size European
p. 12), can be achieved with sustainable modes of transport (as cities can provide the basis for designing sustainable transport poli-
well as with sustainable travel behaviour). In general, there is con- cies.3 Newman and Kenworthy (1999, p. 86) explain that journeys to
sensus that the private car does not enhance sustainability, work account for most peak demand on road networks. Commuting
whereas public transport and non-motorised modes, such as walk- is a major component of daily travel demand and an important
ing and cycling, do (Black, 2010). It is therefore important to under- source of congestion and pollution (Antipova et al., 2011, p. 1010;
stand what drives people towards and away from the private car, Habib et al., 2011, p. 588). Santos et al. (2010, p. 84) point out that
public transport and non-motorised modes, respectively. Once this addressing commuter trips is essential for relieving congestion in ur-
is done, appropriate policies can be designed in order to enhance ban areas.
transport sustainability.

2
Solely concentrating on journeys to work may, however, overstate public
transport, since it is particularly strong in that market (Kenworthy et al., 1999, p. 17).
⇑ Corresponding author at: School of Planning and Geography, Cardiff University, 3
It should be noted that although mode share is not the best indicator for
Glamorgan Building, King Edward VII Avenue, Cardiff CF10 3WA, UK. Tel.: +44 2920 measuring travel-related sustainability, it is still considered among the most
874 462. important indicators. When travel sustainability is mainly concerned with the
E-mail addresses: SantosG@Cardiff.ac.uk (G. Santos), MaohHF@uwindsor.ca (H. environmental impacts that travel causes, other measures such as vehicle kilometres
Maoh), PotoglouD@Cardiff.ac.uk (D. Potoglou), thomasvb@gmx.ch (T. von Brunn). travelled (VKT) can also be favoured. This is because while a lower share of trips by
1
Present address: Ernst Basler + Partner AG, Mühlebachstrasse 11, 8032 Zurich, car may go some way towards limiting the amount of carbon emissions, for example,
Switzerland. it is VKT that determine the level of emissions for a given region.

0966-6923/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.jtrangeo.2013.04.005
128 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137

This paper attempts to answer the following questions: There also seems to be a positive link between income and car
use, although the link between income and other modes (such as
 What are the significant factors most likely to cause an increase public transport or walking and cycling) differs from one study
or decrease in the share of certain travel mode for journeys to to another. Table 1 presents the direction of the link for income
work in medium-size European cities? by mode, according to some authors.
 What type of sustainable transport policies can be advocated As people get older they may have different preferences regard-
based on these significant factors? ing mode choice. Their health, physical and functional abilities
may deteriorate and their confidence in walking and driving, for
To address the above mentioned questions, we develop a set of example, may change as a result of that (Naumann et al., 2009).
discrete choice models using the 2001 and 2004 modal split shares We report the links found by Kim and Ulfarsson (2008) and Sabir
of 112 medium-size European cities. City-level variables mainly (2011) in Table 1.
depicting demographic and socio-economic factors are used in Chen et al. (2008) include ‘household with children’ as a vari-
the specification of the modal-split models. Using a rich data set, able in their model for commuting mode choice and find a positive
Winston and Shirley (1998) develop a discrete choice model, association with cars and a negative one with public transport.
where they combine mode and departure choice, for the largest Dargay and Hanly (2007) also analyse commuting trips and find
116 urbanised areas in the United States. Such type of data are similar results: positive association with car and negative for all
not available for Europe and to our knowledge the type of regional other modes. Kim and Ulfarsson (2008) analyse mode choice on
analysis at European level we present in this paper is novel and has short home-based trips and also find a positive link with cars
not been addressed in previous studies. More specifically, our ef- and a negative one with public transport (and with walking).
forts contribute to the existing transport geography literature by Buchanan (1964) expects a noticeable modal shift from low pub-
highlighting significant factors influencing the observed modal lic transport fares, whereas Asensio (2000) and Cervero (1998) ex-
split in mid-size European cities. By contrast, the majority of the pect higher influence from frequent services. Balcombe et al.
existing studies on mode choice behaviour have been conducted (2004) regard high frequencies and affordable tickets as equally
for a given single city or a small group of cities. influential in the long term and do not consider vehicle comfort as
The remainder of this paper is organised as follows: Section 2 particularly decisive. A number of other authors have also analysed
reviews the most relevant literature for this study; Section 3 pre- the link between public transport fares and public transport use (or
sents the data used and the model specifications; Section 4 pre- share or choice as a transport mode) and between public transport
sents the results and finally, Section 5 summarises the main frequency and public transport use (or share or choice as a transport
findings and offers policy recommendations and suggestions for mode). Their findings are summarised in Table 1.
further research. Curiously, weather variables are not typically included in these
types of study. Sabir (2011) is an exception and we report his find-
2. Previous work ings regarding rain in Table 1.4
Newman and Kenworthy (1989, 2006) and Cervero (1998) point
Modal split is related to a number of factors, ranging from indi- out the importance of population density for the viability of public
vidual mode choice, which in turn depends on individual and mode transport and non-motorised modes. Other studies find that popu-
characteristics, to land use and population density. In this section we lation density plays an important role in people’s mode choice for
summarise previous work relevant to this study. Although the pa- trips to work (for example, Chen et al., 2008; Dargay and Hanly,
pers we cite range in the methods used (literature reviews, qualita- 2007; Kain and Liu, 2002) and for urban trips in general (Souche,
tive analysis of data, individual and aggregate discrete choice 2010). Dargay and Hanly (2007) find a negative association be-
models, OLS regressions, to name the most prominent ones) and tween higher population density and commuting by car and
their geographic and temporal scope, as well as in their objectives, Schwanen (2002) finds a positive association between population
we highlight the issues of interest to the aims of the present re- density and commuting by public transport.
search. The literature reviewed is extensive and in order to save In addition to the variables described above, distance to be trav-
space, we present the main points, in a systematic way, in Table 1. elled and land use mix have also been included in a number of
Population density was not included in our models, even though models whose results are reported in Table 1. Finally, there are
we had reliable data for that variable. The reason for not including studies that concentrate on one or two modes only, such as for
it was that it was not found to be significant or gave counter-intui- example Pucher et al. (2011). Their results are not included in Ta-
tive signs due to multicollinearity with other variables. Trip distance ble 1 but they are compared with ours in Section 4.
and land use mix, the last two categories in Table 1, include variables As it becomes clear from the above, a number of studies have
we were unable to include in our models due to lack of data. All the attempted to identify at least some of the factors affecting either
other rows contain variables we included in our models, although modal split or mode choice. These vary greatly in geographical cov-
we used ‘number of buses (or bus equivalents) operating in the pub- erage and methodology used. Typically, however, microdata from
lic transport per 1000 population’ as a proxy for ‘public transport travel surveys is used or if the data are aggregate, the studies only
service frequency’. cover one or a small group of cities, with the exception of Winston
Schwanen (2002), Scheiner (2010), Susilo and Maat (2007), and and Shirley (1998). The present study, on the other hand, uses
White (2009), all find links between city size and either modal split aggregate data to compare 112 medium size European cities, with
at city level or individual mode choice. In general the share of trips a view to deriving recommendations specific to such European
by car decreases and the share of trips by public transport in- context, in order to inform the on-going transport policy debate
creases with city size. The findings regarding the share of non- at EU level. As highlighted above, modal split analysis at regional
motorised modes, on the other hand, are not clear cut. level for so many cities has not been conducted before for the case
In general, car ownership is found to increase the share of trips of Europe.
by car and decrease the share of trips by public transport (Bal-
combe et al., 2004; Chen et al., 2008; Kain and Liu, 2002; Pinjari
et al., 2007; to name just a few papers from Table 1). Dargay and
Hanly (2007) also conclude that car purchase costs and petrol 4
Pucher et al. (2011) and Sumalee et al. (2011) do take rain into account but their
prices have a negative impact on commuting by car. studies are not empirical studies on modal split like ours.
Table 1
Association of different factors for mode choice and modal split according to the reviewed literature.

