You are on page 1of 14

Environment and Planning A 2002, volume 34, pages 1207 ^ 1220

DOI:10.1068/a349

Urban form, road network type, and mode choice for


frequently conducted activities: a multilevel analysis using
quasi-experimental design data

Danielle Snellen, Aloys Borgers, Harry Timmermans


Urban Planning Group, Eindhoven University of Technology, PO Box 513, 5600 MB Eindhoven,
The Netherlands; e-mail: h.j.p.timmermans@bwk.tue.nl
Received 23 February 2001; in revised form 4 March 2002

Abstract. The question of whether characteristics of urban form influence travel behaviour has a long
tradition in transportation and urban planning. Positive evidence has been found, however, that
serious methodological objections can be raised against many of these studies. Often no explicit
attempt was made to disentangle the effects of spatial characteristics against personal and household
characteristics. In addition, if both kinds of variables were included in the analysis, researchers were
often not aware of the specific statistical problems that relate to the fact that the observations are
made at different levels of aggregation. In this paper, therefore, the authors adopt a multilevel analysis
to examine the relationship between urban form and travel behaviour.

Introduction
One of the most common goals of current transportation policies in the Western world
is to reduce the growth of car use and stimulate the use of public transport. Because
this goal is difficult to achieve, a wide variety of economic, transportation, and spatial
policies have been formulated and implemented. The planning and design of neigh-
bourhoods should also play a role. Planners and urban designers like to believe that the
form concept underlying the neighbourhood, and some other urban characteristics
are strongly related to transportation mode and other activity-location decisions. In
particular, it is believed that the right urban design will stimulate the use of public
transport, thereby reducing car use. This belief is supported by empirical evidence
reported in the literature. For example, several authors have reported positive effects
of neotraditional and related designs on mode choice (Cervero, 1996; Cervero and
Gorham, 1995; Florez, 1998; Friedman et al, 1994; Nasar, 1997). Other studies report
empirical evidence of the relationship between mode choice and higher densities or
mixed land use (for example, Frank and Pivo, 1994; Handy, 1993; Jacobsen, 1997;
Kitamura et al, 1997; N×ss and Sandberg, 1996; N×ss et al, 1995; Schimek, 1996).
Other studies on the influence of urban form on travel behaviour in North America
and Europe include those by Breheny (1995), Gordon and Richardson (1996),
Giuliano and Small (1993; 1999), and Owens (1986).
Unfortunately, however, most of this stream of research potentially has methodolog-
ical flaws. Some studies relate neighbourhood properties directly to transport-mode
choice behaviour. A simple example will show how this may lead to invalid conclu-
sions. In the Netherlands, there are many examples of lower income households in
older, high-density, mixed-land-use neighbourhoods. The finding that mobility in such
neighbourhoods may be lower can also be explained by the lower mobility of this
group, and not just by neighbourhood characteristics such as high density and mixed
land use. As Kitamura et al (1997) have argued, ``Is the observed association between
travel and land use real, or is it an artefact of the association between land use and the
multitude of demographic, socio-economic, and transportation supply characteristics,
which also are associated with travel?'' Hence, the design of these studies may not
1208 D Snellen, A Borgers, H Timmermans

allow one to draw any firm conclusions about the assumed relationship between urban
form and transport-mode choice. The relationships may be spurious.
Other studies, which did include both personal or household and spatial character-
istics may be flawed for statistical reasons. The analysis of variables that relate to different
levels of measurement (for example, individual versus neighbourhoods) creates specific
statistical problems. The aggregation of data results in fewer values for fewer units of
observation, implying that the power of the statistical analysis is reduced. Likewise,
disaggregation implies that the sample size is arbitrarily increased. Consequently, signifi-
cance tests will more often reject the null hypothesis than the alpha level suggests,
implying that the researcher generates spurious results.
To avoid these problems, the present study uses a multilevel model and quasi-
experimental design data to assess the impact of urban form on transportation-mode
choice decisions. Multilevel analysis is very appropriate to identify the contribution of
various variables, specified at different levels, to the variance in the dependent variable.
Thus, the contribution of this paper to the literature is threefold: first, it adds a
European example to the growing list of North American land-use ^ travel studies;
second, it is one of the first applications of quasi-experimental methods of sample
selection to a land-use ^ travel study; and third, to the best of our knowledge, it is the
first application of multilevel analysis to this subject.
To this end, the paper is organised as follows. First, for readers unfamiliar with
multilevel analysis, we will summarise the quintessence of this statistical method. Next,
we discus the concepts of urban form and road network type that underlie our study,
followed by a description of the data collection. Then, we will report the results of our
analyses, and complete the paper by discussing some policy implications of our results.

