Transversal Research · Transversal Research Samuel CARPENTIER, CEPS/INSTEAD, GEODE Eric CORNELIS,...
Transcript of Transversal Research · Transversal Research Samuel CARPENTIER, CEPS/INSTEAD, GEODE Eric CORNELIS,...
Transversal Research
Samuel CARPENTIER, CEPS/INSTEAD, GEODE Eric CORNELIS, FUNDP, GRT
Luc DAL, UCL, GéDAP Thierry EGGERICKX, UCL, GéDAP
Philippe GERBER, CEPS/INSTEAD, GEODE Sylvain KLEIN, CEPS/INSTEAD, GEODE
Xavier PAULY, FUNDP, GRT Philippe TOINT, FUNDP, GRT Fabien WALLE, FUNDP, GRT
SCIENCE FOR A SUSTAINABLE DEVELOPMENT(SSD)
FINAL REPORT T
MOBILITIES AND LONG TERM LOCATION CHOICES IN BELGIUM “MOBLOC”
SD/TA/04A
2
Table of content
SUMMARY ________________________________________4 A. Context ___________________________________________4 B. Objectives ________________________________________4 C. Conclusions _______________________________________5 D. Contribution of the project in a context of scientific support to a sustainable development policy _____________________6 E. Keywords _________________________________________7
1. INTRODUCTION__________________________________8
2. METHODOLOGY AND RESULTS ___________________10 A. Global methodology _______________________________10 B. Accessibility models _______________________________13
Methodology steps for the setting up of the Private Vehicle accessibility modelling_______________________________________________ 13 Observations data sets and accessibility models for comparisons___ 15 Road network modelling ___________________________________ 16 Principles of traffic assignment and model comparisons __________ 18 Off-peak hours accessibility modelling ________________________ 23 Morning peak hours accessibility model _______________________ 25 Communal accessibility indicators ___________________________ 28
C. Travel demand model ______________________________32 Gravity model ___________________________________________ 32 Introduction of border effect ________________________________ 33 Model estimating_________________________________________ 33 Model assessment _______________________________________ 35 Residual analysis ________________________________________ 36 Conclusion _____________________________________________ 38
D. Propensity to move model __________________________39 Residential migration: a concise overview _____________________ 39 Pre-selection of variables for the propensity model ______________ 40 Methodological choices to build the model _____________________ 43 Transformation of variables ________________________________ 45 Sample design __________________________________________ 46 Model(s) _______________________________________________ 47
E. Localization model_________________________________54 Preliminary analysis ______________________________________ 54 Methodology ____________________________________________ 56 Models and results _______________________________________ 63
3. POLICY SUPPORT_______________________________67
4. DISSEMINATION AND VALORISATION______________69 A. MOEBIUS - Mobilities, Environment, Behaviours Integrated in Urban Simulation __________________________________69 B. SimBelgium - NAXYS ______________________________70
3
C. Population and households forecasting of Belgian municipalities _______________________________________71
5. PUBLICATIONS _________________________________73
6. ACKNOWLEDGMENTS ___________________________74
7. REFERENCES __________________________________75 ANNEX 1: COPY OF THE PUBLICATIONS _____________80
ANNEX 2: MINUTES OF THE FOLLOW-UP COMMITTEE MEETINGS ______________________________________109
Annex 2.1. Minutes of the first follow-up committe meeting (3rd of May 2007) _______________________________________109 Annex 2.2. Minutes of the second follow-up committe meeting (12th of June 2008) __________________________________114 Annex 2.3. Minutes of the third follow-up committe meeting (19th of October 2009)________________________________119 Annex 2.4. Minutes of the fourth follow-up committe meeting (31st of January 2011)________________________________123
4
SUMMARY
A. Context Mobility and transport evolve with time and the passing generations. Interactions are
numerous between daily mobility and household migration (here defined as house
moving implying a municipality change). The evolution of the transport system has
deeply modified the barrier of distance and has largely opened the choices in terms
of residence place. The continuing urban sprawl phenomenon resulting from these
modifications has itself resulted in a strengthening of the property and housing
market in certain territories, pushing people (young couples in particular) towards a
residential localization which is further and further away from the traditional, urban
activity centres. The tensions between daily and residential mobility have therefore
increased, notwithstanding the recent rise in energy costs. This in turn generates
unsustainable effects on society and environment.
But these new residential choices have in parallel induced new mobility behaviours,
based on an extensive (and probably excessive) use of the private car in daily trips
(home-work/school, shopping, leisure ...). Social life itself (visits to friend and family)
has become more spatially dispersed. One already knows that the propensity to
change residence is determined by a number of individual or household
characteristics such as age, citizenship or income, but the effects of long-term trends
as population ageing, the evolution of the household/family structure on both
residential choices and mobility behaviours remain so far largely unanticipated.
This research project aimed at analyzing interactions between demographics and the
evolution of mobilities at different time-scales. In particular, localization choice for
household, daily accessibility and internal migrations appear to have strong relations.
B. Objectives The project objective was so to investigate the link between long-term society
evolution, residential choice, transportation demand and the resulting accessibility
evolution. On the societal trends side, particular attention had to be paid to the
population ageing effects, evolution of family cell structure and inter-generational
relationships, but other variables such as land-use and standards of living trends
have also been considered for potential inclusion in the developed models.
On the transportation side, emphasis had to be put on the evolution of transportation
demand (gravity model/mobility demand model more lately described) and resulting
traffic conditions (accessibility models).
5
The project objectives also included simulations of scenarios for the future in Belgium
but as researchers faced a lack of time, such simulations could not be achieved.
Nevertheless, the main models are available for these simulations. Furthermore, the
links between those models are ready to use.
C. Conclusions Major conclusions concern the main models, wihch are, a propensity to move model,
a localization model, a transportation demand model and accessibility models.
(i) The transportation demand model allows the computation of the origins and
destinations of home-work travels flows in order to provide an estimation for the
accessibility model. Indeed, from the margins of an origin-destination matrix, this
model, calibrated on exhaustive data from the national census of 2001, can compute
the flow between each pair of Belgian municipalities. The accuracy could be
improved to better model the long distance trips but without consequences on the
shorter ones being currently well estimated. This transportation demand model can
feed the accessibility models developed in order to get municipal accessibility
indicators. These indicators take account, on one hand, of the accessibility to jobs
during morning peak hours and, on the other hand, of the accessibility to other daily
activity places during off-peak hours. These indicators rely on travel times computed
thanks to an accessibility model calibrated on a highways and national roads
network. This network was characterised with a typology based on the crossing of
urbanized areas from the CORINE land cover GIS layer. Furthermore, a check of the
speed, the length of the travel and the travel time difference between declared
speed, length and travel time of the MOBEL survey in off-peak hours was achieved.
(ii) The propensity to move model shows us that this propensity depends on the life
cycle of individuals, and more particularly their family trajectories. One can observe
this link through the well known correlation with the age and the transition of the
household structure. In fact, the transitions leading to move are break-up situations,
new family-units compositions and leaving parents’ home. It brings out that the more
stable situation concerns people who are in married couple (with or without children);
this situation which is often associated with an owner status, which is another factor
of residential stability. In other words, the likely evolutions of family situations marked
by the rise of less stable households (cohabitation situations, one-parent family…)
should still generate higher propensity to move rates in the coming years.
(iii) The localization model mainly shows us that people tend to settle down in
municipalities where “similar” people live (regarding e.g. the household structure, the
age). The relocation often takes place on short distance. People rarely leave their
residential municipalities to find a new dwelling in a far away place. Accessibility
6
indicators are also significantly explanatory for the residential choice, although, as it
turns out, not dominant.
D. Contribution of the project in a context of scientific support to a sustainable development policy
The project's objective was to investigate the links between long-term residential
choice, accessibility and daily mobility, with the ambition to provide better
understanding of the behaviour of Belgian households regarding these issues. In
particular, the respective importance of several categories of factors is crucial for the
establishment of land-use and accessibility policies. The mixing of long-term
decisions such as housing and short-term ones such as daily mobility turns out to be
a challenging issue.
The accessibility model outputs provide a few relevant accessibility indicators; they
allow the characterization and comparison of municipalities and we feel that this
could have direct consequences on municipality management.
The model describing the propensity of the Belgian individuals to move their
residence (which incorporates a number of explanatory factors at the individual and
household levels) shows us that the choice of changing residence is mainly caused
by alterations in the household structure, accessions to a different position in the
household (from child to head e.g.) and age classes. This reinforced the idea that
societal trends (as opposed to material infrastructure evolution) are crucial to explain
internal migration within a country. In particular, population aging and the increase of
”narrow” households may present specific challenges in urban planning and land-use
in general during the forthcoming years.
Finally, the residential localization model is central to the design of suitable land-use
regulations at the regional level. Remarkably, analysis indicates that the dominant
factors are, by decreasing level of importance, the distance between the previous
residence and the new one, the perceived quality of life in the new municipality of
residence, the household structure, and, in fourth position, the accessibility of the
new municipality of residence. Accessibility is therefore less important than expected
at the start of the project.
A merely interpretation of the results from the MOBLOC project is that migration
within the country is less determined by infrastructural factors (within which
accessibility is an important example) than by factors related to societal life in a more
general sense: household structure and its evolution, closeness to one's relations,
age class and quality of life score indeed higher in our results than purely transport
related factors.
7
We clearly believe that these conclusions are important for any forecast on the
development of land-use and transportation systems. They will be notably discussed
in the new regional prospective study group (SRP) established under the leadership
of the Institut Destrée and the Institut Wallon d'Evaluation, de Prospective et de
Statistiques (IWEPS). We also plan to disseminate those conclusions more widely,
via scientific publications but also aiming the municipality managers and the general
public.
E. Keywords Accessibility, daily mobility demand, propensity to move, residential localization,
household structure, municipalities.
8
1. INTRODUCTION Interactions between daily mobility and household migration are complex. The aim of
the MOBLOC project is precisely to study, from one side, the impacts that daily
mobility can have on residential localization (places, accessibility and the way it
influences residential choices), and, from the other side, the repercussions that
household migration can have on daily trips (transport demand and congestion) and
the linking feedbacks. Indeed current residential choices (periurban localization)
induce new mobility behaviours, often based on an extensive (and probably
excessive) use of the private car for daily trips (home-work/school, shopping, leisure
...) which generate unsuitable effects. Social life itself (visits to friend and family) has
also become more spatially dispersed.
One already knows that the propensity to change residence depends on a number of
individual or household characteristics such as age, nationality or income, but the
effects on both residential choices and mobility behaviours of long-term trends such
as population ageing or the evolution of the household/family structure remain so far
largely unanticipated. In this context, modelling can surely help the understanding of
the behaviours and allows simulation. This is the way followed by the MOBLOC
project; the models which have been developed participate to this objective and are
important constitutive elements of this even more ambitious project.
The present report will successively address the global methodology of the project,
the results of different modelling exercices achieved to reach the objectives without
missing out the methodological aspects of these different steps and their respective
conclusions. A first model concerns the accessibility at municipalities’ level and
calculates travel times between each pair of municipalities (during off-peak hours and
morning peak hours). These travel times allow the computation of accessibility
potential to employment and services (later used as explanatory variables in the
residential localization model). Prior to this accessibility model, a transport demand
matrix has have to be calibrated thanks to a gravitational model taking employment
repartition and residential localization of active people into account. Then a two-fold
model addresses the residential changes. First, a propensity to move model models
the fact that an individual decides to settle down in a new municipality between two
successive 1st of January or not. A residential localization model follows and
estimates several parameters modelling the choice of a new residential municipality.
This latest model notably uses the output of the accessibility model. The report will
successively describes all these developed models.
A policy support section follows; this section underlines the main outcomes and their
interests in a frame of sustainable development. As the MOBLOC project can be the
start point for new researches, some future projects will also be briefly described:
9
these will rely on the new expertise available thanks to the MOBLOC project. Useful
references finally conclude the report.
10
2. METHODOLOGY AND RESULTS
A. Global methodology
Let us so first describe the global methodology of the project. This description will
rely on a presentation of the data flows through the inputs and outputs of each model
and the interactions between them. For that purpose, the Figure 1 (Pattern of the
MOBLOC Project) below could be very helpful.
The main objective of MOBLOC is to draw up a link between residential migration
and daily mobility. Thus the two main bricks of the project are the residential
migration model and the accessibility model. As the figure shows, other models are
necessary to reach this goal. Indeed, we need to determine some inputs for these
two main models and to establish the links relating these inputs
The inputs of the residential migration model (1a) are individual data. They include
age, gender, level of education, household evolution, previous migration and the
current residential municipality (at time Y). The final output of this model is the new
residential municipality (one year after, Y+1). This model has been split up into two
sub-models: a propensity to move model and a localization model.
The first one uses some of the individual information to forecast if a migration occurs
(i.e. if an individual moves) from year Y to year Y+1 (from a municipality to another –
not inside a given municipality). If the answer is yes, then the localization model will
simulate the localization choice between the 588 remaining Belgian
municipalitiessince moving individuals may eventually not choose their original
municipality.
The technique used for the propensity model is a binary logistic regression while the
localization model will resort to a more sophisticated discrete choice technique.
These methodological issues will be further detailed in following sections.
At the first step of the transportation models, the travel demand model (2) is based
on two inputs: one from the migration model (1a): the number of working people and
one from the evolution model (1b):.the number of jobs per municipality.
These figures constitute the margins of the O/D matrix to be calculated by the gravity
model. Let us remark that for now, this matrix (demand) only takes into account the
home-work and home-school trips because of the used inputs. Even if this
assumption is very restrictive, it is a first good approximation considering the
available data.
11
The next step of the transportation models is the modal split model (3). Its purpose is
to compute from the “global” (all modi) O/D matrix another one concerning only the
trips made by car. This step is necessary to provide a feasible input for the private
vehicle accessibility model which is an essential brick of the project.
The private vehicle (PV) accessibility model (4) uses Wardrop equilibrium paradigm
to assign the traffic (car trips) on the roads network. It provides a matrix of the travel
times between each municipality’s pair (a first matrix during off-peak hours and a
second one during the morning peak hours on a working day) by car.
At earlier stages of the project a public transport (PT) accessibility model was also
forecast. It would have computed the travel times by public transport. between each
municipality’s pair This model has not been developed due to the unavailability of a
complete dataset for public transport timetables all around Belgium. Whatever, let us
also point out that these times would not have been directly related to the travel
demand because they would have been based on the supply. The PT accessibility
model could not have been updated at each step of the simulation as we do not know
how the public transport level of service will evolve in the next years. So as we did
not manage even to receive complete information on the current supply, this model
was not built, even if it would have been interesting to test its contributions for the
localization model.
At the last step of the transportation model, the PV travel times between
municipalities are used to compute different accessibility indicators (to employment
and to services) per municipality. They are included as covariates of the localization
model (the second step of the residential migration model) to measure the
attractiveness of the municipality.
This chain of models could then be used in a prospective way. Therefore it will be
necessary to update the input values for the residential migration model. Let us
remind that this model works at an individual level which implies to find adapted
techniques. This would be the role of the evolution models (1b). They could include
different kinds of techniques at different levels: aggregate and disaggregate, e.g.
demography trends, synthetic population… Although these developments are clearly
interesting, they are postponed to further research. The research group is aware that
the original program did include steps in these directions, but the determination of the
accessibility model, propensity to move and residential localization turned out to be
enough challenge for the contract duration. This research agenda is however
considered in another research project relying on MOBLOC (see section 4,
Dissemination and valorisation).
12
Figure 1: Pattern of the MOBLOC Project
13
B. Accessibility models This part deals with the setting up of a road accessibility model (step 4 of the
previous figure), at the scale of Belgium, and its validation. For the whole country, all
the required parameters are not available to calibrate and validate the off-peak and
peak hours accessibility models. Considering this lack of data, several goodness-of-
fit statistics were applied on the basis of different comparisons between our model
values and, on one hand, those of observations (MOBEL, 1999), and, on the other
hand, those of other models (Vandenbulcke et al.,2009).
Methodology steps for the setting up of the Private Vehicle accessibility modelling
As shown on Figure 2, several steps have been carried out. At first, a road network
and municipal centroids representation was set up.
As the modelling is to be done throughout the Belgian territory, the first data set up
step consists in choosing a representative point for each of the 589 municipalities. If
the choice of single municipal centroids constitutes an approximation, this appeared
to be necessary for both a methodological consistency reason -the model coupling of
MOBLOC is done at the municipalities level - and a practical reason, this being the
aggregation level at which the origin/destination matrix was available (ESE2001 for
the first run, gravity model for future simulation). This work was accomplished in the
first phase of the MOBLOC project and fully described in its final report (Cornélis et
al. ,2009) and therefore not further described here.
Afterwards, the mathematical representation of the road network was set up as
follows:
G is a directed graph where G = (V, E),
with V: the set of vertices representing the road intersections
E, the set of edges representing the road sections with homogeneous
properties.
It notably consists in describing these edges' length, free-flow speed and capacity.
Since this information is not available for the whole country, we used a network
typology drawn up by crossing with a land-use GIS layer as described further in
subsection “Road network modelling”.
14
Figure 2: Methodology steps for the setting up of the private car accessibility modelling
In a second step of the modelling procedure, two road accessibility models were set-
up, the former for off-peak hours, the later for morning peak hours (7 to 9 A.M.). The
off-peak hours model is based on a free-flow calculation of the shortest path on the
network. Morning peak hour’s model involves a travel demand matrix, that is to say,
an origin-destination exchange matrix of workers and students using a car (as main
driver) for their home to work or home to school trips. This matrix allows us to include
later the new localization outputs from the MOBLOC localization model and is
extracted from the National Socio-economic Survey of the INS (2001) for the first run.
The results of the two models are then compared to observed data from the MOBEL
survey and to results from other available models for Belgium. Thanks to goodness-
of-fit statistics measuring the accuracy of the model according to the observed
values, it is then possible to calibrate its parameters and its accuracy compared to
other models.
Finally, once both models (on-peak and off-peak) are validated, it is possible to
calculate the municipal indicators. Then these indicators can be introduced in the
localization model, therefore modelling the effect of daily mobility on the residential
choices. We assume here, that for their residential choices, workers primarily
consider their travelling to work and related travel-time to ensure the feasibility of
their activities schedule. In other words, people have to maintain their time-budget
within some limits, as assumed by Zahavi's stable travel-time budget (1979)1. The
1 Let us note that this rule about a constant time-budget is subject to discussions (Joly I., 2003).
15
main constraint influencing this choice is therefore the travel-time during the peak
hours of a working day.
Observations data sets and accessibility models for comparisons
In this part we describe the data sets and accessibility models that we used for the
model building and validation.
Demand estimation in peak hours: ESE 2001
Source: The demand matrix is drawn from ESE 2001 National Socio-Economic
Survey.
Use: demand matrix for peak-hours assignment
Data processing: We take the drivers into account for the two following purposes:
home to work and home to school travels. They represent the vehicles flows. The
share of the trips within morning peak hours flows is estimated from MOBEL survey
to 33.7 % (of total daily traffic) between 7 and 9 am.
MOBEL survey (1999)
Source: The available observation data set used for the validation process is taken
from the national survey about mobility in Belgium conducted in 1999 in the
framework of the MOBEL project (Hubert and Toint, 2001). We thus have, at the
scale of the country, observations about travel times during a working day (excluding
school holidays). The number of observations meeting these criteria is 10,036.
Use: This data set will be used as a reference for the comparisons in the morning
peak hours as well as in off-peak hours. It is also used for the network modeling
checking step and to calculate the morning peak-hours ratio of the home to work and
home-to-school trips.
Data processing: From MOBEL observations were removed those out of the field of
our study, such as the observations with location out of the study area (origin or
destination in a foreign country for instance) or with inappropriate values (null values
or with distances which where very lower than the euclidean distance between two
communal centroids, as well as the abnormal values detected through integrity
checks). In off-peak hours, that is to say, out of the two in-peak period between 7 and
9 am and between 3 and 6 pm (Hubert and Toint, 2001) we then have a total of
1,125 origin/destination couples and 598 couples for the morning peak hours (7 to 9
am).
An internet road model: Google Maps (2010)
Source: http://maps.google.com
Use: for the comparison of off-peak hours models and for the network modeling
checking step
16
Data processing: It is possible to query Google Maps servers to calculate road
itineraries and get travel time values for the selected origin-destination couples. At
first, it was necessary to geocode origins and destinations and check their matching
with the municipality centroids of our road model. The road network used by Google
maps comes from the Ministerie van de Vlaamse Gemeenschap (Flemisch Ministry)
and from the Ministère de l'Équipement et des Transports2 (Walloon Ministry). This
model is used during the stage of network checking and for the off-peak hours
comparisons. It will be thereafter labelled GMAP.
Accessibility model from the UCL (2007)
Source: In the project 'Accessibility indicators to places and transports' an
accessibility model was developed by Vandenbulcke et al. (2009) at the scale of
Belgium. Their approach is noticeably different as it involves the calculation of
impedance based on the communal population and jobs densities for each of the
municipalities’ road sections. Their road network contains 627,856 edges. Thanks to
the authors, the travel time calculation for all the municipalities in off-peak hours was
made available for us. On the other hand, the travel times in peak hours were
available only for the 53 top level cities according to an updated urban Belgian
hierarchy (Van Hecke, 1998). This model is thereafter labeled UCL.
Use: for the comparison of off-peak hours and morning peak-hours models
Data processing: selection of the OD couples available in MOBEL as well
After the description of the datasets and models used for the building and
comparison of our model, let us continue with a quick insight about the road network
setting-up.
Road network modelling This section is about the road network setting-up represented on Figure 1. This step
involves the use of several GIS layers to correct and complement the road network
when necessary.
Our original road network comes from the Service public fédéral Mobilité et
Transports. It has a range of roads from highways to the third level of Belgian
national roads and is generally accurate enough for an inter-communal modeling
(Figure 3). There although remained a bit of work to be fully ready for the modeling.
The representative municipal centroids were linked to the network by connectors
when they were not farther than 200 m from the road network. If farther than 200 m,
we supposed that the network was not enough detailed and therefore additional
roads have been digitized. This was done when necessary on the basis of Google
Maps and thanks to urbanized and commercial zone CORINE land cover 2001 layer.
2 It is thus a priori the same network as ours with less detail l. http://www.google.com/intl/fr_fr/help/legalnotices_maps.html
17
The following step consists in setting up the free flow speed which is one of the traffic
assignment model parameter. These speeds are based on a road sections typology
taking into account the number of road lanes, the presence of segregated lanes, or
the type of urbanization crossed by the road.
We present here three road networks discussed thereafter under the names Mob_a,
Mob_b and Mob_c.
The first model Mob_a is built on a road typology taking into account the
crossing of urbanized municipalities according to the communal classification
of Van der Haegen (1996).
The second model Mob_b is built on a road typology based on the crossing of
urbanized or commercial areas from the CORINE land cover GIS layer at a
subcommunal level.
The last model Mob_c is built on the basis of Mob_b after checking the speed,
the length of the travel and the travel time difference with declared speed,
length and travel time of the MOBEL survey in off-peak hours.
The Google Maps values were used as well, in order to compute shortest path, and
faster travel time between two municipalities. We mainly investigated the records for
which the declared distances between two municipalities were not too small. Indeed,
the origins and destinations of MOBEL can be located anywhere in the municipality.
We thus calculated that, in MOBEL trips, the average distance not going outside a
municipality is four kilometres. This gives an idea about the uncertainty range related
to our aggregated level of analysis compared to the intermunicipal distances. The
different corrections varied from the additions of links in the road network, where the
road hierarchy of our data set did not allow a realistic representation of the
municipalities' access, up to changes in the values of road links parameters (i.e
number of lanes, with the help of aerial photograph) or typology.
18
Figure 3: Belgian road network used in the accessibility models
The last important parameter concerns road capacity. According to the Ministry of
Transports, the capacity of a highway is 2000 UVP per hour and traffic lane. This can
also vary according to the road geometry and other factors. As far as roads in urban
environment are concerned, the conventional capacity is 1200 UVP per hour and per
traffic lane. For countryside roads this value ranges between 1400 and 2000 UVP per
hour and per traffic lane.
On the map (Figure 3) one can notice the privileged position of Brussels at the center
of the highway network connecting the major cities of the country: Antwerp up North,
Ghent and Bruges up West. and the major cities of the Walloon Region with from
East to West: Liege, Namur, Charleroi and Mons. The south of the country appears
less equipped in infrastructures, reflecting thus its lower population density.
Principles of traffic assignment and model comparisons After the centroids and network data are gathered and checked, the traffic modeling
principles have to be settled. They will be presented in the following section. First, the
principles of traffic assignment will be presented. Then, we will introduce the statistics
used for the matrix comparison involved in the validation step, as well as the unipolar
mapping of the estimated travel times.
19
Principle of traffic assignment
The main objective of the traffic modelling is to compute estimated travel time values
between the Belgian municipalities for trips with private vehicle in order to calculate
accessibility indicators. To do so, our methodology relies on a classical assumption
of traffic engineering. This assumption specifies the decrease of the speed when the
traffic approaches the maximal capacity of the road. Several functions describe this
relationship and also define the parameters that have to be set. The assignment itself
consist then on the use of a particular algorithm which will distribute the demand
matrix (flows) on the network iteratively, taking into account the congestion effects on
this distribution.
The modeling is done in two stages, the first for off-peak hours and the second for
morning peak hours. During off-peak period, we assume that travels can be done at
free-flow speed. This allows us to apply an algorithm to solve the minimum cost, as
Dijkstra algorithm (1959) does, and thus to determine the shortest cost path. The
cost function used here, is a simple minimization of travel times.
During morning peak hours, we use a model allowing assigning the demand matrix
on the shortest cost paths, taking account of the relationship between the travel time
on a road and the vehicles flow. These shortest-paths being beforehand determined
in free-flow conditions, it is then possible to use the off-peak calibrated model for the
morning peak model's setup. For each road network's edge a speed-flow function
should be defined, specifying maximal speed decay while the number of vehicles
approaches the maximal capacity of the road. A lot of functions were developed and
discussed (Branston, 1976); so are BPR-like functions developed by the American
Bureau of Public Roads:
With T: travel time,
V: flow,
Q: road edge capacity,
T0: free-flow travel time and
α and β : parameters of the function according to the road.
This function has the following curve (Figure 4):
20
Figure 4: BPR curve (Bureau of Public Roads (1964))
The chosen model for the peak hours assignment is based on Wardrop’s first
hypothesis (1952) and specifies the conditions allowing reaching a user equilibrium.