Factor Author Suggested association

Private car Public transport Motorcycle Bicycle Foot
City size (measured by resident population) Schwanen (2002) na + na  
Scheiner (2010)  + na + (for distances under 1.5 km) + (for distances under 1.5 km)
Susilo and Maat (2007)  + na na na
White (2009) na na na + na
Car ownership Balcombe et al. (2004) +  na na na
Chen et al. (2008) +  na na na
Kain and Liu (2002) +  na na na
Kim and Ulfarsson (2008)a +  na  
Kitamura (2009) + = na na na
Paulley et al. (2006) +  na na na
Pinjari et al. (2007) +  na + +
Sabir (2011)a +    +
Scheiner (2010) +  na  

G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137

White (2009) +  na na na
Income Balcombe et al. (2004) +  na na na
Chen et al. (2008) +  na na na
Dargay and Hanly (2007) + na na na na
Kitamura (2009) + + na na na
Paulley et al. (2006) +  na na na
Sabir (2011)a +    +
Age (elderly) Kim and Ulfarsson (2008) + = na  
Sabir (2011)a +    
Households with children Chen et al. (2008) +  na na na
Dargay and Hanly (2007) + na na na na
Kim and Ulfarsson (2008) +  na na 
Public transport fares Asensio (2000) na = na na na
Balcombe et al. (2004) +  na na na
Buchanan (1964) +  na na na
Cervero (1998) + = na na na
Chen et al. (2008) +  na na na
Paulley et al. (2006) +  na na na
Zhang (2004) +  na na na
Public transport service frequency Asensio (2000) na + na na na
Balcombe et al. (2004)  + na na na
Cervero (1998)  + na na na
Kitamura (2009)  + na na na
Paulley et al. (2006)  + na na na
White (2009)  + na na na
Rain Sabir (2011)a + +   +
Population density Balcombe et al. (2004)  + na na na
Cervero (1998)  + na + +
Chen et al. (2008)  + na na na
Dargay and Hanly (2007)  na na na na
Newman and Kenworthy (1989, 2006)  + na + +
Pinjari et al. (2007) na na na + +
Schwanen (2002) na + na  
Souche (2010)  + na  
Susilo and Maat (2007)  + + + +
Zhang (2004)  + na + +

129
(continued on next page)
 130
G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137
Table 1 (continued)

Factor Author Suggested association

Private car Public transport Motorcycle Bicycle Foot
Trip distance Kim and Ulfarsson (2008) na + na na 
Sabir (2011)a + +   
Scheiner (2010)b + + + + and c 
Land-use mix Balcombe et al. (2004) na = na na na
Cervero (1998)  + na + +
Frank et al. (2008)  + na + +
Newman and Kenworthy (1999)  + na + +
Pinjari et al. (2007) na + na na na
Van Acker and Witlox (2011)  + na + na

Key: (+) stands for positive association, () stands for negative association, (=) stands for no association and (na) stands for not applicable, meaning the source cited did not analyse the association.
a
Sabir (2011) does not model ‘motorcycles’ as a mode of transport but includes motorcycles under ‘other’ mode of transport, which he defines as moped, motor, scooter, taxi, lorry and delivery van . He includes three variables
regarding rain in his model: a dummy variable for rain up to 0.1 mm, a dummy variable for rain higher than 0.1 mm, and the duration of rain (in minutes), all at the hour of departure. The results reported on the table above
correspond to the dummy variable rain up to 0.1 mm, which is average (as non-extreme) rain.
b
Scheiner (2010) includes cars and motorcycles in one category.
c
+ up to 1.5 km and  for distances longer than 1.5 km.
 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137 131

Table 2
The sample medium-size cities in the Urban Audit.