Multilevel analysis
Consider the relationship between the number of trips a person makes on a certain
weekday and a single individual-level independent variable (say, work status) within a
single, hypothetical neighbourhood. Assuming this relationship is linear, the regression
equation for this relationship would be:
Yi ˆ b0 ‡ b1 Xi ‡ ri , (1)
where
Yi is the number of trips,
Xi is the work status of individual i (0 means without a job, 1 means with a job),
b0 is the expected number of trips for an individual i without a job,
b1 is the expected change in the number of trips for an individual i with a job,
ri is the error term, representing a unique effect associated with individual i
(normally distributed with a mean of zero and variance s 2).
Now, consider individuals distributed across neighbourhoods. The relationship between
the number of trips and individual work status in each neighbourhood j can then be
described by the following equation:
Yij ˆ b0j ‡ b1j Xij ‡ rij . (2)
The subscript j is added to the equation to allow each neighbourhood to have a unique
intercept and slope.
Let us now assume that the neighbourhoods vary with regard to their mix of land
uses. This can be included in the equation as follows:
b0j ˆ g00 ‡ g01 Wj ‡ u0j , (3)
b1j ˆ g10 ‡ g11 Wj ‡ u1j , (4)
A multilevel analysis using quasi-experimental design data 1209

where
g00 is the mean number of trips for an individual without a job in a monofunctional
neighbourhood (the intercept for level 1),
g01 is the mean difference in number of trips between monofunctional and multifunc-
tional neighbourhoods (the slope for level-1 intercept),
g10 is the average work-status slope in monofunctional neighbourhoods (the intercept
for the level-1 slope),
g11 is the mean difference in work-status slopes between monofunctional and multi-
functional neighbourhoods (the slope for the level-1 slope),
Wj is the land use in neighbourhood j,
u0j is the unique effect of neighbourhood j on the number of trips, with Wj held
constant,
u1j is the unique effect of neighbourhood j on the work status slope, with Wj held
constant.
It is assumed that u0j and u1j are random variables with zero means, variances t00 and
t11 , respectively, and covariance t01 . These variance ^ covariance components are con-
ditional and represent the variability in b0j and b1j remaining after controlling for Wj .
Substituting equations (3) and (4) in equation (2) yield
Yij ˆ g00 ‡ g01 Wj ‡ g10 Xij ‡ g11 Wj Xij ‡ u0j ‡ u1j Xij ‡ rij . (5)
In this equation we consider the first part (g00 ‡ g01 Wj ‡ g10 Xij ‡ g11 Wj Xij ) as the fixed
part of the model, and the second part (u0j ‡ u1j Xij ‡ rij ) as the random part of the
model. This is the basis for multilevel analysis or hierarchical linear modelling.
Maximum likelihood (ML) estimation can be used to estimate the parameters of
the equation (Hox, 1995). Computing the ML estimates requires an iterative procedure.
First, single-level ordinary least squares estimates are derived as starting values for the
various parameters. After one iteration, this gives generalised least squares estimates
and, when the iterative process converges, it results in ML estimates. Two variations
of ML estimation are commonly used in multilevel regression analysis. When full
maximum likelihood (FML) is used the regression coefficients and the variance com-
ponents are both included in the likelihood function. In contrast, if one uses restricted
maximum likelihood (RML), only the variance components are included in the like-
lihood function. FML treats the estimates for the regression coefficients as known
quantities when the variance components are estimated, whereas RML treats them as
estimates that carry some degree of uncertainty. RML is more realistic, and should
lead to better estimates, especially when the number of groups is small (Bryk and
Raudenbush, 1992).
The choice of RML estimation does, however, pose a limitation in terms of model
performance and hypothesis testing (Snijders and Bosker, 1999). The parameter esti-
mation results in a likelihood value. The statistic `minus twice the natural logarithm of
the likelihood', called the deviance, which is w 2 distributed can be regarded as an
indication of the lack of fit between the model and the data. It can also be used to
compare the performance of alternative models. However, this method can be applied
only when the two models are hierarchical. If RML estimation is used, deviance tests
can be used only if the models being compared have the same fixed parts and differ
only in their random parts.
Model performance assessment can also be assessed by calculating how much of
the variation in the dependent variable is explained by the independent variables of the
model. In normal linear regression, the measure for this is R 2, the squared multiple
correlation coefficient. For the hierarchical linear model, however, the assessment of
the explained proportion of variance is more complicated. Snijders and Bosker (1999)
1210 D Snellen, A Borgers, H Timmermans

discuss alternative measures of R 2 for hierarchical linear models. They present separate
measures for the explained proportion of variance at both levels, for a random inter-
cept model (a model that contains only error terms associated with the intercept). The
explained proportion of variance on level 1 (R12 ) is the proportional reduction in the
value of s 2 ‡ t02 due to inclusion of independent variables. Given two models A and B,
the explained proportion of variance of model B over model A is calculated as follows:
sB2 ‡ t0B
2
R12 ˆ 1 ÿ , (6)
sA2 ‡ t0A
2

