Content uploaded by Gabriel Leite Mariante
Author content
All content in this area was uploaded by Gabriel Leite Mariante on Sep 12, 2017
Content may be subject to copyright.
FACULTY OF SCIENCE
Econometric generation of
individual daily travel and
activity pattern
A case study with the cross-border workers in Luxembourg
Gabriel LEITE MARIANTE
Supervisor: Prof. Martina Vandebroek
Katholieke Universiteit Leuven
Co-supervisor: Dr. Tai-Yu Ma
Luxembourg Institute of
Socio-Economic Research
Thesis presented in
fulfillment of the requirements
for the degree of Master of Science
in Statistics
Academic year 2016-2017
i
©Copyright by KU Leuven
Without written permission of the promotors and the authors it is forbidden to reproduce
or adapt in any form or by any means any part of this publication. Requests for obtaining
the right to reproduce or utilize parts of this publication should be addressed to KU
Leuven, Faculteit Wetenschappen, Geel Huis, Kasteelpark Arenberg 11 bus 2100, 3001
Leuven (Heverlee), Telephone +32 16 32 14 01. A written permission of the promotor is
also required to use the methods, products, schematics and programs described in this
work for industrial or commercial use, and for submitting this publication in scientific
contests.
Contents
Acknowledgements iv
Summary v
List of Tables vi
List of Figures viii
1 Introduction 1
1.1 Activity-based travel models . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Casestudycontext ............................... 2
1.3 Thiswork .................................... 2
2 Literature Review 5
2.1 Theoreticalframework ............................. 5
2.2 Methodologies.................................. 6
3 Methodology 10
3.1 Defining activity pattern . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Models...................................... 12
3.2.1 Activitytype .............................. 12
3.2.2 Activityduration ............................ 14
3.2.3 Modechoice............................... 16
3.2.4 Locationchoice ............................. 18
3.2.5 ModelEvaluation............................ 25
3.3 Synthetic population and pattern generation . . . . . . . . . . . . . . . . . 27
3.3.1 Synthetic population . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.3.2 Patterngeneration ........................... 29
4 Case Study 30
4.1 Descriptionofthedata............................. 30
4.2 ModelResults.................................. 35
4.2.1 Activitytype .............................. 35
4.2.2 Travel time and activity duration . . . . . . . . . . . . . . . . . . . 38
4.2.3 ModeChoice .............................. 39
4.2.4 Locationchoice ............................. 42
4.2.5 Overview of selected models . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Synthetic population and activity pattern generation . . . . . . . . . . . . 47
ii
CONTENTS iii
4.4 Visualization .................................. 47
5 Conclusion 51
References 53
A Parameter estimates for the final models 57
B List of Geographical Zones 95
Acknowledgements
This work was supported by the Luxembourg National Research Funds (C14/SR/8330766),
under the CONNECTING project (Consequential Life Cycle Assessment of multi-modal
mobility policies - the case of Luxembourg). The used database relies on the work of
the CABaC 2010-2013 research project (Construction and Analysis of a Knowledge Base
on mobility habits and attitudes towards energy of cross-border workers in Luxembourg,
FNR INTER/CNRS/09/01). The Survey was funded by the Ministry of Higher Educa-
tion of Luxembourg.
Firstly, I would like to thank my parents, Mariana and Armando, for all the support,
encouragement and love. To them, I owe all opportunities I have had in my life.
I thank Prof. Martina Vandebroek for having me as a Thesis student, for the supervision
and for the invaluable feedback on this work. I would also like to thank my readers, Prof.
Katrien Antonio and Prof. Bart de Ketelaere, for their time and support as members of
the reading committee.
A very special thank you has to be given my co-supervisor, Dr. Tai-Yu Ma, for the close
guidance, support and feedback during the elaboration of this Thesis, and for putting me
in close contact with the “real-world” research scenario. Without his invaluable support,
this work would have not been possible.
I thank the Luxembourg Institute of Socio-Economic Research for the opportunity to
develop my research in a pleasant and stimulating work environment. I would especially
like to thank Sylvain Klein, Olivier Bichel, Thimoth´ee Cuignet and Marion Patte for their
contribution to this work.
I also thank my great friends from the Master in Statistics program, who I was lucky to
meet during the past two years in Leuven, for the good times, the pleasant company, the
stimulating discussions and for making such a drastic change in my life easy and enjoyable.
I would equally like to thank all my dear friends across the Atlantic, in Rio de Janeiro
(and spread across the globe), for their companionship, support and for the invaluable
friendship, even at such a large distance. To you all, muito obrigado!
Finally, I would like to thank my girlfriend Robin, for the support, companionship and
love over the past year, and also for directly contributing to this Thesis through reading
and feedback.
iv
Summary
Activity-based travel modeling is an approach to model individual mobility behavior based
on the principle that traveling episodes undertaken by individuals in their daily schedules
are not considered to be ends in themselves, but rather means to reach locations where ac-
tivities are undertaken. According to this hierarchy of decision making, in order to model
traveling behavior, one needs to generate activity patterns form which travel patterns are
derived.
This work applies econometric methodologies to generate travel and activity patterns rep-
resenting a typical working day of the population of cross-border workers in the France-
Luxembourg border region. Using data from a mobility survey conducted by the Lux-
embourg Institute of Socio-Economic Research (LISER) in 2011, we estimate models to
generate the different attributes pertaining to a daily activity pattern.
An activity pattern is defined as a sequence of activity episodes undertaken by an indi-
vidual over the course of one working day. We define each activity episode as having four
attributes: activity type, duration, travel mode and location. In addition, we also gener-
ate the travel time between two activity episodes. Each attribute is generated separately
in a chain of models in which previously generated attributes are used as explanatory
variables to generate the next ones.
For every activity episode, Multinomial Logit models are used to generate the choice of
activity type. After that, Cox Proportional-Hazards regression models are used to esti-
mate both activity duration and travel time. Models which include survey responses and
time-related variables are compared with models in which only survey responses are used.
For mode choice, we compare results obtained using Multinomial and Conditional Logit
models. For location choice, we compare Conditional and Mixed Logit models to estimate
the municipality where individuals undertake activities which do not have a fixed location.
Moreover, when estimating the model for location choice, we tackle the issue of discrete
choice models with large choice sets due to the number of municipalities in the cross-
border region where individuals can conduct their discretionary activities. We compare
usual sampling of alternatives methods presented in the literature with a constraining
method based on attractiveness of the candidate destinations and on a measure of detour
from the axis formed by the fixed home and work locations of each individual, the detour
factor. The latter is based on the principle that individuals tend to conduct their discre-
tionary activities close to their home to work trajectory.
v
CONTENTS vi
For travel time and activity duration, we find that the full model with survey and time-
related variables performs better than a reduced model with only survey variables. For
mode choice, we find that a Conditional Logit model yields similar predictive performance
with a slightly better fit than a Multinomial Logit.
For location choice, without performing sampling of alternatives, we find that using a con-
strained set which is only based on attractiveness of destination has slightly better results
than using a constrained set based on both the attractiveness and the detour factor. On
the other hand, when performing sampling of alternatives, the results are substantially
better if the samples are drawn from a constrained set based both on attractiveness of
destination and on the detour factor than if the samples are drawn from a constrained set
based only on attractiveness of destination.
After estimating and validating the models with sample data, we apply them to a syn-
thetic population generated to represent the entire population of cross-border workers of
interest. The output is a full daily travel and activity pattern for the full population of
French cross-border workers in Luxembourg, which is then subject to visualization tech-
niques.
List of Tables
3.1 Distribution of detour factors of chosen activities for selected percentiles of
the home-work travel time distribution, taken from Ma and Klein (2017) . 21
3.2 Steps to align source and target sequences . . . . . . . . . . . . . . . . . . 26
3.3 Categories used to divide individuals in Benchmark . . . . . . . . . . . . . 27
4.1 Observed type of each activity episode . . . . . . . . . . . . . . . . . . . . 31
4.2 Average observed duration for each activity episode by type . . . . . . . . 31
4.3 Average observed duration for the travel to activity episode by type . . . . 32
4.4 Observed mode for each activity episode . . . . . . . . . . . . . . . . . . . 32
4.5 Comparing the effects of imbalance in the data . . . . . . . . . . . . . . . . 36
4.6 Results for activity type . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.7 Generated type of each activity episode . . . . . . . . . . . . . . . . . . . . 37
4.8 Average generated duration for the trip to each activity episode by type . . 38
4.9 Average generated duration for each activity episode by type . . . . . . . . 38
4.10 Cox and Snell’s pseudo-R2for travel and activity duration . . . . . . . . . 39
4.11 P-values for Likelihood Ratio tests comparing full models with reduced
models for travel time and activity duration . . . . . . . . . . . . . . . . . 39
4.12 Resulting performance measures for different approachs in mode choice
modeling..................................... 40
4.13 Generated mode for each activity episode . . . . . . . . . . . . . . . . . . . 41
4.14 Explanatory variables marked with ×were included in the models for the
different activity categories . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.15 Size of choice sets for the different activity categories . . . . . . . . . . . . 43
4.16 Location choice results for activity category “External meal” . . . . . . . . 44
4.17 Location choice results for activity category “Shopping” . . . . . . . . . . . 44
4.18 Location choice results for activity category “Services and Leisure” . . . . 44
4.19 Location choice results for activity category “Others” . . . . . . . . . . . . 45
4.20 List of models selected for each attribute . . . . . . . . . . . . . . . . . . . 46
A.1 Activity type - Episode 2 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 58
A.2 Activity type - Episode 2 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 59
A.3 Activity type - Episode 2 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 60
A.4 Activity type - Episode 3 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 61
A.5 Activity type - Episode 3 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 62
A.6 Activity type - Episode 3 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 63
A.7 Activity type - Episode 4 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 64
A.8 Activity type - Episode 4 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 65
vii
LIST OF TABLES viii
A.9 Activity type - Episode 4 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 66
A.10 Activity type - Episode 5 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 67
A.11 Activity type - Episode 5 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 68
A.12 Activity type - Episode 5 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 69
A.13 Activity type - Episode 6 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 70
A.14 Activity type - Episode 6 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 71
A.15 Activity type - Episode 6 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 72
A.16 Activity type - Episode 7 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 73
A.17 Activity type - Episode 7 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 74
A.18 Activity type - Episode 7 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 75
A.19 Activity type - Episode 8 (1/3) . . . . . . . . . . . . . . . . . . . . . . . . 76
A.20 Activity type - Episode 8 (2/3) . . . . . . . . . . . . . . . . . . . . . . . . 77
A.21 Activity type - Episode 8 (3/3) . . . . . . . . . . . . . . . . . . . . . . . . 78
A.22 Activity duration - Episode 1 . . . . . . . . . . . . . . . . . . . . . . . . . 79
A.23 Activity duration - Episode 2 . . . . . . . . . . . . . . . . . . . . . . . . . 80
A.24 Activity duration - Episode 3 . . . . . . . . . . . . . . . . . . . . . . . . . 81
A.25 Activity duration - Episode 4 . . . . . . . . . . . . . . . . . . . . . . . . . 82
A.26 Activity duration - Episode 5 . . . . . . . . . . . . . . . . . . . . . . . . . 83
A.27 Activity duration - Episode 6 . . . . . . . . . . . . . . . . . . . . . . . . . 84
A.28 Activity duration - Episode 7 . . . . . . . . . . . . . . . . . . . . . . . . . 85
A.29 Travel time - Episode 1 to Episode 2 . . . . . . . . . . . . . . . . . . . . . 86
A.30 Travel time - Epidose 2 to Episode 3 . . . . . . . . . . . . . . . . . . . . . 87
A.31 Travel time - Episode 3 to Episode 4 . . . . . . . . . . . . . . . . . . . . . 88
A.32 Travel time - Episode 4 to Episode 5 . . . . . . . . . . . . . . . . . . . . . 89
A.33 Travel time - Episode 5 to Episode 6 . . . . . . . . . . . . . . . . . . . . . 90
A.34 Travel time - Episode 6 to Episode 7 . . . . . . . . . . . . . . . . . . . . . 91
A.35 Mode choice - Commuting . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
A.36 Mode choice - Discretionary activities . . . . . . . . . . . . . . . . . . . . . 93
A.37LocationChoice................................. 94
B.1 Geographical zones in the Luxembourg-France border region . . . . . . . . 95
List of Figures
1.1 Spatial distribution of residential and workplace location of the French
cross-border workers in Luxembourg . . . . . . . . . . . . . . . . . . . . . 3
1.2 Block diagram showing the main steps in the process of generating an
activitypattern................................. 4
3.1 Example of activity pattern for one individual . . . . . . . . . . . . . . . . 11
3.2 First principle for constraining the choice set: adequacy of destination . . 19
3.3 Second principle for constraining the choice set: detour factor . . . . . . . 20
3.4 Ellipse containing feasible locations for discretionary acitivities . . . . . . 21
3.5 Combining the two constraining methods results in a signifcantly smaller
choiceset .................................... 22
3.6 Example of distance between two sequences . . . . . . . . . . . . . . . . . 26
4.1 Location of activities in the France-Luxembourg border region . . . . . . . 48
4.2 Location of activities by travel mode in the France-Luxembourg border
region ...................................... 49
4.3 Graphical representation of all trips in the France-Luxembourg cross-border
region ...................................... 50
4.4 Graphical representation of the trips in the France-Luxembourg cross-
border region separated by mode . . . . . . . . . . . . . . . . . . . . . . . 50
ix
Chapter 1
Introduction
1.1 Activity-based travel models
A variety of approaches can be used to model qualitatively and quantitatively individual
travel patterns. The choice of model type depends on the specific objectives, research
questions, resources and data available.
According to Arentze and Timmermans (2004), most traveling activities performed by
individuals in their daily life schedules do not consist an end in themselves, but are rather
a means to reach locations in which one or more activities will be undertaken. Bhat and
Koppelman (1993) define an activity pattern (or program) as an agenda of activities which
an individual has chosen to undertake over a certain time period together with additional
attributes pertaining to the activity episodes, such as duration, destination and travel
mode.
These principles suggest a hierarchy in the decision making process of traveling, in which
the main concept is the sequence of activities. Traveling behavior is a direct consequence
of activity choice. In consistence with these principles, individual travel patterns should
be generated from individual activity patterns. Moreover, the output of activity-based
travel models can have a variety of further applications, such as travel demand prediction
or mobility policy evaluation.
Although behavioral models applied to transportation research can be traced back to a few
decades earlier, the theoretical framework for activity-based travel models was developed
in the 1990’s, mainly in the United States. Interest in this new approach grew rapidly
and, by the end of the decade, the first seminal models started being operational. The
first major conference fully dedicated to activity-based models happened in 1995 in the
Netherlands, which helped spread the new methodology across the international research
community and trigger international interest (Rasouli and Timmermans, 2014).
Activity-based travel models generate and represent every activity and the corresponding
travel choice in an individual’s daily schedule. Several features are taken into account
when modeling daily activity schedule generation in an attempt to approximate the hu-
man behavior when making this type of decision in daily life (Castiglione et al., 2015;
Bhat and Koppelman, 1993). These include, but are not limited to individual needs to
1
CHAPTER 1. INTRODUCTION 2
take part in a given set of activities, available spatial and temporal constraints as well as
social, economic and demographic characteristics at the individual, household and larger
community levels.
Furthermore, activity-based models consider the interdependence between trips made over
the course of a day. According to Castiglione et al. (2015), activity-based modeling is the
most detailed type of travel modeling. The main distinctive feature of this type of mod-
eling is the capacity to chain a series of interrelated trips in a tour, thus more accurately
taking into account the sequence of travels performed by individuals, who often perform
more than one activity in a single tour with several chained trips.
1.2 Case study context
The Grand-Duchy of Luxembourg is a landlocked country in Western Europe bordered by
Belgium (North and West), France (South) and Germany (East). According to the World
Bank (2015), Luxembourg has the highest GDP per capita in the world, with 101050
dollars per inhabitant per year. With a population of approximately 570000 inhabitants
and an area of 2590 square kilometers, it is one of the smallest sovereign nations in Europe.
The economic and demographic characteristics of Luxembourg result in very high wages
and living costs for the population. Such costs, combined with a high demand for work
and excellent transportation connections, make it attractive for workers to live in the
neighboring countries, where life costs are lower, and commute to work on a daily basis.
Every day, around 160000 cross-border workers (“frontaliers”) travel across the border
from Belgium, France and Germany to their workplaces in Luxembourg. This number
is exceptionally high when compared to the total population of Luxembourg. In 2012,
the number of workers who lived in the neighboring countries and worked in Luxembourg
corresponded to 41% of the total working population in the country (Schmitz et al., 2012).
In the case study performed in this work, we apply activity-based travel models to ana-
lyze the mobility behavior of the French cross-border workers in Luxembourg. According
to national public statistics (STATEC, 2015), around 86000 people live in northeastern
France, mainly in the region of Lorraine, and commute for work in Luxembourg on a
daily-basis. Figure 1.1 shows an overview of the distribution of residential locations in
France and workplace locations in Luxembourg for the French cross-border workers.
1.3 This work
The objective of this study is to develop a model to generate a full daily travel and activity
pattern for the French cross-border workers of Luxembourg.
In this work, different methodologies from the literature in activity-based travel models
are applied to generate individual activity patterns for the population of French cross-
border workers in Luxembourg, as described in the previous section. Data collected in a
CHAPTER 1. INTRODUCTION 3
Figure 1.1: Spatial distribution of residential and workplace location of the French cross-
border workers in Luxembourg
mobility survey is used for model estimation and evaluation. The resulting models are
applied to a synthetic population generated to represent all French cross-border workers.
The generated patterns, which are subject to visualization techniques, can be used in a
variety of applications, such as input for travel demand forecast.
Data was collected through a mobility survey undertaken by the Luxembourg Institute
of Socio-Economic Research (LISER) in 2011 and comprises travel and activity episodes
of individuals who work in Luxembourg and live in northern France, commuting every-
day across the border. The survey interviewed 2320 individuals who live in France and
work in Luxembourg. the questionnaire contained questions regarding socio-economic
characteristics of the individual, details about the household and a description of the se-
quence of activities undertaken by each individual during his/her most recent working day.
The final model (or rather, the final sequence of models) generates daily activity patterns
with a list of activities to be undertaken by individuals. Furthermore, for every activity
episode, decisions have to be made regarding its duration, travel mode choice and munic-
ipality of destination.
The work process can be divided in four main stages:
1. Model estimation, in which a sequence of models to generate a full activity pattern
is estimated using survey data.
2. Synthetic population generation, in which a synthetic population is generated
CHAPTER 1. INTRODUCTION 4
to represent the full population of French cross-border workers.
3. Pattern generation, in which the models estimated in stage 1 are applied to the
synthetic population in stage 2, resulting in the generation of a daily activity pattern
for the full population.
4. Visualization, in which visualization techniques are applied to the resulting pattern
obtained in stage 3.
The main steps in the work process to generate a full activity pattern for the population of
French cross-border workers used in this paper are presented in a simplified block diagram
in Figure 1.2.
Figure 1.2: Block diagram showing the main steps in the process of generating an activity
pattern
After this introductory chapter, the following chapter presents a literature review on sev-
eral aspects which are relevant for this work, regarding the theoretical framework and
methodology. The third chapter describes the methodologies used for model estimation,
model evaluation and the procedure of synthetic population generation. The fourth chap-
ter presents the results of the case study, the estimated models, the generated activity
pattern for the synthetic population as well as visualization of the final results. The final
chapter concludes and suggests possible future research on the topic.
The results presented in this work were obtained mainly with the open-source software R.
Other softwares were used in specific parts of the work, such as Octave 4.0.0 to perform
Sequence Alignment, BIOGEME 2.0 to estimate the models for location choice, ArcMap
10.5 to produce the maps in the visualization section and NodeXL to produce the graphs
in the same section.
The discrete choice models for large choice sets applied in the location choice section
presented in Chapter 3 and part of the corresponding results presented in Chapter 4 of
this work are also the subject of the conference paper Ma et al. (2017).
Chapter 2
Literature Review
This chapter presents an overview of selected topics from the literature on activity-based
travel modeling. The literature review is divided in two main parts: theoretical frame-
work, in which main concepts and ideas related to activity-based modeling are discussed,
and methodological framework, in which methodologies used in selected past studies are
discussed. Special attention is given to aspects of previous methodologies which are rele-
vant in the application to the problem of the Luxembourg cross-border workers introduced
in the previous chapter.
2.1 Theoretical framework
As mentioned previously, the main goal of activity-based models is to generate daily ac-
tivity patterns for the individuals of a given target population, taking into account spatial
and temporal constraints. This generation can be used for different purposes, one of the
most important being as input to forecasting of travel demand (Rasouli and Timmermans,
2014).
According to Castiglione et al. (2015), activity-based models are the final step in a com-
plexity scale of travel modeling techniques. The different modeling techniques are, in
increasing order of complexity: sketch planning, strategic planning, trip-based models
and activity-based models. Moreover, the main feature which distinguishes activity-based
models from the previous steps in the complexity scale is the incorporation of detailed
chaining between trips as well as between individuals within a household.
Several ways to define trip chains have been explored in literature. Primerano et al. (2008)
argue that the most appropriate approach for activity pattern generation is the so called
“home-based approach”, in which a trip chain is defined as a sequence of trip segments
which begins and ends at home. The main reasons for this are (i) it is in agreement with
the basic philosophy that travel is merely a means to accomplish activities, (ii) empirical
evidence shows that the vast majority of daily trip schedules begin and end at home and
(iii) many constraints on travel time and, especially, mode choice come into play when
an individual leaves home, therefore, planning a trip chain in its entirety is more practical.
5
CHAPTER 2. LITERATURE REVIEW 6
As mentioned previously, individual schedules are often constrained by household-level
characteristics. Bhat and Koppelman (1993) propose a general conceptual framework for
individual activity schedule generation which is mostly based upon the intra-household
relations. Initially, household needs are identified which have to be fulfilled by activities
that can be broadly divided into subsistence (e.g. work) and maintenance (e.g. grocery
shopping). After that, household and individual attributes (such as vehicle ownership,
income, possession of a driver license, working hours, etc.) are considered to allocate
the household activities to the individuals. Finally, individuals plan their discretionary
activities based on their needs and time availability given the other household members’
schedules.
Furthermore, Arentze and Timmermans (2013) point out that most activity models pro-
posed in literature use one-day travel data as a measure of a typical day of an individual
or a household. This way, most activity models do not consider relationship across days,
possible influence of a history of previously conducted activities or relationships between
individuals/households. This does not necessarily correspond to reality and, thus, may be
seen as a limitation of the classical approach based on one-day data. Arentze et al. (2011)
propose an improvement in this direction through a method which takes into account
stated day of the week preferences to extend the one-day data by generating activity pat-
terns on a multi-day horizon.
2.2 Methodologies
A variety of methodologies to generate individual activity patterns can be found in the
literature. Very broadly, such methodologies can be divided into two main categories
of theoretical approaches: machine learning models and econometric models. This work
focuses primarily on the econometric approach.
Machine Learning approaches
On the machine learning side, Arentze and Timmermans (2004) propose a decision-tree
model to generate daily activity schedule based on heuristic decision making by users
when facing problems with multiple alternatives. Decisions are sequentially modeled,
first generating a so-called “schedule skeleton” with the main activities to be undertaken
in a given day and, at a later stage, specifying different attributes for each schedule (e.g.
transportation mode, travel party, etc.) ordered based on pre-defined priority levels. The
main goal is to mime the decision-making process of actual human beings, which does not
explicitly involve utility maximization, but rather some set of implicit decision rules.
Vanhulsel et al. (2009) use reinforcement learning techniques to generate sequential data
representing a sequence of interconnected activities which are undertaken by an individual
over the course of one day. Allahviranloo and Recker (2013) generate activity patterns
through Markov models, and find evidence that Support Vector Machines have a supe-
rior performance in prediction accuracy when compared to classic econometric techniques.
CHAPTER 2. LITERATURE REVIEW 7
Machine learning approaches have also been used to tackle specific attributes of activity
patterns rather than to generate a pattern itself. Ma et al. (2016) apply Bayesian Networks
to estimate choice of travel mode for the trips of a previously generated activity pattern.
Similarly, Ma and Klein (2017) apply Bayesian Networks to generate activity location
from a constrained choice set, also on previously generated activity patterns.
Econometric approaches
On the econometric side, the most common methodological framework used in generation
of activity patterns is discrete choice modeling due to the discrete nature of several of
the decisions related to choosing types of activities to be conducted and their attributes.
A seminal study in the application of discrete choice econometric techniques for travel
behavior was conducted by Daniel McFadden (1974) using data from mobility surveys
around the San Francisco Bay area to estimate the future demand of a new transporta-
tion network.
Arentze et al. (2011) propose a model for dynamic activity generation based on one-day
travel diaries. The model aims to maximize a utility function across a multi-day time
horizon which is defined as the utility of satisfying a need for a given activity which was
built up since the last time it was conducted. Estimation is performed using an Ordered
Logit approach. Activity patterns for a sequence of days in the future are generated
from observations from one single day based on the principle that “one-day diaries can be
viewed as random draws from long-term activity patterns”. In the practical application
presented here, the same principle is adopted and the one-day travel diary is assumed to
fully represent the working schedule of the individuals of interest.
Goulias et al. (1990) and Shiftan (1998) also apply Multinomial logistic techniques for
household level activity-based trip generation in two different United States-based datasets.
In both cases, work-related trips are initially estimated and, in the following step, non-
work related trips are estimated conditionally to the results of work-trip estimation, in
agreement with the previously mentioned conceptual division proposed by Bhat and Kop-
pelman (1993).
Bhat (2000) uses discrete multilevel models for mode-choice in previously generated pat-
terns. He assumes that, in the case of a spatially spread population, individuals can be
correlated within their geographical units. More specifically, individuals who work and/or
live in the same city/town/neighborhood tend to be more similar to each other.
Travel time and activity duration are usually the main attributes of activity pattern gen-
eration that do not fall within the discrete choice framework. For this type of decision,
Bhat and Koppelman (1999) suggest the use of hazard-based duration models. Practical
applications of such models can be found in Bhat (1996) and Ma et al. (2009). In the
latter, Cox Proportional Hazards models are applied to estimate duration of alternating
consecutive travel and activity episodes.
In addition, Goulias et al. (1990) perform a case study using data from the United States
and compare the results with those from a similar model applied in a previous study using
CHAPTER 2. LITERATURE REVIEW 8
data from the Netherlands. The large differences in results question the often discussed
notion of model transferability across areas, i.e., a model could be estimated using data
from one geographic region and immediately transfered to another region using the same
obtained coefficients.
Moreover, a large part of the time and research involved in this work was dedicated to ex-
ploring the destination choice methodology presented in Chapter 3 to tackle the problem
of excessively large choice sets in discrete choice models. The most common approach in
literature is to perform sampling of alternatives in order to reduce the choice set faced by
each individual. Azaiez (2010) and Guevara and Ben-Akiva (2013) present the method-
ology and discuss the advantages of this approach. Furthermore, they empirically show
that, although selecting different choice sets for different individuals may introduce bias,
this is a minor issue and adding a bias-correction term may in fact lead to less accurate es-
timates and poorer forecasts instead. Scott and He (2012) obtains improved fitting results
when using separate models for different types of activity. To further constrain the choice
set, Ma and Klein (2017) propose a method in which the set of available destinations for
each discretionary activity undertaken by an individual is constrained to locations around
the axis formed by the (fixed) home and work locations. This methodology is based on
the assumption that individuals tend to conduct non-work related activities in locations
which do not involve large detours from the way from home to work.
Other methodological issues
Decision sequencing
When the goal is to generate a full activity-based travel schedule for an individual (or a
household), several interrelated decisions have to be made regarding one single trip (or
trip-chain). An example would be which mode will be used, at what time of the day and
for how long will the activity be conducted. Therefore, an issue that naturally arises is in
which order these decisions are taken and how much one preliminary decision influences
the subsequent ones.
Most studies assume a pre-defined order of priority for the different decisions in which
the lower-priority decisions are taken conditionally on the result of the higher-priority
ones. As an example, Arentze and Timmermans (2004) propose the following sequence of
decisions, in decreasing order of priority: whether or not the activity will be conducted,
duration, travel party, time of day, in which trip chain it will be included, travel mode
choice and location.
Krygsman et al. (2007) tackle this issue by proposing a co-evolutionary model to clarify
the direction of the causal relationship between choice of activity type and of travel mode.