Such equilibrium stipulates that no any road-user can unilaterally improve his travel
time by modifying his itinerary. An approximate of this equilibrium is reached by the
method of successive averages.
At the end of this balancing process, a new speed is calculated for every road
section. The travel time matrix is then calculated.
Validation principle by model comparison
After the setup of the accessibility model raised the question of its validation. We
choose to compare our estimated values with declared travel times from the MOBEL
survey and with results from other models based on distinct methodologies, in order
to validate our model. This proceeding allows us the validation of our model,
throughout goodness-of-fit statistics computed at the different steps of its
development.
In order to test the accuracy of the model and its estimation, many statistics
comparing estimated matrix with observed matrix are available (Knudsen and
Fotheringham, 1986). The following statistics were selected considering their linear
sensitivity to the error level for a complete matrix.
SRMSE, the Standardized Root Mean Square Error:
with pi and qi, the observed and estimated travel times. The lower limit of this
statistic is zero indicating perfectly accurate prediction and its upper limit is
21
generally one, though values higher than one arise whenever the average error is
greater than the mean.
phi statistic:
with pi and qi, the probabilities of the observed and estimated flows. This statistic
has limits of zero and plus infinity.
absolute value of the psi statistic:
with pi and qi the probabilities of the observed and estimated flows, and si = ( pi +
qi ) / 2. This statistic has no known theoretical distribution.
We tested two other statistics:
AED: Absolute Entropy Difference, defined as the absolute value of the
difference of the entropies of the observed and predicted probability values.
with Hp and Hq Shannon's entropy (Shannon (1948)) so that, for example,
the chi-square statistic: χ2
We tested the sensitivity of these statistics on a MOBEL sample for off-peak hours,
introducing errors:
22
with qi, the estimated travel time values, pi, the observed travel time values, δ,
random numbers {-1,1}, rnd, a random number between zero and one and fact, the
percentage of error divided by 100. For each level of error the values of the following
statistics: AED, χ2, SRMSE, φ and ψ, were calculated.
This process was repeated a hundred times and the average of the statistics were
calculated. The sensitivity of each statistic to error level is drawn on Figure 5, where
the horizontal axis represents error levels and the vertical axis represents the
standardized value of the average statistics.
The SRMSE, φ and ψ statistics present a linear relationship with error level, on the
contrary with AED and χ² that underestimate the error rate for low random error
levels.
Figure 5: Error sensitivity of five Goodness-of-fit statistics
After this insight in the theoretical aspects of our method, lets continue with the
results of the modelling, first in off-peak (free-flow) conditions, then during the
morning peak hours.
23
Off-peak hours accessibility modelling
Off-peak hours assignment
The road model was imported in OmniTRANS software3. This software allows us to
perform the traffic assignment for peak hours and off-peak hours.
The three travel time matrices (Mob_a, Mob_b, Mob_c) were calculated for off-peak
hours with an all-or-nothing assignment based on the three road networks typology
presented earlier (see subsection “Road network modelling”). The checking
consisted in cartography of the estimated travel times compared with the observed
travel times, as well as the goodness-of-fit statistics calculation of our model with the
observed values from MOBEL and the statistics from other models.
The cartography of travel times from each municipality to Brussels destination is
presented in Figure 6. We notice that the travel times spread amongst the country in
different ways depending on the models. The Mob_a and UCL models have a large
spread of the classes from Brussels to the periphery. This illustrates quite high
speeds. Models Mob_b and Mob_c show a similar distribution to the one from
Google Maps as well as a more tighten spread corresponding to lower speeds.
Besides, we can see through the spread of the classes along highway axis an higher
sensitivity to the road hierarchy compared to Mob_a and UCL.
3 Version 5.0.28. This software is developed in the Netherlands by OmniTRANS International. http://www.omnitrans-international.com
24
Figure 6: Representation of travel time in off-peak hours from different models
Goodness-of-fit in off-peak hours
The goodness-of-fit statistics of the models for off-peak hours are presented in Table
I.
25
Table I: Goodness-of-fit values in off-peak hours
As far as AED statistic is concerned, Mob_a and GMAP have better values than
Mob_c, UCL and Mob_c. On the contrary,Mob_b, Mob_c and GMAP models have
quite the same magnitude regarding X², SRMSE, PSI and PHI and we notice better
global accuracy than Mob_a and UCL. The last two statistics show quite little
differences between the models.
From the map and the goodness-of-fit statistics from Table 1, we conclude that taking
into account the crossed urbanization type of the road section in models Mob_b and
Mob_c brings a better accuracy level. With respect to these results we decided to
build the peak hours model on the basis of Mob_c.
Morning peak hours accessibility model
Morning peak hours assignment
Two models were tested for the morning peak hours. They were built on Mob_c
model. The parameters of the speed-flow function had to be defined for each road
section. It has been done according to their typology. The demand matrix based on
the ESE 2001 census was brought back to a one hour period and assigned on the
edges to reach user equilibrium4.
Two sets of parameters were tested from Mob_c off-peak model. They are named
Mob_c1 and Mob_c2. These parameters (Table II) were at first set according to the
literature (Barton-Aschman Associates, Inc.; Cambridge Systematics, Inc., 1997) and
then adjusted according to their results.
4 the O/D demand matrix is assigned without taking account of its diagonal
26
Table II: BPR function parameters values of Mob_c1 and Mob_c2
The travel times (from every municipality to Brussels) computed with these
parameters are represented on maps (Figure 7).
To facilitate the comparison between off-peak and morning peak models, the travel-
time classification scheme is kept across the different maps. On the map of Mob_c2
model, the isochrones are quite narrow compared to the situation of Mob_c1. The
travel time in the case of the UCL model in peak-hours have a concentric distribution
around Brussels, on the contrary with the models developed in MOBLOC that more
favour the Antwerp-Brussels axis. The explanation could rely in the presence of the
two highways, A12 and A1, connecting the two cities and therefore better able to
distribute the flows towards Brussels during the peak hours than in the cases of other
municipalities that are mostly connected by only one highway.
27
Figure 7: Travel time to Brussels in rush hours for the different models
Goodness-of-fit in peak hours
The results of the statistics, comparing the MOBEL observations in morning peak
hours with travel times matrices of the different models, are presented in Table III.
This comparison was made on the base of 230 origin/destination couples5.
Table III: Goodness-of-fit values in morning rush hours
The three models in Table III show a close magnitude for the different statistics. The
best results are exhibited by Mob_c1 model. Therefore this model will be used for the
calculation of the accessibility indicators allowing the interaction with the migration
model of MOBLOC.
5 This size is lesser than the one of the off-peak analysis for two reasons. On one hand, the considered period is shorter. On the other hand, the origin/destination couples available in the UCL model don't concern every destination but only those of the municipalities of the three upper classes of Van Hecke urban hierarchy.
28
Communal accessibility indicators Now that the travel times matrices of the traffic models have been validated, it is
possible to calculate the accessibility indicators needed for the localization model.
Jobs potentially accessible by car during morning peak-hours
We decided to use a gravity-based indicator to take the employment into account.
The advantage of this indicator compared to simpler indicators (number of jobs
accessible in 30 min for instance) is to account for the negative influence of travel-
time on the attraction of the destination as well as for the relative importance of the
number of jobs.
We then calculated potential job accessibility for each municipality using the following
formula.
Ai = ∑ Dj.exp(-β.dij)
with Ai, the potential accessibility to jobs of the municipality i,
Dj the jobs in the municipality of destination j in year 2004,
β a parameter calibrated according to MOBEL database, and
dij the travel time between municipalities i and j 6.
The value of beta is calibrated according the distributions of travel time values from
the MOBEL data set during morning peak hours. Its calculated values is 0.0715 and
its Pearson correlation coefficient with cumulative probability of moving is 0.996.
On Figure 8, the potential accessibility to jobs during morning peak hours is
represented. The highest potentials are met in Brussels and Antwerp and have
important consequences for the potential of municipalities linked by motorways to
those major employment centres. A group of municipalities with high potential values
surrounds a North-South axis between Antwerp and Brussels. Here the impact of the
motorways network is clearly visible. On the contrary, the southern part of the country
appears to have a poorer job potential accessibility due its important distance to a
major employment centre7.
6 For computer time consideration dij is less or equal 60 min. 7 We should stress out that this methodology doesn’t take into account jobs available in bordering countries which could have impact the results especially for the Walloony/Luxembourg border.
29
Figure 8: Potential accessibility to jobs during morning peak hours
Accessibility potential to services during off-peak hours
We present here a second accessibility indicator taking into account the services of
Belgian cities. These services are supposed to have to be reached occasionally and
during off-peak hours. The level of equipment is measured by the equipment score
calculated by van Hecke in his work on urban typology. The equipment score
encompasses eight functions (health, leisure and sports, communication, public
service offices, public service administrations, culture, education and retail) by the
mean of several quantitative and qualitative indicators (Van Hecke (1998)).
A gravity measure will put in perspective the weight of nearby equipped cities
according to their equipment score, their distance to the municipality of origin. To do
so, we decided to build a continuous variable on the basis of this equipment score.
Ai = ∑ Dj.exp(-β.dij)
with Ai potential accessibility to jobs of the municipality i,
Dj the equipment score in the municipality of destination j,
β a parameter calibrated according to MOBEL database, and
dij the travel time between municipalities i and j with dij less or equal 60 min.
30
The value of β is calibrated according the distribution of travel time values from the
MOBEL data set in off-peak hours. Its calculated values is 0.0735 and its Pearson
correlation coefficient with cumulative probability of moving is 0.998.
Figure 9: Potential accessibility to services in off-peak hours
This potential is represented on a map (Figure 9). The highest potential values
clearly spots on the well-equipped cities such as Brussels, Antwerp, Gent, Liege as
well as the regional cities of the hierarchy (Van Hecke (1998)) which also show
values in the upper-class. The effect of the motorway network is stronger than for the
accessibility potential to jobs and the attractiveness of the cities is visible far along
the motorways axis. The southern part of Belgium appears once again as having a
weak potential accessibility to services.
The two gravity based accessibility indicators (i.e. peak hours accessibility to jobs
and off-peak hours accessibility to services) reveal similar spatial structures. These
two indicators should thus influence the localization model in a quite similar way.
31
Finally, the accuracy of estimated travel times seems to be correct for the model
coupling and the related simulation.
32
C. Travel demand model The modelling of peak hours accessibility needs simulating journeys to work
distribution between all origins and destinations (i.e. 589*589 Belgian municipalities).
Considering the available data and the accuracy needed for forecasting, the chosen
model is a doubly constrained gravity model. Indeed, in our project, such a model as
to be coherent with other MOBLOC models providing the margins of the
Origin/Destination matrix which justifies the choice of a doubly constrained model.
Moreover, the model has to take into account the diagonal and also local effects
affecting the distance-decay function and reflecting on one hand the major linguistic
dichotomy (Flanders vs Wallonia) and, on the other hand, the strong polarization
around Brussels. To achieve these goals a doubly constrained gravity model with
regional belonging effect and using mixed distance-decay function (combining
exponential and Pareto function) has been developed with MATLAB.
Gravity model From the knowledge of places of residence of the employed people and the number
of jobs per municipality, forming the margins of the flow matrix (Figure 10), it is
possible to model the origin/ destination flows for journeys to work. To do so, we
implement a so-called gravity model, by analogy with the theory of universal
gravitation. This model assumes that the intensity of flows between i and j depends
on the respective masses of the spatial units (jobs/employed people) and is inversely
proportional to the distance between them (thus regarded as an obstacle to the
interaction).
Fij: flow between i and j
Oi: number of employed residents in municipality i
Dj: number of persons working in j
T: total workers/jobs
Figure 10: Data: flow matrix
In order to ensure consistency with other MOBLOC models (including the localization
model), it is necessary to respect the margins of the O/D matrix and thus to constrain
the marginal totals. The doubly constrained gravity model is then conventionally
written as:
1 j n Oi 1 i Fij
n Dj T
33
where f is an exponential or Pareto function or a combination of them; and Ai and Bj
are the margin constraints.
It could be formalized as:
Besides the respect of margins, the case of Belgium also means taking account of
regional belonging and related linguistic specificities (Dujardin, 2001).
Introduction of border effect It is easy enough to take into account known border in the model by introducing
additional parameters. In the Belgian case, where three regions exist, it could be
written as:
where FB, FW and BW are binary variables equal to 1 if the flow crosses the border
between Flanders and Brussels (respectively Flanders and Wallonia, Brussels and
Wallonia) and 0 otherwise.
These barrier parameters measure the eventual brake (or accelerating) of flows from
one area to another: < 1 indicates a braking and conversely, > 1 indicates an
intensification of flows.
Model estimating Given these preliminary choices, the model can then be estimated in two main ways:
either by optimization (under constraints) or by a statistical method.
Estimation by optimization
The estimation by optimization has the advantage of not assuming any hypothesis on
the flows distribution and thus allows introducing additional explanatory variables.
However, it does not provide test statistics.
34
In this case, the criterion to minimize is the weighted chi-square which is written as:
Calculating and
β being a parameter of braking, it is necessarily negative. It is put = -1 for this first
step and by iterations, we calculate
The iterative process continues until the norms and
do not exceed a fixed epsilon.
From these first approximations, β is estimated by minimizing:
.
During the minimization, the parameters and are recalculated at each step.
We can calculate an adjustment factor by comparing the model residuals versus
the independence model: This model does not involve any parameter but simply
respects the conservation of marginal flows:
Then,
Statistical estimation
The equations of the model in its simplest form, can be rewritten as:
which suggests a linear model.
However, this method is not appropriate for several reasons: among them, we
include the non-compliance with the conservation of margins (and therefore of total
flows), failure to take into account the zero flows (since we minimize the least
squares: , or the assumption of normality of errors that is
not verified.
35
Finally, are measures of counts (ie integers) and this suggests a continuous
statistical distribution such as that of Poisson. Instead of choosing a linear or log-
linear regression, it is preferable to turn to a Poisson regression (Flowerdew R.,
Aitkin M., 1982).
Model assessment The choice of the ad hoc model is based on the comparison of several models. In our
case, four models were tested, varying the type of the distance decay and the
possible addition of a barrier/regional belonging effect:
1. Pareto model:
2. Exponential model:
3. Mixed model:
4. Mixed model with barrier effect:
To compare these four models, three goodness of fit statistics were selected,
namely:
1. Deviance:
2. Χ²:
3. AED (absolute entropy difference):
with , , ,
Additionally, two modelling procedure were used and compared for each model: the
GENMOD procedure from SAS software, on the one hand, and DAL-POULAIN
procedure, developed for the project MOBLOC by L. Dal (GéDAP, UCL), on the
other. These two procedures differ in the choice of estimators (optimization
procedure for DAL-POULAIN and statistical method for GENMOD, see Table IV)
36
Table IV: Goodness of fit statistics and parameters
GENMOD (SAS) MOBLOC (DAL-POULAIN)
Χ² 2,452,802.736 2,172,747.563
AED 0.224 0.622
Deviance 1,902,006.870 2,052,654.522
Scale 2.663 2.506
Par
eto
mod
el
β -2.147 -1.959
Χ² 5.9699698E14 16,165,603.353
AED 0.382 1.545
Deviance 3,242,955.556 6,345,416.024
Scale 41553.670 6.837
Exp
onen
tial
mod
el
β -0.128 -0.057
Χ² 5,218,387.423 1,816,077.712
AED 0.167 0.553
Deviance 1,424,267.213 1,675,262.105
Scale 3.885 2.292
Mix
ed m
ode
l
β [-1.5464; -0.0364] [-1.6427; -0.0169]
Χ² 4,283,521.443 1,576,916.186
AED 0.145 0.477
Deviance 1,232,247.132 1,434,702.069
Scale 3.519 2.136
Mix
ed l
mod
el
with
bar
rier
β [-1.5659; -0.0317; 0.0341; -1.4617;
0.7864]
[-1.6643 ; -0.0143 ; 1.0153 ;
0.3391; 1.9152]
Considering the results of this comparison, we chose to use the mixed model (to
estimate correctly both the short and long distance flows) with the regional belonging
effect (to take account of the strong polarization of Brussels and brakes between
Flanders and Wallonia). We also retain the DAL-POULAIN procedure insofar as it
does not involve any assumption on the distribution of flows.
Residual analysis At the end of this stage of modelling the gravity flow distribution, we obtain an
acceptable fit with the reference matrix (ESE 2001). However, some residuals exist,
notably through the overestimation of long-distance intercity flows (Figure 11).
The residuals are calculated as follow:
37
The main residuals of our model consist in the overestimating of long distance trips,
especially to Brussels Region. Furthermore, in the Brussels Region some
underestimating of intra-regional flows appears, highlighting thus the local
employment dynamics in the capital.
Figure 11: Inter-municipalities residuals
Considering, in a second step, the intra-urban residuals (Figure 12), it appears that
while the global fitting appears to be quite good, local specificities lead to some under
and over estimates of intra-urban journeys to work. At a regional scale, Wallonian
intra-municipalities flows are underestimated while Flanders and Brussels region trips
are more often overestimated. Moreover, the major under and over-estimatings
mainly affect the biggest cities.
38
Figure 12: Intra-municipalities residuals
Conclusion The gravity modelling of journeys to work distribution in the Belgian case implies to
take additional parameters into account, namely the regional belonging effect. At the
end of this step, a quite well fitted model was built in order to provide the best
accuracy possible regarding observed data. It seems now acceptable to use such a
model for the forecasting of daily trips in order to provide an estimation of flows for
the accessibility model.
However, some methodological improvements could be foreseen in order to enhance
the accuracy of the model. The challenge is then to better model the long distance
trips while keeping good fit for the shorter ones. It could be achieved by taking into
account additional parameters such as wages level or unemployment rates
depending on the availability of such data at the municipality level.
39
D. Propensity to move model In this section, let us first expose recent trends of residential migration in Belgium
during the past decades and present a set of explanatory variables to study this
event. In other words through this point we will lay the emphasis on the basic but
crucial elements for explaining and modelling residential migration. Then we will
describe our researches on modelling the propensity to move throughout different
steps which lead to the building of a propensity to move model. Finally, main findings
will be highlighted.
Before going deeper in methodologies it should be mentioned that we could access a
database (National Register - NR) describing “states” at successive 1st of January at
an individual level from 2001 till 2006. That means that we cannot observe any real
events but the ones deducted from transitions between two consecutive first of
January.
A second source of data is the national census (Socio-Economic Survey – ESE
2001) achieved in October 2001 giving only one shot information. This remark about
the lack of series data is important for the prospective step as we will see later.
Moreover, it must be clearly defined that by residential migration we only consider
migrations occurring between two municipalities since the data does not allow
investigating migration within a same municipality.
Residential migration: a concise overview First of all, we studied the evolution of residential migrations between municipalities
for the last decades. Figures highlight two main trends. On one hand residential
migration is rising. Since 1988, figures in residential migrations have strongly
increased, growing from 379,000 in 1988 to 481,000 in 2004 and even 532,000 in
2009 (Statbel, SPF Economie, PME, Classes moyennes et Energie, Direction
générale Statistique et Information économique). The calculation of an index which
corrects the effects of differences in age structure (age structure of the Belgian
population has been getting older for the last decades), shows that the propensity to
move has increased at all ages from 0 to 65.
On the other hand we observe that the periurbanization process is ongoing and
getting larger on the Belgian territory whilst young households are looking for
cheaper housing and building plot. Different factors can be exposed to explain these
trends. Among them, we should mention more frequent household transformations
that increase residential mobility and the "deadlock" of ancient periurban areas that
force in particular the younger ones who leave their parent's home to go farther and
farther away (Eggerickx et al., 2008).
40
Pre-selection of variables for the propensity model Let us recall that given the available data, we defined that a residential migration
occurs for an individual when his/her residential place (municipality) is different
between two successive 1st of January. So migration inside the same municipality or
multiple migrations in one calendar year are not detectable in the datasets since we
only have the records of the municipalities for the place of residence at each 1st of
January.
The measured “risk” reflects thus the fact that the residential municipality at a 1st of
January is different from one year to the next one. This kind of "event" is quite rare
since it only concerns 4% of the Belgian population each year.
The explanatory variables were selected according to a literature review (e.g.
Debrand and Taffin, 2005; or Henley, 1998) and to their availability through the
National Register or the Socio-Economic Survey of 2001. We are dealing with
individual characteristics as well as housing and area of residence characteristics.
As individual characteristics we selected age, gender, nationality, household
type, number of individuals per household, position in the household (i.e. the
link with the head of household), the highest education level successfully
completed, the activity status and type of activity, and if a migration occurred
the year before.
As housing characteristics we considered the type of housing (house, flat…)
and housing tenure type.
As regards area of residence characteristic, we took into account the
urban/rural profile of the municipality (downtown, rest of the city, old and
recent periurban areas, rural areas) built by Van der Haegen et al. (1996).
In order to explore what could be the most relevant covariates we built some
contingency tables with these explanatory variables and the binary variable “Has an
individual moved? Yes or no”. We will not expose the results in details but just
indicate which categories of people have the highest (or the lowest) propensity to
move for each covariate we selected (Table V).
41
Table V: Categories having the highest or lowest propensity to move regarding available
explanatory variables
VariablesCategories having the highest
propensity to moveCategories having the lowest
propensity to moveAge 19-29 55-74
Gender Men Women
Nationality Congolese RDC Turk
Household type Cohabitants without children Married couple without childrenNumber of people per household
1 individual 4 individuals
Position in the household Other (than head, partner or child) PartnerPrevious migration (the year before)
Migrated the year before Did not migrate the year before
Place of residence Agglomeration (downtown excluded) Small cities
Highest education level successfully completed
High education Primary or no instruction
Activity status People looking for a job Retired
Housing tenure type Renters in the private sector Owners
Type of housing Apartments 4 front houses
On the basis of the previous observations and of practical considerations we selected
only a group of these variables. We could indeed not keep the whole set for the
following reasons:
On one hand, we recommended the propensity to move model to focus mainly
on individual characteristics; so it excluded de facto the urban/rural profile of
the area of residence and the type of housing.
On the other hand, we found some limits provided by the data for the
upcoming steps. Indeed, in the second phase of this project, we planned to
achieve forecast for municipalities or group of municipalities. So we already
need to know which variables will be available for extrapolation. The problem
is that extrapolation needs individual data at at least two dates, and yet for
some explanatory variables such as education level, occupation status, or
housing tenure type, we know that we only can have data for a single time i.e.
October 2001, from the 2001 Socioeconomic Survey.
We thus had to consider which are the most essential variables from the 2001
Socioeconomic Survey to be retained for the model. Here are explained the most
crucial choices:
Education level versus activity status
Since the highest successfully completed education level and the activity status are
strongly associated, we decided to keep only one from these two variables. We
finally chose the education level for two main reasons:
42
Firstly, the way activity status and activity types have been built in the ESE2001
survey is not satisfying: the variable "activity status" contains modalities of responses
that are not exclusive: one can be student and at the same time in activity; so we
wondered how people answered choosing one rather than the other category. As
regards "activity types", the main critic is that in the 2001 census a distinction is made
between workers and employees in the private sector, but not in the public sector so
that we cannot compare the same positions in the two sectors.
Secondly, education level is a "capital" that one cannot lose contrary to activity
status; hence, it is a variable more stable in time and should be easier to project in a
near future. In comparison, occupation status or activity types might change from one
year to the other according to the conjuncture or socioeconomic policy. Therefore,
these variables are more volatile and then would be more difficult to manage for
forecast.
In this context and given that education level is generally as good as activity status
as a proxy for the socioeconomic profile of individual we preferred selecting
education level.
Housing tenure type
We have also kept in the set of variables the housing tenure type because it is rather
essential to analyze residential migration: this variable largely discriminates people in
migration depending on whether they are tenant or owner, or whether they are tenant
in public or private market. For instance the ratio of migrating people is more than 3
times higher for tenants in a private dwelling than for owners.
Extracted from the Socio-Economic Survey of 2001, housing tenure type at individual
level is only available at one date (October 2001). However trends provided by the
National Institute of Statistics shows that this is stable along time, especially in
Belgium where ownership is traditionally high: between 1991 and 2001, the
percentage of owners increased from 65.4 to 65.9%, and forecast concerning
potential impact of an economic uncertainty on becoming owners do not seem
significant on the Belgian market (Vanneste et al., 2007). In this context, it appears
that this variable would be stable during time, and not too complex to forecast in a
near future.
However an additional disadvantage for the housing tenure type is that it is provided
at the household level. That means that if the head of household answers that he/she
owns his/her housing then each household’s member is as well considered as owner.
This makes problems especially if we are interested in leaving parents-household
phenomenon because if a young adult lives with his parents’ owners in 2001, he is
then also considered as owner. As we have only data at one time, once he leaves his
parents' home he will still be considered as owner whatever the new housing.
Despite these reluctances, we have included this essential variable and tried to deal
with its drawbacks.
43
Here are described the currently pre-selected variables and their modalities:
Age (years over January, the 1st) - 6 modalities: 0-18, 19-29, 30-44, 45-54,
55-74, 75 and more.
Gender - 2 modalities: Male, Female
Nationality - 8 modalities: Belgian, Congolese, Moroccan, Turk, citizens from
the 12 new EU members, citizens from the 15th former EU members and
citizens from West European countries non EU members (Switzerland,
Norway, Iceland,..), rest of East Europe (Balkans, Byelorussia, Russia,
Ukraine), rest of the world.
Household Types - 8 modalities: Married couple without children, Married
couple with children, cohabiting couple without children, cohabiting couple with
children, single-household, one-parent family, collectives and households
"others" (with more than one family kernel).
Position in the household – 4 modalities: head of household, spouse (wife,
husband, partner), child, others. The position is determined in relation to the
head of household.
Number of people per household – 5 or 6 modalities: 1, 2, 3, 4 or more, or 1,
2, 3, 4, 5 or more. The total amount of modalities depends on household types
and except for single-household and couple without children, the last class is
open.
Highest education level successfully completed – 5 modalities: no instruction
or primary, secondary school (inferior), secondary school (superior), higher
education, no answer.
Housing tenure type - 5 modalities: owner, private tenant, social tenant, tenant
of a free housing, no answer.
Residential migration during previous year – 2 modalities: yes, no.