Country Cities
Denmark Aalborg, Aarhus, Copenhagen, Odense
Germany Augsburg, Bielefeld, Bochum, Bonn, Darmstadt, Dresden, Erfurt, Freiburg, Gottingen, Halle an der Saale, Karlsruhe, Kiel, Koblenz, Leipzig, Magdeburg,
Mainz, Moers, Mönchengladbach, Mülheim, Nurnberg, Potsdam, Regensburg, Saarbrucken, Trier, Wiesbaden, Wuppertal
Estonia Tallinn, Tartu
Finland Tampere, Turku, Oulu
Italy Firenze, Bari, Bologna, Catania, Venezia, Verona, Trento, Trieste, Perugia, Ancona, Pescara, Taranto, Reggio di Calabria, Sassari, Cagliari
Netherlands The Hague, Utrecht, Eindhoven, Tilburg, Groningen, Enschede, Arnhem, Apeldoorn, Nijmegen, Breda, Almere
Portugal Porto, Braga, Funchal, Coimbra, Setubal
Slovakia Bratislava, Kosice
Spain Murcia, Las Palmas, Valladolid, Palma de Mallorca, Vitoria / Gasteiz, Oviedo, Pamplona / Iruña, Santander, Badajoz, Logroño, Santa Cruz de Tenerife,
L’Hospitalet de Llobregat, Gijón, Vigo, Alicante/Alacant, Córdoba, Bilbao
Sweden Örebro, Gothenburg, Jönköping, Linköping, Malmö, Umeå, Uppsala
Switzerland Bern, Geneva, Lausanne, Zurich
UK Bradford, Bristol, Cambridge, Cardiff, Coventry, Exeter, Kingston-upon-Hull, Leicester, Liverpool, Manchester, Newcastle, Nottingham, Portsmouth, Stoke-
on-Trent, Wirral, Wolverhampton, Wrexham

3. Data and modelling approach ies, it provides a fair coverage across 12 European countries, as
shown in Table 2.
3.1. Data The data used in the subsequent analysis are based on averages
for the years 2001 and 2004. The averaging is performed to address
Petrol prices were retrieved from ‘Energy Prices and Taxes’ the issue of missing values, which occurred in many variables
(International Energy Agency, 2008). GDP per capita was retrieved across cities. To maintain consistency, the averaging of the 2001
from the International Monetary Fund (IMF) website5. The rest of and 2004 values is applied on both the modal split shares and
the data used in this paper come from the Urban Audit from the the independent variables used in the estimation of the statistical
Directorate-General for Regional Policy and Eurostat (European models. In cases where there are missing values for both 2001 and
Commission, 2004). 2004, the average value of the country for which the city belonged
The Urban Audit provides information on selected urban areas to is used to populate the variable.
across Member States of the European Union and the Candidate Identifying and testing relevant variables was informed by the
Countries. It is coordinated by Eurostat and involves a number of literature review and our prior expectations regarding the relation-
partners including National Statistical Offices, the towns and cities ship between these variables and the observed modal split. This
themselves, existing inter-city cooperation networks, international was further coordinated by the validity and availability of relevant
organisations, the European Commission and national govern- information in the data given the missing or inconsistent values
ments. Much data already exists at the National Statistical Offices across the sample of cities. Table 3 presents the coding and defini-
in their databases or in administrative registers linked with them. tion of the variables used in the analysis, together with our a priori
The remaining part of the data needs to be collected by the very expectations. One variable not included in any of the reviewed
towns and cities. National Urban Audit Coordinators compile this studies but included in ours is the proportion of students in a city.
data and Eurostat amalgamates the data coming from the different This proved to be a significant variable in relation to public trans-
countries.6 port, motorcycle, bicycle and foot, as we explain in Section 4, prob-
The database is freely accessible at: http://www.urbanau- ably because it acts as a proxy for cities which are pedestrian and
dit.org/index.aspx. It covers 357 cities, 336 variables and three spa- cycling friendly and public transport oriented.
tial levels. Core cities refer to areas of local government Table 4 provides descriptive statistics for each of the explana-
responsibility, larger urban zones or functional urban areas usually tory variables.
exceed the core city boundaries and sub-city districts are used for Following Rodrigue et al. (2009, pp. 241, 336) we define modal
comparing disparities within cities (European Commission, 2004, split or mode share as the proportion of trips that is made by each
pp. 5, 9, 12). Although larger urban zones are the most suitable transport mode. This is the dependent variable. The different mode
spatial reference, due to limited data availability (the data set shares include car, motorcycle, bicycle, walking and public transport.
has many missing values) we use core city data instead. The main We concentrate on journeys to work or commuting journeys, which
problem is that the larger urban zones concept ‘is not defined for are defined according to the European Commission (2004, p. 46) as
all cities involved in the Urban Audit, and even where it is defined the ‘shortest trip from place of residence to the work place, includ-
the criteria and principles are not the same’ (European Commis- ing change of transport mode’. These morning commute trips
sion, 2004, p. 11). ignore trip chaining, understood as a stop (for shopping or drop-
Table 2 lists the 112 cities that are extracted from the Urban ping children off at school, for example) during travel between
Audit to form our sample. These are medium-size European cities two anchors, such as home and work. As a consequence, the data-
with population between 100,000 and 500,000. Overall, there are base we used does not take into account trip chaining. This does
394 medium-size cities in Europe, but only 112 of them are ac- not necessarily mean that all the trips used to compile the data
counted for in the Urban Audit. As such, our sample represents in our data set were free from trip chaining. Commuters may drop
28.4% of all medium-size European cities. Although our sample their children at school on their way to work, but this is not re-
represents little over one quarter of all medium-size European cit- corded in the Urban Audit.
The Urban Audit endeavours to guarantee data quality. Accord-
5
ing to the European Commission (2004, p. 77), this is achieved by
http://www.imf.org/external/data.htm
6
The data that we extracted from the Urban Audit and used in our models comes
setting a range for each variable (where possible), checking the
either from different National Censuses or are collected by the town or city in data falls within that range, reviewing the data for anomalies
question. and subsequently validating or correcting it.
132 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137

Table 3
City level variables used in the analysis and expected parameter signs.