Separate models for the two choice-ordering possibilities are estimated (first activity type
and then mode and vice-versa) and a measure of decision uncertainty is computed for
each alternative, with the least uncertain decision being assumed to be made first. In a
case study with data from the Netherlands, results of this model’s application show that
activity type is chosen before mode in 97% of the times, suggesting a clear direction of
CHAPTER 2. LITERATURE REVIEW 9
causality. Indeed, it can be argued that this result agrees with the basic principle previ-
ously mentioned that activity is the main driver of human traveling.
Although not directly, Goulias et al. (1990) also provide some insight on the matter
in their practical application to a case-study. A model of hierarchical decisions is pre-
specified with the result of a higher-level decision being used as explanatory covariate for
a lower-level decision. Specifically in the case study presented, the number of mandatory
activity trips is estimated in the highest level and the result is used as a covariate for the
next level, which is the number of non-mandatory activity trips. The significance of the
coefficient of the higher-level decision in the model for the lower-level decision estimation
can indicate whether the influence is significant.
Furthermore, Arentze and Timmermans (2013) use the results of an experiment to argue
that the hierarchy of attributes decision associated to an activity episode may depend
on the type of activity. For grocery shopping, for example, being on the way from home
to work or having convenient opening hours is the dominating quality, while for clothes
shopping the specific products available at a certain shop are more important than mode
or time related attributes.
Model Evaluation
Standard goodness-of-fit techniques can be used to evaluate performance of models used
to generate travel and activity pattern, such as the appropriate types of pseudo-R2mea-
sures. McFadden’s R2for discrete choice models (McFadden, 1973) and Cox and Snell’s
R2for hazard models (Cox and Snell, 1989) are two of the most commonly employed.
Given the nature of the output of activity pattern generating models, methods for com-
paring a sequence of discrete choices (more specifically, the generated versus the real) are
appropriate to evaluate results. Such technique, usually referred to as Sequence Align-
ment is used by Wilson (1998) and Vanhulsel et al. (2009) to evaluate discrete attributes
of a generated sequence of activities, such as type of activity or transport mode.
Further research on this area led to the proposal of a multidimensional extension to the
Sequence Alignment method, introduced by Joh et al. (2002) and applied to a case-study
in Joh et al. (2006). In Multidimensional Sequence Alignment, a distance measure is cre-
ated to compare two sets of sequences of discrete choices which span over more than one
dimension. A two-dimensional example would be the simultaneous comparison between
two sequences of activity type and travel mode.
Chapter 3
Methodology
This chapter presents the main methodologies used in the case study with the French
cross-border workers of Luxembourg. In the first section, we give a precise definition of
an activity pattern and list all the attributes it encompasses. In the subsequent sections,
the model used to estimate each attribute present in the activity pattern is explained.
Afterwards, we present the methods used for model evaluation and the technique used
to generate a synthetic population. Description of the data and results are presented in
Chapter 4.
3.1 Defining activity pattern
According to the concepts presented in the previous chapters, an activity pattern is defined
as a sequence of activities undertaken by every individual of the population of interest
over the course of a day. Furthermore, each activity episode in an activity pattern has
a number of attributes which need to be sequentially generated. In this work, every
activity episode undertaken by the individuals of our population of interest (the French
cross-border workers of Luxembourg) has the following four attributes:
•Type (e.g. Home, Work, Shopping, ...)
•Duration
•Travel mode (e.g. car, train, ...)
•Location (for activities with non-fixed location)
Moreover, between two activity episodes there is always a trip. The travel time of each
trip between two activities is also present in every individual’s activity pattern and must
also be generated. The order in which each attribute is generated follows the order of
the list above, in accordance with results and assumptions from Krygsman et al. (2007)
and Arentze and Timmermans (2004). Therefore, for each activity, we first generate the
activity type, followed by the duration of the travel time and of the activity itself. In the
next step, the choice of travel mode is generated for all activities. Finally, the location in
which the activity is conducted is chosen in the case of activities without a fixed location.
Activities related to home or work have a fixed location, which is defined as the munic-
ipality where each individual lives and works, both of which are variables on the survey
10
CHAPTER 3. METHODOLOGY 11
which provides the data used here. For all other activity types, models for location choice
are estimated. A complete list of activity types can be found in the next section.
Example of activity pattern
In Figure 3.1, we can see an example of an activity pattern for one individual. It contains
all the activities this individual undertakes over the course of one day, as well as travel
time, duration, mode and location for each activity episode.
Figure 3.1: Example of activity pattern for one individual
The individual’s first activity of the day is “Home”, which happens from the start of the
day (defined as 00h) until 7:30 in the morning, having therefore a duration of 7h30min.
For this activity, mode choice does not apply. Besides, the home location is defined as
fixed to the municipality of Thionville, so no location choice has to be performed.
After leaving home, this individual undertakes a trip to his second activity of the day,
which is of type “Work”. He travels from home to work for 1h15min and stays at work
for 8h45min, from 8:45 until 17:30. His mode choice is car and his workplace is at the
municipality of Luxembourg City, which is also a fixed location.
When leaving work, the individual from our example undertakes his third activity of
the day, which is of the type “Shopping” and lasts for 35 minutes. To reach the third
activity, the individual travels for 45 minutes and uses a car. This time, the location
is not fixed since “Shopping” is a discretionary activity, and a location choice model has
CHAPTER 3. METHODOLOGY 12
to be used. The generated location for this activity is the municipality of Esch-sur-Alzette.
Finally, the individual reaches his fourth and final activity of the day, which is of the type
“Home”. This activity has unspecified duration since it is the final one. To reach his fixed
home location, he travels for 50 minutes and uses a car.
3.2 Models
In this section, we define the models used to estimate a full daily activity pattern. The
results of estimation using sample data in the case-study with the Luxembourg workers
are presented in Chapter 4. We use a sequence of chained models to generate the at-
tributes for the activity episodes in the order defined in the previous section. In the chain
of models, the results of one model are used as explanatory variables for the following, in
addition to exogenous variables present in the dataset, as in Goulias et al. (1990).
Models for activity type, duration and travel mode are estimated and validated using a
5-fold cross-validation procedure. The dataset containing the activity diary of the French
cross-border workers is randomly split in 5 parts of equal size. For each division, the
models are estimated using 80% of the data (training set) and applied to the remaining
20% (test set), thus generating activity patterns on unknown data. For location choice,
simple validation is performed due to computing reasons, also with 80% used for training
and 20% for test.
3.2.1 Activity type
According to the order of decision-making defined in the previous section, the first at-
tribute of each activity episode in an activity pattern to be generated is the activity type.
This is a discrete choice problem since, for each activity episode, the individual has a
finite number of choices. Furthermore, there is no ordering or ranking in the different
activity types.
Data used in the case study with the French cross-border workers define the following
types of activities:
1. Pick up/drop off someone
2. Home
3. Work-related trip
4. External meal
5. Shopping
6. Services (e.g. go to the doctor, to the bank, ...)
7. Visit to family/friends
8. Take a walk
9. Leisure, sports, culture
CHAPTER 3. METHODOLOGY 13
10. Work
11. End
The first ten attributes from the list above are the alternatives that each individual could
choose in the survey item which asked for the type of each undertaken activity. The last
activity type “End” was artificially added after the pattern of each individual. This type
works as a flag to signal the end of the activity pattern for every individual and if the model
for choice of activity type generates the activity type “End”, the previous activity becomes
the last one. In the example provided in Figure 3.1, the model would in fact generate a
sequence of five activities for the individual: “Home”-“Work”-“Shopping”-“Home”-“End”.
The model used to generate the type of each activity episode is a Multinomial Logistic
Model, defined in Equation 3.1:
logP(Yij =n)
P(Yij =ref )=β0n+
K
X
k=1
βknXk i +αnYi,j−1+nij , j = 2...Ni, Ni≤8 (3.1)
where Yij is the type of the jth activity undertaken by the ith individual, nis each of the
possible activity types defined above and βkn (k= 1...K) and αnare alternative-specific
coefficients to be estimated. Xki(k= 0...K) are K coded survey responses by individual
i, Yi,j−1is the type of the previous activity conducted by the same individual and nij is
an extreme-value distributed error.
This model is estimated for all individuals for every activity episode j. Each individual
ihas a different number of activity episodes Ni. Due to the relatively small number of
activities undertaken by the majority of individuals in the data used in our case-study
(see Table 4.1 in Chapter 4), the model allows up to eight activity episodes per individual,
including “End”. Furthermore, for every individual, the first activity episode is fixed as
being “Home” (thus, jstarts at 2) and, therefore, the coefficients αnare only estimated
for j≥3, since for j= 2 the value of Yi,j−1will be the same (“Home”) for every indi-
vidual. The activity type used as a reference category is “End” for Episodes 3 to 8 and
“Pick-up/drop off someone” for Episode 2.
The explanatory variables used in the model are either at the individual level i(Xki) or
at the activity level j(Yi,j−1). There are no alternative-specific variables with attributes
of the different activity types. Therefore, one set of coefficients [αn,βkn] (k= 0...K) is
estimated for each alternative n.
For each of the five iterations in the cross-validation algorithm, model estimation is per-
formed on the training set and predictions are made using the test set. In each episode,
probabilities are estimated for all activity types and the chosen type is the one with the
highest estimated probability.
CHAPTER 3. METHODOLOGY 14
3.2.2 Activity duration
According to the pre-defined order, the second attribute to be estimated for each activity
episode is duration, which can be subdivided in travel duration and activity duration.
For such models, the response is not a discrete choice, but rather a continuous duration
of a travel or activity episode. According to Bhat and Koppelman (1999), “hazard-based
duration models are ideally suited to modeling duration data”. Therefore, such type of
response can be tackled using a survival analysis approach.
Both travel time and activity duration are modeled in a similar way using semi-parametric
Cox Proportional-Hazards models (Cox, 1992). This approach is largely based on the
one used in Ma et al. (2009). Over the course of a day, individuals alternate between
travel and activity episodes. The transition hazard from one state to the next (from one
travel episode to the next activity episode or from one activity episode to the next travel
episode) is estimated as a function of exogenous explanatory variables, results from the
model for activity type and an unspecified baseline hazard function. In a survival analysis
framework, the transition hazard from state a(current travel or activity) to state b(next
activity or travel) as a function of time tcan be defined as:
λab(t) = lim
h→0
P[t≤T < t +h|T > t]
h(3.2)
where Tis the transition time. The Cox proportional hazards regression model, which is
used to estimate the transition hazards and is presented in further detail for each of the
two durations to be estimated in the next subsections, can be written as a function of
time and explanatory variables Xk:
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXk) (3.3)
λ0ab is the unspecified baseline hazard function. In this model, the estimated coefficients
βkrepresent the relative risk against the baseline hazard. If βkis positive (negative), an
increase in Xkwill result in an increase (decrease) in the transition hazard and, therefore,
a decrease (increase) in the travel/activity duration. Furthermore, the use of the expo-
nential function (the most common in Survival literature) implies that covariates have a
multiplicative effect on the hazard function.
For every iteration in the cross-validation algorithm, after fitting the model using the
training set (80% of the data), we use the model to make predictions for unknown data in
the test set (20% of the data). In this study, the prediction is made by using the median
value of the fitted survival function estimated based on the Cox model. The survival
function can be defined as:
Sab(t) = exp(−ˆ
Λab(t)) (3.4)
Where Λab is the cumulative hazard function, defined as the integral of the transition
hazards from state ato state bfor every pair (a, b).
Λab(t) = Zt
0
λab(x)dx (3.5)
CHAPTER 3. METHODOLOGY 15
Travel time to activity
For each episode, we estimate the transition hazard from current travel episode ato the
next activity episode b.
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXki +α1Yi,b−1+α2Yi,b +α3DurActi,b−1+α4tbia)
(3.6)
In this model, besides the already defined baseline hazard λ0ab and the survey responses
Xk, we also include as explanatory variables the predicted activity type of the previous
(Yi,b−1) and next (Yi,b) activity episodes, the estimated duration of the previous activity
episode (DurActi,b−1) and a variable indicating the time-budget of individual iat the
start of the travel episode a(tbia).
For every episode, the time-budget is defined as 1440 (number of minutes in a day) minus
the total duration in minutes of all travel and activity episodes conducted before the start
of the current episode. Given that the day is assumed to start at 00h, the duration of the
first activity (which is always of the type “Home”) is the same as the time of departure
for the trip to the second activity. Therefore, the time-budget variable at every episode
is a linear function of the time of the day in which that episode starts.
Survey variables to be used are selected taking into account data availability and usual
variables in the literature. Moreover, the inclusion of the predicted activity type makes
the link between the different attributes in the chained generation of activity pattern, in
which the results of the model for one attribute are used as explanatory variables for the
next. This idea is applied in the entire modeling process described in this work.
Furthermore, in order to assess the relevance of using time-related variables as predictors,
the full model for travel time represented by Equation 3.6 can be compared with two
reduced models in which each of the time-related variables present in the full model
is removed. The comparisons are performed through Likelihood Ratio tests. The two
reduced models are:
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXki +α1Yi,b−1+α2Yi,b +α3tbia) (3.7)
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXki +α1Yi,b−1+α2Yi,b +α3DurActi,b−1) (3.8)
In Equation 3.7 the duration of the previous activity episode (DurActi,b−1) is removed.
In Equation 3.8, the time-budget is removed (tbia). The results of the Likelihood Ratio
tests for each travel episode are presented in Chapter 4.
Actual activity duration
The model for activity duration is very similar to the one for travel time. The only
difference is that the roles of episodes aand bare now switched. We use a Cox regression
CHAPTER 3. METHODOLOGY 16
model to estimate the transition hazard from current activity episode ato next travel
episode b. The transition from an activity to the subsequent trip marks the end of the
activity episode and the time interval between the start of the activity and the transition
is the activity duration. The model is estimated according to Equation 3.9.
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXki +α1Yi,a +α2DurT ripi,b−1+α3tbia) (3.9)
Again, λ0ab is the unspecified baseline hazard function and Xkare the individual sur-
vey responses. We also include as explanatory variables the predicted activity type of
the current activity episode Yi,a, the estimated duration of the previous travel episode
DurT r ipi,b−1and a variable indicating the time-budget of individual iat the start of the
activity episode a(tbia).
Similarly to the procedure conducted in the model for travel time, we also compare the full
model for activity duration with two reduced models without the time-related variables.
The comparisons are made through Likelihood Ratio tests and the results are reported in
Chapter 4.
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXki +α1Yi,a +α2tbia) (3.10)
λab(t, X1, ...XK) = λ0ab (t)exp(β0+
K
X
k=1
βkXki +α1Yi,a +α2DurT ripi,b−1) (3.11)
Equation 3.10 presents the reduced model without the duration of the previous travel
episode and Equation 3.11 presents the reduced model without the time budget variable.
3.2.3 Mode choice
After estimating the models for activity type and travel and activity duration, we already
have a full daily schedule for each individual with a sequence of activities, their durations
and the duration of the trips between them.
The next step in the pre-defined order of attributes to be generated is the mode choice.
For this attribute, we are back to discrete choice modeling, since every individual has four
non-ordered alternatives for the mode of each trip:
1. Walking
2. Bicycle/Motorcycle
3. Public transportation (train, bus or both)
4. Car
The mode choice problem is divided in two subproblems: (i) generating the commuting
mode used in trips from home to work and from work to home and (ii) generating the
CHAPTER 3. METHODOLOGY 17
mode for discretionary trips, which are not from home to work or from work to home.
Discrete choice models are estimated using explanatory variables at the individual level
(from the survey data) and the resulting estimates coming from the previous models for
activity type and duration. For the commuting mode, in which locations are fixed (home
and work), two approaches are compared: a Multinomial Logit (similar to the one used
in activity type) and a Conditional Logit, in which an alternative-specific attribute rep-
resenting the theoretical travel time for each mode alternative is added to the individual
specific variables. For the discretionary trips, only a Multinomal Logit is estimated since
the data used in the estimation only contain commuting times for the different alterna-
tives, and not travel time for every possible pair of locations.
In fact, data used for model estimation contain, for each individual, the commuting time
by car, the commuting time by public transportation and the commuting euclidean dis-
tance. To calculate the commuting time by foot and by bike/motorcycle, we divide the
commuting distance, in kilometers, by usual estimates of walking and biking speed. For
walking, the speed is assumed to be 5 km/h and for biking 15 km/h.
Multinomial Logit
The first approach is a standard Multinomial Logit for the choice of mode for each trip.
The model equation is represented by Equation 3.12.
logP(Mij =m)
P(Mij =ref )=β0m+
K
X
k=1
βkmXk i +α1mYij +α2mDepartur eT imeij +mij (3.12)
Here, Mij is the mode choice for the trip to the jth activity undertaken by the ith individ-
ual, with the choice being one of the four previously mentioned alternatives represented
by m.Xkare K survey response variables for individual i,Yij is the predicted type of the
jth activity undertaken by the ith individual and Departur eT imeij is a binary variable in-
dicating whether the departure time of the trip to the jth activity by the ith individual was
between 9h and 16h. The alternative specific terms mij are extreme-value distributed.
The reference category is “Bike/Motorcycle”.
Separate Multinomial Logit models are estimated for commuting mode and for mode of
discretionary trips. Since commuting trips are always either from home to work or from
work to home, the variable Yij representing the activity type is not included in the model
for commuting trips.
For each of the five iterations in the cross-validation algorithm, model estimation is per-
formed on the training set and predictions are made using the test set. Probabilities are
estimated for every mode and the chosen mode is the one with the highest probability.
Conditional Logit
For the discretionary activities, the Multinomial Logit model described above is the only
approach used. For commuting mode, we compare the results obtained with the Multi-
CHAPTER 3. METHODOLOGY 18
nomial Logit model with results obtained with a Conditional Logit model, in which an
alternative specific variable for the travel time with each mode is included.
While individual specific variables have alternative specific coefficients, alternative specific
variables have one general coefficient which is the same for all individuals (except in the
case of a Mixed model, such as the one described in the next section for location choice).
The Conditional Model is presented in Equations 3.13 and 3.14 below, using a notation
similar to Rabe-Hesketh and Skrondal (2008).
P(Mij =m) = exp(Vijm )
P4
c=1 exp(Vijc )(3.13)
Vijm =β0m+β1T ravelT imeij m +
K
X
k=2
βkmXk i +mij (3.14)
Here, Mij is the mode choice for the trip to the jth activity undertaken by the ith individ-
ual, with the choice being one of the four previously mentioned alternatives represented
by mand Xkare K survey response variables for individual i. We now have the variable
T ravelT imeijm , which is the duration of the trip to the jth activity undertaken by the
ith individual using mode m. The alternative specific terms mij are extreme-value dis-
tributed. Moreover, β1is a parameter which is the same for all alternatives, while β0m
and βkm (k= 2...K) are sets of alternative specific parameters, one associated with each
alternative mof travel mode.
This model also uses the training set as estimation data and predicts the mode choice
using the test with a 5-fold Cross-Validation algorithm. The chosen mode is the one with
the highest probability or, equivalently, which the highest value for Vijm , which can be
interpreted as a utility function pertaining to each alternative mto be maximized.
3.2.4 Location choice
After generating activity type, travel time, activity duration and mode choice for every
episode in the population’s activity pattern, the last attribute to be generated is the
location choice for those activities which are not undertaken at a fixed location. According
to what we previously defined in the beginning of this chapter, activities of the type
“Home”, “Work” and “Work-related trip” are considered to have a fixed location. For all
other activities, a choice of location has be to be estimated.
Constraining the choice set
Once more, when choosing the location to undertake a given activity, individuals are
facing a discrete choice problem. The goal of the location choice model is to generate a
choice of municipality where a given activity with non-fixed location will be conducted.
However, the size of the choice set of possible locations for each activity poses an extra
problem. The dataset collected in the Luxembourg Mobility Survey records 916 differ-
ent municipalities where activities with non-fixed locations are conducted by individuals.
Presenting all individuals with such a large choice set may result in a model with very
CHAPTER 3. METHODOLOGY 19
little power and very inaccurate prediction results. Therefore, for every activity we apply
an algorithm to constrain the choice set of available locations which is based on two main
principles: adequacy of the destination and detour factor, both of which will be described
next.
The first principle, adequacy of the destination, is based on the idea that, for every ac-
tivity to be conducted by an individual facing a location choice, only a certain number
of destinations from the universal choice set will have some characteristic which makes
it an appropriate option for that given activity. The choice models, which will be de-
scribed in more detail in the coming sections, are estimated separately for the different
activity types, as in Scott and He (2012). For every activity type (or group of types), the
constrained choice set of adequate destinations contains only those municipalities which
either (i) have characteristics which make it appropriate for that specific activity or (ii)
were recorded in the mobility survey as an observed destination where that type of activity
was conducted. As an example, for activities of the type “External meal”, the destination
choice set will only include destinations which (i) have at least one restaurant or (ii) do
not have any restaurants but were the observed destination of an activity of the type
“External meal” in the collected data.
Figure 3.2: First principle for constraining the choice set: adequacy of destination
Figure 3.2 shows a graphical representation of the principle of adequacy of the destination
using the activity type “External meal” as an example. Each square on the grid repre-
sents one of the municipalities of the universal choice set comprising all municipalities
in the Luxembourg-France cross-border region. Municipalities in green have at least one
restaurant. Municipalities in red have no restaurants and no activities of the type “Ex-
ternal meal” observed in the mobility survey. Municipalities in yellow, despite having no
restaurants, were the destination of at least one activity of the type “External meal” in
the survey. The constrained choice set based on this principle will include all green and
yellow municipalities and will discard the red ones.
In spite of significantly reducing the choice set for every activity type, constraining the
choice set based on the adequacy of destination alone does not solve the problem of ex-
cessively large choice sets. In the example of the activity type “External meal”, there are
still 432 municipalities with at least one restaurant. Therefore, the second principle used
in constraining the choice set, to be described next, needs to be applied.
CHAPTER 3. METHODOLOGY 20
The second principle underlying the methodology to constrain the choice set used in
this work is introduced by Ma and Klein (2017), who perform an initial analysis on
the same dataset used here. The main principle is that individuals are more likely to
conduct discretionary activities in locations around the home-work axis. Therefore, for
every possible pair of individual and location, we define a measure named detour factor
(Justen et al., 2013) to quantify how much the individual would have to go out of his/her
home-work trajectory to reach that location. Such measure can be defined as the ratio
between the travel time between the location and the individual’s home plus the travel
time between the location and the individual’s workplace and the travel time between
home and workplace.
Figure 3.3: Second principle for constraining the choice set: detour factor
df =a+b
c(3.15)
Figure 3.3 and Equation 3.15 show how the detour factor is calculated. The further away
a given location is from the home-work axis, the larger its detour factor will be. If a
location is exactly on the home-work axis, its detour factor will be 1. Individuals are
more likely to choose locations with small detour factors.
In order to define the available choice set for each activity, we fist analyze the empirical
distribution of home-work travel times for all individuals of the considered dataset. For
selected percentiles of this distribution, we can then compute the empirical distribution
of detour factors of chosen locations for discretionary activities. For each individual, we
can constrain the location choice set for discretionary activities to include only those lo-
cations which have a detour factor smaller than the 75th percentile of the distribution
of detour factors for this individual’s home-work travel time (approximately). Table 3.1
shows results from Ma and Klein (2017) with the distributions of detour factors for chosen
discretionary activities for selected percentiles of the observed distribution of home-work
travel time.
As an example, if the home-work travel time for a given individual is 39 minutes, his/her
reduced choice set will only contain locations which have a detour factor smaller than
1.18 (highlighted in Table 3.1), which represents the 75th percentile of detour factors of
chosen activities undertaken by individuals with a similar home-work travel time.
CHAPTER 3. METHODOLOGY 21
Table 3.1: Distribution of detour factors of chosen activities for selected percentiles of the
home-work travel time distribution, taken from Ma and Klein (2017)
Home-work Detour factor distribution for each home-work travel time
time (mins) Perc. 5% 10% 25% 50% 75% 90% 95% Mean Std. Dev.
15 5% 1.13 1.14 1.23 1.42 2.11 5.9 9 2.83 5.54
18 10% 1.11 1.11 1.15 1.17 1.34 1.89 2.17 1.36 0.4
23 25% 1.06 1.08 1.09 1.18 1.27 1.61 1.82 1.29 0.48
29 50% 1.05 1.07 1.1 1.19 1.31 1.5 1.71 1.25 0.3
39 75% 1.03 1.04 1.06 1.13 1.18 1.36 1.47 1.19 0.33
47 90% 1.04 1.05 1.09 1.12 1.15 1.28 1.44 1.15 0.13
52 95% 1.02 1.03 1.05 1.09 1.15 1.22 1.28 1.11 0.1
Using this methodology, we can build an ellipse around the fixed home and work locations,
within which the individuals are more likely to conduct their discretionary activities. Fig-
ure 3.4 illustrates the region containing feasible locations for discretionary activities. Ma
and Klein (2017) report a resulting coverage of 94.5% of the observed chosen locations for
discretionary activity when applying this methodology.
Figure 3.4: Ellipse containing feasible locations for discretionary acitivities
The analysis presented in Ma and Klein (2017) shows that the constrained choice set re-
sulting from this methodology is quite often still very large. Therefore, like the principle
of adequacy of destination, the detour factor alone may not suffice to reduce destination
choice sets to feasible sizes. In order to further constrain the choice set of feasible des-
tinations, therefore, we can combine both of the abovementioned principles. For every
activity of every individual, the set of possible destinations contains those which (i) fulfill
the conditions defined in the adequacy principle and (ii) are located within the ellipse
defined by the rule of the 75th percentile detour factor.
Figure 3.5 shows a graphical illustration of the final choice set of feasible locations. The
possible alternatives for a given activity will be the green and yellow squares located inside
the ellipse representing the 75th percentile detour factor.
After restricting the choice set, we compare the performance of different discrete choice
models as well as a combination of the constraining methods with random sampling of
alternatives. These methodologies will be described next.
CHAPTER 3. METHODOLOGY 22
Figure 3.5: Combining the two constraining methods results in a signifcantly smaller
choice set
Choice models
Unlike the previous discrete choice problems described in this work, in which individual
specific variables are available, for destination choice we use only alternative specific vari-
ables which represent attributes of the municipalities which are the alternatives in the
choice sets. The use of individual (or trip) specific variables (as in the models for activity
type and mode choice) implies alternative-specific coefficients. Even with the constrained
choice set, this would require the estimation of one coefficient per available alternative
(for any individual) per variable, which would result in an extremely large number of
parameters, beyond the limitations imposed by our data availability, thus compromising
power and prediction accuracy.
Therefore, we start by estimating Conditional Logit models with alternative specific vari-
ables to estimate location choice. This first approach is compared with a second approach
using a random effects framework in which one of the estimated parameters is treated as
a random variable distributed over the entire population. This will be performed through
the estimation of Mixed Logit models (Rabe-Hesketh and Skrondal, 2008), which are
the most general case of discrete choice models of which all other models are a special case.
Ujd =
K
X
k=1
βjk Xjdk +jd (3.16)
Equation 3.16 describes the utility function to be maximized of destination dfor trip jas
a function of alternative-specific explanatory variables Xjdk. In the more general case of
a Mixed Logit model, the set of kcoefficients βjk is a random vector normally distributed
over the population of trips. In the special case of a Conditional Logit with no random
parameters, each of the k coefficients can be written as βk. We can then construct the
likelihood function of the coefficients βby writing:
Ljd (β) = exp(PK
k=1 βjk Xjdk)
Pl∈Djexp(PK
k=1 βlkXldk)(3.17)
where Djis the alternative choice set for each trip j. For the constraining method in
which only adequacy of destination is used to reduce the number of alternatives, all trips
face the same choice set. When the detour factor is incorporated in the constraining
CHAPTER 3. METHODOLOGY 23
method, each individual faces a different choice set. The probability of choosing each
alternative is then:
Pjd =ZLj d(β)f(β|θ)dθ (3.18)
where f(β|θ) is a specified distribution for the set of random coefficients βwith parameters
θ. The log-likelihood function is represented by 3.19 and must be solved through iterative
Monte-Carlo techniques since it does not have a closed form.
ljd (θ) =
J
X
j=1
log(Pj d) (3.19)
In our case study, as mentioned earlier, we compare results from a series of Conditional
Logit models, with no random parameters, with results from a series of Mixed Logit
models. For the Mixed Logit models, a random parameter is defined for the variable
T ravelT ime, which measures, for every destination in the choice set, the travel time by
car in the off-peak hours from the individual’s current destination. The utility function
of the two models are represented, respectively, by Equations 3.20 and 3.21 below.