Methodological choices to build the model The technique used is a binary logistic regression. The original idea was to use
discrete choice model for the residential migration model. For this purpose, it was
planned to use BIOGEME (software developed by Michel Bierlaire from the EPFL,
Lausanne). Since we faced some problems with the size of the database and given
that, for the propensity model (first part of the migration model) we had a binary
dependant variable and therefore considered to choose a logit function, we decided
to make a binary logistic regression rather than a discrete choice model.
Nevertheless discrete choice modelling remains the technique we used for the
localization model because it offers more possibilities when there are a lot of
alternatives.
There are different ways to conduct a logistic regression. At a time, our ambition for
the propensity model was to deal with panel data methods. Using a five-year period
44
database, we first explored how to build a model on the basis of the years 2001 to
2005. For this purpose, we built a person-year database: it means five observations
per individual. But at the end, this method had some drawbacks for our case so that
we finally opted for a simple binary regression model.
We had two main constraints when building this model:
Annual step for the migration and for the planned projections
When we started to build the model for propensity to move, the question arose
about how to express the temporal unit of the dependent variable (and
explanatory variables). As we had information at each 1st of January it seemed
obvious that the most useful and simplest way to study migrations would be
based on an annual step, which would estimate residential migration within one
year, i.e. between two consecutive 1st of January.
Take advantage of the dynamic dimension: past and anticipation effects.
As we already mentioned, the available database is a sequence of states for the
overall Belgian population during the 5-year period 2001-2005.
By convention in demography, when we refer to a "period", it means a period in
years of time. For instance, the 2001-2005 period is a 5-year period in year of
time; i.e. from 1st January 2001 till 31st December 2005. However, in this section
and in particular here below we also use the terminology of the "transition
period" which refers to the year between 2 firsts of January, e.g. the transition
period 2002-2003. This results from the use of observed differences of states
between two consecutive 1st of January.
Data from the National Register provide the demographic profile of each
Belgian people each year (while data from the 2001 Socioeconomic Survey
supply some other variables, but only for October 2001). Therefore we want to
take advantage of the dynamic dimension conveyed by this database.
Indeed, it seems obvious that past information about the life of an individual can
help to explain his/her migration behaviour. For example a change in the
household’s structure or size (which can reflect events such as separations,
births or unions) can increase (or decrease) the probability for change in
residential municipality. At the same time we also want to test effects from
anticipation. For instance, testing whether an observed migration for a couple in
the transition period Y to Y+1 may be explained by the fact that this couple has
a baby in the transition period Y+1 to Y+2. They could have anticipated the birth
by looking for a bigger housing the year before.
45
While building the model which has to explain migration between years Y and
Y+1, we looked for a suitable way to include this dynamic dimension. Rather
than simply use the state variables at the first of January of the years Y-2, Y-1,
Y, Y+1 and so on, we decided to use the state variables only at time Y and to
create transition variables which describe variation in the state variables
between two successive 1st of January.
These transition variables have been created for variables from the National
Register i.e. household structure and size, position in the household and
nationality. But for anticipation we have only considered household transitions
(variables related to household). Regarding the Socio-Economic Survey of 2001
variables, as they are only available in October 2001, it was not possible to
create transition variables.
These transformed variables are helpful to add in the model a "temporal depth"
of one or more years back or forward. From a statistical point of view, transition
variables act as a substitute of a sequence of temporal successive states and
have the great advantage to handle multicolinearities or autocorrelation, which
could occur when using these states as explanatory variables in the model.
Moreover, they make easier the interpretation of the results since transition
variables implicitly suppose an event (although not observed).
At a last point, it should be said that since dynamic database allows catching
unobserved heterogeneity, we add in the model an autoregressive dimension
which is the dependent variable at the previous year (did an individual move the
year before?).
Transformation of variables Before creating transition variables, we studied and applied specific treatments for
intrinsically highly associated variables such as household types and number of
people per household. Indeed, e.g. single households have all one person living,
while married or cohabitant couples with no children are all composed of two
members. In other cases, the association between household type and household
size is not univocal so that we wanted to keep information from both variables. As a
consequence, we created synthetic variables based on these two variables
describing the household type and the number of people living in it.
Another variable transformation was the setting of transition variables. As already
described, such a variable substitutes for 2 successive state variables. For each
individual, if these two successive state variables are identical then the value of the
transition variable is “no change”. If not, we first concatenated the two successive
states and then we grouped these categories in less numerous modalities that could
explain the migration behaviour.
For instance naturalization variable 0102 was built from two variables: nationality in
2001 (more precisely at 1st January) and in 2002. If this nationality has changed we
46
distinguished two cases: if it was a nationality change towards the Belgian or the UE
citizenship, then we consider it as a change (i.e. naturalization); but if it was a change
towards another extra-European nationality, then we do not. Here, becoming Belgian
or UE citizenship has been only seen as a change because becoming Belgian or a
European citizen might encourage a long stay in Belgium and the adoption of local
residential migration behaviour.
In the same way, we redefined a lot of transitions, especially for the synthetic variable
(household type and size).
Sample design The question to use a sample rather than the entire base arose when we faced some
problems with the size of the database and the calculation time. As there are more
than 10 million citizens in Belgium, we have about the same number of observations,
which is more than necessary. We may indeed reasonably reduce the size of the
sample without losing quality of information for the model.
The sample design was also a constraint and needed technical adaptations. Since
the model deals with residential migration within the country we worked with an
enclosed population which was people who were residents in Belgium at the
considered years.
Furthermore, including the "temporal depth" - of one or two years in the past and one
year after the reference year - implied to add the condition of being resident in
Belgium during at least 3 consecutive years (or 4 consecutive 1st of January).
Furthermore as residential migration is a rare event (4% per year), we wanted to
increase its frequency in order to improve the quality of the coefficients of regression
(Allison, (1999). This is the reason why we oversampled the migrants. As a
consequence the sample is based on a stratified random sampling in the overall
Belgian population. Stratification has been done according to the dependent variable:
for one year observed, we selected all people having changed from municipalities
and, for the non-migrants, we realized a simple random withdraw. The total number
of drawn individuals arises 2 millions and the percentage of annual migrants comes
around 15%. It should be mentioned that we realized several sampling depending on
the years of the studied migration. If we study the migration between the firsts of
January of the years Y and Y+1, then we use the sample in which migrants in the
transition period from Y to Y+1 are oversampled.
Finally from each sample, we created two subsamples:
the first one is composed of 70% of the individuals and is used for the model
calibration,
the second one with the rest 30% to test and validate the models.
47
The propensity model is based at the individual level. If we agree that the decision to
migrate can be individual, it could also be the decision of the household, or at least of
the couple.
At first we wanted to integrate the household and individual units in order to take into
account both units of decision. However we realized that drawing a sample on
household units with a temporal depth was hardly impossible because the numerous
household transformations do not allow the follow-up of every individual in the
sample. Indeed, between two consecutive years one household can be divided into
two households; and one individual can change from a household to another. It is
thus difficult to build a model with households as units and to study their migration
behaviour with a temporal depth. Hence, we chose to limit the propensity to move
model at individual level. However we included explanatory variables related to
household characteristics and positions in the household that could help to take
indirectly this second unit into account.
Model(s) Given that the available data covers the five-year period from begin 2001 to end
2005 (1st January 2001… 2006), we finally concentrated on two transition periods:
2002-2003 and 2003-2004. The reasons for this choice are:
First of all, since variables from the Socio-Economic Survey of 2001
(education level, occupation status, housing tenure type …) are only available
at that time (in our individual database), the closest data from the National
Register (household structure, nationality …) are thus those from the first of
January 2002; that is why we first decide to study 2002-2003.
But we also wanted to introduce, as explanatory variable, transformations in
the household structure occurred in the past of the individuals (such as
marriage, divorce, birth …). Unfortunately, the available data from the National
Register only covers 2001-2005, so the only previous period that we could use
to explain the period 2002-2003 was 2001-2002. In order to test the
improvement of the quality of the model by adding older information, we
decided to also study the transition period 2003-2004. We have to notice that
the data from the Socio-Economic Survey of 2001 are less contemporary and
so less relevant in the model for 2003-2004 than in the one for 2002-2003;
that is why we have not chosen, for example, 2005-2006. If we want to
estimate the propensity to move between 2005 and 2006, information from
2001 start to be out of date. So we chose to concentrate on migration between
2002 and 2003 and between 2003 and 2004. This implied that we fixed the
maximum "temporal depth" at two years back because we had no information
before 2001 from the National Register.
48
Finally, the purpose (of studying two different periods) was also to determine if
the results (parameters) from the models for 2002-2003 and 2003-2004 were
comparable in time and led to the same interpretations. It was actually
important that the model would be independent from the transition period it is
based on, since it had to be used to predict the migration behaviour in the
future (population forecasting).
To sum up, we tested different sets of variables in order to:
compare migration model for two different transition periods : 2002-2003 and
2003-2004,
test the effects/advantages of adding older (delay) or future (anticipation)
information.
The Table VI below summarizes the five models tested and the variables selected
(with their source) in each case:
Table VI: Explanatory variables of the tested models for the propensity to move
Dependant Variable
Models Explanatory Variables (with sources)
Age in 2002 (NR)Gender (NR)Nationality in 2002 (NR)Type and size of the household in 2002 (NR)Evolution of the type of the household between 2001 and 2002 (NR)Evolution of the type of the household between 2002 and 2003 (NR)Link with the household head in 2002 (NR)Evolution of the link with the household head between 2001 and 2002 (NR)Evolution of the link with the household head between 2002 and 2003 (NR)Migration between 2001 and 2002 (NR)Education level successfully completed (ESE 2001)Housing tenure type (ESE 2001)Age in 2002 (NR)Gender (NR)Nationality in 2002 (NR)Type and size of the household in 2002 (NR)Evolution of the type of the household between 2001 and 2002 (NR)Evolution of the type of the household between 2002 and 2003 (NR)Evolution of the type of the household between 2003 and 2004 (NR)Link with the household head in 2002 (NR)Evolution of the link with the household head between 2001 and 2002 (NR)Evolution of the link with the household head between 2002 and 2003 (NR)Evolution of the link with the household head between 2003 and 2004 (NR)Migration between 2001 and 2002 (NR)Education level successfully completed (ESE 2001)Housing tenure type (ESE 2001)
Model 1
Model 2
Migration between 2002 and 2003
49
Dependant Variable
Models Explanatory Variables (with sources)
Age in 2003 (NR)Gender (NR)Nationality in 2003 (NR)Naturalization between 2002 and 2003 (NR)Naturalization between 2003 and 2004 (NR)Type and size of the household in 2003 (NR)Evolution of the type of the household between 2002 and 2003 (NR)Evolution of the type of the household between 2003 and 2004 (NR)Link with the household head in 2003 (NR)Evolution of the link with the household head between 2002 and 2003 (NR)Evolution of the link with the household head between 2003 and 2004 (NR)Migration between 2002 and 2003 (NR)Education level successfully completed (ESE 2001)Housing tenure type (ESE 2001)Age in 2003 (NR)Gender (NR)Nationality in 2003 (NR)Naturalization between 2001 and 2002 (NR)Naturalization between 2002 and 2003 (NR)Type and size of the household in 2003 (NR)Evolution of the type of the household between 2001 and 2002 (NR)Evolution of the type of the household between 2002 and 2003 (NR)Evolution of the type of the household between 2003 and 2004 (NR)Link with the household head in 2003 (NR)Evolution of the link with the household head between 2001 and 2002 (NR)Evolution of the link with the household head between 2002 and 2003 (NR)Evolution of the link with the household head between 2003 and 2004 (NR)Migration between 2001 and 2002 (NR)Migration between 2002 and 2003 (NR)Education level successfully completed (ESE 2001)Housing tenure type (ESE 2001)Age in 2003 (NR)Gender (NR)Nationality in 2003 (NR)Naturalization between 2001 and 2002 (NR)Naturalization between 2002 and 2003 (NR)Naturalization between 2003 and 2004 (NR)Type and size of the household in 2003 (NR)Evolution of the type of the household between 2001 and 2002 (NR)Evolution of the type of the household between 2002 and 2003 (NR)Evolution of the type of the household between 2003 and 2004 (NR)Evolution of the type of the household between 2004 and 2005 (NR)Link with the household head in 2003 (NR)Evolution of the link with the household head between 2001 and 2002 (NR)Evolution of the link with the household head between 2002 and 2003 (NR)Evolution of the link with the household head between 2003 and 2004 (NR)Evolution of the link with the household head between 2004 and 2005 (NR)Migration between 2001 and 2002 (NR)Migration between 2002 and 2003 (NR)Education level successfully completed (ESE 2001)Housing tenure type (ESE 2001)
Model 3
Model 4
Model 5
Migration between 2003 and 2004
We used the stepwise procedure (Allison, 1999) selecting one by one the variables
which are significantly related to the explanatory variable conditionally to the
previously entered variables. This procedure can also eventually remove one of
these covariates thereafter, if it has lost its significance due to the addition of another
variable.
50
In our models, all of the explanatory variables removed from the regression because
of lack of significance concerned naturalization:
naturalizations between 2001 and 2002 and between 2002 and 2003 for
models 1 and 2;
naturalization between 2003 and 2004 for models 4 and 5;
Then we have compared these models thanks to some global criteria such as AIC
and log likelihood. We present here below (Table VII) the different likelihood ratios of
our models:
Table VII: Tests on the likelihood ratio of the 5 tested models
Khi-2 ddl Pr > Khi-2Model 1 502377,9 215 < 0,0001Model 2 505456,7 301 < 0,0001Model 3 511072,5 219 < 0,0001Model 4 511247,3 306 < 0,0001Model 5 506794,7 394 < 0,0001
Tests on the Likelihood Ratios
The differences between the tested models are not large. From these data we can
assess that all our models have a good calibration: indeed, all khi-2 tests on
likelihood ratios are extremely significant.
Another way to assess models performance relies on the discrimination: does the
model well predict the behaviour of the individuals? Is somebody moving also
predicted as a mover? Does the model well separate the migrants from the non-
migrants? If we have a look at the Figure 13 here below, we can see that principal
diagnostics about discrimination show little amelioration between models 1 and 2
and from models 3 to 5. All four statistics have to present high scores if the models
are well fitted.
51
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
Somers' D Gamma Tau-a c
Model 1 Model 2 Model 3 Model 4 Model 5
Figure 13: Comparison of the 5 models of propensity to move according to various
diagnostics (regarding to the predictive power)
The fact that model 3 seems to be worse than the first 2 models (regarding the
predictive powers) can be explained by the loss of relevance of covariates coming
from the census.
The statistic c is an approximation of the area under the ROC Curve which also
provides a measure of discrimination. As a general rule, when this area is between
0.8 and 0.9 one can consider the model have an excellent discrimination (Hosmer
and Lemeshow, 2000). In our case values of c are close to 0.9 which means our
models are really discriminatory.
As a conclusion about the models, we remark that the 5 models are really good. As
the first one seems to be already more than satisfactory, we decided that this one
could be used in the next steps of our project.
We give here below (Table VIII) the four most explanatory covariates of this first
model (in a decreasing order) as they also appear in the following table for the first
model:
evolution of the type and size of the household between Y and Y+1
housing tenure type
evolution of the link with the household head between Y and Y+1
age in Y (six classes)
52
Table VIII: p-values of the explanatory variables of the propensity to move model
Covariates ddl Khi-2 pvalueEvolution of the type of the household between 2002 and 2003 74 59530,5 <,0001Housing tenure type 4 48683,1 <,0001Evolution of the link with the household head between 2002 and 2003 12 32138,8 <,0001Age in 2002 5 12677,9 <,0001Type and size of the household in 2002 17 4079,0 <,0001Migration between 2001 and 2002 1 3563,2 <,0001Evolution of the type of the household between 2001 and 2002 74 2946,7 <,0001Link with the household head in 2002 3 2458,2 <,0001Highest education level successfully completed 5 2384,8 <,0001Evolution of the link with the household head between 2001 and 2002 12 590,0 <,0001Nationality in 2002 7 218,0 <,0001Gender 1 34,2 <,0001
Are the Covariates of Model 1 Significant?(Migration between 2002 and 2003)
The most significant variable is clearly the simultaneous transition of the household
structure (type and size). Compared to the reference category (no change in the
household structure between the two consecutive years), all the odd ratios are
positive which means that any change in the household structure is correlated with
an increasing probability to move (everything else being equal). The importance of
this increase depends on the household structure change. Amongst the most
significant, let us point out the people whose household structure were "married
couple with children" and became "unmarried couple with/without children", "one-
parent family" or "isolate". They have from 20 to 30 times more likelihood to move
than people whose household structure did not change.
Let us remember that family events are univocal in the data. For example, such
people could be a child who left parents' home, or member of a couple which split up,
or etc.
This can be related to another significant variable: the change of relation to the
household head. For example, from “being a child” status to a “head” or “spouse”
status (or inversely) is related to a high increase of the probability to move (between
11 and 16 times more than in the case of no change).
The second most significant variable for predicting the propensity to migrate is the
housing tenure type. Tenants in the private sector are 4 times more likely to move
than owners.
53
Age class also appears amongst the four most significant variables. The less
residentially mobile are 75 years old and more while the most mobile are the
youngest (less than 18 years). The probability to move decreases with age
(everything else being equal).
All these results underline that the propensity to move is linked to the life course of
individuals, and more particularly their family trajectories. These are well observable
through the age and transition of the household structure covariates. To sum up, the
transitions leading to move are break-up situations, new family-units compositions
and leaving parents’ home. One can observe that the more stable situation concerns
people who are in a married couple (with or without children) ; this situation is often
associated with an owner status, which is another factor to stay in the same
municipality of residence. In other words, the likely evolutions of family situations
marked by the rise of less stable households (cohabitation situations, one-parent
family…) will still generate higher propensity to move rates in the coming years.
54
E. Localization model
Preliminary analysis Contrary to the propensity to move model (where we only focus on individual
characteristics to determine if people are going to leave their places of residence),
we considered also municipalities’ attributes to model the choice of a new residential
municipality.
A way to approach the residential migration is to study the distances between the
new and the former municipality. The Figure 14 illustrates this distance (between
municipal centroids as we do not have more accurate localization in the database) for
a sample of 100,000 people having moved between 01/01/2001 and 01/01/2002.
0
10
20
30
40
50
60
less than 10 km 10 to 20 km 20 to 50 km more than 50 km
Distances
%
Figure 14: Intercommunal distances when a change of residence municipality occurs
between 2001 and 2002 (sample of 100,000 heads of household)
One can observe that residential moves in Belgium are often short distance: for
almost one in every two people changing their residential municipality, the
intercommunal distance is less than 10 kilometers, whereas only 10 percents of the
movers decide to go living 50 kilometers or more away from their former municipality
of residence.
Other preliminary analyses were done with the covariates available for the modelling.
For each of them, we estimated how each can bring explanations (Figure 15). To do
so, we estimated models with only one explanatory variable (age of the individuals,
55
distances between municipalities…) under BIOGEME (Bierlaire, M., 2003) and we
present here below (Table IX) the results. Models are sorted by adjusted rho-square
values. They were run with a sample of 60 municipalities and considered residential
migration between 01/01/2001 and 01/01/2002 for head of household.
Among the available variables, the less explanatory variable taken independently is
the property prices indicator. This variable takes into account the average prices for
houses and flats in the municipality. On the contrary, the most explanatory one
concerns the distance between the municipality where the individual lived in 2001
and the municipality where he settled down in 2002.
0,000 0,050 0,100 0,150 0,200 0,250 0,300 0,350 0,400
Accessibility to Employment(according to the age)
Accessibility to Sevices (according tothe age)
Out of district moving
Distance
Living conditions indicator(environment aspects)
Living conditions indicator (dwellingaspects)
Living conditions indicator (socio-economic aspects)
Proprety prices
Population of the new residentialmunicipality
Age (7 classes)
Kind of household (5 classes)
Education level (4 classes)
Nationality (4 classes)
Co
vari
ates
Adjusted Rho²
Figure 15: Adjusted Rho² of each covariate taken as only explanatory variable
Remark that β values can be interpret as follows: if β is higher than 0, it means that
the attractiveness of the municipality increases (its utility raises) whereas if β is lower
than 0 the attractiveness of the municipality decreases (its utility goes down).
56
Table IX: Parameter of each covariate modality (covariate being taken as only explanatory
variable - let us point out that the only significativeless parameter (at α = 0.05) in this table is
the one related to nationality of border countries, although its p-value is 0.10)
Covaraites β valuesProprety prices indicator -0.580Education : other 5.77Education : secondary school (inferior) -4.82Education : secondary school (superior) -6.65Education : higher education 7.19Living conditions indicator (dwelling aspects) 3.73Age : 0 to 18 -36.3Age : 19 to 29 46.4Age : 30 to 44 -13.5Age : 45 to 54 -27.6Age : 55 to 64 30.5Age : 65 to 74 33.4Age : 75 and more 53.8Nationality : other 22.9Nationality : belgian -5.86Nationality : border countries 2.09Nationality : European Union 51.3Accessibility indicator to employment for active people 0.629Accessibility indicator to employment for non-active people 0.507Living conditions indicator (socio-economic aspects) 6.06Accessibility indicator to services for active people 0.0803Accessibility indicator to sevices for non-active people 0.0649Population of the new residential municipality 0.746Kind of household : other 20.0Kind of household : couple (married or not) with children -10.7Kind of household : couple (married or not) without children 13.7Kind of household : single-household 38.4Kind of household : one-parent family 34.1Living conditions indicator (environment aspects) 8.23Staying inside the same district 3.66Distances between the new residential municpality and the former one -0.0665
In general, municipality variables seem to bring more explanatory power than
individuals ones.
Methodology The purpose of the localization model is to determine which “new” municipality will be
chosen by people who decided to migrate (between two consecutive years as
studied in the propensity to move modelling). To do so the model will take into
account the covariates which significantly play a role in this choice and their relative
weight in the decision process.
57
Discrete choice method
The objective for Mobloc and for this part of the project is to understand processes of
residential migration. The model may not be a black box that gives a new situation
but it has to describe some internal mechanisms of decision, to decompose the
decision in some interpretable parameters, so that, it could be possible to act on
some parameters for simulating the results of different changes/evolutions of
behaviour or the effects of some policies.
Even if the macroscopic methods are more common and are useful for some
purposes, our interest was however more compatible with micro-simulations.
Therefore discrete choice method (Train, 2003) was chosen for modelling as well as
possible the behaviour of people moving from a municipality to another. This
technique consists in determining the utility of each alternative and then computing
the probability of choice of each alternative. Nevertheless resorting to such models
was notably ambitious because some aspects of the project and the available data
would lead us to difficult discussion and decision.
To sum up the way discrete choice methods work, we can simplify things as follows:
each agent of decision (in our case, each person leaving his housing to settle down
in another municipality), evaluates the utilities of each alternative (which are all the
other 588 Belgian municipalities in the frame of this project). The estimation
supposes that a subject chooses the alternative with the highest utility. So only the
difference between utilities is important, not the utility itself.
The utility of each alternative is a combination of explanatory variables weighted by
some parameters (being estimated in a calibration phase). As we will develop below,
these variables can be related to the unit of decision (age, type of household…) or to
the alternatives (property prices in the municipality, distances between the alternative
and the previous municipality...). The formulation of the utility can be summarized as
shown on Figure 16:
Figure 16: Formulation of the utility in discret choice methods
With:
- Uin: utility of the alternative i for the individual n
- Ci: alternative specific constant
- PIndij: the parameters for individual covariate j and alternative i
- VarIndjn: covariates related to the individuals
- PMunk: parameters related to the alternatives (municipalities) characteristics
- VarMunk: covariates related to the municipalities
- εin: random term (to model unobserved effects)
58
Let us note that the utility includes a random term which allows the model not to be
deterministic and takes unmeasurable or unknown effects into account.
Different hypotheses can be made for the relation between the utilities and the choice
probabilities. The most-used model is the logit formulation. Another possibility was to
use a probit formulation which is less restrictive regarding the necessary
assumptions but substantially more time consuming so that we choose the logit
formulation.
Concerning the choice of the software for calibrating the model, we opted for
BIOGEME (Bierlaire, 2003). This software is more flexible in the utility formulation
than many other ones. Moreover the last versions of BIOGEME include Biosim, a
package designed to perform simulations from the estimated model.
The choice of the discrete choice method came quite soon and easily in the project
but some points of the model structure required long discussion and many trials
before leading to an acceptable model. This phase of the project clearly benefited
from the multidisciplinary nature of the research team. In particular the definition and
choice of explanatory variables and the choices made to reflect the dynamic nature
(time depth) of migration process were very enlightening.
We summarize here below one of the main considerations we faced:
Unit of decision (individual VS household)
As explained before, the propensity to move model was built with individual as
unit/agent of decision. This choice was quite natural as the members of a
household can have different behaviours. For example, if a child moves from
his parents’ house, only one individual from the household decides to move.
For the localization model, we would intuitively like to consider a group of
individuals. Indeed, if some members of a household move to a new
municipality, they will move together and “choose together” the same
municipality. The unit/entity of decision should be this (new) entity/household.
Nevertheless, we cannot take the initial household for this unit because the
composition of the households changes between two consecutive years.
The solution we adopted was to work with all the individuals who moved and
were household head in their new municipality. It thus includes individuals
who remain the household head (e.g. the whole household moves) and
individuals who become household head (e.g. he/she moves alone…).
59
This unit “head of household” is characterised by some household variables
(as the household type and size) so that the household information can also
be used for the localization decision as it would intuitively be the case. A
disadvantage of working this way is that we only consider that, for example,
the level of education of the household head influences the decision and we
should therefore suppose that this one is representative of the whole
household. However the lack of other types of data prevented us to act
another way.
Number of alternatives
The main challenge for this model is the number of alternatives. As the common
space level defined for Mobloc is the municipalities, the choice set is composed of
the 589 Belgian municipalities (although for a given individual, his original
municipality is not available in the choice set, leading to 588 alternatives for the
model).
So many alternatives in the choice decision is unusual in the literature. The theory
does not set a limit in the number of alternatives. Nevertheless, we could suppose
(and actually faced) that some complications could appear such as:
computing problems due to the amount of necessary data,
difficulty in the validation of some hypotheses, such as the
independence of the alternatives (IIA).
Sampling of alternatives
In order to reduce the time of computation, it is possible to limit the number of
alternatives that are considered as available for each unit of decision during
the estimation process. It means that, for each unit (household head), we
select randomly only some of the alternatives. Then, the algorithm compares
the utility of the chosen alternative (observed in reality) with the utilities of the
selected alternatives for this unit. (The selected alternatives are not the same
for all the household heads).