Variable Description Prior expectation

A SHARE Car share in city k
M SHARE Motorcycle share in city k
B SHARE Bicycle share in city k
F SHARE Walk share in city k
T SHARE Public transport share in city k
BIKE NETWORK Length of bicycle network (dedicated cycle paths and lanes) Positive correlation with B SHARE
per 1000 population
BUS RATE Number of buses (or bus equivalents) operating in the public Positive correlation with T SHARE
transport per 1000 population
CAR RATE Number of registered cars per 1000 population Positive correlation with A SHARE and negative correlation with T SHARE
ELDERLY The proportion of total population aged 65 and over Positive correlation with A SHARE and negative correlation with all other modes
GDP PER CAPITA Gross Domestic Product per capita (in EUROS – 2004 prices) Positive correlation with A SHARE and F SHARE. Correlation with T SHARE has
been found to be positive by one study and negative by other studiesa
HOUSE WITH CHILD Proportion of households with children aged 0–17 Positive correlation with A SHARE and negative correlation with T SHARE
MOTORCYC RATE Number of registered motor cycles per 1000 population Positive correlation with M SHARE
PETROL PRICE Petrol price per litre (in EUROS – 2004 prices) Negative correlation with A SHARE and M SHARE
POPULATION Total resident population Negative correlation with A SHARE and positive correlation with T SHARE
RAIN Number of days of rain per year Negative correlation with B SHARE and positive correlation with A SHARE,
T SHARE and F SHARE
STUDENTS Number of students in universities and further education Positive correlation with B SHARE and F SHARE and negative correlation with
establishments per 1000 resident population A SHARE
TRANSIT FARE Cost of a monthly ticket for public transport (for 5–10 km) Negative correlation with T SHARE
a
Most of the reviewed studies used personal income and were conducted at household level, so comparisons between our and their results should be taken with caution.

3.2. Modelling approach Within the NL approach, the probability of choosing mode i is cal-
culated as the product of two probabilities:
The aim of the modelling approach is to explore which (of the
available) variables are most likely to influence modal split for
commuting in our 112 cities. The five modes considered included PðiÞ ¼ PrðMÞ  Prði=MÞ ð3Þ
C = {car (A), motorcycle (M), bicycle (B), walk (F) and public trans-
port (T)}. Following Greene (2003, pp. 686–687), Ortúzar and Wil-
lumsen (2001, p. 204), and Small and Verhoef (2007, pp. 9–10) a where Pr(M) is the probability of choosing upper level alternative
discrete choice modelling framework is used to model the factors mode M and Pr(i/M) is the marginal probability of choosing mode
that affect modal split during the morning commute period. The i conditional on choosing alternative M. An example of i in Nested
share of a given mode i is modelled by calculating the probability Structure 1 would be car whereas M would be motorised mode.
of choosing that mode P(i) such that: The marginal probability Pr(i/M) is obtained then by a MNL model
estimated across the modes listed under alternative M. On the other
PðiÞ ¼ Prðbzi þ i P bzj þ j Þ for all i–j and i; j 2 C ð1Þ
hand, the probability Pr(M) is estimated by an MNL model with the
where b is a vector of parameters, z is a vector of independent vari- alternatives listed at the upper level. For example, Pr(M) in Nested
ables, and e is an error term with infinite range. Structure 1 is an MNL model across two alternatives: motorised and
Different assumptions about the type of distribution of the error non-motorised, whereas Pr(M) in Nested Structure 2 is an MNL
terms e result in different model formulations. For instance, the model across three alternatives: private, public and other.
probability P(i) conforms to the well-known Multinomial Logit Finally, the Mixed Logit (MXL) model, which is a fairly recent ad-
model (MNL) when ei is assumed to follow a Type I extreme-value vance, offers flexibility when choosing the distribution of the error
Gumbel distribution and is identically and independently distrib- term, e. According to Train (2009), MXL allows the decomposition
uted (iid) across alternatives and observations. That is: of e into a randomised component that captures all the correlation
among the alternatives and/or decision makers, and another unob-
0
eðb zi Þ served component that is iid (i.e. correlation free). The correlated
PðiÞ ¼ P5 ðb0 z Þ ð2Þ component can be assumed to follow any theoretical distribution
j¼1 e
j

/ such as the normal, log-normal or triangular to name a few.

P(i) in Eq. (2) represents the share of mode i. The MNL model in Parameter estimates for MNL, NL and MXL models were ob-
Eq. (2) assumes that the choice of one alternative mode is indepen- tained using NLogit v.4 (Greene, 2008). The mode shares pei from
dent from the choice of any other mode (see Fig. 1a). each of the 112 cities e are used to represent the dependent vari-
If the independence assumption is violated (also known as the able P(i) in NLogit. Here, data for each city e is represented by five
independence of irrelevant alternatives (IIA) property), one can re- records that correspond to the five modes available to commuters
sort to the Nested Logit (NL) model. Such model relaxes the iid for the journey to work. As such, each of the five records is
assumption among the different alternative modes by allowing weighted by the share pei of the mode i that was chosen in that city.
for correlation to exist among certain alternatives. Within the NL A total of 560 data points (5 records times 112 cities) are then used
modelling framework, various nesting structures can be assumed to construct the log-likelihood of the logit models. Various vari-
and tested, as shown in Fig. 1. For instance, if there is reason to be- ables, as described in Section 3.1, are joined to the 560 modal share
lieve that the choice of a given motorised mode (e.g. car, motor- data points based on the city for which a particular record belongs.
cycle or public transport) and non-motorised mode (e.g. walk, The result is a database that contains the modal shares (i.e. depen-
cycle) is correlated then the choice structure in Fig. 1b can be dent variable) along with a number of socio-economic factors (i.e.
tested. Alternatively, if the choice of private modes (car or motor- independent variables) that are used to estimate the models shown
cycle) is correlated, one can test the structure shown in Fig. 1c. in Fig. 1.
 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137 133

A M B F T

(a) MNL Structure

Motorised Non-Motorised Private Public Other

A M T B F A M T B F

(b) Nested Structure 1 (c) Nested Structure 2

Fig. 1. MNL and NL model structures.

Table 4
Descriptive statistics.

Variable Minimum Maximum Median Mean Standard deviation

A SHARE 15.94 91.30 61.38 59.02 15.82
M SHARE 0.00 27.46 1.57 2.96 4.66
B SHARE 0.05 38.85 5.02 9.66 11.33
F SHARE 1.21 35.29 9.83 10.53 6.59
T SHARE 3.05 71.15 14.93 17.83 12.04
BIKE NETWORK 0.00 3.60 0.35 0.71 0.77
BUS RATE 0.17 7.91 0.76 1.36 1.43
CAR RATE 31.05 688.62 394.16 385.25 125.98
ELDERLY 7.12 26.68 16.24 16.66 3.34
GDP PER CAPITA 5432.93 41,083.78 27,131.57 25,228.38 6817.12
HOUSE WITH CHILD 14.53 59.17 26.81 28.08 10.63
MOTORCYCLE RATE 4.06 136.53 33.33 37.13 27.74
PETROL PRICE 0.85 1.42 1.12 1.12 0.17
POPULATION 100,094 500,406 214,788 237,476 9956
RAIN 67.00 239.00 169.50 162.01 46.07
STUDENTS 0.00 263.83 97.09 104.66 55.54
TRANSIT FARE 0.27 108.00 37.13 36.28 21.77

N = 112 Cities.