Ujd =
K
X
k=1
βkXjdk +j d (3.20)
Ujd =βj1T ravelT imej d +
K
X
k=2
βkXjdk +j d
βj1∼N(µ, σ2) (3.21)
Moreover, in our location choice models, the activity types are grouped into four cate-
gories, defined as follows: (i) External meal, (ii) Shopping, (iii) Services and Leisure and
(iv) Others, which includes the types “Pick-up/drop-off”, “Visit” and “Walk”. Activi-
ties of the type “Work-related trip” are also assumed to have a fixed location, which is
estimated as being the same as the work location. The variables used in the definition
of adequacy of destination and as explanatory variables are defined for each category in
Chapter 4.
Sampling of alternatives
In the previous sections, we described a method to constrain an initially large choice set
to a more reasonable number of alternatives based on adequacy of destination and on
the detour factor. When dealing with large choice sets, the most common constraining
method in literature is sampling of alternatives, as in Azaiez (2010), Guevara and Ben-
Akiva (2013) and Scott and He (2012).
In our case study, we combine the constraining methods described in the previous sections
with sampling of alternative techniques and evaluate the fit and the predictive performance
of different constraining methods. For every activity category as defined in the previous
section, the following models are estimated:
•Constraining based on adequacy of destination
CHAPTER 3. METHODOLOGY 24
–Conditional Logit
–Conditional Logit with 5 sampled alternatives
–Conditional Logit with 10 sampled alternatives
–Conditional Logit with 20 sampled alternatives
–Conditional Logit with 30 sampled alternatives
–Conditional Logit with 50 sampled alternatives
–Mixed Logit
–Mixed Logit with 5 sampled alternatives
–Mixed Logit with 10 sampled alternatives
–Mixed Logit with 20 sampled alternatives
–Mixed Logit with 30 sampled alternatives
–Mixed Logit with 50 sampled alternatives
•Constraining based on both adequacy of destination and detour factor
–Conditional Logit
–Conditional Logit with 5 sampled alternatives
–Conditional Logit with 10 sampled alternatives
–Conditional Logit with 20 sampled alternatives
–Conditional Logit with 30 sampled alternatives
–Conditional Logit with 50 sampled alternatives
–Mixed Logit
–Mixed Logit with 5 sampled alternatives
–Mixed Logit with 10 sampled alternatives
–Mixed Logit with 20 sampled alternatives
–Mixed Logit with 30 sampled alternatives
–Mixed Logit with 50 sampled alternatives
For the constraining method based on adequacy of destination alone, model estimation
is performed by selecting the observed chosen alternative plus D−1 (D = 5,10,20,30,50)
randomly sampled alternatives from the feasible locations. For the constraining method
which also includes the detour factor, a similar procedure is performed, with D−1 lo-
cations being sampled from the constrained choice set. However, when using both con-
straining methods it is often the case that certain individuals have a constrained set with
a number of alternatives which is smaller than D (especially for large values of D, such
as 30 and 50). In these cases, the entire constrained choice set is selected and the re-
maining alternatives are sampled from the ones which are out of the ellipse defined by
the detour factor, but in the choice set of feasible locations based adequacy of destination.
Unlike the previous models, which are evaluated with a 5-fold cross-validation approach,
the destination choice models are subject to simple validation due to the elevated compu-
tational time necessary to estimate Mixed models with large choice sets. For each activity
CHAPTER 3. METHODOLOGY 25
category, the observed data is divided into one group with 80% of data used for model
estimation (training set) and the remaining 20% used to evaluate model prediction when
facing unknown data (test set). For the test set, D alternatives are randomly sampled
instead of the D−1 used for model estimation, and the observed chosen alternative is not
necessarily in the constrained set.
The results, which are presented in Chapter 4, allow us to compare the impact of different
sizes of choice set in the prediction performance of discrete choice models, as well as to
assess the validity of the idea behind the detour factor approach, which hypothesizes that
most discretionary activities are conducted around the axis formed by home and work
fixed locations.
3.2.5 Model Evaluation
After estimating the models for activity type, travel and activity duration, mode choice
and location choice, we have a chain of models which is able to estimate a full daily ac-
tivity pattern containing the four specified attributes.
For the Cox regressions used to predict travel and activity duration, we use Cox and
Snell’s pseudo R2to evaluate goodness-of-fit. Evaluation of the discrete choice models
is performed using McFadden’s pseudo R2measure for goodness-of-fit as well as the Se-
quence Alignment methodology, mentioned in the previous chapter and defined next, and
a confusion matrix on the test dataset which allows us to calculate which percentage of
choices were correctly predicted.
Sequence Alignment
In the Sequence Alignment Method, the distance between two sequences of discrete choices
is defined as the sum of the “costs” of operations necessary to transform the source se-
quence into the target sequence. These operations are: deleting an element, adding an
element and replacing an element. Usually, deletion and addition are defined as having
cost 1 and replacement, which can be decomposed as a deletion followed by an addition,
has cost 2. A more thorough definition can be found in Joh et al. (2002). The distance
can be visually calculated with the aid of a table such as the one shown in Figure 3.6.
Here we present an example with sequences of letters, in which the source sequence is
“BEAD” and the target sequence to be matched is “ABCDE”.
In Figure 3.6, a horizontal line arriving at a given cell represents the addition of the tar-
get element at the top of the cell’s column to the source. A vertical line arriving at a
given cell represents the deletion of the source element at the start of the cell’s row. A
diagonal line indicates that the two elements in the positions corresponding to the cell in
which the arrow arrives are identical, so no change in that particular position is necessary.
The distance between the two sequences is defined as the smallest number of horizontal
and vertical arrows which are necessary to achieve the bottom-right cell in the table. In
this example, the distance between the two sequences is 5. The shortest path is the one
represented in the picture, containing 3 horizontal arrows and 2 vertical arrows.
CHAPTER 3. METHODOLOGY 26
Figure 3.6: Example of distance between two sequences
Each arrow in Figure 3.6 is also one of the steps of the algorithm described in Table 3.2,
which presents the sequence of operations performed to transform the source sequence
BEAD into the target sequence ABCDE. Horizontal and vertical arrows, which repre-
sent addition or deletion of an element, have cost 1 and diagonal arrows, which represent
equality, have cost 0. The total cost incurred to transform the source into the target is
the distance between the two sequences, which in this case is equal to 5. An alternative
representation would be to replace one deletion followed by one addition (or one addition
followed by one deletion, as in steps 3 and 4) by a replacement with cost 2. Here, we
decompose the replacement into two simpler operations with cost 1 each.
Table 3.2: Steps to align source and target sequences
Step Action Source Target Cost Total cost
0 Start BEAD ABCDE - -
1 Add “A” at the start of source ABEAD ABCDE 1 1
2 Nothing, second element is equal ABEAD ABCDE 0 1
3 Add “C” in position 3 of source ABCEAD ABCDE 1 2
4 Remove “E” from position 4 of source ABCAD ABCDE 1 3
5 Remove “A” from position 4 of source ABCD ABCDE 1 4
6 Nothing, fourth element is equal ABCD ABCDE 0 4
7 Add “E” to position 5 of source ABCDE ABCDE 1 5
8 End
Benchmark
The results of predictions generated by the discrete choice models for type and mode
proposed in this work are also compared with predictions generated by a Benchmark al-
gorithm in which the choice of response (type or mode) is generated by drawing from an
empirical distribution. In the algorithm, individuals are divided into categories according
to certain socio-economic variables (e.g. age group, income range, education level) and,
for each individual, an activity pattern is drawn from the empirical distribution of activity
patterns for that category. The Benchmark implementation is based on Barth´elemy et al.
CHAPTER 3. METHODOLOGY 27
(2013).
Three Benchmarks were created using, respectively, 4, 5 and 6 socio-economic attributes
to group the individuals based on Barth´elemy et al. (2013). For all models, the results
show that the Benchmark using 6 attributes performs better than the ones using 4 and 5.
Therefore, only the results of the Benchmark with 6 attributes are presented as Benchmark
results in the rest of this work. The method used to group the individuals is based on
socio-economic attributes defined in Table 3.3 below, with the first 4 and 5 attributes
being the ones used, respectively, on the Benchmarks with 4 and 5 attributes.
Table 3.3: Categories used to divide individuals in Benchmark
Attribute Categories
1 Gender Male, Female
2 Age Up to 40 years old, 40-59 years old, 60 years old or older
3 Employment situation One job, two or more jobs, in formation
4 Education level Primary, secondary, higher
5 Children in household Children, no children
6 Monthly income Up to 2000 €, 2001 to 3000 €, 3001 to 4000 €,
4001 to 6000 €, 6001 to 8000 €, more than 8000 €
A total of 2 ×3×3×3×2×6 = 648 categories of individuals are possible. In the
Benchmark method, for every activity episode, activity type and mode are drawn from
the observed distribution for the category to which the given individual belongs. The idea
is to compare the predictive performance of the generated models for activity type and
mode with the Benchmark. The Benchmark is also subject to a 5-fold cross-validation
method, in which the individuals in the training set are used to generate the empirical
distribution from which the types and modes of activities undertaken by individuals in
the test set are drawn.
3.3 Synthetic population and pattern generation
After estimating and validating the models for all attributes which compose a full activity
pattern using data from a sample of French cross-border workers in Luxembourg, the next
step in the analysis is the generation of a synthetic population which represents the entire
population of workers and the application of the final models from the previous section
to generate a full travel and activity pattern.
3.3.1 Synthetic population
The need to generate a synthetic population for microdata is imposed by two practical
constraints which arise when one works with survey data: costs and privacy (M¨uller and
Axhausen, 2011).
Costwise, it is generally unfeasible to obtain data at the individual level for the entire
population of a country or a region. Survey data covering a sample of individuals allows
CHAPTER 3. METHODOLOGY 28
for inference pertaining to the entirety of the population. However, if one needs to gen-
erate and analyze data in the scale of the entire population, it is necessary to generate a
synthetic population which will represent all the individuals of interest.
In terms of privacy, the need to generate a synthetic population arises when one needs
to expose data which, for legal reasons or because it deals with sensitive topics, cannot
be disclosed. Therefore, the original dataset is replaced by a synthetic one with the same
structure, but not containing real information and, therefore, possible to be disclosed.
In this study, both constraints are present. The survey contains information about ac-
tivities undertaken by 2320 French cross-border workers in Luxembourg. However, the
goal of the study is to analyze the activity pattern of the whole population of approxi-
mately 86000 French cross-border workers. Moreover, data used to estimate and validate
the models comes from an official survey undertaken by a governmental organism which,
therefore, cannot be publicly disclosed.
Survey variables
The first step in the generation of a synthetic population is to generate the survey re-
sponses, which are used as explanatory variables in the models.
The method used to generate the population from survey data is the one proposed by
Nowok et al. (2015) and implemented in the R package “synthpop”. The method generates
each variable jby fitting a predictive model using the previously estimated and the other
observed attributes as explanatory variables, according to Equation 3.22.
˜
Xij =f(˜
Xik, Xil ), k < j < l (3.22)
˜
Xrepresents generated variables for the synthetic population and Xrepresented observed
variables in the survey data. The first variable is generated using observed values. The
second variable is generated used the generated values of the first variable and the ob-
served values of the rest. The third variable is generated using the generated values of
the first and second variables and the observed values of all the others, and so on until
all variables from the survey are generated.
The resulting population has a mean and covariance structure equivalent to the original
data. The advantage of this method over other implemented packages is that not only
the marginal distribution of each variable is respected, but also the joint distribution of
every pair of variables is mimed, resulting in a similar covariance matrix.
Home and work locations
For the destination choice models, described previously in this chapter, it is also necessary
to know the fixed home and work location of every individual. This information is present
in the data for the individuals who participated in the survey, but must be generated for
the individuals of the synthetic population. The initial plan was to include the home and
work locations as survey variables and use the predictive algorithm “synthpop” to also
CHAPTER 3. METHODOLOGY 29
generate home-work pairs of locations. However, the excessive number of municipalities
present in the dataset results in a categorical variable with too many levels to be correctly
predicted, causing computational problems.
Therefore, the home and work locations of each individual in the synthetic population
are generated in a separate step. Given that the models for type, duration and mode use
only survey variables and the models for destination choice, which use home and work
locations, do not use survey variables, the separation of the two steps does not pose any
problem for the models.
For the individuals in the sample, the empirical distribution of pairs of home and work
location is computed. For every individual in the generated synthetic population, a pair
of home and work locations is drawn from the empirical distribution.
It is important to highlight that, when considering the empirical distribution, the pair
of home and work locations is treated as a single attribute rather than two separate at-
tributes (home and work) to construct a joint empirical distribution with the survey data
and draw from it to generate the synthetic population. As an example, if an observed
individual lives in the municipality of Thionville and works in Luxembourg City, the pair
“Thionville - Luxembourg City” is computed as one occurrence in the distribution, and
not as one occurrence for Thionville as home location and one occurrence for Luxembourg
City as work location.
3.3.2 Pattern generation
After generating a synthetic dataset which represents the entire population of 86000
French cross-border workers in Luxembourg, the next step in the analysis is to apply
the models estimated and validated using the survey data to the generated population.
For activity type, travel and activity duration and mode choice, in which a 5-fold cross-
validation method is applied, the models are re-estimated using the entire survey data
(2320 individuals). The resulting coefficients are applied to the synthetic population to
generate sequences of activity type, travel and activity duration and mode choice. For
location choice, in which a simple 80%-20% validation method is used, the same coeffi-
cients from the estimation are used to generate locations for the activities conducted by
the individuals in the synthetic population.
The output of this step is a full travel and activity pattern for the entire synthetic popula-
tion of French cross-border workers. Every generated individual has a sequence of activity
types, travel and activity durations and mode choices. Furthermore, every discretionary
activity has a generated destination, while home and work activities have a fixed location.
Chapter 4
Case Study
In this chapter, we present the results of the implementation of the methodology described
in Chapter 3 to the dataset pertaining to the Luxembourg cross-border workers to gen-
erate a full individual travel and activity pattern for the population of French workers.
We start by describing the dataset used in this work with descriptive statistics and a list
of all variables and their definitions and, in the subsequent sections, present the results
obtained when applying the models for activity type, travel and activity duration, mode
choice and location choice.
4.1 Description of the data
Two datasets are used in our case study. The first (and most important) one, used in the
models for activity type, duration and mode, comes from a Mobility Survey conducted
by the Luxembourg Institute of Socio-Economic Research in 2011 with a sample from the
population of cross-border workers in Luxembourg. In the survey, respondents were asked
several questions regarding socio-economic, work-related and household-related character-
istics. Furthermore, they were asked to describe every trip they undertook in their most
recent working day and provide informations such as what was the goal of the trip (ac-
tivity type), the departure and arrival time, the mode choice and the municipality. The
original dataset contains individuals who work in Luxembourg and live in all three bor-
dering countries (Belgium, France and Germany). This work deals only with the French
cross-border workers, thus, only those who reported to live in France are used for the
analysis. In total, 2320 individuals are considered.
For location choice, the last attribute in the activity pattern generation, a second dataset
is used in combination with the Mobility Survey. This dataset, also collected by the
Luxembourg Institute of Socio-Economic Research, presents characteristics related to the
population (density) and presence of different types of amenities (supermarkets, restau-
rants, services) of the 916 municipalities in the cross-border region between Luxembourg
and France. Those are the alternative-specific variables used for constraining the desti-
nation choice set and included in the discrete choice models for location choice.
30
CHAPTER 4. CASE STUDY 31
Descriptive statistics of observed activity patterns
In this subsection, we present some descriptive statistics for the observed activities in
the Mobility Survey dataset used in the case-study with the French cross-border work-
ers. Tables 4.1, 4.2, 4.3 and 4.4 summarize the observed activity types, average activity
duration and average trip duration per type and observed travel mode for every activity
episode. From the tables for activity type and travel mode, we can see that the number
of individuals undertaking activities decreases sharply after Episode 3. Due to the small
amount of individuals undertaking a large number of activities, our models generate a
maximum of 8 activity episodes for every individual.
Table 4.1: Observed type of each activity episode
Episode 1 2 3 4 5 6 7 8 9 10
Pick-up 0 341 113 161 63 55 25 17 2 0
Home 2320 68 1366 318 447 163 121 48 27 11
Work-trip 0 22 118 100 67 42 21 10 5 2
Meal 0 1 212 67 17 11 5 1 2 0
Shopping 0 1 85 58 36 24 6 1 3 0
Services 0 7 32 26 12 9 2 1 1 0
Visit 0 5 10 16 11 4 4 2 0 0
Walk 0 2 11 8 2 6 3 2 0 1
Leisure 0 2 14 19 8 14 4 3 0 1
Work 0 1846 341 266 106 28 11 7 1 0
Total 2320 2310 2302 1039 769 356 202 92 41 15
Table 4.2: Average observed duration for each activity episode by type
Episode 1 2 3 4 5 6 7 8 9 Mean
Pick-up - 16.9 17.3 10.6 13.5 8.0 32.2 10.1 - 13.6
Home 423.9 568.3 106.4 122.4 81.1 72.9 126.1 75.3 175.0 194.6
Work-trip - 368.1 132.0 154.2 107.3 92.3 90.6 85.0 86.3 139.5
Meal - 5.0 75.3 72.2 65.9 69.9 75.0 - 178.0 77.3
Shopping - 172.6 56.6 62.1 32.6 47.0 48.3 5.0 11.7 54.5
Services - 70.0 29.4 69.4 25.3 24.7 51.5 15.0 60.0 43.2
Visit - 161.3 98.5 84.7 81.0 101.3 170.3 85.0 - 111.7
Walk - 27.5 165.0 263.3 55.0 0.0 25.0 - - 89.3
Leisure - 197.5 110.1 129.3 86.3 72.9 136.3 90.0 - 117.5
Work - 466.5 431.6 274.9 267.0 266.0 213.5 208.8 - 304.0
Mean 423.9 205.4 122.2 124.3 81.5 75.5 96.9 71.8 85.2
CHAPTER 4. CASE STUDY 32
Table 4.3: Average observed duration for the travel to activity episode by type
Episode 2 3 4 5 6 7 8 9 10 Mean
Pick-up 12.8 37.4 31.0 33.9 32.3 25.0 18.2 25.0 - 26.9
Home 36.5 56.8 38.1 39.2 34.0 29.4 25.1 18.1 43.8 35.7
Work-trip 61.3 42.2 24.5 22.7 19.2 25.9 64.5 70.0 - 41.3
Meal 22.5 12.4 11.7 20.2 13.0 12.5 45.0 22.0 - 19.9
Shopping 34.8 26.7 17.9 23.5 17.7 15.0 20.0 26.7 - 22.8
Services 30.8 20.4 20.8 25.9 8.7 6.0 50.0 15.0 - 22.2
Visit 41.0 45.5 13.5 22.8 28.8 10.0 20.0 15.0 - 24.6
Walk 18.3 23.5 35.0 38.3 24.0 7.5 47.5 - 30.0 28.0
Leisure 57.5 33.2 13.3 33.0 11.1 11.7 11.3 15.0 - 23.2
Work 58.0 48.6 17.3 17.3 18.0 13.5 32.5 - - 29.3
Mean 37.4 34.7 22.3 27.7 20.7 15.6 33.4 25.8 36.9
Table 4.4: Observed mode for each activity episode
Episode 2 3 4 5 6 7 8 9 10
Bike/Moto 296 230 225 89 57 8 2 0 0
Car 1809 1697 1703 833 636 318 180 83 37
Public Transp 121 121 109 33 17 3 1 2 1
Walk 18 167 173 81 52 25 15 6 3
Total 2244 2215 2210 1036 762 354 198 91 41
Variables
Mobility survey
The following variables are present in the Mobility Survey dataset and, thus, available to
be selected as explanatory variables for the models.
•Work hours - number of weekly working hours (continuous)
•Type prof - type of profession
1. Directors, higher rank personnel
2. Intellectual and scientific personnel
3. Armed forces, police, firefighters
4. Technical and intermediate rank personnel
5. Administrative personnel
6. Sales personnel, other services
7. Agriculture-related personnel
8. Industry and manufacture
9. Machine operators
10. Manual labor and non-qualified personnel
•Work define - who is responsible for defining the working hours
CHAPTER 4. CASE STUDY 33
1. the worker him/herself
2. the employer
3. joint decision by the worker and the employer
•Transp subs - whether there is some work-related subsidy for transportation
1. No subsidy
2. Availability of a company car
3. The employer pays transportation fees (partially or totally)
•Gender - self-declared gender of the respondent
1. Male
2. Female
•Age - age at the time of the survey (continuous)
•Age cat - age at the time of the survey, categorized
1. <30
2. 30-39
3. 40-49
4. 50-59
5. 60+
•Marital - respondent’s marital status
1. Single
2. Married
3. Divorced
4. Widow
•Education - respondent’s highest education level achieved
1. Primary
2. Secondary
(a) General/Technical
(b) Professional
3. Bachelor level
(a) General Technical
(b) Professional
4. Short higher education
5. Long higher education
•Income - monthly income range of respondent’s household
1. Up to 2000 €
CHAPTER 4. CASE STUDY 34
2. 2001 to 3000 €
3. 3001 to 4000 €
4. 4001 to 6000 €
5. 6001 to 8000 €
6. More than 8000 €
•Children - whether or not there are children living in the household (binary)
•N cars - Number of cars in the household
1. 1 car
2. 2 cars
3. 3 cars
4. 4 cars
5. 5 or more cars
6. No cars
•Location type - Type of the respondent’s home location, according to official gov-
ernment definition
1. Hamlet
2. Village
3. Town
4. City
5. Large city
•Home occ - Status of the respondent regarding home ownership
1. Owner
2. Tenant
3. Free lodging
•Home parking - How easy it is to find parking spaces close to the respondent’s home
1. Easy
2. Difficult
3. Impossible
•DepartureTime - Starting time of each trip
1. between 9h and 16h
2. before 9h or after 16h
CHAPTER 4. CASE STUDY 35
Attributes of municipalities
The following variables are present in the dataset containing the attributes of the mu-
nicipalities and, thus, available to be selected as explanatory variables as well as for
constraining the choice sets in the location choice models.
•Supermarket - Number of supermarkets in each municipality
•Restaurants - Number of restaurants in each municipality
•Service - A score created to quantify the attractiveness of each municipality in terms
of services availabiltiy, which considers the number of amenities such as banks, post
offices and hospitals.
•Density - Demographic density of each municipality, defined as the population di-
vided by the area in km2
Furthermore, a full matrix containing the average travel-time by car in off-peak hours
between every pair of municipalities is available as a complement to this dataset.
4.2 Model Results
In this section, we present the main results of the models described in Chapter 3 applied
to the dataset containing information about the French cross-border workers, using the
variables described in the previous section. A slightly different set of explanatory variables
is selected for each model. Different sets of explanatory variables were tested in each model
and the ones reported here are those considered to yield the best performance.
4.2.1 Activity type
The first model in the chain of models used to generate a full daily activity pattern is the
discrete choice model which estimates a sequence of activity types. The selected explana-
tory variables used in this model are Work hours,Type prof,Transp subs,Gender,Age,
Marital,Education,Location type,Home occ,N cars,Income and Children.
The last column in Table 4.5 shows that almost 80% of the observed individuals have
“Work” as a second activity of the day (after the initial activity “Home”) and a little more
than half of the individuals in the dataset have a “Home-Work-Home” pattern, undertak-
ing no discretionary activity. The strong imbalance in the second activity episode makes
the discrete choice model over-estimate the predominant type and, as a consequence, pre-
dict an exaggerate number of “Home-Work-Home” patterns, affecting the entire chain of
models for activity types.
In order to tackle this issue, we compare the sequence of models fitted with the original
dataset with no changes and the sequence of models in which the model for the type
of second activity episode was fitted with a semi-balanced dataset constructed using an
oversampling technique, as described and implemented by Lunardon et al. (2014) in the
R package ROSE. Using this technique, we add new trips to the dataset for the second
CHAPTER 4. CASE STUDY 36
activity by randomly sampling activities who were not of the type “Work” until the pro-
portion of observed activity types in the second episode is 50% “Work” and 50% of other
types. The resulting percentages of “Work” as type of second activity and “Home-Work-
Home” patterns for both sequence of models are shown in Table 4.5 below.
Table 4.5: Comparing the effects of imbalance in the data
Model (Unbalanced) Model (Balanced) Observed
% Work in Episode 2 96.20% 68.65% 79.60%
% Home-Work-Home 91.76% 64.82% 51.42%
The results of the previously described model evaluation techniques for activity type are
presented in Table 4.6, in which we compare the two model sequences (unbalanced and bal-
anced for second activity type) and the Benchmark. For the McFadden’s pseudo-R2and
for the percentage of correctly predicted episodes in the test data, the numbers presented
in the table are the average value of the five iterations performed in the cross-validation
algorithm. For the Sequence Distance, the generated pattern for every individual is com-
pared with the observed sequence of activities and the reported number is an average
value for all 2320 individuals.
Table 4.6: Results for activity type
Model (Unbalanced) Model (Balanced) Benchmark
Episode R2% Correct R2% Correct % Correct
2 0.269 77.3% 0.344 65.8% 66.73%
3 0.429 56.6% 0.429 51.6% 41.58%
4 0.552 51.3% 0.552 46.4% 36.51%
5 0.619 29.3% 0.619 28.9% 27.12%
6 0.704 37.3% 0.704 34.3% 33.79%
7 0.984 35.0% 0.984 20.9% 30.53%
8 0.993 46.7% 0.993 44.0% 16.67%
Seq. Distance 3.068 3.251 3.364
Both model sequences outperform the Benchmark by having mostly larger percentage
of correctly predicted activity types. Furthermore, the unbalanced model apparently
performs better than the balanced model by predicting more activity types correctly.
However, this high number, especially in the initial episodes, is mostly due to overesti-
mating the “Home-Work-Home” activity pattern. For severely unbalanced data, it is very
often the case that the predominant class is overestimated which, despite having good fit
indexes and prediction performance, yields unrealistic results.
Furthermore, we can see that the McFadden’s pseudo-R2for all activity episodes is satis-
factory, indicating a good model fit. The R2of the models for activity episodes 3 to 8 are
the same since the only difference in model construction is the balancing of the data for
activity episode 2. For episode 2, the two values are hardly comparable because they are
fitted with different datasets. By definition, McFadden’s pseudo-R2compares the fitted
CHAPTER 4. CASE STUDY 37
model with a null model, in which only an intercept is included. On data with severe
imbalance, a model with only the intercept does not perform terribly, since guessing the
most common category (which is, effectively, what a null model does) will fit the data in
a decent way. Therefore, we can expect that models which use datasets in which there is
not much variability, which is the case especially in the unbalanced approach, will have
low values of McFadden’s pseudo R2.
Finally, the last row of Table 4.6 shows that both models have a lower value for aver-
age Sequence Distance than the Benchmark, calculated using the Sequence Alignment
Method. Thus, we can conclude that the sequences of activity types generated by them
are closer to the observed sequences than those generated by the Benchmark, which indi-
cates a better performance. Once more, the unbalanced model has a better score due to
overpredicting the “Home-Work-Home” sequence. Given that around half the individuals
have this sequence in the observed data and the model predicts that almost everyone will
do nothing more than go to work and come back home, close to half of the individual
sequences have a distance of zero between the observed pattern and the pattern generated
by the unbalanced model.
All in all, the final chosen model to generate the activity pattern is the balanced one.
Despite having mostly worse performance indexes, when generating a pattern which aims
to imitate the behavior of people, it is preferable and more realistic to have variability
rather than to assume that the most common behavior will be repeated by almost the
entirety of the population.
Table 4.7 presents a frequency table for all generated activity episodes using the final
model. When comparing the generated with the observed frequency table, (Table 4.1),
we can see that the model tends to somewhat underestimate the number of activities
undertaken by each individual.