From the theory (Ben Akiva and Lerman, 1985) we know that such a process
does not change the value of the estimated global parameters if the
hypotheses are respected (IIA, for example). Nevertheless, the precision is
reduced and the confidence interval is larger.
Working this way allowed us to reduce the computing time. We performed
estimations with 10 and 588 alternatives and results showed that we could
gain substantial time for the estimation process using sampling procedure.
Model structure
There are different ways to express the link between the utilities and the
choice probabilities. Different forms of discrete choice models exist: logit,
60
probit, and more sophisticated structures. For example, the nested logit model
takes into account similarities between alternatives: it considers that some
alternatives belong to a same group by the way of additional parameters.
In our case, the nested logit was a promising way to explore. With so many
alternatives, we could suppose that some similarities between municipalities
would not be explained by the variables. We considered two sources of such
similarities: the geographic proximity and the municipality type (rural, urban,
etc.).
No automatic method exists to form the nests of alternatives; the question on
how to group the 589 alternatives was thus also a big issue. Two main ideas
were explored:
geographic proximity,
municipality type (rural, urban, etc.) : on basis of Van der Haegen
typology or Van Hecke classification,
both of them.
Variables
Another issue is related to the variables to be taken into account in the model
formulation. In our case, here are the main available variables:
Individual variables
o age, education level, nationality, household type.
Source : National Register coupled with the Socio-Economic Survey
(ESE2001).
Communal variables :
o Living condition indicator(s), property price, intercommunal distance.
Source : previous project results.
o Potential accessibilities (services, job opportunities).
Source : accessibility model.
Utility formulation
Another issue regarding the number of alternatives concerns the formulation
of the utility. The utility term for a variable can take different forms. We can try
to estimate a model with a generic parameter for a variable. It supposes that
this variable influences all the alternatives on the same way. Another method
is to use alternative specific parameters, which means that this variable has a
different parameter for each alternative. This can improve the model if the
estimated parameters are significantly different from each others. Intermediate
forms may also be considered.
In our case, performing tests with alternative specific parameters for a variable
means that 589(588) parameters should be estimated (per variable). Working
this way would see the number of parameters exploding. Moreover, most of
61
these would be insignificant or inaccurate, especially for the alternatives
(municipalities) for which we do not have many observations (small
municipalities). (It would require a lot more of data).
In our modelling, we can give as an example the case of the age covariate: we
decomposed this variable in 7 classes (less than 18, 19 to 29, 30 to 44, 45 to
54, 55 to 64, 65 to 74 and 75 and more); so if we kept the same way of
formulating as what is usually achieved, 4,123 parameters (7 * 589) would
have had to be estimated for the consideration of only this covariate in our
utility. Problems would rapidly occur in the estimation process if working like
that (unidentifiable model and/or too high time running).
To avoid this difficulty, we decided to use a generic parameter for all the
alternatives. To come back to the example of the age, we reduced so the
number of parameters to 7.
As explained before, it was unrealistic to use alternative specific parameters
and we decided to use generic parameters. Such a technique reduces the
number of parameters to estimate but it makes the individual variables non-
explanatory. Each of the 589 alternatives would indeed have the same utility
term for the variables of this kind (e.g. the value of the age variable would be
the same whichever municipality is considered). As the choice lies on the
difference between alternatives utilities, this term may not be explanatory (the
model could not distinguish amongst alternatives relatively to such a variable)
and lead to unidentifiable models.
In order to get different values for the explanatory variables and to avoid
problem of constancy between the alternatives, we had recourse to a
“contextualization approach”: for categorical individual variable; the utility term
of an individual variable is computed as the product of a class specific
parameter, a binary variable indicating whether an individual belongs to this
class and the class proportion in the municipality (alternative) (e.g. for the
“less than 18” age class, the utility term is the product of the associated
parameter (giving the weight of this variable in the utility), a binary variable
having 1 as value if and only if the decider is less than 18 years old and a
number (between 0 and 1) giving the part of people less than 18 years in the
population of the considered municipality). Formally, it implies one term per
class, but for an individual every but one of these terms are null: the term of
the class that the individual belongs to. So, for an individual j, the utility term
for such a variable concerning alternative i is actually the product of a class
specific parameter with the class (the individual belongs to) proportion in the
municipality.
62
For example, for the individual j and the alternative i, utility terms for the 7 age
classes are written as follows:
7
1_,___, ...Pr*)(*...
kialtkclasskclassageji opkIsInClassPU
where
- Page_class_k is the parameter for the age class k,
- IsInClass(k) is a binary variable indicating if the individual i is in class k,
- Propclass_k, alt_i is the proportion of the class k in the alternative (municipality) i
For all individuals in class k=r, this expression reduces in:
...Pr*... _,___, ialtrclassrclassageji opPU
The interpretation of the parameters is less evident but can be explained this
way : a positive parameter would indicate that people having the same profile
tend to settle down in a same place while a negative parameter would
represent a trend to a dispersion of the concerned people.
Link with former residence municipality
We mainly tested two ways to take into account the link with the former
residence municipality:
the intercommunal distances (distances between the communal
centroids),
a change of district via a binary variable (1 for municipalities in the
same districts as the former one).
Accessibility indicators
We tested different forms for the utility terms related to the accessibility
indicators.
a unique parameter,
distinct parameters
for each age class (7 classes),
for active and non active people (2 parameters per indicator).
The idea of distinct parameters was to try to detect different influences of
accessibility according to individual characteristics such as age class. For
example, we can suppose that the job opportunities indicator could be less
explanatory for elderly people.
Practical estimation
Even with sampling of alternatives, estimation remains time consuming. With
BIOGEME, only data reading was already memory space and time consuming.
63
Therefore, in order to spare time, we first worked with a sub-set of data and
alternatives to test several model structures. Then, we performed estimation on
bigger datasets and all the alternatives for the models that appeared the best ones
with the partial data.
The way we chose this subset of alternatives was not random. We wanted to have a
representative subset of Belgian municipalities, i.e. municipalities from different types
(rural, urban, etc.) and from different zones. We thus fixed 5 zones (Brussels Region,
two in Flanders, two in Wallonia) and the proportion of municipality types we wanted
for each zone (according to the real proportion).
As a summary, the Table X below summarizes the main process in the model
building. Table X: Practical estimation of the localization model
1.1. Sampling or not
1.2. Structure : logit or nested(different ways to build the nests)1.3. Variables : selection andforms (contextualisation...)
1. Tests with 60 municipalities
2. Estimation with 589 municipalities
Models and results
Goodness-of-fit measures
A discrete choice model simulation gives for each individual probabilities of choosing
each alternative. Comparing the observed choice with the predicted one is not
relevant for discrete choice model. Indeed, the model do not predict "the choice" of
an individual but gives probabilities for each alternative to be chosen by this
individual.
In order to measure how well a model fits the data, Train (2002) recommends the use
of the likelihood ratio index.
)0(
)ˆ(1
LL
LL
where LL(β) is the value of the log-likelihood function at the estimated parameters
and LL(0) is its value when all the parameters are set equal to zero (i.e. a purely
random model).
64
Its value is comprised between 0 and 1. A value of 0 means that the model is not
better than no model (more exactly than a purely random one) . A value of 1 means
that the model predicts perfectly each sampled decision maker's choice.
The index looks similar to R² used in regression but its interpretation is completely
different and values between 0 and 1 of have no intuitive interpretation. So, the
index can only be used to compare models built on the same data and with the same
set of alternatives. In this case, the higher is the index , the better is the model.
Another possibility to compare model is the hypothesis testing. It can be used to test
whether two parameters are significantly different from each other or if a unique
parameter is sufficient.
Model results
Here are exposed the main conclusions of the models testing and the results of the
finally retained model.
Sampling of alternatives
Models were tested without and with sampling of alternatives (10% of the
alternatives used for estimation). As expected, the confidence intervals are
larger for estimation with sampling which use less information to estimate the
parameters. But the parameters values were not quite different, i.e. the
confidence intervals (95%) "have a common part of range". Generally, the
values with sampling belong to the confidence interval of the parameter
without sampling.
Model structure
The retained model is a nested model with 4 nests based on the communal
typology (and not on the geographic proximity).
The four nests regroups the municipalities into:
agglomerations,
suburbs,
commuters municipalities,
rural municipalities.
according to the Van der Haegen typology (1996).
Let us note that the geographic proximity (other type of nests that was tested)
is more efficiently taken into account thanks to the distance variable (see
65
further). More sophisticated nests (crossing both geographic and typological
criteria) lead to model with many unsignificant parameters.
Covariates
The most significant variable is the distance from the former municipality
(distance between the two centroids). The negative sign of the distance
parameter expresses that the utility of a municipality decreases with the
distance from the former municipality (see Table XI). As explained in the
preliminary analysis, (intermunicipal) residential moves are mostly short
distance moves.
Table XI: Parameters of the retained model for the residential localization (calibration on all
municipalities without sampling) Parameter Value Std err t-test p-value
Accessibility indicator to employment for active people
-0.241 0.00573 -42.09 0.00
Accessibility indicator to employment for non-active people
-0.318 0.0134 -23.66 0.00
Distances between the new residential municpality and the former one
-0.0663 0.000385 -172.28 0.00
Living conditions indicator (environment aspects) 0.949 0.0483 19.66 0.00Living conditions indicator (dwelling aspects) -1.67 0.0699 -23.84 0.00Living conditions indicator (socio-economic aspects)
3.57 0.0639 55.93 0.00
Living conditions indicator (services aspects) -1.73 0.0757 -22.90 0.00Property prices indicator 0.407 0.0262 15.55 0.00Population of the new residential municipality 0.506 0.00534 94.85 0.00Age : 0 to 18 -4.27 1.77 -2.41 0.02Age : 19 to 29 13.1 0.289 45.20 0.00Age : 30 to 44 4.59 0.365 12.58 0.00Age : 45 to 54 0.480 0.887 0.54 0.59Age : 55 to 64 19.8 0.967 20.43 0.00Age : 65 to 74 10.8 1.07 10.09 0.00Age : 75 and more 1.30 1.36 0.96 0.34Education : other 2.22 0.168 13.23 0.00Education : secondary school (inferior) 5.45 0.292 18.63 0.00Education : secondary school (superior) 4.41 0.300 14.73 0.00Education : higher education 4.19 0.109 38.31 0.00Kind of household : other 0.234 1.09 0.21 0.83Kind of household : couple (married or not) with children
2.36 0.128 18.36 0.00
Kind of household : couple (married or not) without children
2.78 0.256 10.89 0.00
Kind of household : single-household 8.11 0.152 53.42 0.00Kind of household : one-parent family 4.03 0.248 16.27 0.00Nationality : other 5.94 0.278 21.38 0.00Nationality : belgian 2.71 0.0745 36.46 0.00Nationality : border countries 4.23 0.362 11.68 0.00Nationality : European Union 4.45 0.312 14.28 0.00
Regarding the accessibilities variables, only one of the two accessibility
variables was kept. The accessibility to services indicator and the accessibility
66
to employment indicator are indeed highly correlated (r=0.89). The decision to
keep the accessibility to employment indicator came from the consideration
that the accessibility to services indicator is more complicated to make evolve.
Moreover, attractiveness of services can be taken into account via one of the
four components of the living conditions indicator (services aspects).
In the utility formulation, the best results for the accessibility to employment
indicator are obtained with specific parameters for active and non active
individuals. People were considered as active between 19 and 64 year old,
and non active outside of these limits. Considering these two distinct
parameters lead to a better model than considering a unique parameter for
employment indicator. Introducing one parameter per age class (7 classes)
does not improve significantly the quality of the model.
Both parameters for accessibility to employment indicator are negative.
Regarding the other municipal characteristics:
the four components of the living condition indicator are significant : Note
that these indicators have a range between 0 and 1 and the 1 value is the
less interesting so that the parameters must be carefully interpreted;
the average property price at municipal level is also significant.
Regarding the individual characteristics, let us recall that these variables have
been defined as contextualisation variables, so that a positive sign means that
individuals tend to settle down in the municipalities where they find similar
individuals. In the retained model, most of these parameters are significant.
Amongst them, only the parameter for 0-18 age class is negative. All other
parameters are positive (for age classes, household types, nationalities and
levels of education). From this point of view, the model reflects well the
behaviour of tending to live with one's own socio-economic group.
67
3. POLICY SUPPORT The project's objective is to investigate the links between long-term residential
choice, accessibility and daily mobility, with the ambition to provide better
understanding of the behaviour of Belgian households regarding these issues. In
particular, the respective importance of several classes of factors is crucial for the
establishment of land-use and accessibility policies. The mixing of long-term
decisions such as housing and short-term ones such as daily mobility turns out to be
a challenging issue.
The project developed three main models: an accessibility model, a model of the
propensity of an individual to move his/her residence and a model of localization
choice for a new residence.
The first of these models (accessibility) involves a specific assignment procedure to
compute travel times matrices from OD matrices at the municipal level; the flows are
assigned on a simplified road network that was built to this purpose. The model
outputs both travel time in free-flow for non-peak hour and travel times for peak
hours, taking congestion into account. Finally, a few relevant accessibility indicators
were also produced; they allow the characterization and comparison of municipalities
and we feel that this has direct consequences on municipality management.
The model describing the propensity of the Belgian individuals to move their
residence incorporates a number of explanatory factors at the individual and
household levels. Not unexpectedly, the parameters which turned out to dominate
the individual's choice of changing residence are changes in the household structure,
change in the position in the household and age class.
This reinforced the idea that societal trends (as opposed to material infrastructure
evolution) is crucial to explain internal migration within a country. In particular,
population aging and the growth in importance of smaller household structures may
present specific challenges in urban planning and land-use in general during the
forthcoming years.
Finally, the residential localization model is central to the design of suitable land-use
regulations at the regional level. Remarkably, our analysis indicates that the
dominant factors are, by decreasing level of importance, the distance between the
previous residence and the new one, the perceived quality of life in the new
municipality of residence, the household structure, and, in fourth position, the
accessibility of the new municipality of residence. Accessibility is therefore less
important than expected at the start of the project.
68
A simple interpretation of the results obtained in the course of the MOBLOC project is
that internal migration within the country is less determined by infrastructural factors
(within which accessibility is a important example) than factors related to societal life
in a more general sense: household structure and its evolution, closeness to one's
relations, age class and quality of life score indeed higher than purely transport
related factors in our results.
We believe that these conclusions are important for any prospective view on the
development of land-use and transportation systems. They will be discussed in the
new regional prospective study group (SRP) which is being established under the
leadership of the Institut Destrée and the Institut Wallon d'Evaluation, de Prospective
et de Statistiques (IWEPS). We also intend to dissiminate those conclusions more
widely, via scientific publications but also aiming the municipality managers and the
general public.
69
4. DISSEMINATION AND VALORISATION In this section, we list different research projects which rely on the outcomes of the
MOBLOC project our use its main results, but we can also make reference to
different informal dissemination in the frame of contacts abroad (e.g. meeting
focusing on the residential localizations at the LET, Lyon, in January 2009) or coming
seminar in the frame of the NAXYS seminars (Namur Center for Complex Systems).
Three first projects are presented here after: as explained, they can be seen as
prolongations of the MOBLOC Project.
A. MOEBIUS - Mobilities, Environment, Behaviours Integrated in Urban Simulation
CEPS/INSTEAD coordinates a research project (MOEBIUS) funded by the FNR
(Fonds National de la Recherche Luxembourg) that deals with some of the issues
addressed in the MOBLOC project. The outcomes of our project (and the expertise
derived of our research) will be helpful for the several partners involved in this
MOEBIUS project to reach the objectives planed. A brief description of this new
project follows here below.
Luxembourg emerges as a very attractive cross-border regional metropolis leading to
increasing residential migration (topic modelled in the MOBLOC project) and longer
commutes. Empirical evidence shows that current urbanisation trends toward
suburban and more remote periurban areas favour urban sprawl and car
dependence. In Luxembourg, the awareness of the role of spatial planning for the
implementation of sustainable development has led to the developing of a planning
concept, called IVL, involving several Ministries. It promotes an integrated spatial
development which joins number of assumptions of the New Urbanism Theory. The
main objective of the MOEBIUS project is then to assess the sustainability of this
planning scenario (IVL / “New Urbanism”) comparing to the current urban sprawl.
Further understanding the social and environmental impacts of both the residential
mobility and the daily mobility of households is at the core of our research project.
We will deal with the complexity of this issue by using spatial simulation tools that are
hybrids from agent-based and cellular-automata. This simulation platform prototype
will integrate land use, residential mobility (well studied in the frame of the MOBLOC
project) and daily mobility, within the cross-border metropolitan area of Luxembourg.
Because of its complexity, this platform will be devoted to a particular population (i.e.
workers employed in the Grand-Duchy and living inside or outside the country) and
involve expertise from different disciplines (urban planning, geography, psychology,
economics, mathematics, and computer sciences). The development of a small,
simple and transparent model is planed: this model will include new theoretical
70
components, by using belief theory, residential choices and determinants of daily
commuting specific to Luxembourg.
The aim is to simulate (i) the future urbanisation (based on demographic forecast and
residential choice modelling) in the commuting area of Luxembourg, and (ii) the
future daily mobility (commuting pattern and travel mode choice) for different
planning scenarios. When assessing those scenarios, particular attention will be put
on how they can provide a good trade-off between, economic growth (via the
provision of attractive and affordable living places) and environmental and social
sustainability (modal split, land take, accessibility…).
Because the simulation model is spatially explicit, the specificities of Luxembourg can
be taken into account, e.g. the urban network and the cross-border setting. In
addition, unique datasets are accessible in the frame of this project (e.g. the cross-
borders transport & mobility survey (2010) or Social Security longitudinal data) and
building on the empirical knowledge based on the MOBILLUX outcomes is possible
(FNR project on the link between daily and residential mobilities in Luxembourg).
B. SimBelgium - NAXYS The substantial amount of work invested during the MOBLOC project into the design
and calibration of models for internal migration of individuals and households within
Belgium has resulted in the creation of a new research project in the framework of
NAXYS, the Namur Center for Complex Systems. After a short description of the
NAXYS interests, we briefly describe this projet below.
NAXYS is a new research center at FUNDP (established in the fall 2010) whose
purpose is the analysis, modelling and general scientific approach of complex
systems in society and nature. It has research interests in several complementary
directions, including natural chaotic systems (in celestial mechanics and elsewhere),
large networks, opinion propagation, weather forecasting and, more importantly in
this context, spatio-temporal dynamics of large populations. It is truly multi-
disciplinary group, combining mathematicians, economists, geographers, computer
scientists, engineers, etc. Its initial activity (which can be found on
http://www.fundp.ac.be/en/sci/naxys) reflects its scientific dynamism.
The NAXYS project which is extending the work of MOBLOC is (temporarily) named
"SimBelgium" and aims at producing a detailed microsimulation of the behaviour of
the Belgian population with a special focus on the interaction of internal migration
(the MOBLOC input is crucial here), population dynamics and evolution of the
participation to the job market. This microsimulation is based on the availability of a
71
virtual population for the complete Belgian territory, whose (synthetic) individuals
belong to (synthetic) households, themselves localized in the 589 Belgian
municipalities. The technique of using virtual individuals is necessary in order to
avoid privacy issues in the research program. Daily activity chains (obtained from the
MOBEL and, when available, BELDAM projects) are associated with each individual.
This in turns is intended to allow a dynamic simulation of the aggregate effect of the
combined activity demand.
The research group associated with this project includes participants of the MOBLOC
projects, but extends well beyond this group. In particular, direct collaborations have
been established with the FUNDP Department of Geography in order to secure
additional expertise in the spatial analysis of internal migration and of the housing
market and their dynamics.
The project has already started and has, in the past three months, established a first
version of the population dynamics over a period of 30 years. The association of
activity patterns with individuals and households is currently proceeding. The
incorporation of the results obtained by the MOBLOC project is expected to follow
shortly. Of particular interest in this context are the MOBLOC models for the
propensity of individuals to migrate within the country and the localization model at
the municipality level. Because a large part of the methodology used in the new
project is based on modelling the many individuals' choices by using tools derived
from the random utility theory, and because this approach has also been central in
the MOBLOC project, the integration of the results of the latter in the former is
expected to proceed reasonably smoothly.
It is very clear that this new ambitious project would not have been possible without
the significant contribution of the MOBLOC research.
This project is also at the basis of several discussions that have taken place in the
context of the SRP (Service Régional de Prospective), a working group established
in the Région Wallonne under the leadership on the Institut Destrée and Institut
Wallon d'Evaluation, de Prospective et de Statistiques (IWEPS).
C. Population and households forecasting of Belgian municipalities
The aim of this project is to provide population and households forecasting across
municipalities of Belgium. It will be multi-state projections.
A special method taking into account the problem of "small numbers" will be
developed.
72
This project will provide for each municipality, in 2020, the number of population,
population distribution by age and sex and household characteristics.
Several scenarios will be considered and the assumptions of internal migration
trends will be based on the results of the propensity to move model developed in the
project MOBLOC
Finally, we must also mention that links were established with the team in charge of
the SIMBAD projet at LET (Laboratoire d’Économie des Transports), Lyon (F). Their
goal is to apply a LUTI (land Use and Transport Integrated) model to the Lyon
agglomeration. These researchers are quite interested in the methodologies
developed and used for this MOBLOC project. Therefore they have invited Eric
CORNELIS to become member of their accompying committee.
73
5. PUBLICATIONS
KLEIN S., OMRANI H., CARPENTIER S., GERBER P., VANDENBULCKE G.,
Validation d’un modèle d’accessibilité par recoupement de données multi-sources.
Application aux communes de Belgique, 2ème Journée de Recherche « Mobilité,
Transport et Logistique » (MTL 2010), Lyon, Mercredi 23 juin 2010 (copy in annex)
PAULY X., CORNELIS E., WALLE F., MOBLOC : links between residential mobility
and accessibility, in “Proceedings of BIVEC Transport Research Day 2011”, E.
CORNELIS ed., pp. 126-137, University Press, 2011
BAHRI A., DAL L., CORNELIS E., EGGERICKX T., PAULY X., WALLE F.,
Propensity to move and residential localization choice in Belgium, modelling of the
individual behaviours, Namur Center for Complex Systems (NAXYS) Report, 2011 (in
preparation)
DAL L., CARPENTIER S., CORNELIS E., EGGERICKX T., KLEIN S., GERBER P.,
PAULY X., WALLE F., TOINT P., Interactions between accessibility and residential
choices in Belgium: modelling at a municipal level, Namur Center for Complex
Systems (NAXYS) Report, 2011 (in preparation)
74
6. ACKNOWLEDGMENTS
In order to perform our calibrations for the accessibility model, the national network
was necessary. This network was provided by the National Ministry of Mobility and
Transports (SPF MT – Service Public Fédéral Mobilité et Transports).
Concerning the residential choices, numerous variables were necessary for the
calibration of models. Those were available thanks to by the National Register and
Statbel (Service Public Fédéral Economie, P.M.E., Classes moyennes et Energie,
Direction générale Statistique et Information économique).
75
7. REFERENCES
ALLISON P., 1999, Logistic Regression Using SAS: Theory and Application, SAS
Press
ANDAN O., POCHET P., ROUTHIER J.-L., SCHEOU B., 1999, Stratégies
résidentielles des ménages et mobilité domicile-travail, Programme de recherche
Transports Terrestres 1996-2000, Direction de la Recherche et des Affaires
Scientifiques et Techniques
BARTON-ASCHMAN ASSOCIATES, Inc.; CAMBRIDGE SYSTEMATICS, Inc., 1997,
Model Validation and Reasonableness Checking Manual, 118 p.
BEN AKIVA M., LERMAN S., 1985, Discrete choice analysis, MIT Press
BERTRAND M., DUPONT V., GUÉRIN-PACE F., 2003, Espaces de vie. Une revue
des concepts et applications, Paris, INED, Documents .de travail, Paris
BIERLAIRE M., 2003, BIOGEME: A free package for the estimation of discrete
choice models , Proceedings of the 3rd Swiss Transportation Research Conference,
Ascona, Switzerland,
BLACK W. R., 1998. Sustainability of Transport. In: Hoyle, B. S., Knowles R., (Eds),
Modern Transport Geography, 2nd Edition. Chichester: Wiley, 337-351.
BLIJIE B., 2005, The impact of accessibility on residential choice: empirical results of
a discrete choice model, Delft University of Technology, 45TH Congress of the
European Regional Science Association, Vrije Universiteit Amsterdam
BRANSTON, D., 1976, Link capacity functions : a review. Transportation Research
10, 223-236
BRÜCK L., MAIRY N., HALLEUX J.-M., MERENNE-SCHOUMAKER B.,
SAVENBERG S., VAN HECKE E., 2001, Les comportements résidentiels des
ménages face à la problématique du développement durable
BUREAU OF PUBLIC ROADS,1964, Traffic Assignment Manual. Urban Planning
Division, US Department of Commerce, Washington D.C.
76
CORNÉLIS E., BAHRI A., EGGERICKX T., CARPENTIER S., KLEIN S., GERBER
P., PAULY X., WALLE F., TOINT P., 2009. Mobilities and long term location choices
in Belgium "MOBLOC", Final report for the Belgian Science Policy, Phase 1. 47 p.
COURGEAU D., BACCAINI B., 1997,Analyse multi-niveaux en sciences sociales,
Population (French Edition), 52e Année, No. 4, Nouvelles approches
méthodologiques en sciences sociales, pp. 831-863.
DEBRAND T., TAFFIN C., 2005, Les facteurs structurels et conjoncturels de la
mobilité résidentielle depuis 20 ans, Economie et statistique n° 381-382, pp. 125-146
DELAUNAY D., 2001, L’inscription dans l’espace des biographies individuelles.
Communication présentée à la XXIVe Conférence Générale de l’IUSSP, Salvador
(Brésil). (téléchargée sur le site de l’IUSSP).
DES ROSIERS F., THÉRIAULT M., KESTENS Y. and VILLENEUVE P., 2002,
Landscaping attributes and house prices: looking at interactions between property
features and homeowners’ profiles, Paper presented at the 18th Annual ARES
Conference, Naples, Florida.
DES ROSIERS F., THÉRIAULT M., VILLENEUVE P-Y. and KESTENS Y., 2001,
Landscaping and house values : an empirical investigation, Laval University, Paper
submitted for presentation at the 17th Annual ARES Conference, Coeur D’Alene,
Idaho
DIJKSTRA E. W., 1959, A note on two problems in connexion with graphs.