4. Results and discussion et al. (2004), Chen et al. (2008), Kain and Liu (2002), Kim and Ulf-
arsson (2008), Kitamura (2009), Paulley et al. (2006), Pinjari et al.
Table 5 presents statistically-significant parameters of explana- (2007), Sabir (2011), Scheiner (2010), and White (2009). As ex-
tory variables using the three model structures described above: pected, car share is also likely to increase with GDP per capita.
multinomial logit, nested logit and mixed logit. The results are con- Although our study is at city and not individual level, this seems
sistent across the three models in terms of our a priori expectations to be in line with Balcombe et al. (2004), Chen et al. (2008), Dargay
(described in Table 3) and level of significance. The results further and Hanly (2007), Kitamura (2009), Paulley et al. (2006), and Sabir
suggest that the estimated parameters can shed light on some of (2011).
the factors that might influence the modal split of commuting trips Not surprisingly, higher number of students in universities and
in mid-size European cities. further education per 1000 resident population (STUDENTS) is pos-
The MNL parameter estimates show that city-level share of car itively associated with all modes of transport except the car, which
is likely to increase as the number of registered cars per 1000 res- is an intuitive result. Other things being equal, cities with a higher
idents (CAR RATE) increases, as expected and in line with Balcombe proportion of students are prone to increased shares of public
134 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137

Table 5
Estimation results of mode share models.

MNL NL MXL
Alternative Beta t-Stats p- Beta t-Stats p- Beta t-Stats p-
Value Value Value
CONSTANT A 8.686 9.6 0.000 8.758 9.4 0.0000 6.502 4.5 0.0000
CAR RATEa A 2.175 6.7 0.000 2.247 6.3 0.0000 4.790 5.1 0.0000
GDP PER CAPITAb (+) A 0.608 10.4 0.000 0.671 5.9 0.0000 1.370 5.4 0.0000
STUDENTS M, B, F, T 0.006 6.6 0.000 0.006 6.1 0.0000 0.016 5.6 0.0000
CONSTANT M 9.640 9.9 0.000 9.723 9.7 0.0000 5.722 3.5 0.0005
MOTORCYC RATE M 0.171 7.9 0.000 0.174 7.7 0.0000 0.232 6.6 0.0000
PETROL PRICE (+) M 0.965 2.5 0.013 0.830 1.9 0.0605 3.349 3.2 0.0016
CONSTANT B 9.517 10.9 0.000 8.650 6.0 0.0000 6.532 4.2 0.0000
BIKE NETWORK B 1.003 19.5 0.000 0.992 18.4 0.0000 1.080 15.1 0.0000
CONSTANT F 9.325 10.6 0.000 8.429 5.6 0.0000 6.451 4.1 0.0000
GDP PER CAPITAb F 0.271 4.1 0.000 0.256 3.7 0.0002 0.346 5.2 0.0000
POPULATIONc T 0.427 5.2 0.000 0.422 5.0 0.0000 0.561 4.8 0.0000
GDP PER CAPITAb T 0.513 5.4 0.000 0.575 4.2 0.0000 0.635 5.5 0.0000
ELDERLY T 0.302 13.9 0.000 0.302 13.6 0.0000 0.248 6.8 0.0000
TRANSIT FARE T 0.033 9.3 0.000 0.032 8.9 0.0000 0.036 7.9 0.0000
BUS RATE T 0.211 6.6 0.000 0.203 5.8 0.0000 0.330 6.8 0.0000
RAIN (+) T 0.005 3.9 0.000 0.005 3.9 0.0001 0.001 0.4 0.7255
HOUSE WITH CHILD T 0.147 12.5 0.000 0.147 12.2 0.0000 0.118 6.2 0.0000
Inclusive values in NL model
IV(MOTORISED) 0.882 5.4 0.0000
IV(NON-MOTORISED) 1.000 () Fixed
Parameter
Derived standard deviation of randomised
parameter distribution in MXL model
GDP PER CAPITA (+) A 1.879 3.6 0.0003
PETROL PRICE (+) M 1.835 4.8 0.0000
RAIN (+) T 0.005 3.0 0.0023
L(0) 19150.82 19150.82 19150.82
L(B) 12816.43 12816.20 12782.17
Rho-square 0.331 0.331 0.333

(+) Randomised parameters.

() IV(Non-Motorised) is set to a fixed value of 1.000 to make sure it remains within the boundaries of discrete choice theory.
a
Parameter scaled by 1000.
b
Parameter scaled by 10,000.
c
Parameter scaled by 100,000.