where
sA2 is the variance of rij in model A,
sB2 is the variance of rij in model B,
2
t0A is the variance of u0j in model A,
2
t0B is the variance of u0j in model B.
The explained proportion of variance on level 1 (R22 ) is the proportional reduction in
mean squared prediction error. R22 is estimated as the proportional reduction in the
value of (s 2 =n) ‡ t02, where n is a representative value for the group size. Again, given
the models A and B to compare, the explained proportion of variance is calculated as
follows:

sB2 sA2
R22 ˆ 1 ÿ . (7)
n ‡ t0B2
n ‡ t0A2

A complicating factor here is the assessment of n. Sometimes a representative value of n


is easily available. For example, in the case of the performance of schoolchildren grouped
into classes, and a normal class size of 30 children, n ˆ 30 even if the values of nj in the
data set are not on average 30. In the case of varyingP group sizes, and lack of clarity on a
representative group size, the harmonic mean [N= j (1=nj )] can be used.
For a model that contains random slopes, the calculation of R12 and R22 is more
complicated. Snijders and Bosker (1999) propose a method that closely approximates
the exact values of R12 and R22. One reestimates the random slopes model as a random
intercept model with the same fixed part. Then R12 and R22 are calculated in the usual way.
Introductory explanations to multilevel analysis can be found in Bullen et al (1997),
Goldstein (1995), and Kreft and de Leeuw (1998).

Data
As indicated in the introduction, the goal of this research project was to analyse the
relationship between urban form, road network type, and indicators of mobility. A
potential danger of conventional data collection in this area of research is that char-
acteristics of cities and neighbourhoods may be strongly correlated with personal and
household characteristics. Consequently, the explanatory variables of the models may
be strongly correlated, leading to bias in the estimates and even under extreme circum-
stances to invalid and misleading conclusions. Ideally, to avoid such problems, one
would like to use an experimental design. However, because we are analysing real
neighbourhoods, this ideal is impossible to realise. The next best solution is to use
quasi-experimental design data. This means that one selects the neighbourhoods in
such a way that the correlations between the explanatory variables are relatively small
and that the selection reflects differences in urban form and transportation networks.
Selection of neighbourhoods for data collection was done in two stages. First, a
number of cities were selected so as to reflect differences in morphological urban form
and network types for the main road system for motorised transport. Within these
cities neighbourhoods were selected.
A multilevel analysis using quasi-experimental design data 1211

Typology of urban form


Based on the literature on historic and more recent urban forms (for example,
de Klerk, 1980; Rottier, 1980), we identified six different forms (figure 1).
The concentric city represents the most common morphological form in the
Netherlands. It is the typical form of a city that has grown from a small (historic)
centre along a number of radial exit roads. Over the years the areas between the
radial roads have been filled and the city has received its concentric form. This city
shape is often associated with a radial road network, based on the historic routes, but
combinations with ring or grid networks also exist. Often these cities are quite compact
and have a strong centre with a mixed supply of facilities.

(a) (b) (c) (d) (e) (f )


Figure 1. Morphological urban forms: (a) the concentric city, (b) the lobe city, (c) the linear
polynuclear city, (d) the concentric polynuclear city, (e) the linear city, (f) and the grid city.

The lobe city often has a similar history. The main differences are that the city has
developed between some radial roads and not between others. The lobe shape can
result from roads stretching out in some directions or from intentional city planning.
The linear polynuclear and concentric polynuclear cities have a number of things in
common. The morphologies of these cities can develop in two very different ways.
Either a number of smaller settlements, located close to each other, start to function
as one city, or a city is actually designed as a polynuclear city. The only example of a
true polynuclear city in the Netherlands is Almere, which can best be described as
a concentric polynuclear city.
Grid cities is the collective noun for cities with a more or less rectangular shape.
The linear city can be viewed as an extreme version of the city form. These cities often
have a grid-type transportation network for the car, but this is not necessarily the case.
In the selection of locations for data collection these four types (concentric city, lobe
city, polynuclear city, and grid city) were included.
Typology of transportation networks
The second basic element of an urban concept is the transportation network. In theory,
networks for motorised transport (and road network for the car), nonmotorised trans-
port (network for bicycles and/or pedestrians), and for public transport can be
distinguished. However, because not all cities have separate networks for nonmotorised
transport and the public transport networks in all cities are very similar, only networks
for motorised transport were considered in the choice of cities. To identify different
network types, Bolt's (1982) study on urban form and transportation is very useful.
Bolt distinguishes between two principle network forms, the square network and the
triangular network (a corrupted depiction of the `ideal' circular network). On basis of
these two principles, he derived five elementary networks (figure 2, see over).
The radial network is suitable for all transportation modes. In the Netherlands it is
the classic network for urban public transport, where buses radiate to and from the city
centre or railway station. When used as a network for the other transport modes it
offers direct accessibility to the city centre. However, this network type is also prone to
congestion problems. In theory, it can be applied to all morphological urban forms as
described in the previous section.
1212 D Snellen, A Borgers, H Timmermans

(a) (b) (c) (d) (e)


Figure 2. Elementary transportation networks: (a) the linear network, (b) the radial network, (c)
the ring, (d) the grid, and (e) the shifted grid.