Table 4.7: Generated type of each activity episode
Episode 1 2 3 4 5 6 7 8
Pick-up 0 509 25 169 4 28 17 11
Home 2320 21 1270 261 208 75 38 39
Work-trip 0 14 16 33 22 14 8 1
Meal 0 1 8 87 5 10 3 1
Shopping 0 20 0 14 5 3 7 2
Services 0 4 3 9 1 6 3 1
Visit 0 4 7 5 11 0 3 1
Walk 0 4 4 0 0 1 5 4
Leisure 0 3 6 1 3 1 4 6
Work 0 1283 522 15 90 15 8 2
Total 2320 1863 1861 594 349 153 96 68
For every individual present in the survey, the first activity episode was assumed to be
of the type “Home”. For estimation of the subsequent episodes, individuals with missing
data for the explanatory variables were removed, thus the smaller number of individuals
CHAPTER 4. CASE STUDY 38
in Episode 2. Furthermore, in accordance to the observed data, the number of individuals
is smaller for the later activity episodes, with a particularly sharp decrease after Episode
3, mostly due to the predominance of “Home-Work-Home” pattens.
4.2.2 Travel time and activity duration
After generating the type of each activity episode, the Cox Proportional Hazards models
described in Chapter 3 are estimated to predict the duration of each activity episode and
the travel time to each activity. Table 4.8 presents the mean values for the travel time
to every activity episode by activity type and Table 4.9 presents the mean values for the
duration of every generated activity episode by type. Table 4.10 shows the values of Cox
and Snell’s pseudo-R2for each travel time and activity duration for each episode.
The survey variables used as explanatory for both travel time and activity duration models
were Work hours,Type prof,Transp subs,Gender,Age,Marital,Education,Location type,
Home occ,N cars,Income and Children.
Table 4.8: Average generated duration for the trip to each activity episode by type
Episode 2 3 4 5 6 7 Mean
Pick-up 13.6 22.0 25.4 14.0 54.6 22.4 25.3
Home 33.8 53.3 40.8 17.0 52.6 25.4 37.1
Work-trip 70.0 45.3 28.2 26.0 31.4 34.4 39.2
Meal 5.0 19.4 15.0 19.6 18.5 - 15.5
Shopping 41.8 - 28.1 8.0 16.7 29.1 24.7
Services 41.3 15.0 24.1 25.0 15.0 11.0 21.9
Visit 58.0 41.1 9.0 18.6 - 30.0 31.4
Walk 11.3 26.3 - - 10.0 10.2 14.4
Leisure 51.7 42.2 5.0 36.7 10.0 15.0 26.8
Work 55.7 42.3 14.7 16.9 32.1 14.4 29.4
Mean 38.2 34.1 21.1 20.2 26.8 21.3
Table 4.9: Average generated duration for each activity episode by type
Episode 1 2 3 4 5 6 7 Mean
Pick-up - 5.4 16.2 6.0 7.0 5.0 6.5 7.7
Home 416.1 624.7 52.3 97.7 42.4 51.9 104.1 198.4
Work-trip - 459.3 40.7 73.0 28.1 51.5 49.3 117.0
Meal - 1.0 29.8 30.4 6.0 22.9 - 18.0
Shopping - 204.4 - 18.1 10.8 10.7 69.3 62.7
Services - 13.3 21.7 40.6 45.0 16.3 5.0 23.6
Visit - 211.8 87.7 38.8 60.5 - 30.0 85.7
Walk - 9.5 117.5 - - - 63.8 63.6
Leisure - 143.3 104.2 20.0 38.3 42.0 103.3 75.2
Work - 520.8 468.6 196.7 135.6 178.2 153.8 275.6
Mean 416.1 219.3 104.3 57.9 41.5 47.3 65.0
CHAPTER 4. CASE STUDY 39
Table 4.10: Cox and Snell’s pseudo-R2for travel and activity duration
Episode 1 2 3 4 5 6 7
Travel time 0.398 0.381 0.314 0.452 0.506 0.628
Activity duration 0.071 0.515 0.606 0.670 0.649 0.633 0.626
Overall, the values of Cox and Snell’s pseudo-R2for all travel and activity episodes are
satisfactory, which indicates a good model fit. The value for the activity duration of the
first episode, which, given that the first activity is always of the type “Home” and that
the day is assumed to start at 00h, is equivalent to the time at which the individual leaves
home, is somewhat lower than the others. The explanation for this difference is most likely
to be similar to the one presented in the activity type subsection: little variability in the
data leads to a not so large difference between the estimated model and a hypothetical
null model.
Furthermore, as described in Section 3.2.2, we also compare the full models estimated for
travel time and activity duration with two reduced models without the variable related to
duration of previous travel/activity episode and without the time-budget variable. The
comparisons are performed using Likelihood Ratio tests and the resulting p-values can be
found in Table 4.11.
Table 4.11: P-values for Likelihood Ratio tests comparing full models with reduced models
for travel time and activity duration
Episode 2 3 4 5 6 7
Travel time
vs. no previous duration 0.0441 0.0011 <0.0001 0.0764 0.0228 0.0611
vs. no time budget <0.0001 0.8788 0.0719 0.0085 0.2624
Activity duration
vs. no previous duration 0.0183 0.2470 0.0347 0.0299 0.0299 0.3004
vs. no time budget 0.0023 <0.0001 <0.0001 <0.0001 0.1115 <0.0001
The p-values from the Likelihood Ratio tests in Table 4.11 show that, in most of the
cases, both time-related variables are significant. Therefore, both variables are relevant
for predicting the duration of travel time and activity episodes and are kept as explana-
tory variables in the final model.
4.2.3 Mode Choice
After estimating the models for activity type and travel and activity duration, we now
have a full daily schedule for the individuals of the sample. The next step is to estimate
the travel mode choice for the trip to each scheduled activity. The method is the one
presented in the Mode Choice section in Chapter 3. As described, two types of discrete
choice model are applied for commuting: Multinomial Logit and Conditional Logit, with
the difference between them being the inclusion of an alternative-specific variable for each
CHAPTER 4. CASE STUDY 40
trip relating to the travel time between the location of the previous activity and the lo-
cations of next activity. For discretionary activities, only a Multinomial Logit is applied
due to constraints in the dataset.
The selected individual specific survey explanatory variables used in all models are chil-
dren,Home parking,Gender,Transp subs,Location type,N Cars and DepartureTime.
The model for discretionary activities, as mentioned in the model description section,
also includes the activity type as explanatory variable. For the test set, the activity
type used as explanatory variable is the one predicted by the previously estimated model
for activity type. Moreover, for predicting the mode choice in the test set, the variable
DepartureTime is computed by adding the predicted durations of all activities and trips
undertaken by the individual before the trip for which the mode will be estimated. Given
that the day starts at 00h, the duration of the first activity (“Home”) is equivalent to the
time at which the individual home in the morning and, therefore, adding this value to the
duration of all other estimated activities and trips previously undertaken results in the
departure time of the trip.
Furthermore, when estimating the commuting mode choice, we face a similar problem as
when estimating the activity type of the second activity episode: strong imbalance in the
data. A large majority of the French-cross border workers in Luxembourg commute to
work by car. From the survey respondents, this proportion is around 80%. Therefore,
a model without any correction for imbalance will overestimate this proportion to an
extreme extent. To tackle this problem, we use the same approach as in the model for
activity type (Lunardon et al., 2014). We perform oversampling of the data by randomly
selecting individuals who do not commute by car until the proportion of commuting mode
becomes 50% by car and 50% by all other three modes combined (Bike/Motorcycle, Walk-
ing and Public Transportation).
Table 4.12: Resulting performance measures for different approachs in mode choice mod-
eling
Conditional Conditional Multinomial Multinomial Benchmark
Balanced Unbalanced Balanced Unbalanced
Correct pred. Ep. 2 71.7% 78.0% 73.2% 77.9% 65.6%
Correct pred. Ep. 3 67.1% 73.3% 67.9% 73.1% 61.7%
Correct pred. Ep. 4 68.3% 69.7% 68.2% 69.1% 58.8%
Correct pred. Ep. 5 78.6% 80.1% 78.2% 79.7% 63.6%
Correct pred. Ep. 6 67.4% 63.8% 62.4% 62.4% 71.4%
Correct pred. Ep. 7 84.3% 84.3% 91.0% 84.3% 66.7%
R2- Commute 0.131 0.110 0.110 0.090 -
R2- Discretionary 0.180 0.180 0.180 0.180 -
Av. Seq. Distance 3.116 3.033 3.098 3.033 3.395
% Car for commute 71.6% 96.1% 75.6% 97.8% -
Table 4.12 shows, for the Conditional and Multinomial approaches (both using balancing
techniques and not using them) as well as for the Benchmark, the percentage of correctly
CHAPTER 4. CASE STUDY 41
predicted modes for all generated activity episodes in the test dataset, taking into ac-
count all five Cross-validation iterations (thus, the entire test dataset was predicted, 20%
in each iteration). Furthermore, we also present the result of Sequence Distance measure
and the McFadden’s pseudo-R2for the models for commuting mode and for mode for
discretionary activities.
When looking at Table 4.12, we can see that all models outperform the Benchmark, both
in terms of correctly predicted mode (except for activity episode 6) and in terms of Se-
quence Distance. The models in which balancing is used in the dataset for commuting
mode have a somewhat worse performance than their counterparts in both Conditional
and Multinomial approaches in the percentage of correctly predicted modes. Moreover,
despite the unbalanced having worse value of McFadden’s pseudo-R2than the balanced,
they cannot be directly comparable for the same reason which was discussed in the sec-
tion for activity type: the models are fitted with different datasets. For discretionary
activities, as previously mentioned, a Multinomial approach without balancing is used in
all cases, thus the same value for McFadden’s pseudo-R2.
Moreover, the Multinomial and Conditional approaches have very similar performances
in the prediction, with the Conditional having a slightly better model fit. Taking all the
measures into account, as well as the previously discussed issue of the importance of vari-
ability in the generation of activity patterns, the chosen model to be used in the synthetic
population is the Conditional Balanced.
Table 4.13 presents a frequency table for the generated mode choice for each activity
episode. This table is the counterpart of the observed frequency table, shown in Table
4.4.
Table 4.13: Generated mode for each activity episode
Episode 2 3 4 5 6 7
Bike/Moto 283 281 6 6 3 4
Car 1308 1306 537 347 135 74
Public Transp 6 6 0 0 0 0
Walk 6 14 42 8 8 4
Total 1603 1607 585 361 146 82
As in the results for activity type, we can see that the number of individuals in each
episode decreases after episode 3 due to the predominance of “Home-Work-Home” ac-
tivity pattern. Individuals with missing data for any of the covariates used to generate
mode choice are excluded. Moreover, we can see that, when comparing the generated
modes with the observed ones (Table 4.4), the model correctly predicts the preference
for car trips with bike being the second most used mode. However, walking and public
transportation trips are somewhat underrepresented in comparison to the observed data.
CHAPTER 4. CASE STUDY 42
4.2.4 Location choice
The final model to be estimated and validated in the sequence of models to generate indi-
viduals’ daily travel and activity pattern is the one used to predict the municipality where
each predicted activity episode without a fixed location will occur. As we have mentioned
in the previous chapter when defining the models, activities of the types “Home”, “Work”
and “Work-Related Trip” are assumed to take place in fixed locations. All the other ac-
tivity types, which are discretionary activities, are subject to destination choice models
to define in which municipality of the France-Luxembourg cross-border region they will
take place.
Unlike the other three models, the models for location choice are not subject to a 5-fold
cross validation, but rather to a single validation procedure, in which 80% of the data is
used for training and the remaining 20% for test. This is due to the much larger number
of models to be fitted, half of which being Mixed Logit models, which take a much longer
computing time to be estimated, especially with large choice sets
As defined in Chapter 3, discretionary activities are divided in four categories, with the
models listed in the subsection “Sampling of Alternatives” being estimated for all of them
separately. The first principle for constraining the choice set is adequacy of destination,
in which the municipalities present in the choice sets of all individuals who undertake
activities of each of the four categories are selected based on (i) having an attribute which
makes it appropriate for that activity or (ii) having at least one activity of that category
observed in the data. Below, we list the attributes used in the constraining principle of
each activity type.
•External meal:
(i) has at least one restaurant
(ii) one or more activities of type “External meal” observed in the data
•Shopping:
(i) has at least one supermarket
(ii) one or more activities of type “Shopping” observed in the data
•Services and Leisure:
(i) has a score for Services of at least 1.0
(ii) one or more activities of type “Services” or “Leisure” observed in the data
•Others:
(i) one or more activities of type “Pick-up/drop-off”, “Visit” or “Walk” observed
in the data
The explanatory variables used to estimate the model are a combination of the variables of
the second dataset described in the subsection “Variables - attributes of municipalities” in
this chapter and new variables created from individual characteristics and the previously
predicted activity durations, listed next:
CHAPTER 4. CASE STUDY 43
•TravelTime - Travel time by car in the off-peak hours between the location of the
individual’s previous activity and every candidate location, used as a proxy for
distance.
•TimeNextFixed - Travel time by car in the off-peak hours between every candidate
location and the location of the individual’s next activity of type “Home” or “Work”.
•AvailTime - Available time for the activity, defined as: Reported duration of activity
+ Duration of previous recorded trip + Duration of next recorded trip - (TravelTime
+TimeNextFixed). It may be the case that the resulting value for AvailTime is
negative. In those cases, it is set to zero.
Furthermore, the models for the different activity categories were estimated using dif-
ferent explanatory variables, as shown in Table 4.14 below. Different combinations of
explanatory variables were tested for the different models, including others which are not
mentioned here, only those with good predictive and computing performance are reported.
As a general rule, the attributes listed above plus the variable used in the adequacy of
destination constraining method were included. For the activity category “Others”, in
which no obvious destination-specific attribute can be selected to define adequacy, the
demographic density of each municipality was used as a proxy for general attractiveness
and included in the model.
Table 4.14: Explanatory variables marked with ×were included in the models for the
different activity categories
Variable External meal Shopping Service and Leisure Others
Supermarkets ×
Restaurants ×
Service ×
Density ×
TravelTime × × × ×
TimeNextFixed × × × ×
AvailTime × × ×
Table 4.15 provides descriptive statistics regarding the sizes of the choice sets faced by
the individuals for each activity type. The original choice set has 916 municipalities. We
can clearly see that applying the described methods to constrain the choice sets results
in a very substantial decrease in size.
Table 4.15: Size of choice sets for the different activity categories
Adeq. of destination Adeq. of destination + 75% Detour factor
Size Mean Std.Dev. P5 P50 P95
External meal 432 24.51 16.92 6 9 57
Shopping 229 14.02 8.37 4 12 30
Serv. and Leisure 327 18.07 11.62 6.85 16.5 36.3
Others 178 17.72 11.41 4 15 38
CHAPTER 4. CASE STUDY 44
Tables 4.16 to 4.19 show the performances of the compared models for each of the four
types of activity category measured in terms of McFadden’s pseudo-R2and percentage
of correctly predicted locations in the test set. We compare the Mixed and Conditional
approaches as well as the two constraining methods and the different sampling sizes when
using sampling of alternatives.
Table 4.16: Location choice results for activity category “External meal”
Adeq. of destination Adeq. of destination + 75% Detour factor
Conditonal Mixed Conditonal Mixed
R2% Correct R2% Correct R2% Correct R2% Correct
No sampling 0.721 49.21% 0.721 49.21% 0.614 42.86% 0.614 42.86%
5 alt 0.924 3.17% 0.925 3.17% 0.738 0.00% 0.751 0.00%
10 alt 0.898 3.17% 0.911 3.17% 0.720 7.94% 0.730 7.94%
20 alt 0.870 3.17% 0.879 3.17% 0.692 22.22% 0.693 22.22%
30 alt 0.858 4.76% 0.864 4.76% 0.695 26.98% 0.695 26.98%
50 alt 0.840 7.94% 0.847 7.94% 0.685 41.27% 0.685 41.27%
Table 4.17: Location choice results for activity category “Shopping”
Adeq. of destination Adeq. of destination + 75% Detour factor
Conditonal Mixed Conditonal Mixed
R2% Correct R2% Correct R2% Correct R2% Correct
No sampling 0.459 28.89% 0.482 24.44% 0.182 28.89% 0.209 28.89%
5 alt 0.743 4.44% 0.745 4.44% 0.235 0.00% 0.269 2.22%
10 alt 0.681 11.11% 0.694 8.89% 0.272 17.78% 0.293 17.78%
20 alt 0.625 15.56% 0.643 15.56% 0.315 22.22% 0.331 22.22%
30 alt 0.583 13.33% 0.603 13.33% 0.361 26.67% 0.376 26.67%
50 alt 0.550 11.11% 0.573 13.33% 0.410 26.67% 0.424 24.44%
Table 4.18: Location choice results for activity category “Services and Leisure”
Adeq. of destination Adeq. of destination + 75% Detour factor
Conditonal Mixed Conditonal Mixed
R2% Correct R2% Correct R2% Correct R2% Correct
No sampling 0.387 32.26% 0.405 29.03% 0.092 25.81% 0.146 19.35%
5 alt 0.695 3.23% 0.770 3.23% 0.092 0.00% 0.181 0.00%
10 alt 0.672 3.23% 0.750 3.23% 0.096 3.23% 0.168 6.45%
20 alt 0.599 3.23% 0.657 3.23% 0.178 25.81% 0.211 25.81%
30 alt 0.567 6.45% 0.608 6.45% 0.228 25.81% 0.247 25.81%
50 alt 0.526 12.90% 0.579 9.68% 0.290 25.81% 0.307 29.03%
A few conclusions can be drawn from the results presented in the tables. Firstly, for the
models without sampling, constraining using only adequacy of destination tends to yield
a substantially better model fit and an equal or slightly better predictive performance
CHAPTER 4. CASE STUDY 45
Table 4.19: Location choice results for activity category “Others”
Adeq. of destination Adeq. of destination + 75% Detour factor
Conditonal Mixed Conditonal Mixed
R2% Correct R2% Correct R2% Correct R2% Correct
No sampling 0.328 24.18% 0.396 31.32% 0.033 18.68% 0.162 24.73%
5 alt 0.611 1.10% 0.643 1.10% 0.043 2.20% 0.140 1.10%
10 alt 0.547 3.85% 0.588 3.30% 0.064 4.95% 0.150 7.69%
20 alt 0.497 6.59% 0.546 6.04% 0.134 15.38% 0.203 20.9%
30 alt 0.464 7.14% 0.519 6.59% 0.182 21.43% 0.246 28.02%
50 alt 0.430 10.99% 0.488 10.99% 0.248 24.73% 0.311 29.67%
than constraining using both methods. Moreover, it is never the case that sampling al-
ternatives provides a better predictive performance than using the full constrained set.
Besides, using a Mixed Logit instead of a Conditional Logit can yield somewhat better
model fits (McFadden’s R2), but does not drastically improve predictive performance.
Furthermore, meaningful insights can also be drawn when comparing the results of sam-
pling from the set constrained only by adequacy of destination (larger set) with the results
of sampling from the set constrained by both adequacy of destination and detour factor
principles (smaller set). McFadden’s R2is always much higher when the model uses sam-
ples from the larger set than from the smaller one. Besides, it tends to decrease with the
number of alternatives sampled in the larger choice set and increase in the smaller one.
However, the predictive performance is very poor when sampling from the larger choice
set and increases only slightly with a larger number of alternatives, while the predictive
performance in the smaller choice set is much better and increases substantially with a
larger sample.
The reason for these results is the fact that, when fitting the model with the training set,
the chosen alternative is necessarily selected with D-1 (D being the size of the choice set)
other alternatives being sampled, while in the test set, in which we use new data, this
is not the case and D alternatives are randomly selected since the chosen alternative is
assumed to be unknown. When we sample from the larger choice set, the alternatives
are randomly selected without taking into account any individual characteristics which
make alternatives more or less plausible for that particular trip. In this case, the model
is presented with the chosen alternative and D-1 (D = 5,10,20,30,50) randomly selected
alternatives, which may well be very poor choices with extremely low utility functions.
The model, then, has an “easy” task in choosing between the chosen location, which will
probably have a high utility, and a very poor set of other alternatives, resulting in a high
fit. When testing on unknown data, on the other hand, the sampling of alternatives is
done completely at random, which means that, given the large size of the choice set, the
chance that the chosen alternative will even be selected is small. Therefore, samples taken
from the larger choice set result in a good model fit, but very poor predictive performance.
Models which sample from the smaller set, on the other hand, present a very different
behavior. The values for McFadden’s R2are not as high as the ones using the larger
set because the sample of alternatives contains more plausible choices for each particular
CHAPTER 4. CASE STUDY 46
trip, since sampling is made from a choice set which contains only alternatives around
the home-work axis. The model, therefore, has a somewhat “harder” task separating
the chosen alternative from those which were not chosen. When testing with new data,
however, the predictive performance is much higher than when sampling from the larger
set. The reason is the fact that, in this case, the choice set is much smaller, which gives
the chosen alternative a much higher chance of being sampled.
This comparison brings evidence in favor of the principle upon which the detour factor
method is based, which states that individuals tend to conduct discretionary activities
around the home-work axis.
The chosen model to be applied to the synthetic population is the one which provides
the highest prediction performance for each category of activities. In cases in which two
models are tied in yielding the best predictive performance, the one with the highest
McFadden’s pseudo-R2is chosen. If a tie persists, Conditional Logits are preferred over
Mixed Logits and the smaller choice set (adequacy of destination and detour factor)
is preferred over the larger (adequacy of destination only). The chosen models for each
activity category are listed next. Their coefficient estimates are presented in the Appendix
(Table A.37).
•External meal - Conditional Logit using adequacy of destination only.
•Shopping - Conditonal Logit using adequacy of destination only
•Services and Leisure - Conditonal Logit using adequacy of destination only
•Others - Mixed Logit using adequacy of destination only
4.2.5 Overview of selected models
Table 4.20: List of models selected for each attribute
Attribute Selected model Criterion
Activity type Multinomial logit -
Travel time Cox proportional-hazards, full model Likelihood Ratio tests
Activity duration Cox proportional-hazards, full model Likelihood Ratio tests
Mode - Commuting Conditional logit after balancing Better fit and prediction
Mode - Discretionary Multinomial logit -
Location - External meal Conditional logit + adequacy of destination Better fit and prediction
Location - Shopping Conditional logit + adequacy of destination Better fit and prediction
Location - Services and Leisure Conditional logit + adequacy of destination Better fit and prediction
Location - Others Mixed logit + adequacy of destination Better fit and prediction
Table 4.20 presents an overview of the final selected models to generate each attribute of
the final activity pattern and the criterion used for model choice. The final models are
applied to the synthetic population and used to generate a full activity pattern to the
French cross-border workers of Luxembourg.
CHAPTER 4. CASE STUDY 47
4.3 Synthetic population and activity pattern gener-
ation
After choosing the final models used to generate each attribute in the generation of a
full daily travel and activity pattern using sample data, we generate a synthetic popu-
lation using the methodology described in Chapter 3 (Nowok et al., 2015) to apply the
final models and obtain a full pattern for the French cross-border workers in Luxembourg.
From the 2320 individuals, we select only the ones which have no missing information
in the survey responses to use as a base for the synthetic population generation, a total
of 1560 individuals. The number of individuals generated in the synthetic population is
86000, which is the approximate number of French cross-border workers in Luxembourg
according to the official government statistics (STATEC, 2015).
In order to validate the generated population, we can compare the covariance matrix of
the observed survey population with that of the generated synthetic population through
Box’s M test for equivalence of covariance matrices (Box, 1949), which yields a p-value
of almost 1 (χ2statistic of 49.17 with 136 degrees of freedom), meaning that we cannot
reject the hypothesis of equality between the survey and the synthetic covariance matrices.
The final models are then applied to the synthetic population to generate a full daily travel
and activity pattern. For activity type, travel and activity duration and mode choice, in
which model estimation was performed using a 5-fold Cross-validation method, the cho-
sen models are re-estimated using the entire sample and then applied to the synthetic
population. For location choice, in which model estimation was performed using simple
validation, the final models are the same which were estimated using the 80% training
dataset randomly selected in the estimation phase. The parameter estimates of all final
models are presented in the Appendix.
4.4 Visualization
The final activity pattern contains a total of 299758 activities generated for the 86000
individuals from the synthetic population representing the entirety of the French cross-
border workers in Luxembourg. In this section, we visualize the activities’ geographic
distribution as well as the trips between the different geographic areas.
Figure 4.1 shows the distribution of activities by municipality in the border region of
Luxembourg and France. We can clearly see what appears to be a high level of concen-
tration of activities in a few areas. The main point of attraction is Luxembourg City, the
darkest brown area in the south-central part of Luxembourg, where the highest number
of activities is concentrated. Three other municipalities come next, with less activities
than the capital, but still attracting a large number of individuals from the surrounding
regions. They are, from North to South: Esch-sur-Alzette (brown area on the southwest
of Luxembourg, right on the border with France), Thionville (northernmost brown area
in France, a few kilometers from the border) and Metz (one of the southernmost munici-
CHAPTER 4. CASE STUDY 48
Figure 4.1: Location of activities in the France-Luxembourg border region
palities on the map, somewhat further away from the border).
Figure 4.2 shows the same map, but with the activities divided by mode used in the trip
to reach them. We can see that the same concentration dynamics described above appears
to hold, to a certain extent, for every mode. This is especially true for the modes “Car”
and “Public Transportation”, with the four municipalities previously mentioned acting as
attraction poles. The good connections via bus, trains and highways and the large work
and entertainment offer make those cities specially attractive via modes which are usually
taken for longer trips. For “Bike” and “Walk”, although the concentration also exists, it
appears to be somewhat less extreme, most likely due to the shorter distances of the trips
usually undertaken using these modes.
Figure 4.3 presents a graph where the edges are formed by connecting the origins and
destinations of every generated trip. The graph was built using the application NodeXL
(Smith et al., 2010), originally designed to visualize social network connections and also
applied to transportation visualization in Chow and Sayarshad (2014). Due to the ex-
cessive number of municipalities, they were grouped into geographical zones, which are
the vertices of the graph. The list of geographical zones can be found in the Appendix
(Table B.1). The thickness of the connection is proportional to the number of trips made
between the zones. We can see that connections between almost every pair of zones
are present, as well as within-zone trips (represented by a circle). The zone number
114, located in the left side of the graph, is the one with the thickest lines, having the
stronger connections with the rest. Zone 114 is the area around Luxembourg City which
is, as we have mentioned before, the main attraction pole of activities in the border region.
CHAPTER 4. CASE STUDY 49
Figure 4.2: Location of activities by travel mode in the France-Luxembourg border region
Figure 4.4 shows the same graph, but with trips divided by travel mode. We can see that
the connection structure from the first graph (Figure 4.3) is almost replicated in the graph
for trips by car, on the upper right corner. The graph with trips by bike shows somewhat
less connections than the car, with thick lines in a few connections and not so many thin
lines, indicating that most trips are concentrated around some preferred biking routes in
the cross-border region. From the graph for public transportation, there appears to be
a group of areas with worse services in the bottom part, with a much smaller number of
connections. Besides, we see no circles representing trips within the same zone, meaning
that public transportation tends to be used for longer-distance trips. For walking trips, on
the other hand, we can see that there are not so many connections between the different
zones, with more circles on the edges representing trips within the same zone, which is
logical given that walking trips tend to be of shorter distance, thus less likely to change
geographical zones. For better visualization, the thickness of the lines was defined using
different scales for each graph and cannot be directly compared.
CHAPTER 4. CASE STUDY 50
Figure 4.3: Graphical representation of all trips in the France-Luxembourg cross-border
region
Figure 4.4: Graphical representation of the trips in the France-Luxembourg cross-border
region separated by mode
Chapter 5
Conclusion
In this study, we develop, estimate and validate predictive models to generate daily travel
and activity patterns. We define the attributes pertaining to an activity pattern (namely
type, duration, mode and location) and use different statistical and econometric tech-
niques to model, estimate and validate each one of them. We apply the results to a
case-study in which we analyze the mobility behavior of the French cross-border workers
in Luxembourg and generate a full activity pattern for a synthetic population representing
the 86000 workers of our population of interest.
Overall, the results obtained seem to be acceptable with good evaluation indexes which
account for goodness-of-fit and predictive performance. The predictive patterns generated
by chaining the models for each attribute, in which results of the previous model are used
as inputs for the next, appear to appropriately resemble the observed behavior in terms of
its diversity of activities and geographical distribution. As seen in the chapters “Method-
ology” and “Case Study”, however, it was often the case that further adjustments had to
be made to the observed data in order to generate better patterns, such as the balancing
technique applied to the models for type and mode.
Furthermore, in this study we focus on the use of the estimated models to predict and
generate valid activity patterns rather than on the explanatory power of each model. A
different approach for the same problem and case study, which can possibly be the subject
of future works, would be to analyze and discuss the estimates of each model, interpreting
their magnitude and significance in order to gain more insight on the drivers and causes
of mobility behavior of our population of interest.