transport and green and active modes such as bicycle, foot and car for their household mobility needs. When such an invest-
motorcycles. This is intuitive since university and college cities in ment in owning a car occurs, the household is more likely to
Europe tend to be more pedestrian oriented in terms of their land rely on the car for commuting, among other travel activities,
use development and normally benefit the existence of reliable and as such will be less likely to use public transport. Although
public transport systems. Therefore, it is not surprising to find that we are only interested in commuting trips and the data does not
cities with higher proportion of students will be associated with explicitly include trip chaining, there may be trip chaining. Drop-
commuting trips that are less dependent on the car, ceteris paribus. ping children at school or day care on the way to work would
The share of motorcycle (M SHARE) as a travel mode to work is help explain why modal split for commuting trips seems to be
more likely to be higher in cities with high numbers of registered affected by the proportion of households with children. Dargay
motorcycles per 1000 population (MOTORCYC RATE) and lower and Hanly (2007) also find a positive association between com-
with increasing petrol prices (PETROL PRICE). As expected, bicycle muting by car and the presence of children, even though they do
share (B SHARE) is positively associated with length of the bicycle not take trip chaining into account either. Chen et al. (2008)
network (BIKE NETWORK), in line with Pucher et al. (2011). Walk does include trip chaining in his commuting trips and also finds
share (F SHARE) is positively associated with GDP per capita, in line a positive association between the presence of children and the
with Sabir (2011), although Sabir (2011) estimates a model at use of car for commuting purposes. Kim and Ulfarsson (2008)
household level. model home based short trips and find the presence of children
The share of public transport (T SHARE) increases with popu- to be positively linked to the car. We did not include HOUSE
lation size, in line Schwanen (2002) and Susilo and Maat (2007) WITH CHILD in the specification of car utility because it returned
and decreases with the cost of a monthly ticket (TRANSIT FARE), a non-significant parameter, unlike the case of public transport
in line with Balcombe et al. (2004), Cervero (1998), Chen et al. utility, where it returned a significant parameter.
(2008), Paulley et al. (2006), and Zhang (2004). Public transport The number of buses (or bus equivalents) per 1000 popula-
shares are also negatively associated with the proportion of tion (BUS RATE) is positively associated with public transport
households with children aged between 0 and 17 years (HOUSE share. This is in line with Asensio (2000), Balcombe et al.
WITH CHILD). The intuition behind this is that families may (2004), Cervero (1998), Kitamura (2009), Paulley et al. (2006),
not find public transport a convenient mode of transport, as it and White (2009). GDP per capita is positively associated with
typically entails walking to and from a stop or station and wait- increased shares of public transport at the city level. Kitamura
ing, on top of the actual journey, which entails a number of (2009) is the only study that finds a positive association between
stops and sometimes a detour relative to the final destination. income and public transport, although he does so for a model at
Also, families with young children are more likely to own a household level.
 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137 135

While the elderly may not necessarily be part of the labour force icance of the estimated standard-deviation parameters, the results
and commute to work on a daily basis,7 share of public transport is suggest that GDP PER CAPITA in the car alternative, PETROL PRICE in
found to decrease with the proportion of elderly population (EL- the motorcycle alternative, and RAIN in the public transport alter-
DERLY), in line with Sabir (2011), who finds a negative association native are random parameters that follow the normal distribution.
between people over 60 and public transport as a mode to travel The estimated mean values and corresponding standard deviation
to work in the Netherlands, and Kim and Ulfarsson (2008), who also of these parameters are listed in Table 5. The estimates are in line
find a negative association for people over 65 and public transport with the results obtained from the MNL and NL models as far as
for short home-based trips in the Puget Sound region of Washington the a priori expectation and level of significance. However, the
State in the US. Interestingly, public transport share also seems to MXL model as expected corrects serial-correlation effects.
decrease with the number of days of rain per year (RAIN). This result Compared to the MNL results, the t-statistic values for most of
is not in line with the findings by Sabir (2011). However, Sumalee the estimated parameters in the MXL model are smaller, thus effi-
et al. (2011, p.339) point out that ‘anecdotal evidences also suggest cient. Furthermore, the MXL is also a significant improvement over
that people are less willing to travel by public transport or walk un- the MNL specification as inferred from the log-likelihood ratio test
der adverse weather. Sumalee et al. (2011) propose a theoretical [v2ð3Þ ¼ 2ðLLMNL  LLMXL Þ ¼ 268:52 vs: 7:81 at 5% level of
multi-modal transport network assignment model taking into ac- significance].
count uncertainties due to adverse weather conditions and find that It should be noted that some Urban Audit variables other than
the demand for bus travel decreases and the demand for under- those reported in Tables 2–4 were also tested. For instance, average
ground travel increases with rain, and also the demand for bus travel size of households was included in the utility of car (A SHARE) to
increases with rain when there are weather proof pedestrian facili- test if larger households would be more inclined to select car over
ties. The data we use do not differentiate between bus and under- other modes. Also, median disposable annual household income
ground and we do not have data on weather proof pedestrian was tried instead of GDP per capita in all equations. A proxy vari-
facilities. Our negative association therefore may be signalling that able for parking cost (namely, maximum charge of on-street park-
our case study cities do not have weather proof walking facilities ing in the city centre per hour) was also used to test if increased
and bus dominates over underground. We did not include RAIN in cost of parking in central areas was negatively associated with
the specification of bicycle utility because it returned a non-signifi- the shares of car and motorcycle. The length of public transport
cant parameter. Nonetheless, the evidence regarding the link be- network per inhabitant (km/capita) was also tried in T SHARE to
tween rain and cycling is not strong. Sabir (2011) finds a negative examine if a more developed public transport infrastructure had
association while Pucher et al. (2011) do not report any association. a positive association with the share of commuting trips by public
Further to the MNL model, we explore different NL model spec- transport. The same was done with the variable number of stops of
ifications, including the two structures shown in Fig. 1b and c. The public transport per square-km, RAIN for B SHARE and HOUSE
purpose for developing NL models is to capture potential violations WITH CHILD for A SHARE. Population density was also tested for
of the IIA property. Our a priori expectation is that some of the all equations and petrol prices for A SHARE. All the above variables
alternatives may be more related to each other. For example, car, were insignificant and in some cases had an opposite sign to the
motorcycle and public transport, being motorised modes, may expected one and as such were dropped from the specification of
share more similarities than walking and cycling as being non- the model. We believe the insignificance and/or counter-intuitive
motorised modes (Fig. 1b); similarly car and motorcycle may share signs were due to either correlation with some of the existing vari-
more similarities as being private modes vs. public transport and ables reported in Table 5 or due to too many missing values in
non-motorised modes (Fig. 1c). From all the tree specifications some of these tested variables.
mentioned, the nested structure 1 in Fig. 1b is the only specifica-
tion that results in statistically significant inclusive value (IV)
parameters. The NL-model estimation in Table 5 is based on the 5. Conclusions and policy recommendations
RU1 normalisation (Hunt, 2000).
The results from the IV parameters suggest that the NL is no dif- This paper identifies factors influencing modal split for journeys
ferent than the MNL. All the estimated parameters are consistent to work in 112 European cities with populations between 100,000
across the MNL and NL models. Although the NL implies some cor- and 500,000. To our knowledge, this is the first study that covers
relation among the motorised modes, the estimated inclusive value such a big number of cities for Europe. Previous studies are re-
is relatively high and very close to 1 (i.e. 0.88). Accordingly, the viewed, and city-level data from the Urban Audit (European Com-
correlation among motorised modes is fairly weak. Therefore, the mission, 2004) together with national petrol prices from the
MNL structure in Fig. 1a can still be used to explain the observed International Energy Agency (2008) and GDP per capita from the
mode split shares. However, the estimated parameters in the IMF website are used.
MNL model are most likely inefficient due to serial correlation.8 Three models are estimated: a multinomial logit, a nested logit
The latter is caused by the repeated city-level values used to con- and a mixed logit. Virtually all the results are in line with expecta-
struct the independent variables across the 5 records within a given tions and findings reported in the literature. Car ownership is pos-
city (see Section 3.2). To account for any potential serial correlation itively correlated with the share of commuting trips by car.
in the data, a mixed logit (MXL) model was also estimated. The MXL Therefore, policies aimed at discouraging car ownership, such as
also accounts for any unobserved heterogeneity (i.e. random effects) high registration fees or annual excise duties, may help reduce
in the estimated parameters. the share of car use for trips to work. Unsurprisingly, GDP per
All parameters in the MNL model are tested for potential random capita is also positively associated with car share. The number of
effects, and three of those emerge as significant. Based on the signif- students in universities and further education per 1000 resident
population (STUDENTS) is positively associated with all modes of
transport except the car.
7
We note that in Europe, Directive 2000/78/EC (European Commission, 2000), Public transport shares are negatively associated with the cost
establishing a general framework for equal treatment in employment and occupation,
of a monthly ticket, which would point towards the idea of cheaper
does not allow age discrimination. Most EU member states have age discrimination
legislation in place as a result, and many had it even before the directive.
(perhaps, subsidised) fares in order to increase the share of public
8
An inefficient parameter is usually associated with an inflated t-statistics value transport in commuting trips to work. Unfortunately, we do not
since the standard error of the parameter is usually small. have data on public-transport subsidies across the 112 cities exam-
136 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137