Numerische Mathematik, 1, p. 269–271
DUJARDIN C., 2001. Effet de frontière et interaction spatiale. Les migrations
alternantes et la frontière linguistique en Belgique. L’espace géographique 01(4):
307-320.
EGGERICKX T., DEBUISSON M., DAL L., 2008, Les tendances récentes des
migrations entre les arrondissements, In: Perspectives de Populations 2007-2060,
Bureau Fédéral du Plan / DGSIE ed(s), Planning Paper, 105, p. 35-38
EGGERICKX T., SANDERSON J.-P., BAHRI A., 2010, Transformation des ménages
et ségrégation sociale de l’espace. Une analyse de la précarisation des parcours
résidentiels en Belgique (1991-2005), Dynamiques de pauvretés et vulnérabilités.
Mesures et processus explicatifs en démographie et en sciences sociales, Bruno
77
Masquelier, Thierry Eggerickx, ed(s), Chaire Quetelet 2007, Presses Universitaires
de Louvain, Louvain-la-Neuve, pp. 187-222
EGGERICKX T., HERMIA J.-P., SURKIJN J., WILLAERT D., 2011 (à paraître) (avec
la collaboration de DAL L., POULAIN M. et SANDERSON J.-P.), Les migrations
internes en Belgique, Monographie 2 de l’Enquête socioéconomique de 2001,
DGSIE, Bruxelles, 200 p.
ENAUX C., 1997, Essai de modélisation spatio-temporelle des flux de déplacements
de travail. Exemples de la région urbaine strasbourgeoise de 1975 à 1990, Thèse de
doctorat, Université Louis Pasteur, Strasbourg, 265 p. (+ 1 volume d'annexes 107 p.
+ 3 disquettes)
FLOWERDEW R., AITKIN M., 1982. A method of fitting the gravity model based on
the Poisson distribution, Journal of Regional Science, 22(2): 191-202.
FLOWERDEW R., 1991. Poisson regression modelling of migration. In Congdon P.,
Stillwell J., Migrations models : macro and micro approaches, London: Behaven
Press, 92-112.
FOOT D., 1981, Operational Urban Models : an Introduction. London: Methuen & Co
Publishing Company, 230 p.
GOODWIN, P., 1995, Car dependence. Transport Policy, 2(3), p. 151–152.
HALLEUX J.-M., 2001, Évolution des organisations urbaines et mobilités
quotidiennes : espace de référence et analyse des processus, L’Espace
géographique, Vol. 30, n°1, pp. 67-80
HENLEY A., 1998, Residential mobility, housing equity and the labour market, The
Economic Journal, n° 108, pp. 27-414
HOSMER D. W., LEMESHOW S., 2000, Apllied Logistic Regression, 2nd Edition,
HUBERT J.-P., TOINT P., 2002. La mobilité quotidienne des Belges, Presses
Universitaires de Namur, 352p.
IRWIN M., BLANCHARD T., TOLBERT C., NUCCI A., LYSON T., 2004, Pourquoi
certains ne migrent pas : L'impact du contexte local sur la sédentarité aux Etats-Unis,
Population (French Edition), 59e Année, No. 5, pp. 653-680.
78
JOLY I. ,2005, Décomposition de l’hypothèse de constance des budgets temps de
transport, in Mobilités et temporalités Montulet B et al.(dir), Facultés Universitaires
Saint Louis, Bruxelles, 129-150
KNUDSEN D., FOTHERINGHAM S., 1986. Matrix Comparison, Goodness-of-Fit, and
Spatial Interaction Modeling, International Regional Science Review; 10(2): 127-147,
doi:10.1177/016001768601000203.
MURAT C.H., 2010. Sample size needed for calibrating trip distribution and behavior
of the gravity model. Journal of Transport Geography, 18(1): 183-190
doi:10.1016/j.jtrangeo.2009.05.013.
NELDER A., WEDDERBURN R., 1972. Generalized linear models, Journal of the
Royal Statistical Society. Series A (General), 135(3): 370-384.
NIJKAMP P., RIETVELD P., SALOMON I., 1990. Barriers in spatial interactions and
communications. A conceptual exploration. The Annals of Regional Science 24, 237-
252, doi:10.1007/BF01580472.
PUSHKAR A. O., Modelling household residential search processes: methodology
and preliminary results of an original survey, University of Toronto
REYMOND H., CAUVIN C., KLEINSCHMAGER R., (coord.), 1998, L’espace
géographique des villes - Pour une synergie multistrates, Anthropos, collection Villes,
Paris, 557 p.
ROMANI J., SURIN’ACH J., ARTI M., 2003, Are commuting and residential mobility
decisions simultaneous?: The case of Catalonia, Spain, Regional Studies, Vol. 37.8,
pp. 813–826
SHANNON, C. E., 1948, The mathematical theory of communication. Bell System
Technical Journal 27: 379-423 ; 623-56
TAYLOR P., 1983, Distance decay in spatial interaction, CATMOG - Concepts and
Techniques in Modern Geography, n°2, 35 p.
TRAIN K. E., 2003, Discrete Choice Methods with Simulation, Cambridge University
Press
79
VAN DER HAEGEN H., VAN HECKE E., JUCHTMANS G., 1996, Les régions
urbaines belges en 1991, Études Statistiques, n° 104
VAN HECKE E., 1998, Actualisation de la hiérarchie urbaine en Belgique, Bulletin du
Crédit Communal, vol. 205, n° 3, p. 45-76.
VANDENBULCKE G., STEENBERGHEN T., THOMAS I., 2009, Mapping
accessibility in Belgium: a tool for land-use and transport planning? Journal of
Transport Geography, p. 39-53
VANNESTE D.., THOMAS I., GOOSSENS L., 2007, Woning en woonomgeving in
België, Sociaal-economische enquête 2001 , SPF Politique scientifique & SPF
Economie – Direction générale Statistique
WARDROP J.G., 1952, Some theoretical aspects of road traffic research,
Proceedings of the Institute of Civil Engineers. Pt. II, vol. 1, p. 325-378
WILLEKENS F., DE BEER J., VAN DER GAAG N., 2005, MicMac - From
demographic to biographic forecasting, paper prepared for presentation at the Joint
Eurostat-ECE Work Session on Demographic Projections, September 21-23, Vienna.
WILSON A., 1971. A family of spatial interaction models, and associated
developments, Environment and Planning A, 1971(3): 1-32.
ZAHAVI Y., 1979, The ‘UMOT’ Project. Report prepared for the U.S. Department of
Transportation and the Ministry of Transport of Federal Republic Of Germany, 267 p.
ZAHAVI Y., TALVITIE A.,1980, Regularities in Travel Time and Money Expenditure,
in : Transportation Research Record, 750, 13-19.
80
ANNEX 1: COPY OF THE PUBLICATIONS
2ème Journée de Recherche « Mobilité, Transport et Logistique » (MTL 2010)
Lyon, Mercredi 23 juin 2010
Validation d’un modèle d’accessibilité par recoupement de données multi-sources. Application aux communes de Belgique
S. Klein1*, H. Omrani1*, S. Carpentier1*, P. Gerber1*, G. Vandenbulcke²
* Consortium de recherche du projet MOBLOC (coord. GRT) cofinancé par la Politique Scientifique Belge (Belgique) et le FNR (Luxembourg)
1 Centre d’Études de Populations, de Pauvreté et de Politiques Socio-Économiques (CEPS/INSTEAD) BP 48, Differdange, Luxembourg .Tél : +352 58 58 55 310
² CORE et Département de Géographie, Université Catholique de Louvain (UCL) Voie du Roman Pays, 34, Louvain-la-Neuve B-1348 (Belgique)
Résumé Dans le cadre du projet MOBLOC (Mobilities and Long Term Location Choice in Belgium), l'exploration à un niveau d'agrégation communal des interactions entre les mobilités résidentielle et quotidienne a nécessité la construction d'un modèle d'accessibilité routière à l'échelle de la Belgique. Dans un premier temps, un modèle de trafic en heures creuses affectant les flux selon un mode tout-ou-rien, est comparé à une base d'observations (MOBEL) ainsi qu'à deux modélisations (Google Maps et un modèle développé à l'UCL par Vandenbulke et al., 2009). Le modèle routier en heures de pointe du matin procède à l'affectation d'une matrice de demande de déplacement domicile-travail et domicile-études sur la base d'une recherche d'un équilibre utilisateur. En complément de l'analyse cartographique, les statistiques AED, chi-carré, SRMSE, psi-absolu et phi sont mises à contribution pour le calibrage et la validation du modèle d'accessibilité MOBLOC.
1. Contexte et étapes de la recherche
1.1 Le projet MOBLOC
Depuis plusieurs décennies, la croissance urbaine, et notamment la périurbanisation, a favorisé une utilisation massive de l’automobile (Wiel, 1999) conduisant, dans les zones périphériques notamment, à une véritable dépendance automobile (Goodwin, 1995). Ainsi, à l’instar de nombreux pays européens, la Belgique connaît une forte périurbanisation (Brück et al., 2001 ; Halleux et al., 2002) associée à un usage intensif de la voiture privée (Hubert et Toint, 2002).
81
Dans ce contexte, le projet MOBLOC (Mobilities and Long Term Location Choice in Belgium) vise à modéliser et simuler les interactions entre les évolutions à long terme de la société (croissance démographique, choix de localisation des ménages) et les comportements de mobilité quotidienne (Cornélis et al., 2009). Pour ce faire, plusieurs modèles sont développés pour analyser l’impact de ces évolutions sur l’accessibilité et, en retour, sur la croissance urbaine (Figure 17). Ainsi, un modèle de migration résidentielle – articulant propension à migrer et choix de localisation – est couplé à un modèle de trafic – lui-même décomposé en une estimation des flux (modèle gravitaire), un choix modal et une affectation des flux sur le réseau (modèle d’accessibilité en heures creuses et en heures pleines du matin). La construction des modèles, ainsi que leur couplage, est réalisé à l’échelle des 589 communes belges.
Figure 17 : Présentation des modèles en interaction du projet MOBLOC
82
La question à laquelle nous nous intéressons dans cet article est celle de la mise en place d’un modèle d’accessibilité routière (étape 4) à l’échelle de la Belgique et de sa validation. Sachant qu’à l’échelle de toute la Belgique, nous ne disposons pas de l’ensemble des paramètres nécessaires au calibrage et à la validation du modèle d'accessibilité en heures creuses et en heures pleines, nous appliquons une méthode de validation basée sur différentes statistiques permettant de comparer celui-ci avec des observations d’une part, ainsi qu’avec d’autres modélisations d’autre part.
1.2 Étapes méthodologiques du projet
Plusieurs étapes ont été mises en œuvre (Figure 18). La modélisation étant effectuée pour l’ensemble du territoire belge, une première étape consiste à choisir, pour chacune des 589 communes, un point « représentatif ». Si le choix d’un centroïde communal unique constitue sans doute une approximation, cette étape s'est révélée nécessaire, à la fois pour des raisons de cohérence méthodologique (le couplage des modèles de MOBLOC s’effectue à l’échelle communale) et des raisons pratiques, liées au niveau d'aggrégation pour lequel les données sont disponibles. En second lieu, il s’agit de mettre en place la représentation mathématique du réseau routier sous forme d’un graphe orienté G = (V, E), avec V : l’ensemble des sommets représentant les intersections, et E : l’ensemble des arcs représentant un tronçon de route aux propriétés homogènes. Il s’agit ensuite de qualifier ces arcs, notamment par leur longueur, leur vitesse en flux libre et leur capacité. Ces attributs n’étant, la plupart du temps, pas disponibles à l’échelle du pays, nous recourrons à une typologie établie grâce au recoupement des couches SIG (Système d'Information Géographique) de l'occupation du sol avec le réseau digitalisé.
Figure 18 : Étapes méthodologiques de modélisation de l'accessibilité VP
Dans un deuxième temps, deux modélisations de l’accessibilité routière sont mises en œuvre, l'une concernant une période en heures creuses (HC), l'autre la période de pointe du matin (HP). Le modèle en heure de pointe fait intervenir une matrice de demande de déplacements,
83
à savoir une matrice d’échanges origine/destination (O/D) des actifs et des étudiants utilisant la voiture (comme conducteur principal) pour leur trajet domicile/travail ou domicile/école. Celle-ci permet d'inclure des résultats des nouvelles localisations du modèle de migration résidentielle de MOBLOC. Notons que, si dans le cadre du couplage des modèles la demande HP et HC est simulée, pour la première itération à t0, cette matrice est issue de l’Enquête Socio-économique de L’INS (2001). Les résultats des deux modèles sont ensuite confrontés avec une base d’observations provenant de l’enquête MOBEL et comparés avec d’autres modèles disponibles pour la Belgique. Grâce aux statistiques mesurant l'ajustement de ce modèle par rapport aux valeurs observées, il est alors possible d'en calibrer les paramètres et de le situer par rapport aux autres modèles. Une fois les deux modèles (heures creuses et heures de pointes) validés, il est dès lors possible de procéder au calcul des indicateurs communaux à introduire dans le modèle de migration résidentielle pour modéliser ensuite l’impact de la mobilité quotidienne sur les choix résidentiels. Le postulat est ici que lors d’un choix résidentiel, les actifs tiennent compte de leur trajet domicile/travail pour garantir la faisabilité de leur programme d’activité (en contenant leur budget-temps de déplacement dans des limites acceptables, soit une heure en moyenne selon la conjecture Zahavi, 1979)8. La principale contrainte qui influence ce choix est alors le temps d'accès aux heures de pointe, un jour de semaine.
2. Mise en place du réseau routier
2.1 Choix des points représentatifs des communes
Afin de garantir la plus grande cohérence au niveau du couplage des modèles, ces derniers doivent avoir la même précision spatiale, à savoir ici les 589 communes de Belgique. À cette échelle, et compte tenu de la densité du réseau de transport belge, des simplifications sont opérées pour des raisons liées à la capacité de calcul et à l'indisponibilité des données. Concrètement, cette simplification consiste à choisir, pour chaque commune, un point considéré comme représentatif de la centralité du réseau ou du bâti. Cela revient donc à assimiler l'accessibilité de ce point à celle de toute la commune ; sur de courtes distances, cette simplification est certes contraignante, en revanche, à l’échelle d’un pays, ce postulat est acceptable. Les communes étant issues, pour la plupart, de la fusion de plusieurs communes (en 1977), le choix d’un point unique représentant la commune n’est pas chose aisée. Pour autant, dans un souci d’objectivation, des règles systématiques ont été établies en fonction du type de commune. Pour ce faire, cette étape repose sur une typologie communale des « régions urbaines » (Van der Haegen, 1996). Cette dernière distingue quatre types de communes : des agglomérations, des banlieues, des zones résidentielles de migrants alternants et d’autres communes non polarisées. Pour chaque centre d’agglomération (soit 17 communes), nous retenons l'hypercentre (par exemple, pour Bruxelles : la Grand Place). Dans le cas des autres communes, et en absence de couche SIG plus précise, nous nous sommes appuyés sur un géocodage du nom de la commune réalisé grâce aux API9 de Google Maps. Après vérification, nous avons constaté que celui-ci se faisait la plupart du temps sur l’agglomération principale de chaque commune. Dans le cas contraire, lorsque le géocodage
8 Notons cependant que cette conjecture postulant un budget-temps constant fait l'objet de critiques (Joly I., 2003). 9 Application Programming Interfaces : bibliothèque de fonctions utilisables dans un programme, dans notre cas, servant à faire des requêtes spécifiques vers les serveurs de Google Maps.
84
se trouvait en-dehors d’une zone urbanisée, nous avons procédé à un repositionnement du centroïde en fonction du réseau et de l’agglomération la plus proche.
2.2 Représentation du réseau routier
Outre le choix des points communaux, une autre simplification est nécessaire, celle du réseau routier. Le réseau routier utilisé, issu du Service public fédéral Mobilité et Transports, contient une hiérarchie de voies allant des autoroutes aux routes nationales, ce qui est généralement suffisant dans le cas d’une modélisation intercommunale (Figure 3). Par ailleurs, les points représentatifs des communes doivent être raccordés par des connecteurs. Lorsque ces points sont éloignés de moins de 200 m, ils sont reliés au réseau routier par un connecteur. Dans le cas contraire, il se peut que le réseau soit insuffisamment détaillé, ce qui implique de digitaliser une route du réseau secondaire sur base des cartes routières disponibles sur internet ainsi que des zones urbanisées ou commerciales définies à partir d’une couche SIG (couche CORINE land cover 2001)10. L’étape suivante consiste à fixer la vitesse en flux libre, qui est un des paramètres du modèle d'affection du trafic. Ces vitesses ont été fixées en fonction d’une typologie des arcs tenant compte du nombre de voies de circulation, de la séparation ou non des sens de circulation par une bande centrale, ainsi que par le type d’urbanisation traversée.
10 Le paramétrage du réseau implique une vérification SIG des nœuds (intersection de deux arcs) qui doivent correspondre à une possibilité de changement de direction sur le terrain. Du fait de leur impact sur l’accessibilité, une attention particulière a été portée à la vérification des intersections des autoroutes avec le réseau de nationales. Ces vérifications ont été effectuées grâce à des orthophotos.
85
Le dernier paramétrage important concerne la capacité des routes. D’une manière conventionnelle selon le Ministère de l'Équipement et des Transports, la capacité d’une autoroute est de 2000 unités de voiture particulière (UVP) par heure et par bande de circulation. Celle-ci peut cependant varier en fonction de la géométrie et de l'ampleur du trafic. En ce qui concerne les routes en milieu urbain, la capacité conventionnelle est de 1200 UVP par heure et par bande de circulation. Hors agglomération, cette valeur varie entre 1400 et 2000 UVP par heure et par bande. Sur la carte (figure 3), on note la position privilégiée de Bruxelles au centre d'un réseau autoroutier reliant les grandes villes du pays, Anvers au Nord, Gand, puis Bruges à l'est. Les principales villes de Wallonie sont également bien reliées avec, d'ouest en est, Liège, Namur, Charleroi et Mons. Le sud du pays apparaît comme moins bien desservie en infrastructures.
3. Principe de l'affectation de trafic et de la comparaison des modèles
3. 1 Principe d'un modèle d'affectation de trafic
On aborde la modélisation en deux étapes selon que l'on se situe en heures creuses ou en heures de pointe. En effet, en heures creuses, on peut faire l’hypothèse que les trajets s’effectuent sur les tronçons à la vitesse en flux libre. Cela revient donc à appliquer un
Figure 19: Réseau routier de Belgique
86
algorithme de minimisation de coût, tel l’algorithme de Dijkstra (1959), pour déterminer le chemin présentant le coût minimal. La fonction de coût ici retenue est celle d’une minimisation du temps de parcours. En heures de pointe, nous utilisons un modèle permettant d’affecter les flux sur les chemins de coût minimaux, en tenant compte de la relation qui existe entre le temps de parcours sur une route et le débit de véhicules. Ces chemins minimaux étant préalablement déterminés dans des conditions en flux libres, cela signifie qu'il est possible de partir du modèle calibré en heures creuses pour paramétrer le modèle en heures de pointe. Pour chaque arc du modèle routier doit être définie une fonction débit-vitesse, spécifiant la dégradation de la vitesse maximale qu’il est possible d’atteindre à mesure que le nombre de véhicules approche la capacité maximale de la route. Un grand nombre de fonctions ont été développées et discutées (Branston, 1976), parmi lesquelles les fonctions de type BPR (1) développée par le American Bureau of Public Roads :
T T 0 1 α V Q β (1)
avec T : le temps de trajet, V : le flux, Q : la capacité de l’arc routier, T0 : le temps de trajet en flux libre et α and β : les paramètres de la fonction fixés selon la typologie des routes. Elle évolue de la façon suivante (Figure 20) :
Figure 20 : Représentation graphique d'une fonction de type BPR (Bureau of Public Roads, 1964)
Le modèle retenu pour l'affectation en heures de pointe repose sur la première hypothèse de Wardrop (Wardrop, 1952) qui spécifie les conditions permettant d’atteindre un équilibre utilisateur. Un tel équilibre stipule qu'aucun usager ne peut seul améliorer son temps de parcours en modifiant son itinéraire. On fait une approximation de cet équilibre par la méthode des moyennes successives (Method of Successive Averages). A l'issue de cette procédure d'équilibrage, une nouvelle vitesse est affectée à chaque tronçon routier. On calcule alors la matrice des temps de trajet entre chaque commune grâce à un algorithme de plus court chemin (Dijkstra, 1959).
3.2 Principe de la validation par comparaison du modèle
3.2.1 Statistiques d'ajustement
Dès lors que le modèle d’accessibilité est paramétré, se pose la question de la validation de ce dernier. Comme nous l’avons souligné, en l’absence de base de référence suffisamment exhaustive et précise, nous avons opté pour une comparaison de nos modèles avec une base de temps déclarés (MOBEL, 1999) ainsi qu’avec d’autres modèles basés sur des méthodologies différentes afin de les valider. Cette procédure permet à la fois la validation, la vérification et l’ajustement de nos modèles à différentes étapes de son élaboration. Pour tester l'ajustement des modèles et leurs estimations, il existe plusieurs indices statistiques permettant la comparaison de matrices observées ou estimées (Knudsen et Fotheringham, 1986). Les indices suivants ont en particulier été relevés pour leur sensibilité linéaire au niveau d’erreur pour une matrice complète de couple.
87
SRMSE : Standardized Root Mean Square Error
SRMSE
i 0
n
t i t i m n2
i 0
n
t i n (2)
avec ti et �t i , les temps de parcours observés et estimés. La limite inférieure de cette statistique est zéro, indiquant une prédiction parfaite et sa limite supérieure est généralement 1, bien que des valeurs supérieures à un, puissent survenir quand l'erreur moyenne est supérieure à la moyenne.
la statistique phi
i
pi ln pi qi (3)
avec pi et qi les probabilités des flux observés et estimés. Cette statistique a pour limites zéro et l'infini.
la valeur absolue de la statistique psi
i
pi ln pi sii
qi ln qi si (4)
avec pi et qi les probabilités des flux observés et estimés, et si = ( pi + qi ) / 2. Cette statistique n'a pas de distribution théorique connue. Nous avons testé deux autres statistiques :
AED : Absolute Entropy Difference, définie comme la valeur absolue de la différence des entropies entre les valeurs de probabilités observées et prédites.
AED H p H q (5)
avec Hp et Hq les entropies de Shannon (Shannon, 1948) telles que H p
i
pi ln pi.
le chi carré : χ2
χ 2
i
pi qi2
qi (6)
3.2.2 Analyse de sensibilité
Nous avons testé la sensibilité de ces indicateurs sur l'échantillon de MOBEL en heures creuses en introduisant un terme d'erreur (7) :
qi pi pi rnd fact (7) avec qi : les valeurs estimées du temps de parcours, pi : les valeurs observées du temps de parcours, δ : nombres aléatoires {-1,1}, rnd : nombre aléatoire [0,1] et fact : pourcentage d'erreur. Pour chaque niveau d'erreur, les valeurs des indices statistiques AED, χ2, SRMSE, φ et ψ, ont été calculés.
88
Cette procédure a été répétée plusieurs fois (n=100) et les moyennes des indices ont été calculées. La sensibilité de chaque indice selon la variation des niveaux d'erreur est reportée en figure 5, avec en abscisse les niveaux d’erreur et en ordonnée la valeur standardisée de l'indice statistique moyen. Les statistiques SRMSE, φ et ψ présentent une relation linéaire au taux d'erreur, au contraire de l'AED et χ² qui sous-estiment le taux d'erreur pour les faibles taux d'erreur aléatoire.
4. Modèle d'accessibilité en heures creuses
4.1 Affectation en heures creuses
Le modèle routier a été importé dans le logiciel OmniTRANS11. C’est également dans ce logiciel que sont réalisées les affectations de trafic en heures creuses comme en heures pleines. Trois modèles ont été mis en œuvre et trois matrices des temps ont été générées en heures creuses selon une affectation tout-ou-rien (AON assignment) en fonction des typologies de routes et de leurs vérifications successives, avec une cartographie des temps issus des observations d’une part, et par le calcul de leur ajustement avec les autres sources d’autre part. Le premier modèle (Mob_a) base sa typologie de routes selon la classification communale de Van der Haegen et al. (1996). Le modèle Mob_b s’appuie sur une typologie différente tenant compte de la traversée des zones urbaines ou commerciales à un niveau infracommunal.
11 Version 5.0.28, développé aux Pays-Bas par la société OmniTRANS International. http://www.omnitrans-international.com
Figure 21: Test de sensibilité à l'erreur pour cinq statistiques d'ajustement
89
Le modèle Mob_c, quant à lui, est une variante de Mob_b, pour laquelle on a cherché à corriger les écarts de vitesse s’éloignant de plus de 20% par rapport aux temps déclarés MOBEL et par rapport aux temps de trajet estimés par le modèle Google Maps, pour les distances supérieures à 10 km. Nous avons choisi de ne pas retenir le cas des distances trop faibles. En effet, les origines et destinations dans MOBEL peuvent se situer partout dans le périmètre d’une commune. Ainsi, nous avons pu calculer que la distance moyenne des trajets domicile-travail et domicile-études intracommunaux fait 4 km, ce qui donne une idée de l’erreur possible sur les distances intercommunales. Les corrections apportées ont varié, allant de l’ajout de liaisons routières, là où la hiérarchie routière retenue ne permettait pas une modélisation réaliste de la desserte des communes, jusqu’à la correction des paramètres des liaisons routières (correction du nombre de voies) à l’aide de photos aériennes. La représentation cartographique des temps à partir de chaque commune à destination de Bruxelles est présentée en Figure 6. On constate que les classes de temps de trajets vers Bruxelles s’étalent différemment selon les modèles. Les modèles Mob_a et UCL ont une distribution des classes plus large autour de Bruxelles, ce qui correspond à des vitesses plus importantes. Les modèles Mob_b et Mob_c révèlent une distribution similaire à celle du modèle Google Maps, et montrent une distribution plus resserrée, correspondant à des vitesses plus réduites. Par ailleurs, on peut voir à travers la distribution des classes le long des axes autoroutiers que ces modèles sont davantage sensibles à la hiérarchie des routes par rapport à Mob_a et UCL.