The ring network is frequently used in the Netherlands, mainly as a network for
motorised transport. It offers the opportunity to concentrate a large amount of traffic
on a single road, while other areas are relieved of traffic nuisance. In particular city
centres are often enclosed by a ring structure. This network type is not common for
public transport in medium-sized cities, although it would provide good connections
between city districts while avoiding a trip to and from the city centre. For motorised
and nonmotorised transport this network type can be applied to all morphological urban
forms. For public transport, this network type is especially suitable for concentric
polynuclear cities.
The (shifted) grid network is simple and direct, offers many choices of routes, and
disperses traffic over many streets. Disadvantages of this network type are the many road
crossings in the system. It is used mainly for motorised and nonmotorised transport. In
theory, this network type is flexible and can be applied to all morphological urban forms.
In selecting the cities, we did not consider the linear network as there are no major
cities in the Netherlands with that type of network, and the grid and shifted grid were
combined.
Selected cities
These two basic characteristics were used to choose cities for data collection. Table 1
lists the selected cities and their characteristic urban form and transportation system.
The table shows that not all cells could be represented. Several combinations of urban
shape and main network type for motorised transport, although possible in theory,
are not logical and therefore do not occur in the Netherlands. An example would be a
lobe city with a grid network. Furthermore, there is only one polynuclear city in the
Netherlands (Almere) and this city has a ring network. Consequently, it is not possible
to find other network types in combination with a polynuclear city. The city of
Eindhoven needs further explanation. The main network for motorised transport
in Eindhoven shows distinct characteristics of both a radial and a ring network. In the
southern part of the city, the ring network type is dominant, whereas in the northern
part of the city the radial network type is dominant. This difference will be reflected in
the choice of neighbourhoods within the city. These considerations should be kept
in mind when interpreting the results of the analyses.
Table 1. Selected cities.

Concentric city Lobe city Polynuclear city Grid city

Radial network Enschede Arnhem


Eindhoven
Ring network Zoetermeer Eindhoven Almere Leeuwarden
Grid network Apeldoorn Haarlem
Den Haag
A multilevel analysis using quasi-experimental design data 1213

Once the cities had been selected, the next step involved choosing neighbourhoods
within these cities. Because activity ^ travel patterns are known to be influenced by the
relative location of individuals and households to such elements as the downtown area,
the main intercity train station and, if present, regional shopping centres, neighbourhoods
were selected so that they would have low correlation on these dimensions. In particular,
three central considerations dictated the choice of neighbourhood: (1) two neighbour-
hoods per city; (2) two neighbourhoods in every ring from the city centre and (3) best
possible combinations with regard to the other dominant urban elements. Nineteen
neighbourhoods were selected (table 2). The characteristics of the chosen neighbourhoods
were tested for correlations to check if the chosen neighbourhoods actually allow us to
test the relationship between characteristics of the urban environment and activity and
travel patterns. Correlations were calculated for ten characteristics: morphological urban
form, city-road network type, neighbourhood-road network type, local-street network
type, employment (number of jobs or businesses), location of city within the Netherlands
(inside or outside the Randstad Holland), distance from neighbourhood to city centre,
distance from neighbourhood to an intercity station, distance from neighbourhood to the
subcentre, and density (inhabitants and/or dwellings). It should be noted that urban form
and network type were used as discrete, classifactory variables. One might argue that in
reality they are more like continuous variables.
With this in mind, the results indicated that most characteristics were only (modestly)
correlated (table 3, over). More specifically, table 3(a) shows that only a few spatial
characteristics showed associations exceeding 0.5. Distance to city centre is highly corre-
lated with distance to the intercity train station because the intercity train station is
usually located in the city centre. There is also a significant correlation between urban
form and employment. Furthermore, we see two considerable correlations between
distance to subcentre and network type at the neighbourhood level, and distance
Table 2. Selected neighbourhoods.
City Neighbourhood Distance to Distance to Distance to
city centre (km) station (km) subcentre (km)
Almere Muziekwijk 2±3 3
Regenboogbuurt 5±6 6
Apeldoorn Matenveld/Matendonk 3±4 4
Ugchelen 4±5 5
Arnhem Presikhaaf 3±4 5 1
Alteveer/Cranevelt 2±3 3
Den Haag Centrum 0±1 2
Beresteijn 5±6 4 2
Voorburg Essesteijn 3±4 6 2
Eindhoven Stratum 1±2 2
Achtse Barrier 5±6 6 3.5
Enschede Stadsveld/Bruggert 2±3 3
Helmerhoek Zuid 4±5 5
Haarlem Amsterdamse buurt/ 1±2 2
Slachthuisbuurt
Schalkwijk Zuid 3±4 4.5 1.5
Leeuwarden Centrum 0±1 1
Camminghaburen 3±4 4.5
Zoetermeer Centrum 0±1 2
Rokkeveen 2±3 1
1214 D Snellen, A Borgers, H Timmermans

Table 3. Associations (a) between neighbourhood characteristics and (b) between neighbourhood
characteristics and sociodemographics.