ined and this is an area worth of research. Leaving to one side that Bratzel, S., 1999. Conditions of success in sustainable urban transport policy – policy
change in ‘relatively successful’ European cities. Transport Reviews 19 (2), 177–
there is some economic efficiency justification for public transport
190.
subsidies (Parry and Small, 2009), White (2009) reports a positive Buchanan, C., 1964. Traffic in Towns: The Specially Shortened Edition of the
relationship between public transport subsidies and public trans- Buchanan Report. Penguin Books, Harmondsworth.
port use. Policies aimed at attracting potential public transport Cervero, R., 1998. The Transit Metropolis: A Global Inquiry. Island Press,
Washington, DC.
users, including current car users, with cheap fares are ‘carrot’ pol- Chen, C., Gong, H., Paaswell, R., 2008. Role of the built environment on mode choice
icies. Perhaps they could and should be complemented with ‘stick’ decisions: additional evidence on the impact of density. Transportation 35 (3),
policies, discouraging car ownership (as discussed above) and car 285–299.
Dargay, J., Hanly, M., 2007. Volatility of car ownership, commuting mode and time
use. Policies designed to reduce car use include road pricing and in the UK. Transportation Research Part A: Policy and Practice 41 (10), 934–948.
high parking charges. None of the cities in our sample has road European Commission, 2000. Directive 2000/78/EC of the European Parliament and
pricing and we do not have enough data on parking charges to test of the Council Official Journal of the European Union L303, 02/12/2000, pp. 16–
22.
such hypotheses,9 so more research is needed on this front as data European Commission, 2004. Urban Audit: Methodological Handbook. 2004 edition,
become available. Directorate-General for Regional Policy and Eurostat, Luxembourg: Office for
From our model results, it is clear that policies in favour of pub- Official Publications of the European Communities.
European Commission, 2008. First Thematic Research Summary: Economic Aspects
lic transport, such as reducing fares and increasing the number of of Sustainable Mobility, DG Energy and Transport. <http://www.transport-
buses, are likely to increase the share of public transport in trips research.info/Upload/Documents/201002/20100215_112235_50880_TRS%20
to work. Finally, there is a clear positive association between bicy- Economic%20 Aspects%20i1.2%20011008.pdf> (accessed 18.04.10).
Frank, L., Bradley, M., Kavage, S., Chapman, J., Lawton, T.K., 2008. Urban form, travel
cle share and the length of bicycle network. Reallocating road
time, and cost relationships with tour complexity and mode choice.
space from motorised transport to bicycles (for example by desig- Transportation 35 (1), 37–54.
nating cycle lanes) and even building cycle lanes are relatively Greene, W.H., 2003. Econometric Analysis, fifth ed. Prentice Hall, Englewood Cliffs.
inexpensive policies that are likely to achieve an increase in bicycle Greene, W.H., 2008. NLOGIT Ver. 4, New York & Sydney: Econometric Software Inc.
Habib, K.M.N., Tian, Y., Zaman, H., 2011. Modelling commuting mode choice with
share of trips to work. explicit consideration of carpool in the choice set formation. Transportation 38
There are some caveats in this study. First, it would have been (4), 587–604.
more accurate to conduct the analysis for ‘functional urban areas’ Hunt, G.L., 2000. Alternative nested logit model structures and the special case of
partial degeneracy. Journal of Regional Science 40 (1), 89–113.
rather than for ‘cities’ as defined by the administrative city bound- International Energy Agency, 2008. Energy Prices and Taxes, Quarterly Statistics, 1st
aries. Traffic does not recognise administrative boundaries. Also, Quarter 2008, Paris: OECD/IEA.
traffic problems can be better addressed by considering functional Kain, J.E., Liu, Z., 2002. Efficiency and locational consequences of government
transport policies and spending in Chile. In: Glaeser, E.I., Meyer, J.R. (Eds.), Chile:
urban areas (Bratzel, 1999). Second, since our database (except for Political Economy of Urban Development, John E. Harvard University Press,
petrol prices and GDP per capita) is the Urban Audit, we are unable Kennedy School of Government, pp. 105–195.
to include some variables which may help explain modal split be- Kenworthy, J.R., Laube, F.B., Newman, P.W.G., Barter, P.A., Raad, T., Poboon, C., Guia,
B., 1999. An international sourcebook of automobile dependence in cities,
cause there are either too many missing values or the values are 1960–1990. University Press of Colorado, Boulder.
missing altogether. These omitted variables include average dis- Kim, S., Ulfarsson, G.F., 2008. Curbing automobile use for sustainable
tance between home and work for each city, average monetary transportation: analysis of mode choice on short home-based trips.
Transportation 35 (6), 723–737.
(out-of-pocket) costs of each mode in each city, public transport
Kitamura, R., 2009. A dynamic model system of household car ownership, trip
subsidies, parking charges, some measure of pedestrian zones, generation, and modal split: model development and simulation experiment.