The main output of this study can be used as input for several models which predict,
evaluate and improve public policies in the field of mobility in Luxembourg. Moreover,
the chain of models used to generate an activity pattern can be used in simulations in
which different potential inputs are submitted to models to see how exogenous changes
could affect mobility behavior.
Future research
Several topics approached in this work can be the subject of further research in the field
of prediction of mobility behavior. Regarding the general methodological framework, as
51
CHAPTER 5. CONCLUSION 52
mentioned in the chapter “Literature Review”, there is a duality in methodologies for
activity pattern generation between econometric and machine learning techniques. Here,
all approaches used to generate the activity pattern are on the econometric side. Further
research can be undertaken by performing the same case study using machine learning
methods for the attributes of the activity patterns instead of econometric ones, and com-
paring results and conclusions.
Moreover, as described in the chapter “Literature Review”, even within the econometric
framework, there is a large variety of modeling approaches which can be applied to the
type of problems tackled in this work. For the discrete choice problems, for example, lim-
itations of our data made us choose to use a Multinomial approach on the model used to
generate activity type, with no alternative specific characteristics, while for mode choice
and location choice, those were incorporated in the modeling. In a different case study,
the Conditional approach can also be applied to model activity type. Another example
would be the role played by random effects in improving model fit and prediction. Here,
we only introduced them for location choice models, in which Mixed models were com-
pared with Conditional Logit models. It is also possible to investigate their performance
when compared to the fixed-effects only models for activity type and for travel mode.
Moreover, the balancing method used in the datasets for activity type and travel mode
can also be the subject of further study, with a more in depth analysis of the effects of
different balancing techniques in the generated patterns.
A particularly fertile field for further research among the ones investigated in this work
is the estimation of discrete choice models with large choice sets, such as the one used
here to estimate location choice. The constraining methodology presented here has sev-
eral possibilities still to be explored, such as, for example, the incorporation of estimated
travel time to construct a reachable region for every trip, constraining feasible locations
to those within the region. Another possibility is to conduct a sensitivity analysis on the
75% detour factor rule to investigate how choosing a different percentile affects model fit
and prediction performances.
Finally, this entire study was built on the foundations of a survey on mobility behavior
undertaken in our geographic region of interest. In order to further validate and gen-
eralize the conclusions and outputs of the study, applying the sequence of models used
here to similarly shaped data from other geographic regions would be a valuable future
step. Furthermore, when estimating behavioral models such as the ones tackled here, the
larger and more diverse the data about the subjects of interest, the better. Therefore,
further data collections, surveys and interviews which can enrich the database used in this
study are likely to contribute to better models and to a more accurate pattern generation.
Bibliography
Allahviranloo, M. and W. Recker
2013. Daily activity pattern recognition by using support vector machines with multiple
classes. Transportation Research Part B: Methodological, 58:16–43.
Arentze, T. and H. Timmermans
2013. Travel demand modelling: conceptual developments and perspectives. Trans-
portation Letters.
Arentze, T. A., D. Ettema, and H. J. Timmermans
2011. Estimating a model of dynamic activity generation based on one-day observations:
method and results. Transportation Research Part B: Methodological, 45(2):447–460.
Arentze, T. A. and H. J. Timmermans
2004. A learning-based transportation oriented simulation system. Transportation
Research Part B: Methodological, 38(7):613–633.
Azaiez, I.
2010. Sampling of alternatives for logit mixture models. Unpublished masters thesis.
Transport and Mobility Laboratory, ´
Ecole Polytechnique F´ed´erale de Lausanne, Switzer-
land.
Barth´elemy, J., E. Cornelis, L. Hollaert, and P. L. Toint
2013. Virtualbelgium: a simulation platform for the belgian population and an appli-
cation to mobility.
Bhat, C. R.
1996. A generalized multiple durations proportional hazard model with an applica-
tion to activity behavior during the evening work-to-home commute. Transportation
Research Part B: Methodological, 30(6):465–480.
Bhat, C. R.
2000. A multi-level cross-classified model for discrete response variables. Transportation
Research Part B: Methodological, 34(7):567–582.
Bhat, C. R. and F. S. Koppelman
1993. A conceptual framework of individual activity program generation. Transporta-
tion Research Part A: Policy and Practice, 27(6):433–446.
Bhat, C. R. and F. S. Koppelman
1999. Activity-based modeling of travel demand. In Handbook of transportation Science,
Pp. 35–61. Springer.
53
BIBLIOGRAPHY 54
Box, G. E.
1949. A general distribution theory for a class of likelihood criteria. Biometrika,
36(3/4):317–346.
Castiglione, J., M. Bradley, and J. Gliebe
2015. Activity-Based Travel Demand Models: A Primer, number SHRP 2 Report S2-
C46-RR-1.
Chow, J. Y. and H. R. Sayarshad
2014. Symbiotic network design strategies in the presence of coexisting transportation
networks. Transportation Research Part B: Methodological, 62:13–34.
Cox, D. R.
1992. Regression models and life-tables. In Breakthroughs in statistics, Pp. 527–541.
Springer.
Cox, D. R. and E. J. Snell
1989. Analysis of binary data, volume 32. CRC Press.
Goulias, K. G., R. M. Pendyala, and R. Kitamura
1990. Practical method for the estimation of trip generation and trip chaining. Uni-
versity of California Transportation Center.
Guevara, C. A. and M. E. Ben-Akiva
2013. Sampling of alternatives in logit mixture models. Transportation Research Part
B: Methodological, 58:185–198.
Joh, C.-H., T. Arentze, F. Hofman, and H. Timmermans
2002. Activity pattern similarity: a multidimensional sequence alignment method.
Transportation Research Part B: Methodological, 36(5):385–403.
Joh, C.-H., H. Timmermans, and T. Arentze
2006. Measuring and predicting adaptation behavior in multidimensional activity-travel
patterns. Transportmetrica, 2(2):153–173.
Justen, A., F. J. Mart´ınez, and C. E. Cort´es
2013. The use of space–time constraints for the selection of discretionary activity
locations. Journal of Transport Geography, 33:146–152.
Krygsman, S., T. Arentze, and H. Timmermans
2007. Capturing tour mode and activity choice interdependencies: A co-evolutionary
logit modelling approach. Transportation Research Part A: Policy and Practice,
41(10):913–933.
Lunardon, N., G. Menardi, and N. Torelli
2014. Rose: a package for binary imbalanced learning. R Journal, 6(1):79–89.
Ma, T.-Y., J. Y. Chow, and J. Xu
2016. Causal structure learning for travel mode choice using structural restrictions and
model averaging algorithm. Transportmetrica A: Transport Science, Pp. 1–27.
BIBLIOGRAPHY 55
Ma, T.-Y. and S. Klein
2017. Bayesian networks for constrained location choice modeling using structural
restrictions and model averaging. Luxembourg Institute of Socio-Economic Research
working paper.
Ma, T.-Y., G. L. Mariante, and V. Van Acker
2017. Location choice modeling based on mixed logit model and sampling of alterna-
tives. Conference paper submitted to BIVEC-GIBET Transport Research Days 2017
at the HEC - Li`ege.
Ma, T.-Y., C. Raux, E. Cornelis, and I. Joly
2009. Multistate nonhomogeneous semi-markov model of daily activity type, timing,
and duration sequence. Transportation Research Record: Journal of the Transportation
Research Board, (2134):123–134.
McFadden, D.
1973. Conditional logit analysis of qualitative choice behavior.
McFadden, D.
1974. The measurement of urban travel demand. Journal of public economics, 3(4):303–
328.
M¨uller, K. and K. W. Axhausen
2011. Hierarchical ipf: Generating a synthetic population for switzerland.
Nowok, B., G. M. Raab, and C. Dibben
2015. synthpop: Bespoke creation of synthetic data in r. Journal of Statistical Software.
Primerano, F., M. A. Taylor, L. Pitaksringkarn, and P. Tisato
2008. Defining and understanding trip chaining behaviour. Transportation, 35(1):55–72.
Rabe-Hesketh, S. and A. Skrondal
2008. Multilevel and longitudinal modeling using Stata. STATA press.
Rasouli, S. and H. Timmermans
2014. Activity-based models of travel demand: promises, progress and prospects. In-
ternational Journal of Urban Sciences, 18(1):31–60.
Schmitz, F., G. Drevon, and P. Gerber
2012. La mobilit´e des frontaliers du luxembourg: dynamiques et perspectives.
Scott, D. M. and S. Y. He
2012. Modeling constrained destination choice for shopping: a gis-based, time-
geographic approach. Journal of Transport Geography, 23:60–71.
Shiftan, Y.
1998. Practical approach to model trip chaining. Transportation Research Record:
Journal of the Transportation Research Board, (1645):17–23.
BIBLIOGRAPHY 56
Smith, M., C. A., N. Milic-Frayling, B. Shneiderman, E. Mendes Rodrigues, J. Leskovec,
and C. Dunne
2010. Nodexl: a free and open network overview, discovery and exploration add-in for
excel 2007/2010/2013/2016. http://nodexl.codeplex.com/ from the Social Media
Research Foundation, http://www.smrfoundation.org.
STATEC
2015. Travailleurs frontaliers occup´es au luxembourg selon la residance et la na-
tionalit´e (en 1000 personnes) 1974-2015. data retrieved from http://http://www.
statistiques.public.lu.
Vanhulsel, M., D. Janssens, G. Wets, and K. Vanhoof
2009. Simulation of sequential data: An enhanced reinforcement learning approach.
Expert Systems with Applications, 36(4):8032–8039.
Wilson, W. C.
1998. Activity pattern analysis by means of sequence-alignment methods. Environment
and Planning A, 30(6):1017–1038.
The World Bank
2015. The world bank open data. data retrieved from http://data.worldbank.org.
Appendix A
Parameter estimates for the final
models
In this Appendix, we present the coefficient estimates for the final models applied to
the synthetic population to generate the final daily activity pattern. Tables A.1 to A.21
present the results of the models for activity type. Tables A.22 to A.28 present the re-
sults of the models for activity duration. Tables A.29 to A.34 present the results of the
models for travel time to activity. Tables A.35 and A.36 present, respectively, the re-
sults of the model for commuting mode and mode for discretionary activities. Table A.37
presents the results of the models for location choice applied to each category of activities.
The categories of the explanatory variables are numbered according to the list presented
in Chapter 4. The chosen models and explanatory variables are also the ones mentioned
in Chapter 4. For activity type, duration and travel time, it is often the case that, for the
further activity episodes, limitations of the data forced the exclusion of certain variables
due to an excessively large number of categories and to the fact that the majority of the
observed individuals tends to undertake no more than four activities.
Furthermore, in the models for mode choice, two variables were recoded. The variable
Location type became a binary variable equal to 1 if the location is of the type “Town”,
“City” or “Large city” and 0 otherwise. The variable 1 was recoded as follows: 1 if the
household has 0 or 1 cars, 2 if the household has 2 cars and 3 if the household has 4 or
more cars.
57
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 58
Table A.1: Activity type - Episode 2 (1/3)
Dependent variable:
234567891011
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Work hours −0.019 0.244∗∗∗
−0.114∗∗∗ 0.021 0.006 0.096∗∗ 0.076∗∗∗ 0.042 0.035∗∗∗ 0.118∗∗
(0.015) (0.074) (0.032) (0.018) (0.026) (0.040) (0.030) (0.098) (0.006) (0.049)
Type prof 2−0.022 0.187∗∗∗
−0.240∗∗∗ 0.618∗∗∗
−0.092∗
−0.694∗∗∗
−0.426∗∗∗ 1.254∗∗∗
−0.133 0.730∗∗∗
(0.051) (0.054) (0.0002) (0.077) (0.051) (0.0004) (0.003) (0.003) (0.090) (0.017)
Type prof 3−2.436∗∗∗
−1.974∗∗∗
−0.238∗∗∗
−1.856∗∗∗
−1.574∗∗∗
−1.776∗∗∗
−1.047∗∗∗
−0.272∗∗∗
−0.060∗∗∗
−0.093∗∗∗
(0.00005) (0.0001) (0.00000) (0.0001) (0.0001) (0.00003) (0.00001) (0.00004) (0.003) (0.00003)
Type prof 4 0.991∗∗∗
−0.054 −0.643∗∗∗ 0.361∗∗∗
−1.023∗∗∗ 0.220∗∗∗
−0.172∗∗∗
−1.156∗∗∗
−0.095 −0.214∗∗∗
(0.128) (0.101) (0.0005) (0.138) (0.006) (0.005) (0.011) (0.001) (0.071) (0.019)
Type prof 5−0.094 −0.080 1.796∗∗∗ 0.218∗∗
−0.763∗∗∗
−2.752∗∗∗
−0.942∗∗∗
−1.171∗∗∗
−0.008 −0.685∗∗∗
(0.099) (0.088) (0.001) (0.101) (0.056) (0.001) (0.011) (0.001) (0.073) (0.010)
Type prof 6 2.166∗∗∗
−0.148∗∗∗ 0.649∗∗∗ 0.566∗∗∗
−3.731∗∗∗
−0.817∗∗∗
−3.316∗∗∗
−0.277∗∗∗ 0.268∗∗∗ 2.042∗∗∗
(0.065) (0.009) (0.0002) (0.024) (0.001) (0.003) (0.0003) (0.0004) (0.100) (0.019)
Type prof 7 2.782∗∗∗
−1.070∗∗∗ 0.125∗∗∗
−2.017∗∗∗
−1.793∗∗∗
−1.146∗∗∗
−1.885∗∗∗ 0.229∗∗∗
−0.216∗∗∗
−1.254∗∗∗
(0.008) (0.0001) (0.00000) (0.0001) (0.0001) (0.00003) (0.00003) (0.00002) (0.008) (0.0001)
Type prof 8 1.582∗∗∗
−1.655∗∗∗ 1.133∗∗∗
−0.404∗∗∗
−2.801∗∗∗
−1.499∗∗∗
−3.496∗∗∗
−0.516∗∗∗
−0.519∗∗∗ 1.122∗∗∗
(0.097) (0.005) (0.001) (0.077) (0.017) (0.003) (0.0003) (0.001) (0.085) (0.057)
Type prof 9 1.199∗∗∗ 0.978∗∗∗
−0.241∗∗∗
−3.692∗∗∗
−2.514∗∗∗
−0.677∗∗∗
−3.640∗∗∗
−0.300∗∗∗
−0.523∗∗∗
−1.607∗∗∗
(0.049) (0.020) (0.0001) (0.0004) (0.011) (0.002) (0.0002) (0.0002) (0.121) (0.0003)
Type prof 10 2.529∗∗∗
−4.669∗∗∗ 1.230∗∗∗
−3.401∗∗∗
−6.703∗∗∗ 0.816∗∗∗
−2.563∗∗∗ 0.347∗∗∗ 0.318∗∗∗ 2.611∗∗∗
(0.051) (0.0004) (0.0003) (0.0003) (0.0003) (0.008) (0.0003) (0.0005) (0.119) (0.007)
Transp subs 2−1.439∗∗∗ 0.882∗∗∗
−0.934∗∗∗ 0.105∗∗
−0.496∗∗∗
−1.840∗∗∗
−0.874∗∗∗ 1.208∗∗∗
−0.333∗∗∗
−3.554∗∗∗
(0.017) (0.068) (0.0004) (0.050) (0.039) (0.001) (0.002) (0.003) (0.078) (0.002)
Transp subs 3 0.567∗∗∗ 0.474∗∗∗
−0.938∗∗∗
−0.785∗∗∗ 0.884∗∗∗
−2.647∗∗∗ 0.144∗∗∗
−0.627∗∗∗
−0.112 1.199∗∗∗
(0.151) (0.023) (0.0003) (0.010) (0.016) (0.0004) (0.003) (0.0004) (0.091) (0.038)
Gender 0.070 0.972∗∗∗
−2.594∗∗∗ 0.109 −0.129∗∗∗ 0.784∗∗∗ 1.238∗∗∗ 1.442∗∗∗
−0.234∗∗∗ 0.147∗∗∗
(0.139) (0.091) (0.001) (0.105) (0.047) (0.004) (0.007) (0.003) (0.067) (0.044)
age 0.045∗∗∗ 0.071∗∗∗ 0.088∗∗ 0.023∗∗ 0.055∗∗∗
−0.016 0.032 0.071 0.040∗∗∗ 0.095∗∗∗
(0.009) (0.012) (0.043) (0.011) (0.016) (0.024) (0.022) (0.048) (0.004) (0.015)
Marital 2−0.334∗∗∗
−0.766∗∗∗
−0.624∗∗∗
−1.252∗∗∗ 1.035∗∗∗ 3.244∗∗∗
−1.754∗∗∗ 0.846∗∗∗
−0.692∗∗∗
−0.392∗∗∗
(0.110) (0.091) (0.001) (0.101) (0.035) (0.005) (0.004) (0.001) (0.093) (0.038)
Marital 3−1.438∗∗∗
−1.419∗∗∗
−1.519∗∗∗
−0.926∗∗∗ 1.872∗∗∗ 2.789∗∗∗
−3.937∗∗∗
−1.204∗∗∗
−1.364∗∗∗
−3.650∗∗∗
(0.050) (0.030) (0.0003) (0.046) (0.017) (0.005) (0.0002) (0.0002) (0.126) (0.001)
Marital 4−0.235∗∗∗
−4.203∗∗∗
−0.893∗∗∗
−5.189∗∗∗
−2.298∗∗∗
−1.278∗∗∗
−3.394∗∗∗
−0.515∗∗∗
−0.917∗∗∗
−4.832∗∗∗
(0.008) (0.0001) (0.00002) (0.00003) (0.00004) (0.00004) (0.00004) (0.00002) (0.012) (0.00003)
Akaike Inf. Crit. 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 59
Table A.2: Activity type - Episode 2 (2/3)
Dependent variable:
234567891011
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
Education 2a −3.387∗∗∗
−3.251∗∗∗
−0.254∗∗∗ 1.759∗∗∗
−7.762∗∗∗
−1.876∗∗∗
−2.290∗∗∗
−0.808∗∗∗
−0.997∗∗∗
−2.674∗∗∗
(0.014) (0.012) (0.0002) (0.007) (0.001) (0.0002) (0.0003) (0.0001) (0.128) (0.0001)
Education 2b −1.771∗∗∗
−4.017∗∗∗ 0.627∗∗∗ 2.674∗∗∗
−3.175∗∗∗ 2.184∗∗∗ 1.752∗∗∗
−0.337∗∗∗
−0.446∗∗∗ 1.626∗∗∗
(0.118) (0.016) (0.001) (0.097) (0.051) (0.005) (0.007) (0.001) (0.074) (0.011)
Education 3a −1.222∗∗∗
−3.213∗∗∗
−0.417∗∗∗ 2.464∗∗∗
−6.748∗∗∗ 1.621∗∗∗ 0.666∗∗∗ 1.140∗∗∗
−0.782∗∗∗ 3.008∗∗∗
(0.130) (0.034) (0.0004) (0.065) (0.002) (0.003) (0.005) (0.002) (0.076) (0.036)
Education 3b −1.676∗∗∗
−7.937∗∗∗
−0.516∗∗∗ 2.547∗∗∗
−2.588∗∗∗ 2.842∗∗∗ 2.022∗∗∗
−0.153∗∗∗
−0.417∗∗∗ 1.872∗∗∗
(0.029) (0.0001) (0.0002) (0.017) (0.023) (0.003) (0.008) (0.0002) (0.090) (0.008)
Education 4−1.355∗∗∗
−3.888∗∗∗ 1.186∗∗∗ 1.014∗∗∗
−5.760∗∗∗ 0.153∗∗∗
−1.424∗∗∗
−0.388∗∗∗
−0.909∗∗∗ 3.572∗∗∗
(0.122) (0.080) (0.001) (0.057) (0.005) (0.002) (0.003) (0.001) (0.061) (0.037)
Education 5−2.340∗∗∗
−3.460∗∗∗
−0.202∗∗∗ 1.912∗∗∗
−3.526∗∗∗
−0.976∗∗∗
−1.372∗∗∗
−0.408∗∗∗
−0.991∗∗∗ 1.869∗∗∗
(0.111) (0.108) (0.0004) (0.098) (0.096) (0.001) (0.006) (0.003) (0.065) (0.023)
Location type 2 0.732∗∗∗ 3.559∗∗∗
−0.444∗∗∗
−0.379∗∗∗
−0.740∗∗∗ 2.368∗∗∗
−1.878∗∗∗ 0.034∗∗∗
−0.116∗∗
−2.606∗∗∗
(0.078) (0.106) (0.001) (0.130) (0.025) (0.004) (0.006) (0.002) (0.056) (0.060)
Location type 3−0.232∗∗∗ 3.116∗∗∗
−0.322∗∗∗ 0.438∗∗∗ 0.309∗∗∗ 0.226∗∗∗
−4.526∗∗∗
−1.094∗∗∗
−0.518∗∗∗
−5.038∗∗∗
(0.017) (0.003) (0.0001) (0.058) (0.016) (0.0001) (0.0002) (0.0004) (0.087) (0.0003)
Location type 4 1.034∗∗∗ 4.426∗∗∗ 1.491∗∗∗
−0.833∗∗∗
−3.358∗∗∗ 2.268∗∗∗
−2.581∗∗∗ 0.472∗∗∗
−0.247∗∗∗
−2.431∗∗∗
(0.084) (0.106) (0.001) (0.061) (0.002) (0.004) (0.002) (0.002) (0.060) (0.042)
Location type 5 0.698∗∗∗ 4.627∗∗∗ 0.876∗∗∗ 0.922∗∗∗ 0.719∗∗∗ 1.249∗∗∗ 0.085∗∗∗
−0.035∗∗∗ 0.451∗∗∗
−2.150∗∗∗
(0.036) (0.012) (0.0002) (0.028) (0.015) (0.0002) (0.004) (0.0005) (0.090) (0.006)
Home occ 2 1.271∗∗∗
−0.410∗∗∗
−0.141∗∗∗ 0.127∗∗ 0.675∗∗∗
−0.009∗∗
−2.347∗∗∗ 0.818∗∗∗ 0.471∗∗∗ 1.325∗∗∗
(0.145) (0.026) (0.001) (0.055) (0.036) (0.004) (0.001) (0.001) (0.081) (0.039)