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

(a) Associations between neighbourhood characteristics


Urban form ns ns ns e 2 ˆ 0:77 ns
2
City-road l ˆ 0:42 ns e ˆ 0:07 ns
network type (2)
Neighbourhood- l ˆ 0:15 e 2 ˆ 0:35 ns
road network
type (3)
Local-street e 2 ˆ 0:43 ns
network type (4)
Employment (5) e 2 ˆ 0:40
Location (6)
Distance to
city centre (7)
Distance to
intercity
station (8)
Distance to
subcentre (9)
Degree of
urbanisation of
neighbourhood (10)
Degree of
urbanisation
of city (11)

(b) Associations between neighbourhood characteristics and sociodemographics


Age e 2 ˆ 0:01 e 2 ˆ 0:02 e 2 ˆ 0:04 ns e 2 ˆ 0:00
Household composition l ˆ 0:04 ns ns e 2 ˆ 0:08 ns
Household income l ˆ 0:04 ns ns e 2 ˆ 0:03 ns
2
Dwelling type l ˆ 0:08 ns l ˆ 0:11 e ˆ 0:03 ns
Gender ns ns ns e 2 ˆ 0:00 ns
2
Educational level ns ns ns e ˆ 0:00 ns
Personal income ns ns l ˆ 0:04 e 2 ˆ 0:01 ns
2
Job situation ns ns ns e ˆ 0:01 ns
Car ownership ns ns ns e 2 ˆ 0:01 ns
Note. ns, not statistically significant.

to subcentre and degree of urbanisation of the neighbourhood. Because distance to


subcentre is included only for those neighbourhoods that are closer to a subcentre than
to the city centre itself, the number of cases for this correlation is very low (N ˆ 5). This
is the most likely cause of these strong correlations. Most importantly, however, there is
no evidence of any significant correlation between the selected physical attributes and the
sociodemographic characteristics [table 3(b)]. In fact, all these correlations are very low
indeed.
A multilevel analysis using quasi-experimental design data 1215

Table 3 (continued).

(7) (8) (9) (10) (11)

(a) Associations between neighbourhood characteristics


Urban form e 2 ˆ 0:06 e 2 ˆ 0:06 e 2 ˆ 0:06 l 2 ˆ 0:39 l 2 ˆ 0:43
City-road e 2 ˆ 0:12 e 2 ˆ 0:17 e 2 ˆ 0:06 ns l ˆ 0:48
network type (2)
Neighbourhood- e 2 ˆ 0:14 e 2 ˆ 0:26 e 2 ˆ 0:86 ns ns
road network
type (3)
Local-street e 2 ˆ 0:48 e 2 ˆ 0:44 e 2 ˆ 0:10 e 2 ˆ 0:38 ns
network type (4)
Employment (5) ns ns ns e 2 ˆ 0:04 e 2 ˆ 0:32
Location (6) e 2 ˆ 0:01 e 2 ˆ 0:02 e 2 ˆ 0:06 ns ns
Distance to 0.845 ns e 2 ˆ 0:25 e 2 ˆ 0:10
city centre (7)
Distance to ns e 2 ˆ 0:19 e 2 ˆ 0:09
intercity
station (8)
Distance to e 2 ˆ 0:95 e 2 ˆ 0:06
subcentre (9)
Degree of ns
urbanisation of
neighbourhood (10)
Degree of
urbanisation
of city (11)

(b) Associations between neighbourhood characteristics and sociodemographics


Age ns 0.105 ns e 2 ˆ 0:06 e 2 ˆ 0:00
Household composition e 2 ˆ 0:05 e 2 ˆ 0:03 e 2 ˆ 0:08 ns ns
2 2
Household income e ˆ 0:01 e ˆ 0:02 e 2 ˆ 0:05 ns ns
2 2 2
Dwelling type e ˆ 0:09 e ˆ 0:05 e ˆ 0:06 l ˆ 0:10 ns
Gender e 2 ˆ 0:00 e 2 ˆ 0:00 e 2 ˆ 0:01 ns ns
Educational level e 2 ˆ 0:03 e 2 ˆ 0:02 e 2 ˆ 0:04 ns ns
2 2
Personal income e ˆ 0:01 e ˆ 0:01 e 2 ˆ 0:03 ns ns
2 2 2
Job situation e ˆ 0:02 e ˆ 0:01 e ˆ 0:02 ns ns
Car ownership e 2 ˆ 0:02 e 2 ˆ 0:02 e 2 ˆ 0:06 ns ns