and some measure (perhaps an index) of land use mix. In addition Transportation 36 (6), 711–732.
to that, the very dependent variable, mode share, does not include Newman, R., Dellinger, A., Anderson, M., Bonomi, A., Rivara, F., Thompson, R., 2009.
‘Preferred modes of travel among older adults: What factors affect the choice to
any trip chaining, and this would have been useful to consider. walk instead of drive?’. Journal of Safety Research 40 (5), 395–398.
Third, as already highlighted in the introduction, modal share is Newman, P.W.G., Kenworthy, J.R., 1989. Cities and Automobile Dependence: A
an important issue for analysis and policy, but total vehicle kilome- Sourcebook. Gower, Aldershot.
Newman, P.W.G., Kenworthy, J.R., 1999. Sustainability and Cities: Overcoming
tres travelled by the different modes would have complemented Automobile Dependence. Island Press, Washington, DC.
and enhanced the study, especially from a sustainability Newman, P.W.G., Kenworthy, J.R., 2006. Urban design to reduce automobile
perspective. dependence. Opolis: An International Journal of Suburban and Metropolitan
Studies 2 (1), 35–52.
Ortúzar, J.D., Willumsen, L.G., 2001. Modelling Transport, third ed. John Wiley &
Sons Ltd., Chichester.
Acknowledgements Parry, I.W.H., Small, K.A., 2009. Should urban transit subsidies be reduced?
American Economic Review 99 (3), 700–724.
Paulley, N., Balcombe, R., Mackett, R., Titheridge, H., Preston, J., Wardman, M.,
We are grateful to the School of Planning and Geography at Car- Shires, J., White, P., 2006. The demand for public transport: the effects of fares,
diff University for financial assistance. We are also indebted to two quality of service, income and car ownership. Transport Policy 13 (4), 295–306.
anonymous reviewers for helpful comments on earlier versions of Pinjari, A.R., Pendyala, R.M., Bhat, C.R., Waddell, P.A., 2007. Modeling residential
sorting effects to understand the impact of the built environment on commute
this paper. mode choice. Transportation 34 (5), 557–573.
Pucher, J., Buehler, R., Seinen, M., 2011. Bicycling renaissance in North America? An
update and re-appraisal of cycling trends and policies. Transportation Research
References Part A: Policy and Practice 45 (6), 451–475.
Rodrigue, J.P., Comtois, C., Slack, B., 2009. The Geography of Transport Systems,
second ed. Routledge, London.
Antipova, A., Wang, F., Wilmot, C., 2011. Urban land uses, socio-demographic
Sabir, M., 2011. Weather and Travel Behaviour. PhD Thesis, Department of Spatial
attributes and commuting: a multilevel modelling approach. Applied
Economics, Faculty of Economics and Business Administration, Vrije
Geography 31 (3), 1010–1018.
Universiteit Amsterdam. <http://dare.ubvu.vu.nl/bitstream/handle/1871/
Asensio, J., 2000. The success story of Spanish suburban railways: determinants of
19500/dissertation.pdf?sequence=1>.
demand and policy implications. Transport Policy 7 (4), 295–302.
Santos, G., Behrendt, H., Teytelboym, A., 2010. Part II: policy instruments
Balcombe, R., Mackett, R., Paulley, N., Preston, J., Shires, J., Titheridge, H., Wardman,
for sustainable road transport. Research in Transportation Economics 28, 46–
M., White, P., 2004. The Demand for Public Transport: A Practical Guide. TRL
91.
Report 593. <http://www.demandforpublictransport.co.uk/TRL593.pdf>
Scheiner, J., 2010. Interrelations between travel mode choice and trip
(accessed 09.06.10).
distance: trends in Germany 1976–2002. Journal of Transport Geography 18
Black, W., 2010. Sustainable Transportation: Problems and Solutions. The Guilford
(1), 75–84.
Press, New York.
Schwanen, T., 2002. Urban form and commuting behaviour: a cross-European
perspective. Tijdschrift voor Economische en Sociale Geografie 93 (3), 336–343.
9
As explained above, we tested parking charges but the variable had a counter- Small, K., Verhoef, E., 2007. Economics of Urban Transportation. Routledge, London.
intuitive sign and too many missing values.
 G. Santos et al. / Journal of Transport Geography 30 (2013) 127–137 137

Souche, S., 2010. Measuring the structural determinants of urban travel demand. Van Acker, V., Witlox, F., 2011. Commuting trips within tours: how is commuting
Transport Policy 17 (3), 127–134. related to land use? Transportation 38 (3), 465–486.
Sumalee, A., Uchida, K., Lam, W.H.K., 2011. Stochastic multi-modal transport White, P., 2009. Public Transport: Its Planning, Management and Operation, fifth
network under demand uncertainties and adverse weather condition. ed. Routledge, London.
Transportation Research Part C: Emerging Technologies 19 (2), 338–350. Winston, C., Shirley, C., 1998. Alternate Route: Towards Efficient Urban
Susilo, Y., Maat, E., 2007. The influence of built environment to the trends in Transportation. Brookings Institution Press, Washington DC.
commuting journeys in the Netherlands. Transportation 34 (5), 589–609. Zhang, M., 2004. The role of land use in travel mode choice: evidence from Boston
Train, K., 2009. Discrete Choice Methods with Simulation, second ed. Cambridge and Hong Kong. Journal of the American Planning Association 70 (3), 344–360.
University Press, Cambridge.