Home occ 3 0.044 1.758∗∗∗
−0.163∗∗∗ 0.661∗∗∗ 1.305∗∗∗ 2.703∗∗∗
−1.904∗∗∗ 3.649∗∗∗ 0.364∗∗∗ 1.594∗∗∗
(0.028) (0.028) (0.0003) (0.036) (0.009) (0.006) (0.0004) (0.004) (0.133) (0.040)
Akaike Inf. Crit. 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 60
Table A.3: Activity type - Episode 2 (3/3)
Dependent variable:
2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
N cars 2−1.865∗∗∗
−0.509∗∗∗ 0.276∗∗∗ 0.339∗∗∗ 2.529∗∗∗
−0.120∗∗∗
−0.056∗∗∗
−0.466∗∗∗
−0.209∗∗∗ 0.041
(0.119) (0.121) (0.001) (0.117) (0.022) (0.007) (0.006) (0.002) (0.078) (0.049)
N cars 3−1.102∗∗∗ 0.293∗∗∗
−0.290∗∗∗ 1.063∗∗∗ 0.066∗∗∗
−2.161∗∗∗
−0.099∗∗∗ 1.119∗∗∗ 0.215∗∗
−0.425∗∗∗
(0.090) (0.049) (0.0005) (0.075) (0.001) (0.001) (0.004) (0.003) (0.104) (0.004)
N cars 4−0.182∗∗∗
−0.043∗∗∗ 0.346∗∗∗ 1.883∗∗∗ 4.144∗∗∗
−2.692∗∗∗
−3.090∗∗∗
−0.852∗∗∗ 0.285∗∗∗ 2.267∗∗∗
(0.019) (0.008) (0.0001) (0.026) (0.005) (0.0001) (0.0002) (0.0001) (0.068) (0.047)
N cars 5−2.306∗∗∗ 0.625∗∗∗ 0.942∗∗∗
−0.025∗∗∗ 0.865∗∗∗ 0.611∗∗∗ 0.780∗∗∗ 1.041∗∗∗ 5.847∗∗∗ 0.746∗∗∗
(0.00000) (0.00001) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00004) (0.00001)
Ncars 6−8.009∗∗∗
−2.547∗∗∗ 1.327∗∗∗
−1.709∗∗∗ 5.746∗∗∗ 4.649∗∗∗ 1.014∗∗∗ 1.198∗∗∗ 1.239∗∗∗
−0.546∗∗∗
(0.00001) (0.0001) (0.00003) (0.0002) (0.011) (0.002) (0.00003) (0.0002) (0.012) (0.0001)
Income 2 1.235∗∗∗ 3.607∗∗∗
−0.032∗∗∗ 0.806∗∗∗ 0.041 −2.181∗∗∗
−0.497∗∗∗ 0.140∗∗∗ 0.389∗∗∗
−1.400∗∗∗
(0.099) (0.067) (0.0004) (0.082) (0.048) (0.006) (0.0005) (0.001) (0.077) (0.005)
Income 3 1.703∗∗∗ 3.665∗∗∗ 0.265∗∗∗ 0.510∗∗∗
−2.448∗∗∗
−1.484∗∗∗ 1.210∗∗∗ 0.104∗∗∗ 0.302∗∗∗
−0.131∗∗∗
(0.116) (0.052) (0.0004) (0.142) (0.019) (0.005) (0.003) (0.001) (0.061) (0.017)
Income 4 1.830∗∗∗ 2.596∗∗∗ 1.436∗∗∗
−0.447∗∗∗
−2.149∗∗∗
−2.700∗∗∗ 1.976∗∗∗ 0.236∗∗∗ 0.187∗∗∗
−0.651∗∗∗
(0.123) (0.046) (0.001) (0.092) (0.008) (0.002) (0.009) (0.001) (0.050) (0.018)
Income 5 2.521∗∗∗ 3.043∗∗∗ 0.209∗∗∗ 0.072 −1.878∗∗∗
−3.241∗∗∗ 3.246∗∗∗ 0.962∗∗∗ 0.029 −3.434∗∗∗
(0.092) (0.037) (0.0003) (0.070) (0.035) (0.001) (0.009) (0.003) (0.062) (0.0005)
Income 6 2.951∗∗∗ 3.415∗∗∗ 1.077∗∗∗
−0.470∗∗∗
−1.159∗∗∗
−0.625∗∗∗ 1.268∗∗∗
−0.154∗∗∗ 0.034 1.751∗∗∗
(0.013) (0.025) (0.0002) (0.009) (0.038) (0.002) (0.001) (0.0003) (0.089) (0.019)
Children −0.351∗∗ 0.877∗∗∗ 0.957∗∗∗
−0.546∗∗∗ 0.306∗∗∗ 1.339∗∗∗
−0.412∗∗∗
−0.070∗∗∗
−0.056 −0.143∗∗∗
(0.140) (0.015) (0.001) (0.169) (0.017) (0.002) (0.007) (0.001) (0.059) (0.016)
prev dur −0.013∗∗∗
−0.013∗∗∗
−0.004 −0.004∗∗∗ 0.001 −0.006∗∗∗
−0.008∗∗∗
−0.016∗∗∗
−0.002∗∗∗
−0.009∗∗∗
(0.001) (0.002) (0.002) (0.001) (0.001) (0.001) (0.002) (0.004) (0.0003) (0.002)
tb 0.001∗∗∗
−0.015∗∗∗
−0.003 −0.004∗∗∗
−0.003∗∗∗
−0.011∗∗∗
−0.004∗∗∗
−0.007∗∗ 0.0004 −0.008∗∗∗
(0.001) (0.002) (0.002) (0.001) (0.001) (0.001) (0.001) (0.003) (0.0003) (0.001)
Constant −0.00001∗∗∗
−0.00002∗∗∗
−0.00000∗∗
−0.00001∗∗∗
−0.00000 −0.00001∗∗∗
−0.00001∗∗∗
−0.00002∗∗∗
−0.00000∗∗∗
−0.00001∗∗∗
(0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000) (0.00000)
Akaike Inf. Crit. 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910 12,671.910
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 61
Table A.4: Activity type - Episode 3 (1/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Work hours 0.013 0.056 0.056 0.077 0.014 −0.012 0.157 0.144 0.107 0.029 0.082
(0.073) (0.072) (0.074) (0.074) (0.073) (0.075) (0.103) (0.098) (0.093) (0.074) (0.089)
Type prof 2−0.147 0.251 0.666 0.080 0.772∗
−0.262 2.253∗2.981∗∗∗ 0.083 −0.016 −0.481
(0.450) (0.353) (0.418) (0.397) (0.453) (0.566) (1.337) (1.051) (0.715) (0.506) (1.308)
Type prof 3−3.692∗∗∗ 2.149 −3.535∗∗∗
−4.574∗∗∗ 3.481∗∗ 4.108∗∗∗
−0.767∗∗∗
−0.182∗∗∗
−1.178∗∗∗ 4.680 −0.414∗∗∗
(0.012) (1.319) (0.022) (0.015) (1.486) (1.419) (0.020) (0.004) (0.040) (3.404) (0.016)
Type prof 4 0.423 0.787∗0.320 0.785∗1.089∗∗
−0.289 1.272 3.867∗∗∗ 0.650 0.388 2.283∗∗
(0.512) (0.443) (0.491) (0.473) (0.520) (0.613) (1.297) (0.994) (0.753) (0.556) (0.900)
Type prof 5 0.166 0.587 −1.445∗∗ 0.527 1.000 −0.352 1.353 3.818∗∗∗
−0.519 0.024 1.530
(0.647) (0.585) (0.691) (0.615) (0.649) (0.732) (1.328) (1.056) (0.895) (0.690) (1.029)
Type prof 6 0.703 0.552 −1.704∗∗ 0.292 0.818 0.052 2.187 −0.051 −0.096 0.848 2.132∗∗
(0.572) (0.445) (0.815) (0.561) (0.592) (0.737) (1.360) (2.870) (1.218) (0.656) (1.086)
Type prof 7−2.357∗∗∗ 2.558∗2.934∗∗
−3.210∗∗∗
−2.365∗∗∗
−1.931∗∗∗ 0.031∗∗∗
−0.145∗∗∗
−0.136∗∗∗ 5.087∗∗
−0.415∗∗∗
(0.017) (1.416) (1.478) (0.019) (0.015) (0.031) (0.001) (0.002) (0.009) (2.536) (0.008)
Type prof 8−0.937 −0.727 −1.597 −2.046∗
−0.661 −2.315∗0.609 1.601 −1.095 −1.268 0.325
(1.032) (0.963) (1.030) (1.058) (1.040) (1.191) (1.655) (1.414) (1.540) (1.056) (1.409)
Type prof 9−1.979∗
−1.421 −1.277 −3.359∗∗
−1.803 −8.503∗∗∗ 0.374 0.963 −4.564∗∗∗
−2.885∗∗∗
−0.097
(1.111) (0.943) (1.037) (1.389) (1.203) (0.067) (1.786) (1.510) (0.449) (1.101) (1.585)
Type prof 10 1.005∗1.165∗∗∗ 1.405∗∗∗ 1.377∗∗ 0.941 −1.136 3.503∗∗
−0.315∗∗
−2.789∗∗∗ 0.966 −2.598∗∗∗
(0.605) (0.377) (0.539) (0.578) (0.660) (1.061) (1.523) (0.148) (0.031) (0.688) (0.045)
Transp subs 2 0.299 0.452∗2.154∗∗∗ 0.817∗∗∗ 0.750∗∗ 0.080 1.039 −0.737 −0.934 0.214 0.830
(0.344) (0.257) (0.303) (0.290) (0.324) (0.488) (0.717) (0.990) (0.981) (0.354) (0.560)
Transp subs 3 0.731∗∗ 0.733∗∗∗ 0.726∗0.844∗∗∗ 0.136 0.474 1.225∗1.457∗∗∗ 0.105 0.821∗∗ 0.741
(0.330) (0.249) (0.378) (0.297) (0.358) (0.490) (0.672) (0.514) (0.735) (0.365) (0.573)
Gender 1.849∗∗∗ 1.322∗1.805∗∗ 1.284∗1.849∗∗∗ 1.524∗∗ 2.539∗∗∗ 1.641∗∗ 3.093∗∗∗ 1.746∗∗ 1.414∗
(0.713) (0.694) (0.717) (0.705) (0.710) (0.750) (0.865) (0.826) (0.823) (0.723) (0.799)
age −0.128∗
−0.053 −0.052 −0.074 −0.061 −0.060 −0.037 −0.005 −0.046 −0.087 −0.079
(0.067) (0.066) (0.067) (0.066) (0.066) (0.068) (0.072) (0.071) (0.071) (0.067) (0.070)
Marital 2 0.940 0.385 0.650 0.259 −0.037 0.383 −0.996 −0.391 −0.669 0.149 −0.302
(0.855) (0.806) (0.848) (0.827) (0.836) (0.909) (1.009) (1.027) (1.025) (0.866) (0.966)
Marital 3 0.441 −1.049∗
−1.762∗∗
−1.148∗
−0.956 −1.909∗∗
−2.287∗
−0.535 −1.928 −0.285 −1.389
(0.720) (0.623) (0.771) (0.697) (0.683) (0.965) (1.306) (0.986) (1.239) (0.741) (1.009)
Marital 4−1.639∗∗∗ 5.005∗∗∗
−2.374∗∗∗
−2.226∗∗∗
−3.502∗∗∗
−1.642∗∗∗
−0.204∗∗∗
−0.414∗∗∗
−0.071∗∗∗ 7.275∗∗∗
−0.970∗∗∗
(0.003) (1.919) (0.008) (0.010) (0.004) (0.008) (0.021) (0.021) (0.012) (1.811) (0.020)
Akaike Inf. Crit. 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 62
Table A.5: Activity type - Episode 3 (2/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Education 2a −2.061 −2.265∗∗
−0.250 −1.843 −1.914 −2.128 −0.087 −5.388∗∗∗
−5.574∗∗∗
−3.551∗∗∗
−0.195
(1.375) (1.072) (1.601) (1.600) (1.228) (1.392) (1.707) (0.652) (0.497) (1.348) (1.235)
Education 2b −0.162 0.324 1.864 1.112 −0.379 −0.693 −4.466 0.750 −0.516 −0.192 1.019
(1.172) (0.844) (1.436) (1.373) (1.013) (1.187) (3.199) (1.218) (1.229) (1.105) (0.985)
Education 3a 0.449 0.550 2.488∗1.082 0.589 −1.328 2.507∗∗
−2.601 2.030∗∗∗
−0.310 2.151∗∗∗
(0.999) (0.570) (1.311) (1.238) (0.800) (1.148) (1.175) (2.555) (0.670) (0.916) (0.648)
Education 3b 0.686 0.298 2.449∗1.250 −0.244 −1.296 1.463 1.386 0.564 −0.762 1.392∗
(0.977) (0.524) (1.303) (1.223) (0.796) (1.135) (1.168) (0.961) (0.879) (0.911) (0.726)
Education 4 0.198 0.243 1.998 1.665 0.095 −0.105 −0.024 1.253 0.586 −0.180 1.989∗∗∗
(1.007) (0.579) (1.317) (1.232) (0.809) (1.021) (1.314) (0.896) (0.621) (0.929) (0.598)
Education 5 0.652 0.303 2.034 1.863 0.451 −0.158 −0.420 1.329 1.474∗∗ 0.105 1.439∗∗
(1.050) (0.647) (1.349) (1.263) (0.860) (1.073) (1.449) (0.956) (0.590) (0.979) (0.725)
Location type 2 0.111 0.281 0.572 0.501 0.405 −0.261 4.568∗∗∗ 3.427∗∗ 0.785 −0.387 −0.047
(0.941) (0.891) (0.953) (0.932) (0.938) (0.991) (1.538) (1.389) (1.359) (0.977) (1.108)
Location type 3 0.545 0.437 0.661 0.740 0.752 0.404 −1.397∗∗∗ 0.466 1.916 −0.469 0.282
(0.584) (0.442) (0.612) (0.557) (0.576) (0.706) (0.118) (2.830) (1.199) (0.656) (0.995)
Location type 4−1.732∗
−1.448 −1.218 −1.006 −1.336 −1.909∗2.310 1.653 −1.821 −2.323∗∗
−2.467∗∗
(1.000) (0.946) (1.012) (0.988) (0.995) (1.053) (1.558) (1.453) (1.478) (1.035) (1.197)
Location type 5 0.559 0.606∗0.654 0.982∗∗ 1.069∗∗
−8.714∗∗∗
−0.848∗∗∗ 3.727∗∗∗ 1.759 0.357 0.044
(0.524) (0.324) (0.543) (0.451) (0.474) (0.001) (0.108) (1.305) (1.122) (0.694) (0.942)
Home occ 2−2.629∗∗
−2.017∗
−1.820 −2.141∗
−2.097∗
−1.855 −1.847 −2.152∗
−1.473 −2.143∗
−1.671
(1.165) (1.139) (1.158) (1.151) (1.155) (1.189) (1.301) (1.287) (1.245) (1.171) (1.222)
Home occ 3−0.461 0.129 −0.542 0.392 −0.166 −0.825 1.191 0.998 −0.658 −0.162 −0.710
(0.484) (0.256) (0.611) (0.340) (0.421) (0.935) (0.799) (0.722) (1.061) (0.516) (1.059)
N cars 2−0.317 −0.443 −0.949 −0.621 −0.462 −0.726 −0.509 0.653 −0.097 −0.337 −0.805
(1.201) (1.174) (1.196) (1.188) (1.194) (1.239) (1.349) (1.343) (1.354) (1.212) (1.292)
N cars 3 0.508 0.525 0.426 0.400 −0.003 0.359 −0.337 1.418∗1.087 0.652 −0.401
(0.464) (0.343) (0.431) (0.401) (0.457) (0.601) (0.958) (0.822) (0.850) (0.508) (0.874)
N cars 4 0.847 1.411∗∗∗ 1.443∗∗ 1.262∗∗ 1.201∗
−5.368∗∗∗ 1.459 −1.379∗∗∗ 3.135∗∗∗ 1.509∗∗
−5.723∗∗∗
(0.765) (0.362) (0.580) (0.540) (0.626) (0.001) (1.164) (0.026) (1.153) (0.712) (0.003)
N cars 5 3.337∗∗∗ 2.530∗∗∗ 3.415∗∗∗
−4.687∗∗∗ 2.657∗∗∗
−3.095∗∗∗
−2.159∗∗∗
−0.139∗∗∗
−0.646∗∗∗ 0.470∗∗∗
−1.684∗∗∗
(0.860) (0.521) (0.736) (0.0001) (0.876) (0.001) (0.031) (0.003) (0.002) (0.002) (0.003)
N cars 6−5.479∗∗∗ 1.673∗∗∗ 0.747 2.012∗∗∗ 0.402 2.295∗∗
−1.339∗∗∗
−0.800∗∗∗
−2.298∗∗∗ 0.186 3.063∗∗∗
(0.003) (0.433) (0.994) (0.666) (0.970) (1.000) (0.007) (0.012) (0.004) (1.264) (1.006)
Akaike Inf. Crit. 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 63
Table A.6: Activity type - Episode 3 (3/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Income 2−1.004 −1.126 −2.177∗∗
−0.354 −1.464∗
−0.825 3.991∗∗∗ 1.660 1.772∗
−0.043 −1.900∗
(0.922) (0.814) (0.898) (0.977) (0.878) (1.038) (1.042) (1.403) (1.016) (0.924) (1.058)
Income 3−1.114 −1.719∗∗
−2.505∗∗∗
−0.880 −1.751∗
−1.750 4.510∗∗∗
−1.356 1.156 −0.140 −2.298∗∗
(0.938) (0.834) (0.914) (0.996) (0.899) (1.078) (1.033) (1.590) (0.951) (0.951) (1.095)
Income 4−0.621 −0.923∗
−1.799∗∗∗ 0.398 −0.881 −0.932 4.948∗∗∗ 0.990 2.287∗∗∗ 0.701 −1.575∗
(0.686) (0.524) (0.653) (0.755) (0.631) (0.879) (0.749) (1.253) (0.620) (0.706) (0.951)
Income 5−0.330 −1.165∗∗∗
−2.368∗∗∗ 0.378 −1.087∗∗
−1.532∗6.890∗∗∗ 1.582 2.301∗∗∗ 0.976 −1.285
(0.589) (0.352) (0.556) (0.657) (0.527) (0.872) (0.691) (1.213) (0.582) (0.630) (1.009)
Income 6 0.133 −0.807 −1.976∗∗∗ 0.541 −0.412 −0.855 1.220∗∗∗
−1.936 1.846∗1.454∗1.057
(0.727) (0.509) (0.699) (0.770) (0.670) (1.028) (0.056) (3.355) (0.945) (0.794) (1.033)
PrevAct type 2 0.197 −12.684∗∗∗ 3.597∗1.556 3.221∗∗∗ 0.676 2.278∗∗ 2.403∗∗
−1.048∗∗∗
−0.054 8.084∗∗∗
(0.607) (0.0002) (2.066) (1.272) (0.868) (0.907) (1.048) (1.222) (0.012) (0.542) (0.908)
PrevAct type 3−7.806∗∗∗ 2.799∗∗∗ 7.286∗∗∗ 3.988∗∗∗ 3.076∗∗∗ 2.082∗∗
−3.566∗∗∗
−2.462∗∗∗
−1.465∗∗∗
−1.879∗∗
−1.487∗∗∗
(0.003) (0.574) (1.789) (1.012) (1.129) (1.041) (0.024) (0.035) (0.001) (0.805) (0.0001)
PrevAct type 4−1.373∗∗∗
−1.654∗∗∗
−0.052∗∗∗
−0.214∗∗∗
−0.301∗∗∗
−0.590∗∗∗ 0.048∗∗∗ 0.121∗∗∗ 0.052∗∗∗ 4.001∗∗∗
−0.021∗∗∗
(0.0002) (0.0002) (0.0001) (0.0005) (0.0003) (0.0004) (0.00003) (0.001) (0.00000) (0.001) (0.00000)
PrevAct type 5−7.934∗∗∗ 2.367∗∗∗ 5.469∗∗∗ 3.313∗∗∗ 2.788∗∗
−4.727∗∗∗ 3.296∗∗∗
−1.253∗∗∗
−1.795∗∗∗
−0.029 −1.252∗∗∗
(0.004) (0.614) (1.938) (1.183) (1.149) (0.010) (1.250) (0.049) (0.002) (0.611) (0.0001)
PrevAct type 6 1.634∗∗ 1.970∗∗∗
−1.127∗∗∗
−1.441∗∗∗ 5.207∗∗∗
−3.044∗∗∗
−1.356∗∗∗
−1.091∗∗∗
−0.684∗∗∗ 1.189∗∗
−1.289∗∗∗
(0.669) (0.677) (0.011) (0.006) (0.809) (0.003) (0.005) (0.009) (0.001) (0.467) (0.0001)
PrevAct type 7−5.927∗∗∗ 1.904∗∗
−0.909∗∗∗
−0.612∗∗∗ 5.821∗∗∗
−1.051∗∗∗
−0.468∗∗∗
−0.249∗∗∗ 0.074∗∗∗ 2.049∗∗∗
−0.591∗∗∗
(0.0004) (0.825) (0.011) (0.004) (0.915) (0.004) (0.004) (0.003) (0.0001) (0.655) (0.0001)
PrevAct type 8−4.819∗∗∗
−4.071∗∗∗
−0.527∗∗∗
−0.540∗∗∗ 7.490∗∗∗
−0.958∗∗∗
−0.640∗∗∗
−0.038∗∗∗ 0.068∗∗∗ 3.707∗∗∗
−0.051∗∗∗
(0.0003) (0.0003) (0.003) (0.001) (0.642) (0.001) (0.002) (0.001) (0.0001) (0.645) (0.00001)
PrevAct type 9−2.880∗∗∗
−3.787∗∗∗
−0.574∗∗∗ 7.848∗∗∗
−0.560∗∗∗
−0.673∗∗∗
−1.291∗∗∗
−0.085∗∗∗
−0.220∗∗∗ 2.383∗∗∗
−0.018∗∗∗
(0.001) (0.0004) (0.002) (0.852) (0.001) (0.001) (0.002) (0.0003) (0.00004) (0.854) (0.00000)
PrevAct type 10 −2.488∗∗∗ 1.050 3.890∗∗ 2.205∗∗ 1.509 −1.135 −1.200 −0.063 4.131∗∗∗
−9.633∗∗∗ 6.062∗∗∗
(0.722) (0.717) (1.893) (1.003) (0.939) (0.797) (1.006) (1.148) (1.362) (1.252) (0.911)
PrevAct type 11 0.780 0.695 −0.939∗∗∗
−1.451∗∗∗
−2.071∗∗∗
−3.342∗∗∗
−1.865∗∗∗
−0.834∗∗∗
−0.514∗∗∗
−0.026 9.541∗∗∗
(0.670) (0.686) (0.049) (0.023) (0.013) (0.010) (0.013) (0.029) (0.002) (0.500) (0.947)
Children 0.276 0.386 0.394 0.575 0.400 0.195 0.175 0.113 0.215 0.312 0.420
(1.261) (1.250) (1.261) (1.257) (1.259) (1.279) (1.336) (1.316) (1.320) (1.264) (1.308)
Constant 10.655∗∗∗ 6.883∗∗∗
−0.250 1.185 5.086∗∗∗ 10.134∗∗∗
−13.073∗∗∗
−11.313∗∗∗
−8.835∗∗∗ 11.303∗∗∗
−3.600∗∗
(1.485) (0.858) (2.503) (1.806) (1.416) (1.902) (1.679) (1.973) (1.268) (1.565) (1.819)
Akaike Inf. Crit. 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331 8,384.331
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 64
Table A.7: Activity type - Episode 4 (1/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Work hours −0.047∗∗∗
−0.035∗∗ 0.050 −0.039 −0.035∗
−0.009 0.029 −0.046 0.008 −0.023 0.041
(0.017) (0.017) (0.039) (0.030) (0.019) (0.032) (0.064) (0.045) (0.037) (0.020) (0.062)
Type prof 2 0.966∗∗ 0.019 0.705 0.349 1.477∗∗∗ 0.358 −19.681∗∗∗
−0.538 −0.773 0.236 1.603
(0.466) (0.389) (0.512) (0.508) (0.573) (0.834) (0.00000) (1.454) (0.800) (0.413) (1.343)
Type prof 3 1.107 0.012 −1.217∗∗∗
−2.626 −1.636 −1.373∗∗∗
−0.568 −0.055 −1.350 0.769 −0.854∗∗∗
(1.721) (1.707) (0.436) (3.619) (6.453) (0.431) (6.862) (0.496) (6.208) (1.850) (0.072)
Type prof 4 0.993∗∗
−0.201 0.270 −0.424 0.624 0.362 −0.309 0.735 −0.518 0.188 1.169
(0.432) (0.352) (0.472) (0.504) (0.558) (0.749) (0.782) (1.086) (0.660) (0.376) (1.349)
Type prof 5 1.228∗∗∗ 0.269 −0.561 0.437 1.363∗∗ 0.580 −0.312 0.085 0.494 0.186 1.726
(0.461) (0.382) (0.656) (0.575) (0.564) (0.774) (0.832) (1.178) (0.648) (0.404) (1.406)
Type prof 6 1.076∗0.242 −0.316 −0.295 0.297 −0.045 −9.628∗∗∗
−3.253 0.110 −0.951 1.918
(0.586) (0.524) (0.947) (0.899) (0.765) (1.083) (0.001) (3.569) (0.886) (0.622) (1.626)
Type prof 7−2.886∗∗∗
−4.015 1.241 −2.309 −1.946 −1.092∗∗∗
−2.636 −0.821 −1.393 −0.860 −0.762∗∗∗
(0.398) (4.457) (1.633) (6.601) (6.643) (0.231) (5.439) (4.299) (5.989) (5.026) (0.118)
Type prof 8 2.591∗∗∗ 0.517 0.732 −0.570 1.958∗∗∗ 1.839∗∗
−0.212 −0.930 −0.486 0.457 2.399
(0.504) (0.474) (0.704) (0.988) (0.635) (0.879) (1.053) (1.575) (0.975) (0.513) (1.547)
Type prof 9 2.737∗∗∗ 1.489∗∗ 2.158∗∗ 1.240 1.690∗∗ 1.116 −0.156 −2.584 −5.192∗∗∗ 0.530 2.959∗
(0.622) (0.682) (0.929) (1.308) (0.849) (1.292) (1.434) (5.198) (0.022) (0.809) (1.762)
Type prof 10 1.819∗∗∗ 0.384 1.263 −6.247∗∗∗ 1.014 1.040 −4.984∗∗∗
−3.393 −5.863∗∗∗
−1.309 2.148
(0.674) (0.680) (0.883) (0.012) (0.850) (1.137) (0.086) (6.313) (0.018) (0.942) (1.819)
Transp subs 2−0.261 0.076 0.487 0.417 −0.037 −0.567 0.066 −0.021 −0.192 0.779∗∗∗ 0.691
(0.307) (0.275) (0.364) (0.383) (0.372) (0.672) (0.747) (0.870) (0.575) (0.284) (0.743)
Transp subs 3−0.154 −0.398 −0.841 −1.910∗∗
−0.613 −0.167 1.202∗∗ 0.659 −1.032 −0.665∗0.079
(0.277) (0.296) (0.523) (0.852) (0.382) (0.495) (0.578) (0.697) (0.714) (0.341) (0.685)
Gender −0.133 0.207 0.355 −0.245 0.484∗0.969∗∗ 0.424 0.384 0.421 0.307 1.099∗
(0.220) (0.215) (0.341) (0.347) (0.260) (0.389) (0.557) (0.620) (0.383) (0.234) (0.575)
age −0.030∗∗
−0.010 0.009 −0.032 −0.026∗
−0.015 0.044 −0.006 −0.012 −0.020 0.005
(0.012) (0.013) (0.019) (0.022) (0.015) (0.023) (0.031) (0.037) (0.023) (0.014) (0.032)
Marital 2 0.650∗
−0.205 0.426 0.129 −0.170 −0.409 −0.283 −0.792 −0.223 −0.323 1.005
(0.336) (0.314) (0.512) (0.511) (0.363) (0.502) (0.883) (0.807) (0.535) (0.336) (0.838)
Marital 3 1.008∗∗
−0.177 1.151∗0.246 0.298 −10.299∗∗∗
−0.214 −8.646∗∗∗
−1.402 −0.296 −9.935∗∗∗
(0.441) (0.433) (0.662) (0.674) (0.496) (0.0002) (1.117) (0.002) (1.121) (0.483) (0.0003)
Marital 4 0.402 −1.221 1.046 −3.507∗∗∗ 0.092 −2.227 2.263 −2.023 1.179 1.512 −1.631
(1.134) (1.364) (1.662) (0.329) (1.244) (6.820) (1.428) (6.501) (1.197) (1.110) (6.332)
Akaike Inf. Crit. 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 65
Table A.8: Activity type - Episode 4 (2/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Education 2a 2.637∗∗ 0.623 −1.008 −5.727∗∗∗
−0.712 0.694 −6.848∗∗∗
−2.572 1.207 1.464 −5.138∗∗∗
(1.310) (0.980) (1.282) (0.004) (0.847) (1.207) (0.020) (5.249) (1.020) (1.344) (0.009)
Education 2b 1.541 0.245 −1.658 0.980 −1.261∗
−0.950 1.535 0.894 1.254∗0.869 2.624∗∗
(1.260) (0.882) (1.050) (0.623) (0.689) (1.165) (2.491) (1.578) (0.655) (1.269) (1.204)
Education 3a 1.141 0.325 −1.356 0.969∗
−1.006 −0.666 1.798 0.833 0.167 0.848 2.103∗
(1.288) (0.903) (1.098) (0.569) (0.728) (1.215) (2.539) (1.553) (0.791) (1.286) (1.248)
Education 3b 2.117 0.742 −1.227 2.020∗∗∗
−0.723 −0.500 1.766 0.533 0.723 1.567 −12.348∗∗∗
(1.288) (0.919) (1.142) (0.607) (0.751) (1.263) (2.585) (1.698) (0.795) (1.290) (0.00002)
Education 4 2.100 0.594 −1.320 1.708∗∗∗
−0.734 −0.348 1.678 0.024 1.391∗∗ 1.094 2.843∗∗
(1.279) (0.895) (1.089) (0.442) (0.725) (1.198) (2.531) (1.537) (0.569) (1.272) (1.123)
Education 5 2.087 0.869 −1.120 2.299∗∗∗
−1.290∗
−0.559 0.503 0.711 1.452∗∗ 1.218 3.039∗∗∗
(1.287) (0.901) (1.093) (0.453) (0.752) (1.222) (2.529) (1.540) (0.586) (1.277) (1.135)
Location type 2 0.145 −0.116 0.127 4.534∗∗∗
−0.792∗∗
−0.282 −0.260 2.101 −0.254 −0.346 −0.012