90
Figure 22: Représentation des temps de trajet en heures creuses selon différents modèles
91
4.2 Enquête et modèles d'accessibilité pour la Belgique
4.2.1 Enquête MOBEL
La base d’observation disponible est issue de l’enquête nationale réalisée en 1999 sur la mobilité des personnes en Belgique dans le cadre du projet MOBEL (Hubert et Toint, 2001). Nous disposons ainsi, à l’échelle du pays, de données de temps de parcours pour un jour ouvrable moyen hors vacances scolaires, ce qui correspond à 10036 observations brutes. Cette base servira de référence pour les comparaisons en heures de pointe du matin et en heures creuses pour une journée type. Des observations de MOBEL, il faut néanmoins soustraire les observations hors du champ de notre étude. C’est le cas, par exemple, pour des localisations hors du périmètre d’étude (origine ou destination à l’étranger, par exemple) ou encore des valeurs aberrantes (valeurs nulles ou distances très inférieures à la ligne droite entre deux centroïdes de communes). En regroupant les observations restantes en période creuse (soit en dehors des deux périodes de pointe entre 7 et 9h et entre 15 et 18h), nous avons ainsi un total de 1125 couples origine/destination, et 598 en heures de pointe du matin (7-9 h).
4.2.2 Modèle routier internet : Google Maps
Il a été possible, grâce aux API mises à disposition par Google, de procéder à la résolution des itinéraires entre les couples O/D en heures creuses présents dans la base MOBEL. Pour cela, une première étape a consisté à géocoder les origines et destinations afin de vérifier leur adéquation avec les centroïdes retenus dans modèle routier. Les données routières utilisées dans Google Maps sont fournies par le Ministerie van de Vlaamse Gemeenschap (Ministère de la Communauté flamande) et par le Ministère de l'Équipement et des Transports.12 Il s'agit donc a priori du même réseau que celui dont nous disposons, à un niveau de détail moindre. Nous utiliserons ce modèle pour les comparaisons en heures creuses uniquement. Ce modèle sera par la suite noté GMAP.
4.2.3 Modèle développé à l'UCL
Dans le cadre du projet ‘Accessibility indicators to places and tranports’ un modèle d’accessibilité a été développé par G. Vandenbulcke et al. (2009) à l’échelle de la Belgique. L’approche est sensiblement différente en ce qui concerne le calcul de l’accessibilité, puisqu’elle implique une impédance sur les tronçons d’une commune selon sa densité d’emploi et de population. Les calculs sont disponibles pour toutes les communes en heures creuses. En heures pleines en revanche, seuls sont disponibles les temps des communes à destination des 53 villes des trois niveaux supérieurs de la hiérarchie urbaine belge actualisée (Van Hecke, 1998). Ce modèle sera par la suite noté UCL.
4.3 Qualité de l'ajustement en heures creuses
Les statistiques d'ajustement des modèles en heures creuses sont résumés dans le Tableau 1. Modèle AED X2 SRMSE PSI PHI Mob_a 0,026 12670 0,84 0,30 0,33 Mob_b 0,037 5779 0,68 0,27 0,30 Mob_c 0,036 5789 0,67 0,27 0,29 UCL 0,048 11400 0,80 0,30 0,33 GMAP 0,024 6281 0,76 0,28 0,29
12 http://www.google.com/intl/fr_fr/help/legalnotices_maps.html
92
Tableau 1: Valeurs d'ajustement des modèles en heures creuses
En ce qui concerne l'AED, les modèles Mob_a et GMAP sont mieux positionnés que Mob_b, UCL et Mob_c. Par contre, pour χ2, SRMSE, ψ et φ les modèles Mob_b et Mob_c et GMAP présentent des ordres de grandeurs similaires devant Mob_a et le modèle UCL. Il semble donc que la prise en compte du type d'urbanisation traversée des modèles Mob_b et Mob_c apporte un niveau de précision intéressant, si on se base sur les comparaisons cartographiques avec GMAP et les quatre dernières statistiques du tableau 1. A la lumière des résultats de ces statistiques, le modèle Mob_c a été conservé plutôt que Mob_b pour mettre en œuvre le modèle en heures de pointe.
5. Modèle d'accessibilité en heures pleines
5.1 Affectation en heures de pointe du matin
Trois modèles ont là aussi été testés. Ils ont été construits sur le modèle Mob_c. Les paramètres de la fonction débit-vitesse doivent être définis pour les différents arcs routiers. Là encore, nous les définissons en respectant la typologie définie lors de l'étape de mise en place du réseau routier. Les flux issus de l'enquête socio-économique de 2001 ont été rapportés à une périodicité d'une heure et affectés sur les arcs routiers par un script recherchant un équilibre utilisateur13. Deux jeux de paramètres ont été testés à partir du modèle Mob_c défini en heures creuses, notés Mob_c1 et Mob_c2 et leurs temps à destination de Bruxelles cartographiés (Figure 7). On a conservé la discrétisation des temps de trajets en heures creuses dans un but de comparabilité. Dans le Mob_c1, les couronnes des communes d'une même classe temporelles sont très resserrés par rapport à la situation de Mob_c2. Les temps d'accès dans le cas du modèle UCL en heures de pointe ont une distribution concentrique autour de Bruxelles, au contraire des modèles développés dans MOBLOC qui privilégient un axe Anvers-Bruxelles. L'explication réside probablement dans la présence des deux autoroutes qui relient les deux villes, et donc mieux à même de répartir les flux entrants vers Bruxelles en heures de pointe que dans le cas des autres communes périphériques, la plupart ne bénéficiant que d'une seule autoroute.
13 La matrice O/D de la demande de déplacements est affectée sans tenir compte de la diagonale.
93
5.2 Qualité de l'ajustement en heures pleines
Les résultats des statistiques comparant les observations de MOBEL en heures de pointe du matin avec les matrices de temps des différents modèles sont présentés dans le Tableau 2. Précisons que cette dernière comparaison est faite sur un effectif de 230 couples O/D. Cet effectif est bien plus réduit que dans l’analyse en heures creuses pour deux raisons. D’une part, la période considérée est plus courte, et d’autre part, les données disponibles dans le modèle UCL ne concernent pas toutes les destinations, mais seulement les communes des trois premières classes de la hiérarchie urbaine de Van Hecke. Modèle AED X2 SRMSE PSI PHI Mob_c1 0,019 801 0,312 0,234 0,241 Mob_c2 0,024 883 0,376 0,276 0,279 UCL 0,068 872 0,420 0,262 0,272
Tableau 2: Valeurs d'ajustement des modèles en heures de pointe du matin
Des statistiques d'ajustement des trois modèles sont présentés dans le tableau 2 et ont des ordres de grandeurs similaires. Le modèle Mob_c1 présente les meilleurs résultats. Nous
Figure 23: Temps de trajet vers Bruxelles en heure de pointe du matin selon les différents modèles
94
considérons ainsi que c'est sur ce modèle que doivent être calculés les indicateurs d'accessibilité permettant le retour vers le modèle de migration de MOBLOC (Figure 1).
6. Conclusion L'objet de cette communication était double : d'une part présenter la mise en place d'un modèle basé sur un réseau routier simplifié à l'échelle des communes de Belgique, et d'autre part de définir des critères de validation de ce modèle s'appuyant sur deux sources, le modèle Google Maps et le modèle développé à l'UCL (Vandenbulcke et al. 2009). Concernant le premier point, on peut dire que malgré les simplifications opérées au niveau du réseau routier, simplifications rendues nécessaires pour des raisons pratiques de temps de calcul et de disponibilité des données, il a été possible de mettre en place un modèle d'accessibilité en HC et en HP à l'échelle de toutes les communes de Belgique. Enfin, concernant la méthode de validation retenue avec l'emploi des statistiques d'ajustement, elles-mêmes, on constate une certaine variabilité dans les résultats, qui favorisent parfois l'un ou l'autre modèle. Certaines précautions doivent néanmoins être soulevées, quant à l'utilisation des temps de trajets déclarés sur une base d'enquête. D'une part, concernant les trajets en période creuse, ceux-ci sont assimilés à des conditions de trafic fluide, ce n'est pas forcément le cas en milieu de journée, même en dehors des heures de pointe. Néanmoins, il est nécessaire de prendre en compte de tels intervalles pour disposer d'un effectif suffisant. En second lieu, il faut garder à l'esprit l'incertitude liée à la précision des réponses. Ainsi on peut supposer que les temps de trajets courts seront plus fréquemment arrondis (à 5 ou 10 min) ce qui crée une erreur proportionnellement plus importante par rapport au temps réellement nécessaire. Des relevés terrain, en éliminant la part de subjectivité de la base d'enquête, permettraient-ils un meilleur ajustement ?
95
Références bibliographiques Branston, D. (1976) Link capacity functions : a review. Transportation Research 10, 223-236 Brück L., Halleux J.-M., Mérenne-Schoumaker B., Savenberg S. et Van Hecke E., 2001.
« L’intervention de la puissance publique dans le contrôle de l’étalement urbain - Première partie : état de la question en Belgique », SSTC - Leviers d’une politique de développement durable, 154 p. - http://www.ulg.ac.be/geoeco/segefa
Bureau of Public Roads (1964) Traffic Assignment Manual. Urban Planning Division, US Department of Commerce, Washington D.C.
Cornélis E., Bahri A., Eggerickx T., Carpentier S., Klein S., Gerber P., Pauly X., Walle F., Toint P., 2009. Mobilities and long term location choices in Belgium "MOBLOC", Final report for the Belgian Science Policy, Phase 1. 47 p.
Dijkstra E. W. (1959) A note on two problems in connexion with graphs. Numerische Mathematik, 1, p. 269–271.
Goodwin, P. (1995). Car dependence. Transport Policy, 2(3), p. 151–152. Halleux J.-M., Brück L., Mairy N., (2002). « La périurbanisation résidentielle en Belgique à
la lumière des contextes suisse et danois: enracinement, dynamiques centrifuges et régulations collectives », Belgeo, 4(2002), p. 333-354.
Hubert J-P., Toint P. (2002) La mobilité quotidienne des Belges. Presses Universitaires de Namur. 352 p.
Joly I. (2005) «Décomposition de l’hypothèse de constance des budgets temps de transport», in Mobilités et temporalités Montulet B et al.(dir), Facultés Universitaires Saint Louis, Bruxelles, 129-150
Knudsen D. C., Forthingham A. S. (1986) « Matrix Compararison, Goodness-of-Fit, and Spatial Interaction Modeling », International Regional Science Review, Vol. 10, No. 2, p. 127-147
Shannon, C. E. (1948) The mathematical theory of communication. Bell System Technical Journal 27: 379-423 ; 623-56
Van der Haegen H., Van Hecke E., Juchtmans G. (1996) Les régions urbaines belges en 1991, Études Statistiques, n° 104.
Van Hecke E. (1998). « Actualisation de la hiérarchie urbaine en Belgique ». Bulletin du Crédit Communal, vol. 205, n° 3, p. 45-76.
Vandenbulcke G., Steenberghen T., Thomas I. (2009) Mapping accessibility in Belgium: a tool for land-use and transport planning? Journal of Transport Geography, 17(2009) p. 39-53
Wiel M., (1999). La transition urbaine ou le passage de la ville pédestre à la ville motorisée, Mardaga, 149 p.
Zahavi Y. et Talvitie A. (1980) « Regularities in Travel Time and Money Expenditure », in : Transportation Research Record, 750, 13-19.
96
MOBLOC : links between residential mobility and accessibility
Xavier PAULY14 Eric CORNELIS14 Fabien WALLE14
Abstract:. This paper relies on the results of the MOBLOC project which aimed at analyzing localization choice for household, daily accessibility and their strong interactions. Broadly speaking, a mechanism of this interaction can be described as follows: traffic evolution impacts accessibility, which itself, together with long-term societal changes, affects internal population migration and household localization, which in turn influences mobility (traffic) and accessibility. The analysis of this interaction requires two complementary intertwined modelling approaches: the first centred on the residential migration and the second on the evolution of accessibility. We will particularly focus here on the residential mobility described through two successive models: one dealing with the propensity to move and the second with the localization choice. Keywords: residential mobility, accessibility, migration, model 1. Introduction Mobility and transport evolve with time and interactions are numerous between daily mobility and household migration. The evolution of the transport system has deeply modified the barrier of distance and has largely opened the choices in term of residence place. The continuing urban sprawl phenomenon (Brück et al, 2001 – Halleux et al., 2002) resulting from these modifications has itself resulted in a strengthening of the property and housing market in certain territories, pushing people (young couples in particular) towards a residential localization which is further and further away from the traditional, urban activity centres. The tensions between daily and residential mobility have therefore increased, notwithstanding the recent rise in energy costs. This in turn generates unsustainable effects on society and environment. But these new residential choices have in parallel induced new mobility behaviours, based on an extensive (and probably excessive) use of the private car (Wiel, 1999) in daily trips (home-work/school, shopping, leisure ...). Social life itself (visits to friend and family) has become more spatially dispersed. One already knows that the propensity to change residence is determined by a number of individual or household characteristics such as age, citizenship or income, but the effects of long-term trends as population ageing, the evolution of the household/family structure on both residential choices and mobility behaviours remain so far largely unanticipated. Therefore the MOBLOC research project, funded by BELSPO and undertaken by GRT (University of Namur), GéDAP (UCL) and CEPS-INSTEAD, aimed at analyzing retroactions between demographics and the evolution of mobilities at different time-scales. In particular, localization choice for household, daily accessibility and internal migrations appear to have strong interactions. Broadly speaking, a mechanism of this interaction can be described as follows: traffic evolution impacts accessibility, which itself, together with long-term societal changes, affects internal population migration and household localization, which in turn influences mobility (traffic) and accessibility. The analysis of this interaction requires two
14 University of Namur (FUNDP), naXys, Transportation Research Group (GRT)
97
complementary intertwined modelling approaches: the first centred on the residential migration and the second on the evolution of accessibility. In this paper the focus will be put on the first axe, residential mobility. Due to lack of more spatially accurate data, this residential mobility has to be understood in this research project as a change of municipality. For analyzing the factors having an impact on this kind of mobility, we develop, within the MOBLOC framework, two models, each taking into account one stage of the decision process for changing home location. The first one deals with the propensity to move and determines which are the household characteristics impacting the choice of staying at the same location or to move to another residence. The second one models how the new residential location (municipality) is chosen when the decision to migrate is taken. This paper is structured as follows. The global methodology underlying development of the different models is first presented. In this section, we will also sketch the developed accessibility model. Then the focus is put on the migration model and more especially on its two components. A section will be devoted to the propensity to move model. It will describe the data used, the structure of the model and the obtained results. Another section will then present the localization model. Here also the available data will be described; then we will discuss the different types of models developed in this phase of the project to end with the structure of the discrete choice model which seems the most promising for modelling this location choice. Finally, this paper will be concluded with the possible applications of the developed framework and with the possible further steps. 2. General MOBLOC framework Thus let us first describe the global methodology of the project relying on Figure 24 and examining the data flows through the inputs and outputs of each model and the interactions between them. Since the main objective of MOBLOC is to draw up a link between residential migration and daily mobility, the two main bricks of the project are the residential migration model and the accessibility model. However, as Figure 24 shows, other models are necessary to reach this goal. The inputs of the residential migration model are individual data. It includes age, gender, level of education, household evolution, previous migration and the current (at time/year Y) residential municipality. The final output is the new residential municipality (one year after, Y+1). This model has been split up into two sub-models: a propensity to move model and a localization model. The first one uses some of the individual information to predict if a migration occurs (i.e. if a individual moves) between the years Y and Y+1 (from a municipality to another – not inside a given municipality). If the answer is yes, then the localization model will simulate the localization choice between the 588 other Belgian municipalities. As one of the individual characteristics is the activity status, a data aggregation at the municipal level allows providing the number of working people per municipality which is one of the inputs of the transport model and more precisely of the travel demand model. Travel demand model is indeed based on two inputs: the number of working people and the number of jobs per municipality. These are the margins of an O/D matrix from which a gravity model will build all the inner cells. This model assumes that the intensity of flows between an origin and a destination depends on the respective masses of the spatial units (jobs/employed people) and is inversely proportional to the distance between them (thus regarded as an obstacle to the interaction). Border effects (between regions) are also taken into account.
98
Let us remark that for now, this matrix only takes into account the work-home trips (demand) because of the used inputs. Even if this assumption is very restrictive, it is a first good approximation considering the available data (as we use them for the peak hours model). Nevertheless possible improvements could be forecast. The next step of the transport models is a modal split model. Its purpose is to compute from the global O/D matrix another one concerning only the trips achieved by car. This step is necessary to provide a feasible input for the private vehicle accessibility model which is an essential brick of the project. This model (in fact, two: one for morning peak period, one for off peak) outputs is a matrix of the travel time between each municipalities pair (during the morning peak hours on a working day, off peak respectively) by car. At the end, the travel times between municipalities will be used to compute different accessibility indicators (to employment, services …). These will be included as inputs of the localization model (the second step of the residential migration model) to measure the attractiveness of the municipality. For analyzing the impacts of different evolution scenarios, this chain of models will be used in a prospective way so that it will be necessary, at each step, to update the input values for the residential migration model. This is the role of the evolution models used to determine how data must be updated to incorporate the changes from a year to the next one. They include different kinds of techniques at different levels: aggregate and disaggregate, e.g. demography trends, synthetic population…
Figure 24 : Pattern of the MOBLOC project
To close this section we will briefly describe the accessibility model; more details could be found in (Klein et all, 2010).
99
Within the context of the MOBLOC project accessibility is defined here as the "the greatest or least ease of moving from a Point A to a Point B" (Reymond et al., 1998). Therefore, through the accessibility model, the objective is to provide a measurement of this ease for some specific spatial points. In this project this ease of movement is expressed in the traditional manner of journey time, whilst at the same time being aware that other criteria such as cost or even comfort may be taken into consideration. Given the general objective of the MOBLOC project, which is based on model coupling, there is a large constraint, i.e. the accessibility modelling for all 589 Belgian municipalities. In fact, to guarantee the greatest degree of consistency at the model coupling level, the model must have the same spatial precision, i.e. the 589 municipalities here. On this scale and given the density of the Belgian transport network, simplifications have had to be undertaken in order to enable the calculation to be completed in a reasonable time. More specifically this simplification consists of choosing for each municipality a point considered as representative of its centrality according to the network or the urban areas. This comes back therefore to assimilating the accessibility from this point to that of the whole municipality; over short distances this simplification certainly imposes a constraint, but on the other hand, at country level, this premise seems to be a realistic one. For the accessibility model, the input corresponds to an origin/destination matrix of workers and students using the car as the main mode for their home/work or home/school journey. In the first iteration, this data was sourced from the 2001 INS Socio-Economic Survey. The outputs of the model will be a square matrix of the minimum/optimum access times between each pair of municipalities. These outputs (i.e. travel-time indicators) are then used for the coupling of the models. The large phases of the accessibility modelling consist in traffic assignment on the network to consider congestion in the calculation of access time. Two road accessibility models were set-up, the former for off-peak hours, the later for morning peak hours (7 to 9 A.M.). The off-peak hours model is based on a free-flow calculation of the shortest path on the network. During in-peak hours, we use a model allowing to assign the demand matrix on the shortest cost paths, taking account of the relationship between the travel time on a road and the vehicles flow. It is based on the first hypothesis of Wardrop (Wardrop, 1952) and specifies the conditions allowing reaching a user equilibrium by the method of successive averages. 3. Propensity to move model The first stage in a residential mobility model is to understand why people decide to change their location. This is the goal of our propensity to move model, which will highlight the factors involved in the decision “Shall I move or shall I stay?” as well as their respective weights in this choice. Applied on a synthetic population (Barthelemy and Toint 2010) located at municipality level, it will determine which are the candidates for migration. Before going deeper in methodologies it should be mentioned that we could access to a first database (National Register - NR) describing states at successive 1st of January at an individual level from 2001 till 2006. That means that we cannot observe any real events but the ones deducted from transitions between two consecutive firsts of January. A second source of data is the national census (Socio-Economic Survey – ESE 2001) achieved in October 2001 giving only one shot information. This remark about the lack of series data is important for the prospective step as we will see later. We define that a residential migration occurs for an individual when his/her residential place (municipality) is different between two successive 1st of January. So migration inside the same municipality or multiple migrations in one calendar year are not detectable in the datasets since we only have the records of the municipality for the place of residence at each 1st of January. The measured “risk” reflects
100
thus the fact that the residential municipality at a 1st of January is different from one year to the next one. This kind of "event" is quite rare since it concerns 4% of the Belgian population each year. The explanatory variables were selected according to a literature review (e.g. Debrand and Taffin, 2005; or Henley, 1998) and their availability through the National Register or the Socio-Economic Survey of 2001. We are dealing with individual characteristics as well as housing and area of residence characteristics. As individual characteristics we selected age, gender, nationality, household type, number of individuals per household, position in the household (i.e. the link with the head of household), the highest education level successfully completed, the activity status and type of activity, and if a migration occurred the year before. As housing characteristics we considered the type of housing (house, flat…) and housing tenure type. As regards area of residence characteristic, we took into account the urban/rural profile of the municipality (downtown, rest of the city, old and recent periurban areas, rural areas) built by Van der Haegen et al. (1996). However since the highest education level successfully completed and the activity status are strongly associated, we decided to keep only one from these two variables. We finally chose the education level for two main reasons: firstly, the way activity status and activity types have been built in the ESE2001 survey is not satisfying. Secondly, education level is a "capital" that one cannot lose contrary to activity status; hence, it is a more stable variable in time and should be easier to project in a near future. In this context and given that education level is generally as good as activity status as a proxy for the socioeconomic profile of individual we preferred selecting education level. We have also kept in the set of variables the housing tenure type because it is just essential to analyze residential migration: this variable largely discriminates people in migration depending on whether they are renter or owner, or whether they are renter in a public or private market. For instance the ratio of migrating people is more than 3 times higher for renters in a private housing than for owners. Extracted from the Socio-Economic Survey of 2001, housing tenure type at individual level is only available at one date (October 2001). However trends provided by the National Institute of Statistics shows that it is a stable state through times, especially in Belgium where ownership is traditionally high: between 1991 and 2001, the percentage of owners increased from 65.4 to 65.9%, and forecast concerning potential impact of an economic uncertainty on becoming owners do not seem significant in the Belgian market (Vanneste et al., 2007). In this context, it appears that this variable would be stable during time, and not too complex to forecast in a near future. Once the variables to be taken into account chosen, we had to decide which methodology to be used for building our propensity to move model. After different attempts, we focused on the binary logistic regression. In such a model we also want to take advantage of the dynamic dimension of the available data and analyze past and anticipation effects. Indeed, it seems obvious that past information in the life of an individual can help to explain his/her migration behaviour. For example a change in the household’s structure or size (which can reflect events such as separations, births or unions) can increase (or decrease) the probability for change in residential municipality. At the same time we also want to test effects from anticipation. For instance, testing whether an observed migration for a couple in the transition period Y and Y+1 may be explained by the fact that this couple has a baby within the transition period Y+1 Y+2. They could have anticipated the birth by looking for a bigger housing the year before. Therefore, while building the model which has to explain migration between years Y and Y+1, we looked for a suitable way to include this dynamic dimension. Rather than simply use the state variables at the first of January of the years Y-2, Y-1, Y, Y+1 and so on, we decided to use the state variables only at time Y and to create transition variables which describe variation in the state variables between two successive 1st of January. These transition variables have
101
been created for variables from the National Register i.e. household structure and size, position in the household and nationality. But for anticipation we have only considered household transitions (variables related to household). Regarding the Socio-Economic Survey of 2001 variables, as they are only available in October 2001, it was not possible to create transition variables. These transformed variables are helpful to add in the model a "temporal depth" of one or more years back or forward. From a statistical point of view, transition variables as a substitute of a sequence of temporal successive states have the great advantage to handle multicolinearities or autocorrelation, which could occur when using these states as explanatory variables in the model. Moreover, they make easier the interpretation of the results since transition variables implicitly suppose an event (although not observed).At a last point, it should be said that since dynamic database allows catching unobserved heterogeneity, we add in the model an autoregressive dimension which is the dependent variable at the previous year (did an individual move the year before?). For calibrating the model, the question to use a sample rather than the entire base arose when we faced some problems with the size of the database and the calculation time. As there are more than 10 million citizens in Belgium, we have about the same number of observations, which is more than necessary. We may indeed reasonably reduce the size of the sample without losing quality of information for the model. However as residential migration is a rare event (4% per year), we wanted to increase its frequency in order to improve the quality of the coefficients of regression (Allison, 1999). This is the reason why we oversampled the migrants. As a consequence the final sample is based on a stratified random sampling in the overall Belgian population. Stratification has been done according to the dependent variable: for one year observed, we selected all people having changed from municipalities and, for the non-migrants, we realized a simple random withdraw. The total number of individuals drawn for the sample arises 2 millions and the percentage of annual migrants comes around 15%. The sample was, for the calibration, divided in two sub-samples : 70% of the drawn individuals were used for the calibration itself whilst the other 30% were kept to test and validate the models. Different models were built in order to compare migration model for two different transition periods: 2002-2003 and 2003-2004, but also to test the effects/advantages of adding older (delay) or future (anticipation) information. The Table 12 below summarizes the five models tested and the variables selected (with their source) in each case. We used the stepwise procedure (Allison, 1999) selecting one by one the variables which are significantly related to the explanatory variable conditionally to the previously entered variables. This procedure can also eventually remove one of these covariates thereafter, if it has lost its significance due to the addition of another variable. In our models, all of the explanatory variables removed from the regression because of lack of significance concerned naturalization: naturalizations between 2001 and 2002 and between 2002 and 2003 for models 1 and 2; naturalization between 2003 and 2004 for models 4 and 5. Then we have compared these models thanks to some global criteria such as AIC and log Likelihood. We present in Table 13 the different likelihood ratios of our models. Since the differences between the tested models are not large, we can assess that all our models are well calibrated, all khi-2 tests on likelihood ratios being extremely significant. Another way to assess models performance relies on the discrimination: does the model well predict the behaviour of the individuals? Is somebody moving also predicted as a mover? Does the model well separate the migrants from the non-migrants? If we have a look at the Figure 25 here below, we can see that principal diagnostics about discrimination show very few improvements between models 1 and 2 and amongst models 3 to 5. The fact that model 3 seems to be worse than the first 2 models (regarding the predictive powers) can be explained by the loss of relevance of covariates coming from the census.