Data about activities


To analyse the relationship between urban form and activity patterns, data about
activities and related transport need to be collected. The activity diary survey used in
the present study consisted of four parts. In the first part, information about a number of
frequently conducted activities was collected. These activities included working, shopping,
leisure activities, and taking and/or fetching children to and from various locations. For
each of these activities, respondents were asked to report when and where these activities
are usually performed, the origins of these trips, the transportation modes used, and the
1216 D Snellen, A Borgers, H Timmermans

travel times.. These data constitute the input to the analyses reported in this paper,
together with the data collected in the last part of the survey instrument relating to
sociodemographics: year of birth, gender, level of education, personal income, possession
of a driver's license, ability to use transport modes, type of dwelling, location in relation
to public transport, and possession of car and/or cycle. The actual diary is not used in the
analyses reported in this paper. The results of similar analyses, conducted on the activity
diary will be reported elsewhere.
Response rates
In each neighbourhood 300 randomly selected households received a letter requesting
their participation in this research project, explaining the aims and objectives of
the project, and briefly describing the tasks that respondents were asked to perform. Of the
households 10% responded positively and 594 households received the package with
the questionnaire. The initial response varied widely over the neighbourhoods, with a
minimum of 5% and a maximum of 16%. Every participating household received as
many questionnaires as there were individuals aged 13 years and over in the household.
A total of 344 households returned 586 individual questionnaires. The secondary response
varied between 30% and 74%. Secondary nonresponse was significantly related to the
place of residence of the respondents. Cities within the Randstad Holland (the most
urbanised area) experienced a higher nonresponse, and the same held true for the city
centres and the most peripheral neighbourhoods. Household type and household size
were not significantly related to secondary nonresponse.

Multilevel analysis
Dependent variables
Multilevel models were separately estimated for four types of frequent trip (work,
grocery shopping, other shopping, and leisure). Two levelsöneighbourhood and indi-
vidualöwere used. In principle, a third levelöthe cityöcould be added, but this
option was not pursued because of the relatively low number of cities. Separate models
were estimated for (1) the total number of kilometres, (2) the total number of trips, (3)
the number of kilometres per mode, (4) the number of trips per mode, and (5) mode
shares for both kilometres and trips. Models were estimated with data from respond-
ents who indicated that they actually participated in these activities. The models for
total kilometres and trips (by mode) give an indication of the total mobility, and the
models for mode shares give an indication of the relationship between the choice
of modes for certain trips and the independent variables. Mode share was measured
as the proportion of times a particular transport model was chosen.
Independent variables
Several independent variables were selected for analysis, on the basis of existing
literature and assumptions underlying urban and transportation policies. One group
of variables describes the spatial context in which the travel behaviour was measured,
and constitutes the neighbourhood level in the multilevel analysis. A second group of
variables describes the socioeconomic characteristics of the individuals, and constitutes
the individual level in the multilevel analysis.
The neighbourhood characteristics included the following variables: urban form,
transportation network type at the city level, transportation network type at the
neighbourhood level, local-street network type, location of city within the country
(inside or outside Randstad), distance to the city centre, distance to an intercity station,
employment level in the city (defined as the number of jobs in the city per 1000
inhabitants), land-use mix in the neighbourhood, locally available shopping facilities
(defined in terms of five levels), locally available sports facilities (describing whether the
A multilevel analysis using quasi-experimental design data 1217

neighbourhood has sports facilities), housing density, population density, and degree of
urbanisation (defined in terms of a number of classes).
The variable `transportation network type at the neighbourhood level' describes the
type of network that can be discerned for the neighbourhood as a whole. The range of
basic transportation networks was found inadequate for the description of the differ-
ences between neighbourhoods. Therefore, additional types were distinguished, which
typically are compounds or specific manifestations of the basic forms (figure 3).
The variable `local-street network type' refers to the network type that can be
discerned within a neighbourhood at the street level. This low level gives an indication
of the network type in the immediate proximity of the dwellings. Again additional
network types were identified, some of which are composites (figure 4).

(a) (b) (c) (d) (e) (f )


Figure 3. Transportation network type at the neighbourhood level: (a) ring, (b) loop, (c) radial,
(d) axial, (e) grid, (f) tangential.

(a) (b) (c) (d) (e)

Figure 4. Local-street network types: (a) loop/tree, (b) loop, (c) loop/grid, (d) grid, (e) tree.

The degree of urbanisation was defined in terms of a number of classes. For each
address in a neighbourhood, the average number of addresses per square kilometre
within a 1 km radius was calculated. These numbers were then averaged across
addresses in a specific neighbourhood. Based on this density measure, five degrees of
urbanisation were distinguished: `very strongly urbanised' (address density of 2500 and
over), `strongly urbanised' (address density between 1500 and 2500), `moderately urban-
ised' (address density between 1000 and 1500), `weakly urbanised' (address density
between 500 and 1000), and `not urbanised' (address density below 500). Data were
obtained from Statistics Netherlands and the municipalities in question.
The socioeconomic variables at the individual level included age, gender, level of
education, personal income, driving license, car availability, employment status, house-
hold income, household type, and dwelling type, and the average distance per trip to
the chosen location for the purpose in question.