(0.413) (0.372) (0.611) (0.452) (0.368) (0.656) (0.915) (2.665) (0.663) (0.415) (1.126)
Location type 3 0.970∗0.292 −0.617 4.762∗∗∗
−0.637 −0.578 0.478 2.443 −0.882 0.145 −9.153∗∗∗
(0.512) (0.498) (0.869) (0.639) (0.570) (0.993) (1.106) (2.832) (1.152) (0.537) (0.001)
Location type 4 0.699 0.355 0.468 5.341∗∗∗
−0.697∗
−0.341 −0.330 2.106 −0.204 −0.126 0.733
(0.428) (0.393) (0.635) (0.463) (0.402) (0.710) (0.982) (2.696) (0.700) (0.436) (1.144)
Location type 5 0.006 0.050 0.886 5.468∗∗∗
−0.362 0.033 0.821 2.336 −0.342 −0.541 −8.531∗∗∗
(0.616) (0.551) (0.811) (0.581) (0.586) (0.908) (1.140) (2.850) (0.951) (0.594) (0.001)
Home occ 2 0.101 −0.272 −0.410 −0.234 0.061 0.651 1.469∗∗∗
−1.659 0.079 −0.322 0.426
(0.246) (0.258) (0.404) (0.404) (0.290) (0.432) (0.553) (1.036) (0.452) (0.281) (0.618)
Home occ 3 0.291 −0.425 0.299 0.064 −0.774 0.427 −7.892∗∗∗
−3.376 −0.519 −0.424 1.314
(0.472) (0.496) (0.794) (0.824) (0.677) (0.736) (0.002) (4.136) (0.931) (0.518) (0.899)
N cars 2 0.332 0.424 0.309 −0.015 0.748∗∗ 0.548 0.664 0.174 −0.427 0.010 −0.052
(0.280) (0.283) (0.424) (0.455) (0.346) (0.532) (0.762) (0.767) (0.487) (0.305) (0.718)
N cars 3−0.426 −0.298 −0.118 −0.291 −0.056 0.359 0.673 −3.933 −0.337 −0.195 0.016
(0.394) (0.388) (0.560) (0.613) (0.514) (0.665) (0.974) (4.783) (0.671) (0.414) (0.909)
N cars 4−0.725 −0.672 −0.341 −0.588 0.649 0.381 1.317 −1.523 0.120 −0.685 1.016
(0.775) (0.772) (1.061) (1.357) (0.779) (1.235) (1.474) (4.627) (1.144) (0.841) (1.339)
N cars 5 0.025 −5.752 −0.489 −1.822 1.542 1.154 −0.565∗∗∗
−0.657∗∗
−1.173 −4.442 −0.720
(1.441) (6.541) (2.014) (5.590) (1.171) (1.615) (0.058) (0.288) (6.542) (4.376) (5.776)
N cars 6 0.494 1.882∗0.902 2.800∗0.605 1.356 −2.077∗∗∗
−1.889 −3.878∗∗∗ 0.389 2.145
(1.108) (1.121) (1.696) (1.475) (1.133) (1.225) (0.260) (6.916) (0.011) (1.132) (1.535)
Akaike Inf. Crit. 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 66
Table A.9: Activity type - Episode 4 (3/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Income 2−0.342 −0.096 −1.047 −0.235 0.210 3.614∗∗∗ 0.480 1.435 0.427 −0.779 1.602
(0.502) (0.551) (0.761) (1.009) (0.604) (0.651) (2.263) (1.901) (1.129) (0.586) (2.582)
Income 3 0.075 −0.117 −1.277 −0.589 0.062 3.453∗∗∗ 1.584 1.463 0.742 −0.170 1.349
(0.503) (0.559) (0.788) (1.018) (0.624) (0.632) (2.206) (1.873) (1.144) (0.583) (2.593)
Income 4 0.141 0.404 −0.813 −0.578 −0.103 3.810∗∗∗ 1.388 0.911 0.959 0.157 0.889
(0.532) (0.583) (0.832) (1.061) (0.659) (0.607) (2.243) (1.892) (1.180) (0.609) (2.639)
Income 5 0.469 0.203 −1.079 −0.465 0.397 4.385∗∗∗ 1.907 −0.310 1.086 −0.097 0.699
(0.581) (0.630) (0.914) (1.116) (0.711) (0.660) (2.302) (2.084) (1.256) (0.663) (2.736)
Income 6 0.187 0.333 −0.630 −0.077 0.918 4.090∗∗∗ 1.159 −4.637∗∗∗
−8.383∗∗∗ 0.068 2.153
(0.717) (0.712) (1.004) (1.188) (0.796) (0.879) (2.491) (0.282) (0.002) (0.747) (2.775)
PrevAct type 2−6.383∗∗∗
−11.481∗∗∗
−4.682∗∗
−9.444∗∗∗
−4.943∗∗∗
−4.300∗∗∗ 0.763 −5.554∗∗∗
−3.919∗∗∗
−7.901∗∗∗
−5.323∗∗∗
(0.611) (0.981) (1.967) (1.449) (0.721) (1.151) (1.198) (1.145) (1.199) (0.593) (1.201)
PrevAct type 3−0.785 1.499 8.296∗∗∗ 3.315 2.464 3.872∗
−6.293∗∗∗
−3.017∗∗∗
−5.298∗∗∗
−0.480 −6.584∗∗∗
(2.229) (1.996) (2.561) (2.060) (2.078) (2.249) (0.0001) (0.841) (0.035) (2.027) (0.011)
PrevAct type 4−1.271 0.263 4.486∗∗
−0.731 1.791 1.244 −8.421∗∗∗ 0.719 1.077 2.393∗0.779
(1.581) (1.401) (2.239) (1.806) (1.500) (1.957) (0.00002) (1.969) (1.979) (1.396) (2.021)
PrevAct type 5−0.245 0.713 2.029 −0.105 0.170 1.634 −7.518∗∗∗
−3.642 0.380 0.191 −7.345∗∗∗
(1.157) (1.101) (2.284) (1.335) (1.276) (1.567) (0.0001) (6.095) (1.804) (1.108) (0.008)
PrevAct type 6−9.100∗∗∗ 2.136∗∗
−0.037 1.507 3.744∗∗∗ 3.478∗∗ 9.861∗∗∗
−1.372 3.069∗∗ 1.738∗
−2.315∗∗∗
(0.003) (0.909) (0.556) (1.369) (1.026) (1.491) (1.117) (1.897) (1.522) (0.928) (0.175)
PrevAct type 7−6.669∗∗∗
−0.827 −1.293 −3.723 −0.462 0.742 6.707∗∗∗ 1.349 −1.589 −1.568 −1.214
(0.493) (1.398) (2.102) (6.215) (1.739) (1.951) (1.787) (2.089) (4.625) (1.463) (3.386)
PrevAct type 8−7.384∗∗∗
−1.711∗
−1.666 −5.047 −4.501 0.180 −0.893∗∗∗
−2.579 −2.711 −2.030∗∗
−1.925
(0.589) (0.964) (5.430) (5.585) (4.787) (1.700) (0.070) (3.550) (3.976) (1.010) (2.933)
PrevAct type 9−3.550∗∗∗ 3.531∗∗∗ 0.638∗∗∗ 3.448∗∗∗
−2.552∗∗∗
−1.558∗∗∗
−0.891∗∗∗
−0.336∗
−0.924∗∗∗ 2.168∗∗∗
−0.971∗∗∗
(0.021) (0.505) (0.157) (0.822) (0.019) (0.025) (0.001) (0.188) (0.056) (0.573) (0.037)
PrevAct type 10 −1.402∗∗
−2.770∗∗∗ 1.579 −0.968 −1.320∗
−0.913 3.562∗∗∗
−2.276∗
−1.648 −7.543∗∗∗
−1.438
(0.612) (0.583) (1.895) (0.758) (0.734) (1.169) (1.185) (1.188) (1.246) (0.981) (1.204)
PrevAct type 11 −5.536∗∗∗ 2.571 −0.043 3.342 2.392 4.741∗∗
−1.117∗∗∗
−0.367 −1.209∗1.223 −1.259
(0.163) (2.034) (0.485) (2.174) (2.252) (2.305) (0.002) (5.030) (0.629) (2.066) (0.889)
Children 0.199 −0.077 0.004 −0.559∗∗ 0.014 −0.229 −0.289 0.939 −0.300 −0.026 −0.190
(0.185) (0.183) (0.286) (0.284) (0.220) (0.329) (0.472) (0.656) (0.334) (0.201) (0.483)
Constant 1.362 4.752∗∗∗
−3.730 −1.627 2.883∗
−4.583∗∗
−12.410∗∗∗
−2.066 −1.591 4.568∗∗
−10.195∗∗∗
(1.838) (1.570) (3.041) (1.396) (1.700) (2.112) (3.531) (2.170) (2.234) (1.899) (3.282)
Akaike Inf. Crit. 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362 7,255.362
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 67
Table A.10: Activity type - Episode 5 (1/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Work hours −0.017 0.015 0.019 −0.023 −0.008 −0.048 −0.134∗∗∗ 0.080 −0.056 0.017 0.015
(0.024) (0.024) (0.049) (0.051) (0.029) (0.041) (0.050) (0.125) (0.038) (0.027) (0.206)
Type prof 2 2.141∗∗∗ 0.744 1.863∗∗∗ 2.332∗∗ 0.710 1.947∗∗ 3.402∗∗ 0.262 −0.472 0.192 2.693
(0.777) (0.529) (0.665) (0.978) (0.683) (0.932) (1.419) (1.561) (0.859) (0.559) (2.824)
Type prof 3−1.361∗∗∗ 3.006∗∗∗
−0.123∗∗∗
−0.047 −2.312∗∗∗
−0.970∗∗∗ 0.002 0.127∗∗∗
−2.479∗∗∗
−2.222∗∗∗ 0.145∗∗∗
(0.009) (0.087) (0.006) (0.052) (0.020) (0.086) (0.020) (0.006) (0.050) (0.027) (0.010)
Type prof 4 2.303∗∗∗ 1.120∗∗ 0.992 1.254 0.958 1.266 3.987∗∗∗ 0.943 0.032 0.190 3.406
(0.734) (0.479) (0.628) (1.032) (0.608) (0.930) (1.296) (1.179) (0.751) (0.501) (2.688)
Type prof 5 2.608∗∗∗ 0.929∗0.456 2.273∗∗ 0.897 1.516 3.782∗∗∗
−0.440 0.144 −0.170 0.834
(0.769) (0.524) (0.848) (1.101) (0.652) (0.996) (1.244) (1.489) (0.774) (0.558) (3.769)
Type prof 6 2.167∗∗
−0.258 0.648 −1.815 −1.532 −6.602∗∗∗ 2.630 −3.381 −1.258 −1.503 1.413
(0.977) (0.803) (1.152) (6.653) (1.282) (0.041) (1.635) (5.577) (1.286) (0.915) (5.053)
Type prof 7−0.808∗∗∗ 2.397∗∗∗ 3.660∗∗∗
−0.014∗∗
−1.427∗∗∗
−0.170∗∗∗ 0.088∗∗∗ 0.046∗∗∗
−0.246∗∗∗
−2.282∗∗∗
−0.086∗∗∗
(0.008) (0.861) (0.839) (0.006) (0.015) (0.009) (0.00002) (0.002) (0.027) (0.012) (0.011)
Type prof 8 2.862∗∗∗ 0.595 1.954∗∗ 2.733∗1.297 0.098 4.029∗∗∗
−0.444 −0.562 0.265 3.010
(0.928) (0.706) (0.977) (1.415) (0.878) (1.568) (1.411) (1.879) (1.116) (0.741) (3.218)
Type prof 9 1.074 −0.705 1.312 −1.071 −7.919∗∗∗ 1.794 −1.847∗∗
−5.110∗∗
−0.700 0.682 −2.121
(1.482) (0.989) (1.262) (7.154) (0.007) (1.526) (0.799) (2.013) (1.462) (0.963) (5.902)
Type prof 10 2.527∗∗ 1.240 1.502 0.409 0.805 −3.390 −2.514∗∗∗
−1.675 −5.480∗∗∗
−0.299 1.679
(1.278) (1.094) (1.414) (7.659) (1.294) (6.911) (0.053) (8.404) (0.040) (1.135) (5.672)
Transp subs 2−0.077 0.017 1.237∗∗∗ 0.967 −0.237 0.395 −4.870 1.903∗
−0.364 0.208 −2.684
(0.455) (0.391) (0.480) (0.608) (0.517) (0.621) (7.083) (1.119) (0.718) (0.416) (3.074)
Transp subs 3−0.238 0.136 0.571 −0.349 −0.082 0.230 0.983 2.827∗∗ 0.748 0.330 −2.277
(0.528) (0.420) (0.689) (1.101) (0.570) (0.749) (0.850) (1.173) (0.536) (0.452) (4.288)
Gender −0.005 −0.041 −0.130 −0.482 0.255 −0.810 −2.140∗∗∗
−0.180 −0.958∗0.095 −2.797
(0.359) (0.317) (0.460) (0.623) (0.396) (0.573) (0.808) (1.003) (0.503) (0.346) (2.139)
age −0.001 0.007 0.015 −0.048 0.003 −0.009 −0.003 0.003 −0.010 0.003 −0.116
(0.021) (0.019) (0.026) (0.038) (0.024) (0.033) (0.042) (0.058) (0.031) (0.021) (0.087)
Marital 2 1.458∗∗ 0.921∗∗ 1.199∗0.495 0.293 0.353 0.384 0.168 −0.165 1.317∗∗ 0.707
(0.582) (0.461) (0.704) (0.878) (0.589) (0.804) (0.944) (1.423) (0.619) (0.520) (2.895)
Marital 3 0.078 0.174 0.471 1.845∗0.294 1.588 0.225 −3.750 −0.028 0.703 3.186
(0.874) (0.642) (0.931) (1.082) (0.799) (0.980) (1.446) (6.460) (0.877) (0.713) (3.171)
Marital 4−3.254∗∗∗ 1.185 −2.728∗∗
−0.873 2.047 −1.730∗∗∗
−2.008∗∗
−0.436∗
−3.108∗∗∗ 2.122 0.437∗∗∗
(0.413) (2.512) (1.211) (7.819) (2.651) (0.177) (0.815) (0.247) (0.819) (2.551) (0.021)
Akaike Inf. Crit. 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 68
Table A.11: Activity type - Episode 5 (2/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Education 2a 0.818 −2.728 −2.307 −2.084 1.776∗∗
−4.265∗∗ 2.576∗∗
−2.807 1.217 −0.616 −0.974
(1.089) (1.696) (1.864) (1.741) (0.839) (2.011) (1.025) (6.350) (1.128) (1.708) (5.489)
Education 2b 2.300∗∗∗
−1.390 −1.574 0.113 2.010∗∗∗
−5.507∗∗∗ 0.124 0.231 2.014∗∗∗
−0.820 1.252
(0.628) (1.589) (1.668) (2.110) (0.627) (1.937) (1.130) (1.938) (0.750) (1.604) (2.513)
Education 3a 2.553∗∗∗
−2.279 −2.841∗1.417 2.255∗∗∗
−4.909∗∗∗
−0.247 −0.183 1.207 −1.605 −1.445
(0.564) (1.603) (1.717) (1.923) (0.553) (1.843) (1.213) (1.801) (0.743) (1.638) (3.442)
Education 3b 2.433∗∗∗
−1.856 −3.200∗
−3.960 1.482∗∗
−4.231∗∗ 0.580 1.252 1.690∗∗
−0.755 −0.527
(0.642) (1.630) (1.807) (5.844) (0.716) (1.880) (1.129) (1.901) (0.761) (1.663) (4.070)
Education 4 3.023∗∗∗
−1.514 −1.691 1.189 1.739∗∗∗
−4.003∗∗ 1.366∗
−1.217 1.754∗∗∗ 0.005 0.991
(0.509) (1.598) (1.691) (1.907) (0.513) (1.792) (0.723) (1.848) (0.613) (1.624) (1.942)
Education 5 2.453∗∗∗
−1.743 −3.030∗0.754 1.510∗∗∗
−4.655∗∗ 0.762 −1.280 1.588∗∗
−0.761 −0.171
(0.543) (1.607) (1.721) (1.915) (0.525) (1.810) (0.909) (1.830) (0.622) (1.637) (2.076)
Location type 2 0.820 −0.146 −0.079 2.632 −1.113∗0.553 −1.361 3.002 −0.984 0.309 −0.709
(0.848) (0.580) (0.834) (3.629) (0.637) (1.158) (1.043) (2.992) (0.733) (0.685) (3.038)
Location type 3 0.426 −0.194 −0.370 3.378 −0.728 −0.036 −5.415 3.236 −0.455 −0.235 1.422
(1.013) (0.748) (1.151) (3.680) (0.825) (1.548) (8.528) (3.239) (0.980) (0.916) (3.164)
Location type 4 0.294 −0.366 −0.446 2.381 −1.181∗1.021 −1.840∗
−0.139 −0.921 0.316 0.525
(0.876) (0.604) (0.866) (3.646) (0.673) (1.176) (1.118) (3.786) (0.775) (0.707) (3.219)
Location type 5−1.392 −0.436 −1.428 2.528 −1.701∗
−4.570 −2.765 3.432 −1.128 −0.528 1.120
(1.378) (0.833) (1.189) (3.734) (1.013) (4.853) (1.714) (3.230) (1.062) (0.999) (3.312)
Home occ 2 0.246 0.240 0.411 −0.874 0.027 −0.009 0.955 −3.440 0.954∗
−0.351 −2.611
(0.448) (0.393) (0.533) (0.821) (0.511) (0.674) (0.850) (3.499) (0.528) (0.434) (3.268)
Home occ 3−0.395 1.383∗
−6.309∗∗∗ 1.002 0.781 1.145 2.842∗∗
−1.631 1.288 1.159 −1.297
(1.193) (0.728) (0.023) (1.356) (0.892) (1.057) (1.209) (4.974) (0.978) (0.747) (5.175)
N cars 2−0.332 −0.322 −0.623 −0.454 −0.269 0.115 −2.281∗∗
−1.708 −0.584 −0.287 −1.207
(0.479) (0.409) (0.553) (0.752) (0.535) (0.765) (0.899) (1.252) (0.609) (0.446) (1.644)
N cars 3−0.219 0.098 −0.085 −1.343 0.305 0.704 −6.953 −0.039 −0.507 −0.728 0.147
(0.673) (0.591) (0.747) (1.353) (0.707) (0.982) (7.369) (1.429) (0.997) (0.704) (2.272)
N cars 4−2.115 −1.705∗
−3.039∗∗
−5.968∗∗∗
−9.566∗∗∗
−0.267 −1.522 −5.029 −6.053∗∗∗
−0.360 −5.951
(1.373) (0.929) (1.451) (0.701) (0.006) (1.584) (1.834) (8.962) (0.049) (1.006) (5.076)
N cars 5−1.434∗∗∗ 5.687∗∗∗
−1.201∗∗∗ 0.776∗∗∗
−1.631∗∗∗
−0.158∗∗∗
−0.644∗∗∗
−0.144∗∗∗ 0.380∗∗∗
−2.111∗∗∗ 0.141∗∗
(0.002) (0.088) (0.006) (0.002) (0.002) (0.001) (0.022) (0.024) (0.002) (0.003) (0.055)
N cars 6−4.038∗∗∗ 0.963 1.273 −0.760 1.659 −3.218 −3.574∗∗∗ 0.042 −2.683∗∗∗
−0.247 0.702∗∗∗
(0.092) (1.527) (1.760) (7.802) (1.393) (8.348) (0.455) (1.152) (0.231) (1.671) (0.120)
Akaike Inf. Crit. 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 69
Table A.12: Activity type - Episode 5 (3/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Income 2 1.044 0.930 1.420 −2.008 −0.366 −0.427 4.162∗∗∗ 2.030 0.539 1.413 0.250
(1.021) (0.810) (1.204) (5.275) (1.008) (1.230) (1.005) (3.353) (1.182) (1.038) (3.499)
Income 3 0.605 0.695 1.597 3.075 0.213 −1.185 5.115∗∗∗ 1.254 0.934 1.157 1.511
(1.029) (0.807) (1.215) (2.012) (0.967) (1.329) (0.830) (3.533) (1.165) (1.047) (2.150)
Income 4 0.601 0.395 1.417 2.412 0.337 −0.875 5.006∗∗∗ 1.067 0.949 0.117 0.767
(1.056) (0.838) (1.279) (2.019) (1.000) (1.296) (0.863) (3.562) (1.219) (1.085) (2.237)
Income 5 0.820 0.200 1.048 2.571 0.576 −0.527 6.430∗∗∗ 1.338 −5.336∗∗∗ 0.419 3.106
(1.106) (0.893) (1.348) (2.063) (1.080) (1.383) (0.982) (3.734) (0.684) (1.128) (2.209)
Income 6 0.769 −0.226 1.449 2.032 −0.020 −0.835 6.436∗∗∗ 2.174 −4.643∗∗∗
−0.301 0.606
(1.204) (0.997) (1.445) (2.234) (1.226) (1.553) (1.245) (3.700) (0.173) (1.247) (4.689)
PrevAct type 2−6.607∗∗∗
−16.669∗∗∗
−1.541 −6.647∗∗∗
−6.695∗∗∗
−3.820∗∗∗
−5.136∗∗∗
−2.547 −4.947∗∗∗
−5.189∗∗∗
−3.531
(0.874) (1.719) (1.100) (1.231) (0.914) (1.315) (1.265) (2.459) (1.026) (0.890) (2.401)
PrevAct type 3−1.477 −1.741 7.851∗∗∗ 0.446 −1.143 −0.535 −3.147 −1.892 −1.390 −0.714 3.159
(1.252) (1.131) (1.169) (1.401) (1.253) (1.899) (5.076) (4.488) (1.626) (1.227) (2.710)
PrevAct type 4 2.784 0.753 10.356∗∗∗ 2.942 2.428 −1.432∗4.460∗0.702 −2.416 6.275∗∗∗ 2.023
(2.043) (2.011) (1.675) (2.287) (2.108) (0.864) (2.382) (2.834) (2.196) (2.018) (5.618)
PrevAct type 5−0.162 −0.704 4.769∗∗∗
−0.772 −1.522 1.480 1.679 4.487∗
−4.630 1.723 0.802
(1.548) (1.510) (1.627) (1.922) (1.685) (1.927) (1.834) (2.705) (7.105) (1.550) (4.102)
PrevAct type 6−8.475∗∗∗ 2.717 0.791∗∗∗
−0.438 3.517 6.569∗∗∗ 5.663∗∗ 0.836∗∗
−1.340 4.550∗∗ 2.210∗∗∗
(0.007) (2.261) (0.059) (7.121) (2.312) (2.475) (2.456) (0.371) (2.240) (2.300) (0.221)
PrevAct type 7−2.959∗
−2.953∗∗
−1.629∗∗∗
−3.709 −8.446∗∗∗
−2.666 0.865 −2.043 −0.287 −2.230 4.700
(1.670) (1.319) (0.204) (6.638) (0.175) (5.078) (1.839) (6.473) (1.577) (1.667) (3.023)
PrevAct type 8−4.681∗∗∗
−6.221∗∗∗
−5.356∗∗∗
−8.714∗∗∗
−11.759∗∗∗
−6.006 −6.757∗∗∗
−5.554∗∗∗
−9.078∗∗∗
−4.466∗∗∗
−4.172∗∗∗
(1.205) (1.014) (0.037) (0.238) (0.038) (8.096) (0.767) (1.042) (0.205) (1.410) (0.487)
PrevAct type 9−7.698∗∗∗ 2.441∗∗∗ 0.784∗∗∗ 3.973∗∗∗ 1.826∗
−1.072∗∗∗
−1.434∗∗∗ 1.448∗∗∗ 3.163∗∗∗ 3.776∗∗∗ 2.005∗∗∗
(0.002) (0.503) (0.023) (1.084) (0.959) (0.052) (0.046) (0.279) (1.044) (0.684) (0.238)
PrevAct type 10 −1.105 −1.850∗∗ 5.319∗∗∗
−0.440 −1.143 0.234 −1.604 −0.359 −1.346 −9.487∗∗∗
−0.142
(0.919) (0.881) (1.018) (1.107) (0.944) (1.364) (1.401) (2.604) (1.113) (0.283) (2.636)
PrevAct type 11 −8.409∗∗∗
−2.966∗∗
−1.509∗∗∗
−4.281 −1.136 −2.520 −2.713 −1.353 −4.818 −1.232 −0.227
(0.566) (1.351) (0.299) (7.196) (1.492) (5.314) (6.021) (5.771) (7.152) (1.469) (7.250)
Children −0.113 0.334 0.151 0.034 0.107 −0.078 −0.668 0.100 −0.151 −0.159 1.168
(0.296) (0.266) (0.375) (0.521) (0.338) (0.461) (0.630) (0.798) (0.420) (0.286) (1.407)
Constant −3.001 4.907∗∗
−6.095∗∗
−3.282 0.832 6.060 1.391 −8.506∗∗ 4.337∗0.467 −0.275
(1.902) (2.447) (2.814) (3.713) (1.905) (3.763) (2.283) (3.524) (2.446) (2.694) (3.908)
Akaike Inf. Crit. 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822 4,232.822
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 70
Table A.13: Activity type - Episode 6 (1/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Work hours −0.103∗∗∗
−0.095∗∗∗
−0.223∗∗∗
−0.101 −0.052 −0.129∗∗ 0.041 −0.033 −0.072∗
−0.026 −0.143
(0.030) (0.032) (0.054) (0.062) (0.042) (0.054) (0.089) (0.066) (0.043) (0.052) (0.625)
Type prof 2 0.283 0.236 0.269 −1.411 −0.643 −0.516 −0.401 −1.523 0.193 −0.199 −1.079
(0.641) (0.594) (0.841) (1.008) (0.954) (1.360) (9.159) (1.329) (1.074) (0.862) (8.028)
Type prof 3−3.706∗∗∗
−0.033 0.135∗∗∗
−1.416∗∗∗
−2.926∗∗∗
−2.203∗∗∗ 0.145∗∗∗
−1.510∗∗∗
−2.282∗∗∗
−1.231∗∗∗ 0.091∗∗∗
(0.023) (0.021) (0.018) (0.082) (0.040) (0.217) (0.0005) (0.041) (0.026) (0.053) (0.00001)
Type prof 4 0.009 0.259 −0.302 −1.886∗∗
−0.144 −2.336 3.911 −1.014 0.773 −0.384 4.648
(0.608) (0.571) (0.817) (0.961) (0.815) (1.548) (3.582) (1.038) (0.954) (0.854) (5.536)
Type prof 5−0.904 0.612 −1.298 −2.541∗∗
−0.854 −1.444 5.635 −0.225 0.148 −0.171 2.142
(0.696) (0.648) (1.158) (1.290) (0.947) (1.476) (3.575) (1.187) (1.048) (0.976) (7.479)
Type prof 6−1.329 0.632 0.470 −7.295∗∗∗ 0.189 −0.590 4.409 −6.824∗∗∗ 0.210 0.845 1.119∗∗∗
(1.103) (0.970) (1.522) (0.047) (1.252) (2.433) (3.680) (0.011) (1.569) (1.343) (0.029)
Type prof 7−0.376 −5.926∗∗∗
−8.239∗∗∗
−0.638 −3.949∗∗∗
−0.577 −0.771∗∗
−5.102∗∗∗
−0.184∗∗∗
−6.257∗∗∗ 0.109∗∗∗
(2.272) (0.109) (0.064) (3.707) (0.192) (0.875) (0.389) (0.404) (0.054) (0.093) (0.0001)
Type prof 8−0.573 0.779 −0.834 −9.989∗∗∗
−0.223 0.792 3.115 −2.931∗0.102 0.547 1.136
(0.829) (0.783) (1.205) (0.012) (1.110) (1.844) (3.686) (1.645) (1.316) (1.123) (3.768)
Type prof 9−1.404 −0.335 −2.737 −9.707∗∗∗
−0.288 −2.783∗∗∗ 5.077 −1.007 −5.528∗∗∗
−0.253 −2.602
(1.164) (1.023) (1.691) (0.007) (1.350) (0.320) (3.771) (1.727) (0.015) (1.621) (1.707)
Type prof 10 −0.527 −0.182 −2.656 −4.862∗∗∗ 0.751 3.326 −3.586∗∗∗
−8.224∗∗∗
−6.401∗∗∗ 0.868 −0.222
(1.170) (1.347) (2.506) (0.278) (1.385) (4.342) (0.060) (0.015) (0.006) (1.608) (1.309)
Transp subs 2 0.056 0.237 1.331∗∗ 0.914 −0.311 −7.667 1.452 −1.326 0.297 1.113∗
−4.961
(0.459) (0.443) (0.618) (0.732) (0.747) (12.215) (0.938) (1.165) (0.682) (0.587) (5.437)
Transp subs 3−0.581 −0.662 −0.723 0.244 0.315 −0.039 −5.472∗∗∗ 0.537 0.725 −0.107 −1.403
(0.551) (0.466) (1.002) (1.030) (0.601) (1.278) (0.153) (0.925) (0.573) (0.748) (3.544)
Gender −0.182 0.113 −0.634 0.601 0.026 0.664 −0.785 −1.368 0.437 0.716 1.141
(0.387) (0.363) (0.646) (0.747) (0.507) (0.979) (0.880) (0.949) (0.557) (0.543) (5.655)
age −0.056∗∗
−0.010 −0.024 −0.097∗
−0.015 0.073 −0.046 0.075 0.010 −0.070∗∗ 0.221
(0.024) (0.022) (0.038) (0.053) (0.032) (0.058) (0.055) (0.049) (0.032) (0.036) (0.500)