102
All four statistics present high scores which could be understood as a proof that the models are well fitted. For example, the c statistic which is an approximation of the area under the ROC Curve also providing a measure of discrimination must have a value between 0.8 and 0.9 to consider that the model have an excellent discrimination (Hosmer and Lemeshow, 2000). In our case, values of c are close to 0.9 which means our models are really discriminatory.
Dependant variable Models Explanatory variables (with sources) Age in 2002 (NR) Gender (NR) Nationality in 2002 (NR) Type and size of the household in 2002 (NR) Evolution of the type of the household between 2001 and 2002 (NR) Evolution of the type of the household between 2002 and 2003 (NR) Link with the household head in 2002 (NR) Evolution of the link with the household head between 2001 and 2002 (NR) Evolution of the link with the household head between 2002 and 2003 (NR) Migration between 2001 and 2002 (NR) Education level successfully completed (ESE 2001)
Model1
Housing tenure type (ESE 2001) As in Model 1 + Evolution of the type of the household between 2003 and 2004 (NR)
Migration between 2002 and 2003
Model 2 Evolution of the link with the household head between 2003 and 2004 (NR) As in Model 1 (but with a one year shift (e.g. 2003 in place of 2003) + Naturalization between 2002 and 2003 (NR) Model 3
Naturalization between 2003 and 2004 (NR) As in Model 3 + Naturalization between 2001 and 2002 (NR) Evolution of the type of the household between 2001 and 2002 (NR) Evolution of the link with the household head between 2001 and 2002 (NR)
Model 4
Migration between 2001 and 2002 (NR) As in Model 4 + Evolution of the type of the household between 2004 and 2005 (NR)
Migration between 2003 and 2004
Model 5 Evolution of the link with the household head between 2004 and 2005 (NR)
Table 12 : Explanatory variables of the tested models for the propensity to move
Tests on the Likelihood Ratios
Model 1 Model 2 Model 3 Model 4 Model 5 Khi-2 502377.9 505456.7 511072.5 511247.3 506794.7 Ddl 215 301 219 306 394 Pr > Khi-2 < 0.0001 < 0.0001 < 0.0001 < 0.0001 < 0.0001
Table 13 : Test on the likelihood ratios for the 5 tested models
103
Figure 25 : Comparison of the 5 models of propensity to move according to various diagnostics (regarding to the predictive power)
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
Somers' D Gamma Tau-a c
Model 1 Model 2 Model 3 Model 4 Model 5 As a conclusion about the propensity to move models, we remark that all the five models seem quite relevant. As the first one, which is also the model requiring the less data, seems to be already more than satisfactory, we decided that this one could be used in the next steps of our project. The four most explanatory covariates of this model (in a decreasing order) are the evolution of the type and size of the household between Y and Y+1, the housing tenure type, the evolution of the link with the household head between Y and Y+1 and the age in Y . The most significant variable is clearly the simultaneous transition of the household structure (type and size). Compared to the reference category (no change in the household structure between the two consecutive years), all the odd ratios are positive which means that any change in the household structure is correlated with an increasing probability to move (everything else being equal). The importance of this increase depends on the household structure change. Amongst the most significant, let us point out the people whose household structure were "married couples with children" and became "unmarried couple with/without children", "one-parent family" or "isolate". They have from 20 to 30 times more likelihood to move than people whose household structure did not change. Let us remember that family events are univocal in the data. These people can be a child who left parents' home, or members of the couple which split up, or etc. This can be related to another significant variable: the change of relation to the household head. For example, being a child status to a head or spouse status (or inversely) is related to a high increase of the probability to move (between 11 and 16 times more than in the case of no change). The second most significant variable for the prediction of the propensity to migrate is the housing tenure type. Renters in the private sector are 4 times more likely to move than owners. Age class also appears amongst the four most significant variables. The less residentially mobile are 75 years old and more while the most mobile are the youngest (less than 18 years). The probability to move decreases with age (everything else being equal). All these results underline that the propensity to move is linked to the life course of individuals, and more particularly their family trajectory. These are well observable through the age and transition of the household structure covariates. To sum up, the transitions leading to move are break-up situations, new family-units compositions and leaving parents’ home. One can observe that the more stable situation concerns people who are in married couple (with or without children) ; this situation is often associated with an owner status, which is another factor to stay in the same municipality of residence. In other words, the likely evolutions of family situations marked by the rise of less stable households (cohabitation situations, one-parent family…) will still generate higher propensity to move rates in the coming years.
104
4. Localization model Contrary to the propensity to move model, the localization model has to consider also municipalities’ attributes to model the choice of the new residential municipalities. Therefore preliminary analyses were undertaken to highlight the most relevant factors related to municipalities. From a sample of 100,000 people moving between 01/01/2001 and 01/01/2002, one can observe that residential moves in Belgium are often short: for almost one in every two people changing their residential municipality, the intermunicipal distance is less than 10 kilometres, whereas only 10 percents of the movers decide to go living 50 kilometres or more away from their former municipality of residence. Other analyses were achieved on the covariates available for the modelling. For each of them, we estimated how each can explain change of residential location. To do so, we estimated models with only one explanatory variable under BIOGEME (Bierlaire, 2003). Among the available variables, the less explanatory variable taken independently is the property prices indicator (which takes into account the mean prices for houses and flats in the municipality). The most explanatory one concerns the distance between the municipality where the individual lived in 2001 and the municipality where he settled down in 2002. In general, municipality variables seem to bring more explanatory power than individuals ones. Since the purpose of the localization model is to determine a new municipality for people who decided to migrate (between two consecutive years as studied in the propensity to move modelling), each individual who decides to move has to choose between 588 alternatives (municipalities). To model such a behaviour, discrete choice methods (Train, 2003) seems the most relevant tool. This technique consists in determining the utility of each alternative and then computing the probability of choice of each alternative. As explained before, the propensity to move model was built with individual as unit/agent of decision. This choice was quite natural as the members of a household can have different behaviours. For the localization model, we would intuitively like to consider households. Nevertheless, we cannot take the initial household for this unit because the composition of the households can change between two consecutive years. The solution we adopted was to work with all the individuals who moved and were household head in their new municipality. Since the head of household is nevertheless characterised by some household variables (as the household type and size), some household information can also be used for the localization decision as it would intuitively be the case. The main challenge of this model is the number of alternatives. As the meshing defined for Mobloc is the municipalities level, the choice set is composed of the 588 other (than current one) Belgian municipalities. Dealing with so many alternatives is unusual in the literature. However it is possible to limit the number of alternatives that are available for each unit of decision during the estimation process. It means that, for calibrating the model, we could deal with randomly selecting only some of the possible alternatives. During our developments for building the model, different kinds of discrete choices models were considered: Logit but also nested Logit. With so many alternatives, we could suppose that some similarities between municipalities would not be explained by the variables. We considered two sources of such similarities: the geographic proximity and the municipality type (rural, urban, etc.) on basis of Van der Haegen typology or Van Hecke classification, For building the model, we settle up a utility with individual variables (age, education level, nationality, household type) and with communal variables (living condition indicator(s), property price, intermunicipal distance, .potential accessibilities (services, job opportunities) coming from the accessibility model). Since it was unrealistic to use alternative specific parameters, a generic parameter is associated with each variable but it makes the individual variables non-explanatory. Indeed,
105
each of the 588 alternatives (municipalities) would have the same utility term for this variable. As the choice lies on the difference between alternatives utilities, this term may not be explanatory and lead to unidentifiable models. In order to get different values for the explanatory variables and to avoid problem of constancy between the alternatives, we had recourse to a “contextualization approach”: for categorical individual variable. This means that the utility term of an individual variable is computed as the product of a class specific parameter, a binary variable indicating whether an individual belongs to this class and the class proportion in the municipality (alternative). Formally, it implies one term per class, but for an individual every but one (he/she belongs to) of these terms are null. So, for an individual j, the utility term of this variable for the alternative i is actually the product of a class specific parameter with the class proportion in the municipality. The interpretation of the parameters is less evident but can be explained this way: a positive parameter would indicate that people having the same profile tend to settle down in a same place while a negative parameter would represent a trend to a dispersion of the concerned people. We also tested two ways to take into account the link with the former residence municipality: the intermunicipal distances (distances between the communal centroids) or the change of district via a binary variable (1 for municipalities in the same district as the former one). Regarding accessibility indicators we tested different forms for the associated utility terms: a unique parameter or distinct parameters per age class or per active/non active status. The idea of distinct parameters was to try to detect different influences of accessibility according to individual characteristics such as age class. For example, we can suppose that the job opportunities indicator could be less explanatory for elderly people. In order to spare time, we first worked with a sub-set of data and alternatives to test several model structures. Then, we performed estimation on bigger datasets and all the alternatives for the models that were the best ones with the partial data. The way we chose this subset of alternatives was not random. We wanted to have representative municipalities of Belgium, i.e. from different type (rural, urban, etc.) and from different zones. We thus fixed 5 zones (Brussels Region, two in Flanders, two in Wallony) and the proportion of municipality types we wanted for each zone (according to the real proportion). Table 14 here below summarizes the main steps followed in the model building.
1.1. Sampling alternatives or not 1.2. Logit or nested Logit (with different ways of building nest) 1. Tests on a subset of 60 municipalities 1.3. selecting different variables and using contextualisation or not
2. Estimation on the all set (589 municipalities)
Table 14 : How practically the localization model was estimated
Now we can present the main conclusions drawn from the testing of these different models as well as the results of the finally retained model. Concerning the sampling of alternatives, as expected, the confidence intervals are larger for estimation with sampling which use less information to estimate the parameters. But the parameters values were not very different, i.e. the confidence intervals (95%) "have a common part of range". Generally, the values with sampling belong to the confidence interval of the parameter without sampling. Regarding the model structure, the retained model is a nested model with 4 nests based on the communal typology (and not on the geographic proximity) according to the Van der Haegen typology (Van der Haegen et al., 1996). More sophisticated nests (crossing both geographic and typological criteria) lead to model with many unsignificant parameters.
106
The most significant variable is the distance from the former municipality (distance between the two centroids) but Table 4 gives the details for all significant variables.
Parameter Value Std err t-test p-value
Accessibility indicator to employment for active people
0.0417 0.0208 2.01 0.04
Accessibility indicator to employment for non-active people
-0.175 0.0421 -4.16 0.00
Distances between the new residential municpality and the former one
-0.0565 0.000433 -130.38 0.00
Living conditions indicator (environment aspects) 1.08 0.201 5.37 0.00 Living conditions indicator (dwelling aspects) -3.30 0.247 -13.38 0.00
Living conditions indicator (socio-economic aspects) 4.72 0.195 24.27 0.00
Living conditions indicator (services aspects) -3.45 0.285 -12.11 0.00 Property prices indicator -0.373 0.163 -2.28 0.02 Population of the new residential municipality 0.271 0.00987 27.43 0.00 Age : 0 to 18 -17.7 6.49 -2.73 0.01 Age : 19 to 29 11.7 1.05 11.15 0.00 Age : 30 to 44 11.4 1.26 9.03 0.00 Age : 45 to 54 5.77 2.48 2.33 0.02 Age : 55 to 64 31.9 2.44 13.07 0.00 Age : 65 to 74 10.7 2.75 3.88 0.00 Age : 75 and more -2.20 3.57 -0.62 0.54 Education : other 1.09 0.577 1.88 0.06 Education : secondary school (inferior) 3.18 1.20 2.66 0.01 Education : secondary school (superior) 4.28 0.852 5.02 0.00 Education : higher education 4.05 0.581 6.98 0.00 Kind of household : other 3.20 2.59 1.24 0.22
Kind of household : couple (married or not) with children
1.28 0.332 3.87 0.00
Kind of household : couple (married or not) without children
5.95 0.637 9.34 0.00
Kind of household : single-household 12.0 0.472 25.38 0.00 Kind of household : one-parent family 7.78 0.716 10.87 0.00 Nationality : other 3.55 0.818 4.34 0.00 Nationality : belgian 4.07 0.324 12.54 0.00 Nationality : border countries 1.33 1.52 0.87 0.38 Nationality : European Union 6.04 1.83 3.31 0.00
Table 15: Parameters of the retained model (calibration on 60 municipalities)
Regarding the accessibilities variables, only one of the two accessibility variables was kept since accessibility to services indicator and accessibility to employment indicator were indeed highly correlated (r=0.89). In the utility formulation, the best results for the accessibility to employment indicator are obtained with specific parameters for active and non active individuals. People were considered as active between 19 and 64 year old and non active outside of these limits. Considering these two distinct parameters lead to a better model than considering a unique parameter for employment indicator. Introducing one parameter per age class (7 classes) does not improve significantly the quality of the model. Regarding the other municipal characteristics, the four components of the living condition indicator are significant; the municipal property price is also significant.
107
Finally, regarding individual characteristics, defined as contextualisation variables (i.e. a positive sign means that individuals tend to settle down in the municipalities where they find similar individuals), most of these parameters are significant in the retained model. Amongst them, only the parameter for 0-18 age class is negative. All other parameters are positive. From this point of view, the model reflects well the behaviour of tending to live with one's own socio-economic group. 5. Conclusion The framework of MOBLOC models allows now to study the links between accessibility and residential mobility. Applied on the Belgian case, it could be used to analyze the trends within the residential choices. Coupled with a built synthetic population for each Belgian municipality and with evolution models, the MOBLOC models provide a tool for forecasting internal migrations in the future. However these models could still be refined. For example, the accessibility model could take into account mobilities for other purposes than work or school (leisure, shopping, etc.). On another hand, if more spatially disaggregated data could be available, considering residential changes within a same municipality could also be a suitable improvement for our residential mobility model. Bibliography Allison P., Logistic Regression Using SAS Theory and Application, SAS Press, 1999 Barthelemy J, Toint Ph., Synthetic Population Generation in Presence of Data Inconsistencies, naXys Technical Report 12-2010, 2010 Bierlaire M., BIOGEME A free package for the estimation of discrete choice models, Proceedings of the 3rd Swiss Transportation Research Conference, 2003 Brück L., Halleux J.-M., Mérenne-Schoumaker B., Savenberg S. et Van Hecke E.,. L’intervention de la puissance publique dans le contrôle de l’étalement urbain - Première partie : état de la question en Belgique, SSTC - Leviers d’une politique de développement durable, 2001 Debrand T, Taffin C., Les facteurs structurels et conjoncturels de la mobilité résidentielle depuis 20 ans, Économie et statistique, 381-382, pp. 125-146, 2005 Halleux J.-M., Brück L., Mairy N., La périurbanisation résidentielle en Belgique à la lumière des contextes suisse et danois: enracinement, dynamiques centrifuges et régulations collectives , Belgeo, 4, pp. 333-354, 2002 Henley A., Residential mobility, housing equity and the labour market, The Economic Journal, 108, pp. 27-414, 1998 Hosmer D.W., Lemeshow S., Applied Logistic Regression, Wiley, 2000 Klein S., Omrani H., Carpentier S., Gerber P., Vandenbulcke G., Validation d’un modèle d’accessibilité par recoupement de données multi-sources : application aux communes de Belgique, MTL, 2010 Reymond H., Cauvin C., Kleinschmager R. (coord.), L’espace géographique des villes – pour une synergie multistrates, Anthropos, collection Villes, 1998 Train K.E., Discrete Choice Methods with Simulation, Cambridge University Press, 2003 Van Der Haegen H., Van Hecke E., Juchtmans G., Les régions urbaines belges en 1991, Études Statistiques, 104, 1996 Vanneste D., Thomas I., Goossens L., Woning en woonomgeving in Belgie, Sociaal-economische enquête 2001, SPP Politique Scientifique & SPF Économie – Direction générale Statistique, 2007
108
Wardrop J.G., Some theoretical aspects of road traffic research, Proceedings of the Institute of Civil Engineers, Pt II, vol.1, pp 352-378, 1952 Wiel M., La transition urbaine ou le passage de la ville pédestre à la ville motorisée, Mardaga, 1999
109
ANNEX 2: MINUTES OF THE FOLLOW-UP COMMITTEE MEETINGS
Annex 2.1. Minutes of the first follow-up committe meeting (3rd of May 2007)
Personnes présentes :
BAHRI Amel (UCL – GéDAP)
BEX Marie-Carmen (Politique Scientifique Fédérale)
BIERNAUX Olivier (RW – IWEPS)
CARPENTIER Samuel (CEPS/INSTEAD– GEODE)
CORNELIS Eric (FUNDP – GRT)
DEBUISSON Marc (RW – IWEPS)
EGGERICKX Thierry (UCL – GéDAP)
GERBER Philippe (CEPS/INSTEAD– GEODE)
HOORNAERT Bruno (Bureau Fédéral du Plan)
LAMBRECHT Micheline (Bureau Fédéral du Plan)
PAULY Xavier (FUNDP – GRT)
SERBUYNS Martine (Ministerie van de Vlaamse Gemeenschap)
TOINT Philippe (FUNDP – GRT)
VAN DUYSE Dominique (MET)
Personnes excusées :
BEHEYDT Koenraad (Ministerie van de Vlaamse Gemeenschap)
HOFMAN Peter (Ministerie van de Vlaamse Gemeenschap)
La réunion commence à 14h30.
E. CORNELIS informe de l’ordre du jour et propose un tour de table permettant la
présentation de chacun des participants. Suite à cela, la parole est donnée à M-C.
BEX qui présente le programme « La science pour un développement durable »
(2005-2009). Elle informe de la pluri annualité de celui-ci, de son budget, et des
différents projets qu’il connaît, dont une part concerne de la recherche transversale.
Parmi ces projets ISEEM et MESsAGE portent aussi sur la problématique des
transports.
La réunion avec le comité d’utilisateur doit se faire 2 fois par an.
110
Les équipes de recherche impliquées dans le projet MOBLOC font ensuite l’objet
d’une présentation spécifique ; E. CORNELIS présente brièvement le GRT, T.
EGGERICKX, le GéDAP et P. GERBER, le département Géode du CEPS/INSTEAD.
E. CORNELIS reprend la parole pour exposer les objectifs du projet avant de donner
quelques informations quant au plan de travail. Sa présentation se termine par
l’explication d’un schéma résumant le projet MOBLOC avec les interactions entre les
diverses variables.
Réactions du comité :
Les membres du comité d’utilisateurs sont alors invités à réagir et à poser leurs
questions relatives au projet.
M. DEBUISSON se demande à quelles références géographiques la population
synthétique s’applique ?
E. CORNELIS lui répond le GRT travaille à l’échelle communale pour créer une
population synthétique d’individus qui est ensuite regroupée en ménages.
M. LAMBRECHT interroge sur le recours éventuel aux variables relatives aux
activités économiques et au statut d’activité issues de l’Enquête Socio-économique
de 2001 (ESE2001). Elle fait allusion à l’enquête « force de travail » qui pourrait être
employée. Elle demande si l’on est en possession de données précises sur le
revenu. Est-ce une variable disponible et que les chercheurs envisagent d’exploiter.
T. EGGERICKX lui répond que pour le revenu, il existe un proxy réalisé par la VUB
qui permet de l’apprécier car il n’y a pas de variable « revenu » dans l’ESE2001. On
pourrait aussi exploiter les caractéristiques du logement, le niveau d’éducation sous
la forme d’un proxy car ces variables sont corrélées avec le niveau du revenu.
M. LAMBRECHT signale qu’il existe un ensemble de caractéristiques pour la
population (au niveau individuel) mais s’interroge sur de possibles variables macro ?
Celles-ci sont aussi très importantes à considérer !
E. CORNELIS et T. EGGERICKX lui signalent que c’est prévu, et qu’ils souhaitent
prendre en compte - par le biais d’un modèle d’analyse à choix discret – les attributs
pertinents des ménages mais aussi des communes de départ et de destination.
M.-C. BEX rappelle que le projet est à l’échelle nationale. Il lui est signalé que le
GéDAP dispose d’une base de données intéressante à ce niveau pour le volet
démographique, mais que, concernant le pan ‘mobilité et accessibilité’, il n’en est pas
de même et qu’il faudra avoir recours à différentes sources de données, et
notamment les réseaux GIS.
111
E. CORNELIS ajoute qu’il manque de données sur l’accessibilité en région
Flamande. Il questionne sur l’obtention éventuelle des réseaux digitalisés de la part
des deux régions (Flamande et Wallonne). Existe-t-il par ailleurs une digitalisation
des réseaux TEC ?
D. VAN DUYSE intervient pour informer que la RW ne travaille plus avec Téléatlas,
mais avec Navteq désormais. P. TOINT signale à celui-ci que le GRT serait intéressé
de disposer d’une licence d’exploitation pour mener à bien le projet. Concernant le
réseau TEC, D. VAN DUYSE fait remarquer qu’il est digitalisé jusqu’à un certain
niveau et qu’il serait possible d’utiliser ces données.
Toujours pour ce qui est des données relatives à la mobilité, P. GERBER informe
qu’il est possible de travailler avec des comptages sur les voies de communication
pour quantifier les déplacements sur différents tronçons. Mais il faut alors disposer
de tronçons digitalisés.
Selon B. HOONAERT, il serait peut-être utile de voir l’équipe BELSPO, car des
projets antérieurs ont sans doute travaillé sur les réseaux et pourraient avoir déjà
quelques données ou réseaux disponibles.
M. SERBUYNS fait remarquer que pour le territoire couvert par la Région Flamande,
il semble qu’il n’y ait a priori pas de problème à travailler avec Téléatlas pour obtenir
les données du réseau routier. Le réseau DeLijn semble également pouvoir être
obtenu et exploitable.
M. SERBUYNS émet un doute quant à l’utilité de travailler à l’échelle des 589
communes car il lui semble, à juste titre, que pour des communes d’une certaine
taille, il serait peut-être plus opportun de travailler au niveau des quartiers. E.
CORNELIS reconnaît que dans le cas de grandes villes, cela n’est peut être pas une
échelle pertinente. Mais nous n’avons pas trop le choix ! On s’en tient à la
disponibilité des données qui dans la plupart des cas se limitent à l’échelle
communale. Qui plus est, le découpage en secteurs statistiques ne correspond pas
toujours à quelque chose de très homogène dans les faits. Il n’est pas non plus
envisageable de travailler en utilisant les codes postaux qui ne représentent rien de
très rationnel.
T. EGGERICKX pense que dans le cas des grandes agglomérations, il serait
intéressant de descendre à l’échelle du quartier, mais pour la Flandre, il n’existe pas
de découpages au niveau des quartiers, mais seulement les secteurs statistiques. Il
112
lui semble que la commune d’Anvers dispose d’un découpage en quartiers. Il faudrait
peut-être rechercher qui s’en est chargé et voir éventuellement avec ces personnes
s’il est possible de travailler à partir de ce découpage. Il y a aussi la possibilité de
recourir aux anciennes communes, mais c’est à voir car l’usage des codes postaux
n’est apparemment pas toujours fiable!
P.TOINT confirme que les codes postaux ne sont pas fiables ! On a notamment
souvent un même code pour des anciennes communes. Pour travailler au niveau
infracommunal et définir un découpage de zones pertinent vis-à-vis de variables
contextuelles (i.e. variables socio-économiques macro), un travail de recherches
basé sur les annonces immobilières permettrait d’obtenir des données sur les
valeurs foncières et immobilières et définir ainsi ces zones, mais, ce serait fastidieux
et il semble que l’on s’éloigne de l’objectif de départ.
E. CORNELIS et T. EGGERICKX font remarquer que de toute façon, même si l’on
prenait un niveau de désagrégation défini à partir des prix de ventes immobilières, on
est toujours limité : on n’a pas le droit de descendre en deçà de 50 ménages par
zone (pour des raisons de respect de la vie privée). Les statistiques foncières et
immobilières reposent sur une base communale, les statistiques sur le revenu
reposent sur le secteur statistique, mais lorsque ce secteur compte moins de 50
ménages, on remonte à l’échelle supérieure.
Selon M. SERBUYNS, la mobilité ne devrait peut-être pas être envisagée de manière
quotidienne, mais plutôt hebdomadaire vu le comportement des gens : week-end à la
campagne pour les personnes vivant en ville, mobilité vers les secondes résidences,
loisirs en dehors de la ville,…mais le coût d’enquête pouvant mener à de tels
résultats semble être substantiel selon P. TOINT qui ajoute que les cycles ‘naturels’
de la mobilité semblent être effectivement plutôt hebdomadaires. Il s’agira donc
d’une des limites du projet dont les chercheurs sont bien conscients.
La notion d’espace de vie qui pourrait justement aider à prendre en compte ces
déplacements vers une résidence secondaire pour le week-end, d’après E.
CORNELIS
M. SERBUYNS demande d’il sera possible de dégager des tendances sur
l’accessibilité. P. TOINT émet un doute car les séries temporelles pour l’accessibilité
sont peu fournies.
M. LAMBRECHT interroge sur d’éventuelles enquêtes sur la manière dont les
individus font des arbitrages dans le choix du transport. P. TOINT répond par
l’affirmative, mais ces enquêtes n’apportent pas, ou alors peu, de profondeur
113
temporelle. Selon lui, et de façon générale pour ce projet MOBLOC, il faut avoir
conscience qu’il y aura plusieurs imperfections car les données (pour l’accessibilité)
sont plus ou moins complètes, plus ou moins récentes. Mais pour autant, il y aura
sûrement moyen d’en dégager quelques interactions fortes.
D. VAN DUYSE demande quelques informations sur des scénarios d’évolution.
Parmi ceux-ci, y en aura-t-il un reposant sur une rupture de tendance ?
P. TOINT lui répond qu’un des buts du travail est de donner une idée des scénarios
futurs. Les résultats seront sans doute intéressants, mais exploratoires. Les équipes
de recherche vont être notamment confrontées à des problèmes de calibration vu
qu’elles ne disposeront pas de beaucoup de données dans le temps dans le
domaine de la mobilité. Des difficultés de validation seront également rencontrées.
L’intérêt des modèles qui vont être mis en place résidera entre autre dans la
possibilité de mesurer les poids respectifs des différents facteurs. Ces modèles
permettront également d’observer certaines interactions entre le choix résidentiel et
l’accessibilité, ce qui représentera une des parties intéressantes des résultats
attendus.
Enfin, O. BIERNAUX propose d’envisager aussi les coûts des transports pour la mise
en place de scénarios futurs. Il faudrait utiliser des valeurs agrégées de manière
macroscopique. Cependant, ces coûts peuvent connaître une forte variabilité qu’il est
très difficile d’estimer.
La réunion se termine à 16h10.