Analyses and results


This section presents the results of the multilevel analyses. Discussion of the results is
limited to the independent spatial characteristics the urban form and the transporta-
tion networks at the different levels. However, the models estimated to obtain these
results were based on all the independent variables discussed above. To avoid a very
large number of tables and coefficients, only results significant at the 5% probability
level are presented.
Table 4 (see over) shows the significant effects of urban form for each activity and
type of dependent variable. The results indicate that the effect of urban form on the
various indicators of mobility is negligible. Significant parameters were found only for
1218 D Snellen, A Borgers, H Timmermans

home-to-work trips, not for any of the other activities, and then only for the lobe city.
The share of motorised kilometres travelled for home-to-work trips by inhabitants of
lobe cities is reduced 0.8% more with an increasing distance to work than in a grid city.
The results indicate that in lobe cities public transport is used more for home-to-work
trips. Given that most medium-sized Dutch cities have a radial public transport
network, this result is as one would expect.
Table 5 reports the results of city network type. It demonstrates that network type
has no significant effect on any mobility indicator in the case of grocery shopping and
leisure. In the case of other shopping, there is an effect on the number of kilometres
travelled by car. The number of motorised kilometres in cities with a ring network is
significantly higher for those individuals who always have a car at their disposal. This
first of all suggests that people who have a car at their disposal will actually use it
more for other shopping, and/or that a ring network involves travelling a longer
distance, at least in the cities included in our sample. A comparison of travel distance
across city network types supported this interpretation.
Moreover, the number of trips per year tends to be higher in cities with a radial
network as compared with cities with a grid network. It is not readily evident whether
this reflects some causal mechanism. It is also found that cities with a radial network
show an increased use of nonmotorised transport modes, both in number of kilometres
and in number of trips for the home-to-work commute.
Table 6 reports the results of the transportation network type at the neighbourhood
level. It is found that the network type of the neighbourhood has some influence on
both types of shopping trips, but not on leisure or home-to-work trips. For grocery
shopping, there is a positive parameter for share of motorised kilometres in loop-
structured neighbourhoods. For other shopping trips, radial neighbourhoods networks
Table 4. Significant parameters for urban form (base: grid).

Activity Dependent variable Concentric Lobe Polynuclear

Home-to-work Share of motorised kilometres ÿ0.8a


a In combination with average distance from home to work.

Table 5. Significant parameters for transportation network type at the city level (base: grid).

Activity Dependent variable Ring Radial

Other shopping number of motorised kilometres 49.4a


number of trips 15.8
Home-to-work number of nonmotorised kilometres 26.9
number of trips 2.8
number of nonmotorised trips 5.3
a In combination with `car always available'.

Table 6. Significant parameters for transportation network type at the neighbourhood level
(base: tangential).

Activity Dependent variable Loop Grid Radial Axial Ring

Grocery shopping share of motorised kilometres 30.1


Other shopping number of motorised trips 68.01a ÿ23.4
a In combination with `car sometimes available'.
A multilevel analysis using quasi-experimental design data 1219

reduce the number of motorised trips, whereas loop networks increase this number (for
those people who sometimes have a car available). In summary, from this simple
perspective the results of this study suggest that loop networks are not to be preferred
at the neighbourhood level, because they induce more motorised travel. Of course,
there might be other reasons, such as a reduction of through traffic on minor roads
to choose a loop network.
Table 7 shows the results for the local-street network type. The grid and the loop
network reduce the share of motorised kilometres for home-to-work trips. Loop/tree
local-street networks are associated with a higher share of public transport trips for
other shopping. Loop/grid networks reduce the share of nonmotorised trips for other
shopping, for those who always have a car available.
Table 7. Significant parameters for local-street network type (base: tree).

Activity Dependent variable Loop Loop/tree Loop/grid Grid

Other shopping share of nonmotorised trips ÿ40.0a


share of public transport trips 18.6
Home-to-work share of motorised kilometres ÿ26.5 ÿ28.3
a In combination with `car always available'.

Conclusions and discussion


The relationship between land use and travel has been the subject of a considerable
amount of research. Because new neighbourhoods are designed so as to stimulate the
use of public transportation and nonmotorised transportation, a better understanding
of the assumed link between urban form, network type, and mode-choice decision is
paramount to public transportation planning and development. Many previous
research endeavours may be flawed because researchers did not try to reduce the effects
of near multicollinearity in their data. Moreover, the fact that disaggregation proce-
dures are often used implies that the actual significance of the results may be less than
reported.
In the present study, therefore, we adopted the principles of quasi-experimental
design to collect the data, which were examined with a multilevel model. The results
of the present analysis indicate that urban form and network type have only a modest,
yet present, effect on mode-choice decisions for frequently conducted activities. All
effects not discussed in any detail in this paper were nonsignificant. Moreover, it
should be emphasised that, for statistical reasons only, one will always find some
effects to be significant, even though they have no apparent substantial meaning or
interpretation. Obviously, these conclusions relate only to the explanatory variables
included in the present study that focused on urban form characteristics. Future studies
may examine the impact of level of service and price on mobility patterns (for example,
Anderson and Jorgenson, 1995; Cervero, 1995; 1998; Fullerton and Knowles, 1991).
These findings imply that indeed the positive findings reported elsewhere may be
an artefact of the flawed methodology used. Further studies örepetitive studies using
state-of-the-art methodologiesöare therefore required to understand better and assess
the relationship between urban form, network type and travel behaviour.
1220 D Snellen, A Borgers, H Timmermans