114
Annex 2.2. Minutes of the second follow-up committe meeting (12th of June 2008)
Personnes présentes :
DEBUISSON Marc, IWEPS
GANY Bernadette, MET, Direction Générale des Transports
GOOSENS Wilfried, Vlaams Gewest, Dep. MOW
JUPRELLE Julien, IWEPS
VAN STEENBERGEN Alex, Bureau Fédéral du Plan
WILLEMS Michel, DG SIE
BEX Marie-Carmen, SPP Politique Scientifique
BAHRI Amel, GéDAP
EGGERICK Thierry, GéDAP
GERBER Philippe, CEPS/INSTEAD - GEODE
CORNELIS Eric, FUNDP – GRT
PAULY Xavier, FUNDP – GRT
TOINT Philippe, FUNDP – GRT
WALLE Fabien, FUNDP – GRT
Personnes excusées :
MAYERES Inge, Bureau Fédéral du Plan
TINDEMANS Hans, MObiliteitsRAad van Vlaanderen
VAN DUYSE Dominique, MET, Direction Générale des Transports
Les exposés présentés sont joints en annexe de ce PV. Ce compte rendu fera état
des réactions, questions et discussions qui ont accompagné et suivi les
présentations, les différentes interventions ayant été reprises de manière
chronologique. Alex VAN STEENBERGEN s’interroge quant à la non-utilisation de la variable «
statut professionnel » dans le modèle de propension à migrer. Fabien WALLE lui
répond que cette variable est très corrélée avec le « niveau d’instruction » et qu’il
fallait faire un choix. Lors de la construction du modèle, il est apparu plus judicieux
de garder la variable « niveau d’instruction ». En effet, celle-ci s’est avérée
115
statistiquement plus significative que le « statut professionnel » pour expliquer la
propension à migrer dans les différents modèles testés. De plus, considérer la
variable « statut professionnel » apporterait des difficultés pour les perspectives
(suite du projet). En effet, il s’agit d’une variable difficile à faire évoluer alors que la
variable « niveau d’instruction », qui est quelque chose d’acquis, posera moins de
problème.
Michel WILLEMS comprend bien que les deux modèles d’accessibilité développés
(voiture conducteur et transports en commun) représentent la commune par un
même point, mais se demande pourquoi il existe dès lors différents temps de
parcours vers une destination donnée pour une même commune. Philippe GERBER
lui répond qu’il s’agit uniquement d’un problème de représentation cartographique.
Une interpolation par triangulation entre différents points est réalisée, mais par la
suite, la commune se verra bien affecter un seul temps de parcours vers une
destination précise. Philippe GERBER insiste sur le fait que le point choisi pour
représenter une commune (hormis les 18 régions urbaines) est le point optimal pour
cette commune, selon le critère de choix, à savoir l’arrêt de transport en commun
pour lequel la fréquence de passage est maximale. Au passage, Philippe TOINT fait
remarquer qu’il ne s’agit pas de mettre au point un modèle de trafic précis, mais bien
d’un modèle approprié à la thématique étudiée, soit la mobilité résidentielle. Il est
important de respecter l’échelle utilisée pour les modèles de migration, à savoir les
589 communes du pays.
Concernant la demande de mobilité (matrice O/D pour déplacements habitation-
travail/école en voiture conducteur), Michel WILLEMS se demande où l’on a pu
obtenir des informations. Philippe GERBER lui indique qu’il s’agit de données issues
de l’exploitation par le GéDAP des résultats de l’ESE2001. Ce dernier insiste sur le
fait que l’indice d’accessibilité communal qu’il sera possible d’obtenir à partir des
modèles développés entrera directement dans le modèle de localisation résidentielle
en tant que variable explicative du choix de la commune de résidence par les
migrants, ce que comprend bien Michel WILLEMS.
Le fait que le modèle d’accessibilité relatif au mode voiture conducteur pourra donner
des résultats variables en fonction de la demande de mobilité (qui évoluera suite à la
nouvelle répartition de la population) est bien compris par Julien JUPRELLE. Il lui
semblerait plus intéressant d’également pouvoir faire évoluer les résultats du modèle
d’accessibilité transport en commun. Philippe GERBER est tout à fait d’accord avec
lui, mais insiste sur la nécessité de disposer de données futures. Philippe TOINT
signale que, sans nouveaux horaires, il est impossible de prédire une quelconque
évolution. Il existera toujours la possibilité aux chercheurs de pondérer les résultats,
116
mais comme le fait remarquer Eric CORNELIS, il s’agira alors d’émettre des
hypothèses sur cette évolution de l’offre de transport. Ces hypothèses pourront alors
être liées à divers scénarios (qui doivent être réalistes), telle une augmentation de
l’offre de transport.
Un des facteurs importants quant aux choix de localisation résidentielle pourrait être
l’augmentation des prix des carburants selon Julien JUPRELLE. Pourrait-on en tenir
compte dans les perspectives ? Une fois encore, il s’agit de scénarios à imaginer
d’après Eric CORNELIS. Michel WILLEMS fait remarquer que de toute façon,
l’accessibilité est assimilée à un temps de parcours pour lequel le prix n’entre pas en
considération. Cette remarque est correcte, mais il est tout à fait possible d’affecter
un certain poids aux variables.
S’agissant des données infotec utilisée pour les calculs d’accessibilité, Julien
JUPRELLE est informé qu’il s’agit de l’exploitation des bases de données au format
HASTUS reprenant les horaires des différentes lignes des sociétés de transport.
Concernant le choix de destination de la commune de résidence, il a été signalé lors
de la présentation que les chercheurs envisagent d’utiliser un modèle suivant une
hiérarchie de choix (dans un premier temps un groupe de communes – bassin
d’emploi, de vie – et ensuite une commune précise). Marie-Carmen BEX s’interroge
sur la solution qui sera choisie. Philippe TOINT lui répond qu’il s’agit de pouvoir
décrire les choix de la manière la plus souple possible et que, de toute façon, ces
possibilités devront faire l’objet de tests quant à la qualité des modèles développés.
Eric CORNELIS précise que cette construction du modèle de localisation a pour but
de hiérarchiser les choix. Au passage, Philippe GERBER et Philippe TOINT font
remarquer la différence qui existe entre la paramétrisation d’un modèle et sa
calibration. Dans le cas présent, il s’agit d’exploiter une quantité de données
importante afin de faire ressortir au mieux les facteurs influençant la décision des
individus lors d’une migration résidentielle. Il ne s’agit donc pas d’un choix arbitraire
des modélisateurs.
A propos de la typologie de Van der Haegen, Marc DEBUISSON souhaite savoir si
cette variable a été introduite dans le modèle de propension à migrer, et si elle a été
rejetée par le modèle, faute de significativité statistique, vu qu’elle ne fait pas partie
des variables explicatives retenues dans le modèle exposé. Amel BAHRI lui indique
que cette variable n’a pas été reprise dans le modèle de propension à migrer
puisqu’on ne s’intéresse qu’aux caractéristiques individuelles, Toutefois elle sera
considérée dans le cadre du modèle de localisation résidentielle. Philippe TOINT
signale que les chercheurs sont bien conscients de l’a priori que représente le
117
découpage du modèle de migration résidentielle en deux sous-modèles, mais il en
est ainsi pour des raisons pragmatiques.
Toujours concernant la typologie de Van der Haegen, Marc DEBUISSON questionne
quant à la version utilisée : nouvelle typologie de Van der Haegen et Van Hecke ou
bien ancienne typologie (mais adaptée par le GéDAP). Thierry EGGERICKX lui
répond qu’entre les deux versions de la typologie de Van der Haegen, les évolutions
sont parfois surprenantes, voire aberrantes (voir l’exemple cité de La Louvière). On
préfère alors se référer à la typologie adaptée par le GéDAP qui identifie des
catégories plus détaillées que celles proposées initialement par Van der Haegen
(exemples : petits bourgs ruraux, centre urbain).
Dans son exposé, Thierry EGGERICKX a fait référence à une standardisation
indirecte concernant les types de ménage. Marc DEBUISSON lui demande ce qu’il
faut comprendre par ce terme. Thierry EGGERICKX lui fait savoir qu’il s’agit de
passer outre l’effet de l’âge dans les caractéristiques des ménages, et cite l’exemple
de la structure par âge différente des CAE (couples mariés avec enfants) des CoSE
(cohabitant sans enfant) ; on retrouvera plus que probablement plus de personnes
de 45 ans dans la première catégorie citée alors que les 25 ans seront plus présents
dans la seconde.
Marc DEBUISSON demande si les projections se feront à l’échelle de la commune.
Amel BAHRI signale que les projections démographiques classiques permettent
d’obtenir des évolutions communales pour un nombre restreint de variables (âge,
sexe, type de ménage), Pour ce qui concerne l’unité géographique de ces
perspectives, ce sera plus probablement des groupes de communes. Philippe TOINT
informe que la méthodologie des populations synthétiques permet de disposer
d’informations relativement détaillées. Cette méthodologie ne requiert que des
données sur les totaux marginaux de croisements entre variables ainsi qu’un
échantillon de la population.
Concernant ces variables à faire évoluer, Alex VAN STEENBERGEN fait remarquer
que des modèles ont été développés au sein du Bureau Fédéral du Plan pour des
évolutions probables des taux de chômage futurs. Ceux-ci pourraient donc aider les
chercheurs dans l’évolution de la variable « statut professionnel ». Eric CORNELIS
lui indique que, dans le cadre du projet, il faudrait alors disposer d’évolution
individuelle et non globale. Amel BAHRI en profite pour faire remarquer la
désagrégation poussée des variables explicatives. Philippe TOINT explique qu’il
serait possible d’utiliser ces modèles à la condition de disposer d’une distribution
empirique dans laquelle on pourrait alors effectuer un tirage aléatoire simple. Amel
118
BARHI réinsiste sur la forte corrélation qui existe entre cette variable et celle du
niveau d’instruction. Il est donc préférable de ne considérer qu’une seule de ces
deux variables, et le niveau d’instruction présente l’avantage d’être plus facile à
manipuler dans le cadre de projections. Qui plus est, celui-ci donne comparativement
de meilleurs résultats en termes de pouvoir prédictif.
Marie-Carmen BEX lui demande alors si les seuils de significativité sont satisfaisants
pour toutes les modalités (du niveau d’instruction) lors la calibration du modèle de
propension à migrer. Amel BAHRI lui répond par l’affirmative. Ce n’est par contre pas
le cas pour la variable nationalité pour laquelle le pouvoir discriminant est moins
prononcée (la variable naturalisation n’étant quant à elle pas significative). A cette
occasion, Philippe TOINT signale qu’il en était de même pour ces variables dans les
analyses qui ont fait suite à l’enquête nationale de mobilité (MOBEL).
Concernant le modèle de localisation résidentielle, Wilfried GOOSSENS fait
remarquer que si l’accessibilité influence le choix de la commune de résidence, il faut
alors que cette donnée soit réellement perçue et considérée par les individus
amenés à porter un choix. Or ce ne doit pas être le cas pour tous. Philippe TOINT
confirme ses dires en soulignant l’intérêt considérable que représente le fait de
calibrer un modèle sur la réalité. Les aspects perceptifs et cognitifs agissent partout
et le modèle pourra le faire ressortir.
Michel WILLEMS demande si le prix de l’immobilier pourra être une variable
explicative comme mentionné dans la présentation de Thierry EGGERICKX. Eric
CORNELIS lui répond que cette variable sera testée lors de la phase de calibration.
Les membres du Comité de Suivi n’ayant plus de questions, Eric CORNELIS les
remercie pour leur participation à cette réunion, et leur donne rendez-vous pour une
prochaine réunion dans le courant du mois de novembre ou décembre.
119
Annex 2.3. Minutes of the third follow-up committe meeting (19th of October 2009)
Personnes présentes :
BERLIER Jacqueline, SPW – Programmation de la Mobilité
DEBUISSON Marc, IWEPS
JUPRELLE Julien, IWEPS
LAMBRECHT Micheline, Bureau Fédéral du Plan
JAMART Georges, SPP Politique Scientifique
DAL Luc, UCL - DEMO
EGGERICKX Thierry, UCL - DEMO
CARPENTIER Samuel, CEPS/INSTEAD - GEODE
GERBER Philippe, CEPS/INSTEAD - GEODE
CORNELIS Eric, FUNDP – GRT
PAULY Xavier, FUNDP – GRT
WALLE Fabien, FUNDP – GRT
Personnes excusées :
MAYERES Inge, Bureau Fédéral du Plan
VAN DUYSE Dominique, SPW – Programmation de la Mobilité
KLEIN Sylvain, CEPS/INSTEAD - GEODE
La présentation réalisée pour cette réunion est jointe en annexe de ce PV. Le
schéma général du projet figurant à la dia 4 est également annexé pour une
meilleure visibilité de celui-ci. Ce compte-rendu fait état des réactions, questions et
discussions qui ont suivi la présentation.
Concernant le modèle gravitaire
Les chercheurs ont parlé d’actifs résidents. Micheline LAMBRECHT demande des
précisions par rapport à la signification de ces termes. Philippe GERBER lui répond
que ce sont des personnes qui disposent d’un emploi et qui résident dans une des
589 communes du pays. Cela n’inclut donc pas les personnes à la recherche d’un
emploi. Les personnes concernées se déplacent pour travailler (au sein de leur
120
propre commune ou vers une commune autre). Samuel CARPENTIER en profite
pour insister sur le fait que les marges de la matrice de demande de transport
doivent être cohérentes ; le nombre d’emplois occupés (=destinations) équivaut à
l’effectif d’actifs occupés (=origines). Une remarque est également faite pour rappeler
que les données sources de cette matrice sont issues des résultats de l’ESE2001.
Jacqueline BERLIER s’interroge sur le calcul de la distance intra-communale, celle-ci
étant basée sur les « centres » des anciennes communes : existe-t-il une
pondération par les effectifs de population des distances entre ces centroïdes et les
anciennes communes constituant chaque commune ? Samuel CARPENTIER lui
répond que ce n’est pas le cas. En effet, vu l’échelle de travail et donc les ordres de
grandeur, cela ne changerait que très peu les résultats utiles du modèle.
Concernant le modèle de propension à migrer
Marc DEBUISSON souhaite revenir sur les variables explicatives de ce modèle.
Fabien WALLE reprend l’ensemble des 12 variables retenues. Il fait remarquer les
valeurs de Khi² qui donnent une idée de l’apport relatif de chaque variable dans le
caractère explicatif de ce modèle. Il faut cependant relativiser ces valeurs en fonction
du nombre de degrés de liberté de ces variables. La p-valeur tient compte de ces
deux paramètres ; dans le cas présent, cette valeur est inférieure à 0,0001 pour
l’ensemble des variables. Fabien WALLE repasse en revue les tableaux de la dia 13
de la présentation en expliquant que les personnes déclarant avoir obtenu un
diplôme de l’enseignement supérieur (classe de référence) comme niveau
d’éducation le plus élevé sont celles les plus enclines à changer de lieu de résidence
; les coefficients des autres modalités étant négatifs , cela réduit l’utilité de la non-
migration et donc la probabilité de migrer.
Concernant le modèle de localisation résidentielle
Micheline LAMBRECHT comprend bien l’intérêt de l’introduction des prix de
l’immobilier comme variable explicative du choix résidentiel des individus amenés à
déménager, mais elle demande dans quelle mesure il ne faudrait pas tenir compte
également des coûts énergétiques : selon elle, l’économie réalisée sur le prix de
l’immobilier ne doit pas être inférieure au surplus de dépenses engendrées par de
plus longs déplacements quotidiens (déménagement en périphérie par exemple).
Eric CORNÉLIS lui signale que cette augmentation des distances est en quelque
sorte intégrée aux indicateurs d’accessibilité (autres variables explicatives qui seront
ajoutées au modèle par la suite) vu que ceux-ci sont calculés à partir des temps de
parcours qui eux-mêmes sont fonction des distances à parcourir.
121
Concernant le modèle d’accessibilité « voiture privée »
En plus de la congestion prise en compte actuellement dans le modèle, Julien
JUPRELLE demande si un coût plus élevé d’utilisation de la voiture pourra être
considéré. Samuel CARPENTIER lui répond que de telles influences pourront
éventuellement entrer en ligne de compte dans la détermination du partage modal.
Concernant le « modèle » d’accessibilité « transports en commun »
Marc DEBUISSON demande ce qu’il va résulter de ce modèle et à quel niveau il se
trouve dans le projet. Eric CORNÉLIS fait savoir que les chercheurs ont fait une
demande pour obtenir les données relatives à l’offre de transport des diverses
sociétés actives sur le territoire et qu’ils n’ont à l’heure actuelle pas encore reçu ces
informations. Samuel CARPENTIER fait remarquer que pour ce modèle, il n’y aura
pas de pan « simulations ». Les résultats se baseront uniquement sur l’offre. Il serait
en effet beaucoup plus incertain de faire évoluer les temps d’accès entre communes
via le transport public car cela impliquerait de faire des suppositions sur l’évolution de
l’offre en transport en commun, ce qui dépend de la volonté des opérateurs.
L’hypothèse faite est l’adaptation de l’offre à la demande par les divers opérateurs :
s’il existe de plus en plus d’utilisateurs des transports en commun, l’offre sera
adaptée de telle sorte qu’il n’y aura pas de modifications des temps d’accès. Cet
élément aura également son importance dans la modélisation du partage modal. Les
outputs seront utilisés pour déterminer quelle fraction de la population utilise les
transports en commun pour tel ou tel trajet.
Julien JUPRELLE signale que la mise en place d’un parc relais, par exemple,
pourrait avoir son importance dans l’évolution de l’attrait pour les transports publics
et donc du partage modal. Samuel CARPENTIER lui répond que de telles
modifications seraient difficilement modélisables dans le cadre du projet MOBLOC. Il
faut en effet garder à l’esprit l’échelle à laquelle le travail est réalisé (agrégation
communale). Si l’on voulait modéliser le partage modal au niveau d’une ville, il est
évident que de telles modifications devraient être considérées. Mais vu que, dans le
cadre de ce projet, le travail se fait au niveau de l’ensemble du pays, des
simplifications de la réalité sont indispensables.
Micheline LAMBRECHT questionne quant à la présence systématique d’une
connexion vers les transports en commun pour toutes les communes. Samuel
CARPENTIER lui répond que toute commune se voit desservie par au moins une
société de transport en commun et que, par conséquent, il est admis que tout
122
habitant a accès à au moins un point de transport en commun (celui où l’offre est
maximale au sein de sa commune).
Concernant les modèles d’évolution
Marc DEBUISSON souhaite savoir si des perspectives démographiques telles que
celles du Bureau du Plan seront utilisées. Thierry EGGERICKX signale que des
tendances générales pourront être utilisées (évolution de l’âge par exemple).
Cependant, certaines variables comme la transformation des ménages devront faire
l’objet de simulations propres au projet MOBLOC. Il s’agira en effet de faire évoluer
une variable pour laquelle des modèles prospectifs n’existent pas auprès de sources
externes telles que le Bureau du Plan.
Les membres du Comité de Suivi n’ayant plus de questions, Eric CORNÉLIS les
remercie pour leur participation à cette réunion, et leur donne rendez-vous pour une
prochaine rencontre dans le courant du printemps 2010. Il signale enfin que si les
membres du comité de suivi avaient la moindre question ou remarque à faire, il leur
est toujours possible de prendre contact avec les chercheurs impliqués dans ce
projet. A cette fin, nous vous communiquons leurs coordonnées ci-dessous :
Pour le GRT des FUNDP :
Eric CORNÉLIS – [email protected] – 081 72 49 22
Xavier PAULY – [email protected] – 081 72 49 84
Fabien WALLE – [email protected] - 081 72 49 43
Pour le département DEMO de l’UCL :
Thierry EGGERICKX - [email protected] - 0 10 47 29 67
Luc DAL - [email protected] - 010 47 29 54
Pour l’Unité GEODE du CEPS/INSTEAD :
Samuel CARPENTIER - [email protected] - +352 58 58 55 302
Philippe GERBER - [email protected] - +352 58 58 55 601
Sylvain KLEIN - [email protected]
123
Annex 2.4. Minutes of the fourth follow-up committe meeting (31st of January 2011)
Personnes présentes :
BERLIER Jacqueline, SPW – DGO Mobilité et Voies Hydrauliques
CLUYTS Ivo, SPF Environnement
DEBUISSON Marc, IWEPS
NAYES Estelle, Bureau Fédéral du Plan
VANDRESSE Marie, Bureau Fédéral du Plan
JAMART Georges, SPP Politique Scientifique
EGGERICKX Thierry, UCL - DEMO
CARPENTIER Samuel, CEPS/INSTEAD - GEODE
GERBER Philippe, CEPS/INSTEAD – GEODE
KLEIN Sylvain, CEPS/INSTEAD – GEODE
CORNELIS Eric, FUNDP – GRT
PAULY Xavier, FUNDP – GRT
WALLE Fabien, FUNDP – GRT
Personnes excusées :
JUPRELLE Julien, IWEPS
LAMBRECHT Micheline, Bureau Fédéral du Plan
VAN DUYSE Dominique, SPW – DGO Mobilité et Voies hydrauliques
WILLEMS Michel, SPF Economie, P.M.E., Classes moyennes et Energie
TOINT Philippe, FUNDP - GRT
La présentation réalisée pour cette réunion est jointe en annexe de ce PV. Ce
compte-rendu fait état des réactions, questions et discussions qui ont suivi la
présentation des chercheurs.
Marc DEBUISSON demande comment l’évolution de tous les individus se fait dans le
modèle de localisation résidentielle. Est-elle fonction des chefs de ménages
concernés par un déménagement ? Fabien WALLE lui répond en signalant que le
modèle est effectivement calibré à partir d’observations relatives aux chefs de
ménage (et ses propres caractéristiques donc) mais que des variables relatives aux
124
ménages (notamment les 5 types de ménages définis) sont aussi considérées. Eric
CORNELIS insiste en faisant remarquer que l’hypothèse faite quant au choix
résidentiel dicté par le chef de ménage semble cohérente.
Marie VANDRESSE revient ensuite sur les résultats obtenus par les chercheurs
dans leurs modèles et le caractère contre-intuitif du paramètre relatif à l’indicateur
des prix de l’immobilier pour le choix résidentiel. Elle se demande dans quelle
mesure l’échelle d’analyse ne peut pas contenir l’explication à ce paramètre positif ;
on pourrait en effet imaginer que les déménagements intracommunaux sont plus
influencés par la composante « prix des logements ». Thierry EGGERICKX
précise que les mouvements qui ont lieu au sein d’une même commune sont
fréquemment le résultat d’ajustements (liés aux différentes étapes du cycle de vie).
Ce qui se concrétise également par le changement de statut d’occupation des
logements ; de locataire à propriétaire.
Thierry EGGERICKX continue en faisant remarquer que ces résultats quant aux prix
de l’immobilier ne le surprennent aucunement. La logique d’attractivité des
communes périurbaines (bien que les logements y connaissent des prix élevés) ne
semble pas s’essouffler ; les communes telles Lasne, Waterloo restent en effet les
plus prisées bien qu’elles soient les plus chères. Vu les mécanismes financiers
disponibles pour les « ménages les moins favorisés » (emprunt à long terme : 30, 40
ans voire sur plusieurs générations) le remplacement de génération peut se produire
dans ses communes puisque le prix des logements ne revêt plus un caractère trop
dissuasif. L’accès aux logements pour tout type de ménages peut être envisagé dans
ses communes aux prix de l’immobilier particulièrement élevés. De plus, les
personnes qui ont tendance à migrer sont celles qui ont les revenus les plus élevés
souligne Thierry EGGERICKX ; cela avait été montré dans le cadre du modèle de
propension à migrer où les chercheurs approximaient le revenu des individus par
leurs niveaux d’instruction.
Marc DEBUISSON estime que le statut d’occupation du logement
(propriétaire/locataire) aurait pu avoir un certain caractère explicatif pour le modèle
de localisation résidentielle. Par exemple, on s’attendrait à retrouver plus de
propriétaires dans les communes périurbaines alors que les taux de locataires
seraient plus élevés dans les centres urbains. Eric CORNELIS confirme cette idée
tout en rappelant que cette variable fut considérée dans le modèle de propension à
migrer précédemment calibré.
Marie VANDRESSE demande à quel niveau du projet peut-on retrouver le « long
terme » présent dans le titre du projet. Eric CORNELIS lui répond qu’en arrivant à un
nombre d’itérations assez important, on pourrait simuler les comportements
125
résidentiels avec leurs impacts sur l’accessibilité (et les indicateurs qui en découlent)
sur le « long terme ». Il pourrait être envisagé de simuler ces comportements et leurs
conséquences en terme de mobilités sur des périodes de 10 ou 20 ans. Les outils
pour se faire ont été développés mais le temps limité n’a pu permettre aux
chercheurs de pouvoir tout coupler. Eric CORNELIS signale que certains paramètres
des modèles pourraient alors être modifiés pour connaître leurs impacts sur les
résultats des simulations (mise en place de scénarios).
Marie VANDRESSE souhaite savoir si la matrice O/D qui provient du modèle
gravitaire de demande de mobilité (domicile-travail et domicile-école) est à chaque
fois calculée. Eric CORNELIS lui répond par l’affirmative en faisant remarquer que
quelques développements seraient cependant nécessaires pour un tel résultat. Par
exemple, une population synthétique pourrait être utilisée pour actualiser les marges
de cette matrice tout en recourant à des projections sur les localisations futures de
l’emploi.
Marie VANDRESSE mentionne qu’une telle matrice de demande O/D pourrait être
très utile pour le Bureau du Plan dans le cadre des travaux menés pour les «
perspectives transport ».
Jacqueline BERLIER s’interroge sur la prise en considération des nouveaux arrivants
dans nos modèles résidentiels ainsi que des travailleurs transfrontaliers pour
l’affectation des flux sur le réseau. Sylvain KLEIN lui répond qu’en effet, uniquement
les déplacements nationaux ont été considérés dans les comparaisons avec les
observations de MOBEL ; pour considérer les déplacements transfrontaliers, il
faudrait pouvoir recourir à des données qui ne sont pas disponibles jusqu’à présent.
Concernant les migrations internationales, Thierry EGGERICKX cite les chiffres de
80 000 entrées et 50 000 sorties annuelles pour la Belgique. Ces mouvements
concernent principalement Bruxelles et la Flandre.
Les membres du comité de suivi n’ayant plus de questions, Eric CORNELIS les
remercie de leur présence à cette réunion tout en ajoutant que le rapport final est en
cours de rédaction et qu’il leur sera transmis dès que possible.