References
Anderson H T, Jorgensen J, 1995, ``City profile: Copenhagen'' Cities 12 13 ^ 32
Bolt D, 1982 Urban Form and Energy for Transportation Research Centre for Physical Planning,
Organization for Applied Scientific Research, Delft
Breheny M, 1995, ``The compact city and transport energy consumption'' Transactions of the
Institute of British Geographers, New Series 20 81 ^ 101
Bryk A S, Raudenbush S W, 1992 Hierarchical Linear Models: Application and Data Analysis
Methods (Sage, Newbury Park, CA)
Bullen N, Jones K, Duncan C, 1997, ``Modelling complexity: analysing between-individual and
between-place variationöa multilevel tutorial'' Environment and Planning A 29 585 ^ 609
Cervero R, 1995, ``Sustainable new town: Stockholm's rail-served satellites'' Cities 12 41 ^ 51
Cervero R, 1996, ``Mixed land-uses and commuting: evidence from the American Housing
Survey'' Transport Research A 30 361 ^ 377
Cervero R, 1998 The Transit Metropolis (Island Press, Washington, DC)
Cervero R, Gorham R, 1995, ``Commuting in transit versus automobile neighbourhoods''
Journal of the American Planning Association 61 210 ^ 225
de Klerk L A, 1980 Op Zoek naar de Ideale Stad [In search of the ideal city] (Van Loghum
Slaterus, Deventer)
Florez J, 1998, ``Urban development patterns and their relationships with travel behaviour'',
paper presented at the 26th European Transport Conference, Loughborough, September;
copy available from the author
Frank L D, Pivo G, 1994, ``Impacts of mixed use and density on utilization of three modes of travel:
single occupant vehicle, transit and walking'' Transportation Research Record number 1466,
Transportation Research Board, National Research Council, Washington, DC, pp 37 ^ 43
Friedman B, Gordon S P, Peers J B, 1994, ``Effect of neotraditional neighbourhood design on
travel characteristics'' Transportation Research Record number 1466, Transportation
Research Board, National Research Council, Washington, DC, pp 63 ^ 70
Fullerton B, Knowles R D, 1991 Scandinavia (Paul Chapman, London)
Giuliano G, Small K A, 1993, ``Is the journey to work explained by urban structure?'' Urban
Studies 9 1485 ^ 1500
Giuliano G, Small K A, 1999, ``The determinants of growth of employment subcentres'' Journal
of Transport Geography 7 189 ^ 201
Goldstein H (Ed.), 1995 Multilevel Statistical Models 2nd edition (Edward Arnold, London)
Gordon P, Richardson H W, 1996, ``Beyond polycentricity'' Journal of the American Planning
Association 62 289 ^ 295
Handy S, 1993,``Regional versus local accessibility: neo-traditional development and its implications
for non-work travel'' Built Environment 18 253 ^ 2678
Hox J J, 1995 Applied Multilevel Analysis (TT Publications, Amsterdam)
Jacobsen N L S, 1997, ``Household density and public transport: a practical and pragmatic study
of urban land use and city transport'', paper presented at the 2nd International Symposium
on Urban Planning and the Environment, Groningen, May; copy available from the author
Kitamura R, Mokhtarian P L, Laidet L, 1997, ``A micro-analysis of land use and travel in five
neighbourhoods in the San Francisco Bay Area'' Transportation 24 125 ^ 148
Kreft I, J de Leeuw (Eds), 1998 Introducing Multilevel Modeling (Sage, Thousand Oaks, CA)
Nasar J L, 1997, ``Neo-traditional development, auto-dependency and sense of community'', in
Space Design and Management for Place Making Eds M S Amiel, J C Vischer, EDRA
refereed papers 28 39 ^ 43
N×ss P, Sandberg S L, 1996, ``Workplace location, modal split and energy use for commuting
trips'' Urban Studies 33 557 ^ 580
N×ss P, Roe P G, Larsen S, 1995, ``Travelling distances, modal split and transportation energy in
thirty residential areas in Oslo'' Journal of Environmental Planning and Management 38 349 ^ 370
Owens S E, 1986 Energy, Planning and Urban Form (Pion, London)
Rottier H, 1980 Stedelijke Structuren [Urban structures] (Coutinho, Muiderberg)
Schimek P, 1996, ``Land use, transit and mode split in Boston and Toronto'', paper presented at the
ACSP/AESOP joint international congress, July, Toronto; copy available from the author
Snijders T A B, Bosker R J, 1999 Multilevel Analysis (Sage, London)

ß 2002 a Pion publication printed in Great Britain

You might also like