Marital 2−0.480 −0.673 0.417 −0.024 −1.015 −1.258 0.038 0.652 −1.030 0.938 −3.589
(0.591) (0.577) (1.010) (1.204) (0.741) (1.915) (1.075) (1.408) (0.803) (0.866) (4.182)
Marital 3−1.634∗
−1.285∗0.509 −0.715 −1.031 −9.032∗∗∗
−0.403 −5.890∗∗∗
−0.564 0.360 −4.191∗∗∗
(0.951) (0.699) (1.226) (1.622) (0.940) (0.057) (1.442) (0.101) (1.007) (1.009) (1.364)
Marital 4−0.101 −10.553∗∗∗
−3.660∗∗∗
−5.056∗∗∗
−10.127∗∗∗
−0.248 −3.807∗∗∗
−2.708∗∗∗
−6.053∗∗∗
−11.586∗∗∗ 0.125
(1.830) (0.005) (0.057) (0.017) (0.006) (2.136) (0.056) (0.039) (0.002) (0.001) (0.607)
Akaike Inf. Crit. 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 71
Table A.14: Activity type - Episode 6 (2/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Education 2a 3.559∗∗∗
−2.476 −6.031∗∗
−4.796∗∗∗
−3.843∗3.225 2.245 −4.391∗∗∗
−6.628∗∗∗
−19.751∗∗∗
−1.204
(0.859) (2.186) (2.740) (0.096) (2.050) (2.778) (1.805) (0.349) (0.001) (0.001) (2.928)
Education 2b 4.813∗∗∗
−0.576 −3.539 4.721 −1.677 −1.032 1.225 3.487 2.815∗∗∗
−3.660∗∗
−3.058
(0.595) (2.067) (2.241) (3.000) (1.615) (3.420) (1.813) (5.271) (0.820) (1.528) (7.252)
Education 3a 3.630∗∗∗
−1.172 −3.626 −2.111 −2.616 3.107 −0.439 −3.143 2.427∗∗∗
−4.166∗∗
−0.249
(0.619) (2.078) (2.328) (8.902) (1.720) (2.433) (1.907) (3.920) (0.848) (1.636) (5.832)
Education 3b 5.311∗∗∗ 0.684 −1.846 0.775 −1.821 −3.563 0.295 −3.325∗2.993∗∗∗
−3.152∗
−0.044
(0.640) (2.108) (2.396) (9.302) (1.769) (2.309) (1.914) (2.011) (0.860) (1.718) (2.480)
Education 4 4.191∗∗∗
−0.062 −4.104∗4.489 −2.049 4.036∗
−0.085 3.391 3.015∗∗∗
−3.165∗∗
−1.519
(0.476) (2.060) (2.299) (2.857) (1.693) (2.373) (1.695) (5.270) (0.613) (1.554) (5.801)
Education 5 3.376∗∗∗
−0.142 −4.549∗2.501 −2.336 3.537 −5.493 2.849 2.451∗∗∗
−2.963∗
−0.285
(0.526) (2.072) (2.359) (2.862) (1.725) (2.424) (7.445) (5.278) (0.663) (1.616) (5.752)
Location type 2−0.075 −0.591 −0.904 −3.535∗∗∗ 6.679∗∗∗ 3.548 0.229 4.326∗∗∗ 0.104 −0.762 −1.241
(0.911) (0.796) (1.095) (1.044) (0.953) (4.080) (1.292) (1.491) (1.124) (0.914) (5.453)
Location type 3 0.311 −0.018 −0.284 −9.766∗∗∗ 7.538∗∗∗ 4.835 −0.502 4.727∗∗∗
−0.037 −1.567 −1.960
(1.081) (1.004) (1.435) (0.005) (1.053) (4.161) (1.803) (1.724) (1.505) (1.343) (3.288)
Location type 4 0.274 −0.174 −0.724 −2.263∗∗ 7.235∗∗∗ 4.332 −0.750 4.832∗∗∗
−0.581 −0.746 −2.307
(0.928) (0.816) (1.114) (1.035) (0.975) (4.096) (1.480) (1.512) (1.210) (0.951) (6.410)
Location type 5−1.144 0.072 −2.432 −2.252 −3.738∗∗∗
−0.740 −5.023∗∗∗ 5.314∗∗∗ 0.540 −0.818 3.017
(1.489) (1.112) (2.000) (1.545) (0.001) (9.190) (0.035) (1.677) (1.337) (1.376) (5.299)
Home occ 2−0.455 −0.167 −0.349 −0.844 −0.650 −9.109∗∗∗
−0.771 −0.370 −0.253 −0.021 −0.292
(0.487) (0.450) (0.749) (1.013) (0.631) (2.471) (1.011) (0.981) (0.749) (0.638) (7.751)
Home occ 3−1.525 −1.441∗
−7.124∗∗∗ 0.621 −1.437 −2.617 −7.269∗∗∗
−6.974∗∗∗ 0.435 0.010 −3.821
(0.928) (0.847) (0.009) (1.539) (1.266) (2.740) (0.004) (0.005) (0.865) (1.013) (3.274)
N cars 2 0.227 0.196 −0.601 −0.497 0.037 0.312 −1.183 0.139 0.065 −0.118 −0.915
(0.520) (0.505) (0.782) (0.963) (0.666) (1.722) (0.912) (1.030) (0.741) (0.719) (5.321)
N cars 3 0.280 −0.114 −0.901 −0.060 0.096 −0.277 −0.971 0.564 0.274 −0.231 −1.084
(0.688) (0.677) (1.009) (1.318) (0.900) (2.129) (1.426) (1.259) (0.868) (1.011) (4.510)
N cars 4−0.362 −1.849 −0.943 −10.351∗∗∗
−5.653∗∗∗ 4.052∗
−7.753∗∗∗
−4.667∗∗∗ 1.652 1.250 −3.910∗∗∗
(1.181) (1.186) (1.766) (0.002) (0.005) (2.207) (0.007) (0.034) (1.392) (1.263) (0.886)
N cars 5−4.924∗∗∗ 0.190∗∗∗
−0.333∗∗∗
−1.956∗∗∗
−4.364∗∗∗
−3.569 −4.707∗∗∗
−3.244∗∗∗
−5.514∗∗∗
−1.428∗∗∗ 0.111∗∗∗
(0.007) (0.050) (0.049) (0.067) (0.010) (11.169) (0.009) (0.011) (0.005) (0.028) (0.001)
N cars 6−11.901∗∗∗
−1.925 −10.705∗∗∗
−2.100∗∗∗
−6.849∗∗∗
−3.907∗∗∗
−3.731∗∗∗
−4.433∗∗∗
−5.817∗∗∗ 0.148 −0.471
(0.001) (1.760) (0.007) (0.102) (0.007) (0.127) (0.012) (0.021) (0.005) (2.056) (1.112)
Akaike Inf. Crit. 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 72
Table A.15: Activity type - Episode 6 (3/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Income 2−1.687∗
−1.279 −2.893∗3.191∗∗
−0.374 −10.412∗∗∗ 2.939 1.338 5.315∗∗∗
−1.491 −3.975
(0.883) (1.019) (1.665) (1.459) (1.091) (0.534) (2.276) (1.750) (0.738) (1.017) (7.136)
Income 3−1.672∗
−0.688 −1.961 4.318∗∗∗
−0.487 −0.802 2.246 1.081 5.475∗∗∗
−2.339∗∗
−1.814
(0.917) (1.049) (1.642) (1.085) (1.178) (3.139) (2.287) (1.665) (0.651) (1.093) (5.894)
Income 4−1.086 −0.348 −2.253 4.014∗∗∗
−0.129 −1.981 1.956 −0.283 5.532∗∗∗
−3.107∗∗∗
−1.978
(0.958) (1.094) (1.741) (1.061) (1.269) (3.286) (2.380) (1.681) (0.636) (1.200) (6.891)
Income 5−0.920 0.063 −1.585 4.847∗∗∗ 0.480 −1.797 3.465 −0.179 4.608∗∗∗
−1.707 −1.480
(1.019) (1.150) (1.804) (1.172) (1.369) (3.414) (2.373) (1.781) (0.853) (1.258) (6.514)
Income 6−1.105 −0.304 −2.720 4.400∗∗∗
−6.075∗∗∗
−0.891 −1.835 −6.411∗∗∗ 6.629∗∗∗
−3.973∗∗ 4.307
(1.172) (1.266) (1.982) (1.322) (0.060) (3.542) (2. 097) (0.502) (0.751) (1.737) (5.655)
PrevAct type 2−5.741∗∗∗
−13.884∗
−10.620 −5.923∗∗∗
−3.966∗∗∗
−2.789∗
−2.114 1.670 −3.921∗∗∗
−3.655∗∗∗
−6.331
(0.655) (7.126) (8.149) (1.165) (0.867) (1.445) (1.433) (4.074) (0.806) (1.202) (7.653)
PrevAct type 3 0.804 1.370 6.046∗∗∗ 3.220∗∗ 0.779 −4.355 1.500 2.134 −4.657∗∗∗ 4.179∗∗ 1.532
(1.352) (1.299) (1.692) (1.622) (1.755) (7.583) (2.256) (8.704) (1.100) (1.676) (7.483)
PrevAct type 4−1.344 −3.159∗∗∗ 0.264 −0.126 −7.648∗∗∗
−5.384∗∗∗
−3.127∗∗∗
−2.299∗∗∗
−9.302∗∗∗ 4.034∗∗∗
−1.516
(1.119) (1.175) (1.826) (1.700) (0.016) (0.098) (0.209) (0.228) (0.005) (1.409) (3.109)
PrevAct type 5−1.706 1.671 1.928 −4.804 2.170 0.971 −1.870 0.730 0.238 4.554∗∗∗ 1.463
(1.662) (1.219) (1.942) (11.807) (1.448) (2.268) (8.227) (9.168) (1.707) (1.618) (3.445)
PrevAct type 6 8.106∗∗∗ 8.423∗∗∗ 0.454∗∗∗ 9.046∗∗∗ 9.516∗∗∗ 12.377∗∗∗ 0.238∗∗∗ 1.034∗∗∗
−5.573∗∗∗ 9.192∗∗∗ 0.497∗∗∗
(0.680) (0.543) (0.011) (1.266) (0.908) (1.444) (0.001) (0.0003) (0.00001) (1.331) (0.0003)
PrevAct type 7−1.065 0.337 −2.213 −7.117∗∗∗
−4.951∗∗∗ 3.335 −5.873∗∗∗
−0.517 −7.762∗∗∗
−3.218∗∗∗
−1.352
(1.628) (1.321) (10.463) (0.060) (0.114) (2.185) (0.056) (0.355) (0.015) (0.311) (1.789)
PrevAct type 8−16.332∗∗∗
−3.786∗∗∗
−0.069 −8.296∗∗∗
−10.862∗∗∗
−4.989∗∗∗
−6.370∗∗∗
−4.629∗∗∗
−12.288∗∗∗
−6.351∗∗∗ 2.061
(0.00004) (1.067) (1.775) (0.024) (0.002) (0.070) (0.025) (0.044) (0.001) (0.030) (6.910)
PrevAct type 9−2.282 −0.170 −4.751∗∗∗
−6.971∗∗∗ 0.231 −2.836 −4.501∗∗∗
−2.744∗∗∗
−10.145∗∗∗
−4.127∗∗∗
−0.766
(1.394) (0.975) (0.343) (0.075) (1.330) (12.928) (0.086) (0.230) (0.003) (0.323) (2.965)
PrevAct type 10 −1.419∗∗
−1.580∗∗ 0.866 −1.854 −1.806∗
−1.393 −1.450 2.148 −3.456∗∗∗
−1.039 2.984
(0.666) (0.618) (1.254) (1.142) (1.032) (1.861) (1.883) (4.163) (1.334) (1.285) (6.237)
PrevAct type 11 −11.038∗∗∗
−1.976 −3.157 −5.708∗∗∗
−7.021∗∗∗
−2.556∗∗∗
−1.379∗∗
−2.504∗∗∗
−6.736∗∗∗ 2.080 −1.569∗∗∗
(0.003) (1.378) (12.616) (0.096) (0.038) (0.240) (0.542) (0.221) (0.026) (1.934) (0.017)
Children 0.029 −0.216 −0.189 −0.174 0.160 −1.101 −0.183 −0.747 −1.158∗∗∗ 0.017 −3.960
(0.329) (0.312) (0.515) (0.619) (0.463) (0.772) (0.773) (0.631) (0.445) (0.467) (6.862)
Constant 6.183∗∗∗ 8.413∗∗∗ 16.049∗∗∗ 3.764 −0.807 −3.992 −6.667∗
−12.532∗∗∗
−5.248∗∗ 6.736∗
−7.527∗∗∗
(2.126) (3.208) (4.627) (3.809) (2.709) (4.012) (3.482) (4.037) (2.352) (4.005) (2.787)
Akaike Inf. Crit. 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514 2,944.514
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 73
Table A.16: Activity type - Episode 7 (1/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Work hours −0.065 −0.075 0.134 −0.403 0.655 −0.803 −0.514 0.216 0.009 −0.063 −0.352
(0.050) (0.053) (0.201) (0.759) (0.413) (0.805) (0.354) (0.204) (0.113) (0.073) (3.107)
Type prof 2 2.444∗∗ 2.287∗∗ 1.494 1.568 14.587∗∗∗ 12.739∗∗∗ 11.896∗19.532∗∗∗
−1.663 0.287 4.133∗∗∗
(1.221) (1.038) (1.615) (7.871) (3.824) (4.655) (6.782) (3.215) (2.382) (1.339) (0.238)
Type prof 4 1.329 0.768 −1.507 2.926 7.516 5.334 −7.353 11.111∗∗∗
−4.665∗0.085 −0.749∗∗∗
(1.281) (1.035) (1.744) (6.199) (5.215) (9.024) (8.967) (0.197) (2.755) (1.328) (0.015)
Type prof 5−0.219 −0.381 1.497 −0.369 10.010∗∗ 10.195 12.100∗∗ 21.627∗∗∗
−2.058 −0.277 0.619∗∗∗
(1.426) (1.164) (2.286) (10.054) (5.051) (12.771) (5.069) (2.975) (2.615) (1.682) (0.023)
Type prof 6 1.700 −1.588 3.499 1.565 10.823∗24.359∗∗
−8.816 −3.356∗∗∗
−22.289∗∗∗
−14.088∗∗∗
−5.412∗∗∗
(1.847) (1.845) (3.070) (4.729) (5.902) (9.629) (10.886) (0.000) (0.00000) (0.00000) (0.00004)
Type prof 7−1.234∗∗∗ 3.821∗∗∗
−0.373∗∗∗ 0.635 −1.378 0.578∗∗∗ 0.555∗∗∗ 0.141 0.225∗∗∗
−0.432∗∗∗ 0.186∗∗∗
(0.001) (0.004) (0.000) (0.000) (0.000) (0.001) (0.001) (0.00000)
Type prof 8 2.433 0.645 4.650∗14.198∗20.656∗∗∗
−3.475∗∗ 11.974∗29.088∗∗∗
−1.401 2.049 1.517∗∗∗
(1.686) (1.551) (2.723) (7.583) (2.911) (1.659) (6.476) (2.360) (2.863) (2.209) (0.138)
Type prof 9 4.619∗∗
−0.638 5.225∗1.867 30.238∗∗∗
−0.027 −2.945∗∗∗ 30.424∗∗∗
−12.275∗∗∗ 4.732∗∗
−1.910∗∗∗
(1.997) (1.922) (2.966) (4.454) (4.800) (8.416) (0.073) (3.156) (0.0001) (2.379) (0.00003)
Type prof 10 −19.836∗∗∗
−5.005∗2.175 1.246∗∗∗ 28.396∗∗∗
−0.270 −9.805∗∗∗
−2.310∗∗∗
−7.770∗∗∗
−18.836∗∗∗
−3.415∗∗∗
(0.00000) (2.934) (4.343) (0.001) (5.819) (8.783) (0.030) (0.000) (0.004) (0.00000) (0.00000)
Transp subs 2−0.774 0.708 2.874∗∗
−6.686 1.945 11.313 5.985 −1.886 −2.516 1.419 2.477∗∗∗
(0.900) (0.720) (1.176) (8.631) (1.940) (8.946) (6.465) (7.203) (1.995) (1.007) (0.169)
Transp subs 3−1.886 0.636 −0.593 9.549 −16.533∗∗∗ 4.881 −26.002∗∗
−11.001∗∗∗ 1.995 0.214 7.596∗∗∗
(1.471) (1.206) (2.049) (10.431) (0.011) (7.743) (12.434) (0.073) (1.850) (1.605) (0.260)
Gender 0.350 0.125 −1.966 0.143 −6.053 −10.216 −7.101 2.417 −0.361 0.985 −3.663∗∗∗
(0.673) (0.670) (1.560) (9.517) (4.112) (10.709) (5.185) (1.963) (1.459) (1.070) (0.325)
age −0.063 −0.014 0.066 −0.144 −0.215 −0.471 −0.982∗∗∗ 0.007 0.155 0.128∗
−0.020
(0.045) (0.043) (0.076) (0.889) (0.166) (0.920) (0.370) (0.142) (0.117) (0.071) (3.112)
Marital 2 0.376 −0.481 −2.691 0.674 4.436 11.533 6.101 −1.947 −3.314 −1.070 4.496∗∗∗
(1.095) (1.044) (2.018) (10.318) (3.857) (9.132) (6.656) (2.566) (2.983) (1.565) (0.239)
Marital 3 2.574∗∗
−0.084 0.158 0.395 −9.113∗∗∗ 7.519∗29.352∗∗
−5.041∗∗∗
−25.132∗∗∗ 0.599 3.034∗∗∗
(1.289) (1.364) (2.138) (1.898) (0.090) (4.125) (12.727) (0.007) (0.00000) (1.884) (0.0003)
Marital 4−6.940∗∗∗
−1.917 −5.562∗∗∗ 0.220∗∗∗ 0.424∗∗∗
−3.757∗∗∗ 0.873∗∗∗
−1.050∗∗∗
−2.373∗∗∗
−16.101∗∗∗ 0.155∗∗∗
(0.038) (6.501) (0.00000) (0.00000) (0.000) (0.00000) (0.00000) (0.010) (0.004) (0.0001) (0.000)
Akaike Inf. Crit. 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 74
Table A.17: Activity type - Episode 7 (2/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Education 2a −11.814∗∗∗ 2.536 −10.027∗∗∗ 3.415∗∗∗ 3.428∗∗∗ 14.247 4.213∗∗∗
−4.805∗∗∗
−13.317∗∗∗
−9.844∗∗∗ 0.643∗∗∗
(0.00000) (4.187) (0.002) (0.045) (0.159) (12.708) (0.153) (0.007) (0.001) (0.00000) (0.002)
Education 2b 3.033∗∗ 0.239 −0.864 −6.196∗∗∗
−24.097∗∗∗ 10.892 3.111 −7.926∗∗∗
−0.429 8.877∗∗∗
−3.502∗∗∗
(1.528) (3.452) (3.944) (1.437) (0.693) (9.687) (5.196) (2.945) (2.514) (1.819) (0.124)
Education 3a 5.892∗∗∗ 2.112 0.906 −12.122∗∗
−5.726 2.182 −0.072 −5.138∗∗∗ 1.229 10.776∗∗∗ 1.383∗∗∗
(1.335) (3.635) (4.074) (5.875) (5.878) (9.782) (4.065) (1.722) (2.328) (1.795) (0.016)
Education 3b 4.775∗∗∗
−1.091 0.732 −1.046 3.487 −2.397 12.293∗
−12.076∗∗∗ 1.894 9.605∗∗∗
−1.269∗∗∗
(1.326) (3.671) (4.400) (5.872) (4.778) (6.835) (6.394) (0.099) (2.173) (1.662) (0.009)
Education 4 4.529∗∗∗
−0.647 0.202 4.610 −1.343 4.956 −0.981 −4.082∗∗
−5.151 9.934∗∗∗ 0.113∗∗
(1.172) (3.606) (4.007) (8.630) (4.564) (8.762) (3.524) (2.071) (6.342) (1.429) (0.051)
Education 5 4.328∗∗∗
−0.520 0.024 14.257∗∗ 6.110 −0.206 −4.376 −2.368 2.147 11.791∗∗∗
−0.689∗∗∗
(1.194) (3.695) (4.356) (6.496) (4.516) (4.494) (7.251) (1.908) (2.393) (1.301) (0.243)
Location type 2−1.868∗
−1.729 0.406 0.812 1.006 20.327∗∗∗
−2.151 4.112∗15.603∗∗∗ 0.246 −2.581∗∗∗
(1.094) (1.148) (2.312) (8.194) (5.181) (7.512) (3.740) (2.227) (1.787) (1.653) (0.050)
Location type 3−1.793 −1.845 −22.529∗∗∗ 4.393 −14.376 12.289∗∗∗ 6.122 −6.992∗∗∗ 14.819∗∗∗ 0.567 −1.559∗∗∗
(1.479) (1.622) (0.000) (9.891) (12.898) (0.023) (4.317) (0.003) (1.719) (2.183) (0.002)
Location type 4−1.874 −0.907 0.950 1.357 −1.113 12.560∗0.550 3.505 15.671∗∗∗ 1.665 0.702∗∗∗
(1.148) (1.169) (2.317) (8.707) (5.524) (6.727) (4.359) (2.726) (1.797) (1.678) (0.235)
Location type 5−1.410 −0.873 −24.166∗∗∗ 7.347 −2.651 4.814∗∗∗ 2.615 11.728∗∗∗
−8.506∗∗∗
−22.331∗∗∗ 2.271∗∗∗
(1.586) (1.607) (0.00000) (6.632) (12.011) (0.074) (10.421) (2.672) (0.000) (0.00000) (0.001)
Home occ 2−0.142 0.415 −0.041 −3.739 0.950 3.714 8.170∗0.679 1.627 −0.900 −3.788∗∗∗
(0.877) (0.912) (1.429) (8.380) (2.172) (9.042) (4.595) (1.716) (2.112) (1.827) (0.0004)
Home occ 3 0.540 0.749 −12.244∗∗∗
−3.887∗∗∗ 13.646 27.267∗∗∗
−20.984∗∗∗
−4.576∗∗∗
−19.247∗∗∗ 2.930 0.157∗∗∗
(1.666) (1.585) (0.0001) (0.008) (11.632) (8.318) (0.081) (0.004) (0.00000) (2.333) (0.002)
N cars 2−0.682 1.369 0.101 7.432 −2.397 −13.225 11.427 1.157 −1.328 2.130 4.328∗∗∗
(0.992) (0.991) (1.525) (6.455) (3.431) (11.966) (9.963) (2.888) (2.782) (1.397) (0.173)
N cars 3 0.515 0.207 −4.050∗17.868∗∗
−10.170 −18.790∗∗∗
−2.196∗∗∗ 2.836 2.462 −27.389∗∗∗ 5.333∗∗∗
(1.170) (1.219) (2.093) (7.404) (15.627) (6.250) (0.078) (3.305) (3.206) (0.000) (0.065)
N cars 4−0.478 −2.028 −21.601∗∗∗
−3.875∗∗∗ 5.988 −33.189∗∗∗ 5.489∗∗∗
−9.244∗∗∗
−9.854∗∗∗
−13.734∗∗∗
−0.743∗∗∗
(1.738) (1.671) (0.00000) (0.002) (11.744) (0.541) (0.759) (0.001) (0.002) (0.00003) (0.0001)
N cars 6−5.896∗∗∗ 3.258 −0.798∗∗∗
−6.510∗∗∗
−0.245∗∗∗
−10.361∗∗∗ 0.248 −4.581∗∗∗
−1.868∗∗∗
−0.975∗∗∗ 0.110∗∗∗
(0.008) (6.354) (0.007) (0.008) (0.0001) (0.059) (0.429) (0.063) (0.000) (0.000) (0.000)
Akaike Inf. Crit. 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 75
Table A.18: Activity type - Episode 7 (3/3)
Dependent variable:
1234567891011
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Income 2 1.435 0.946 3.427 −24.638∗∗∗ 4.661 −9.708 15.873∗∗
−5.338∗2.845 −0.639 8.960∗∗∗
(1.983) (1.709) (4.318) (5.650) (6.208) (12.222) (6.573) (2.984) (2.360) (2.932) (0.137)
Income 3 2.562 0.620 8.360∗
−19.125∗∗∗ 11.024∗∗
−25.002∗∗ 15.003∗∗
−1.842 4.370∗∗ 1.238 −3.159∗∗∗
(1.995) (1.791) (4.795) (7.097) (5.139) (10.676) (5.875) (1.633) (1.864) (2.822) (0.0004)
Income 4 2.607 1.063 8.134 −15.287∗9.947∗∗
−16.553 7.068 −0.571 5.694∗∗∗ 0.361 −3.923∗∗∗
(2.085) (1.929) (4.949) (8.770) (4.568) (12.232) (5.693) (2.018) (1.749) (2.940) (0.014)
Income 5 2.943 1.035 4.768 −21.979∗∗ 7.695 −14.804∗14.880∗∗ 3.055 5.407∗∗∗
−0.715 −3.428∗∗∗
(2.148) (2.029) (5.010) (9.826) (5.174) (8.347) (6.515) (2.645) (1.695) (3.173) (0.002)
Income 6 3.083 0.689 8.897∗
−35.253∗∗∗ 11.593∗∗
−2.237 18.339∗
−5.418∗∗∗
−7.715∗∗∗ 0.532 −1.003∗∗∗
(2.366) (2.243) (5.327) (5.877) (5.032) (9.058) (10.362) (0.023) (0.004) (3.284) (0.060)
PrevAct type 2−7.414∗∗∗
−12.446∗∗∗ 8.121∗∗∗
−5.926 −10.753∗∗∗
−11.509 −5.335 0.532 −6.280∗∗∗
−7.985∗∗∗
−5.166∗∗∗
(1.629) (1.945) (2.492) (12.066) (2.520) (10.333) (5.779) (3.262) (2.376) (1.916) (0.021)
PrevAct type 3−1.016 −1.707 19.850∗∗∗ 2.155 −1.633 −11.521∗∗
−22.865∗∗∗
−8.308∗∗∗
−18.562∗∗∗
−0.754 5.016∗∗∗
(2.030) (1.934) (2.197) (10.372) (2.631) (4.516) (0.003) (0.112) (0.00001) (2.214) (0.165)
PrevAct type 4−29.679∗∗∗
−6.448∗∗∗ 17.574∗∗∗
−9.159∗∗∗
−12.473∗∗∗
−22.099∗∗∗ 3.538 −6.010∗∗∗
−20.665∗∗∗
−0.750 −4.262∗∗∗
(0.000) (1.862) (2.212) (0.00003) (0.00004) (0.048) (7.115) (0.001) (0.00000) (2.113) (0.00001)
PrevAct type 5 13.543∗∗∗ 15.461∗∗∗ 32.514∗∗∗ 4.766∗∗∗ 28.356∗∗∗
−9.456 34.791∗∗∗ 9.021∗∗∗
−2.150∗∗∗ 13.366∗∗∗ 3.596∗∗∗
(2.260) (2.050) (2.768) (0.0004) (5.018) (6.365) (5.359) (0.001) (0.0001) (2.644) (0.000)
PrevAct type 6−1.315 −2.121 −4.559∗∗∗
−3.194∗∗∗ 0.982 −6.198 5.057∗∗∗
−5.272∗∗∗
−19.539∗∗∗
−1.781 0.225∗∗∗
(2.349) (2.223) (0.00000) (1.019) (7.982) (5.898) (0.426) (0.001) (0.00001) (2.675) (0.0001)
PrevAct type 7−24.061∗∗∗
−3.460∗
−20.451∗∗∗ 2.316∗∗∗
−8.030∗∗∗
−16.397∗∗∗
−6.669∗∗∗
−14.611∗∗∗
−15.455∗∗∗
−25.523∗∗∗ 0.022∗∗∗
(0.00000) (1.792) (0.000) (0.006) (0.0004) (0.260) (0.005) (0.00001) (0.00002) (0.00000) (0.002)
PrevAct type 8−4.758∗∗
−50.987∗∗∗
−15.193∗∗∗
−19.070∗∗∗
−16.474∗∗∗
−4.525∗∗∗ 13.935∗∗ 5.238∗
−26.263∗∗∗
−26.315∗∗∗ 2.857∗∗∗
(1.996) (0.000) (0.000) (0.0003) (0.00002) (0.004) (5.974) (2.684) (0.00000) (0.00000) (0.032)
PrevAct type 9−21.400∗∗∗
−1.379 −1.532∗∗∗
−11.692∗∗∗
−15.066∗∗∗
−21.764∗∗∗
−8.787∗∗∗
−3.035∗∗∗
−20.271∗∗∗
−20.083∗∗∗ 2.379∗∗∗
(0.00000) (1.888) (0.00000) (0.014) (0.00004) (1.634) (0.051) (0.052) (0.00000) (0.00000) (0.292)
PrevAct type 10 −4.161∗∗
−4.055∗∗ 13.774∗∗∗
−11.229∗∗∗
−20.620∗∗∗
−16.742∗∗∗ 5.113 −13.498∗∗∗
−28.513∗∗∗
−30.230∗∗∗
−10.001∗∗∗
(1.702) (1.603) (2.468) (0.044) (5.385) (0.824) (4.322) (0.011) (0.00000) (0.000) (0.00000)
PrevAct type 11 4.720 3.565 0.130∗∗∗ 34.392∗∗∗
−0.239∗∗∗ 0.415∗∗∗ 2.106 0.911∗∗∗ 0.285∗∗∗
−13.253∗∗∗ 0.714∗∗∗
(3.244) (3.244) (0.000) (6.039) (0.00000) (0.000) (0.00000) (0.000) (0.00001) (0.00000)
Children −0.552 0.093 1.896 5.296 −0.413 −0.316 −2.076 5.627∗
−3.430∗
−0.022 −3.128∗∗∗
(0.561) (0.559) (1.183) (10.748) (1.410) (10.956) (6.306) (3.114) (2.041) (0.791) (0.123)
Constant 1.872 8.240 −30.347∗∗∗ 1.830 −40.165∗∗∗ 31.589∗∗∗ 11.516 −43.590∗∗∗
−19.340∗∗∗
−14.578∗∗∗
−2.957∗∗∗
(3.709) (5.731) (9.627) (3.322) (8.993) (3.049) (8.664) (4.375) (4.440) (5.216) (0.063)
Akaike Inf. Crit. 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291 1,567.291
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 76
Table A.19: Activity type - Episode 8 (1/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Work hours −0.079 −0.170 −0.681∗
−0.906 0.240 −0.042 −1.723 −2.837 0.622 −2.469∗∗∗
−0.260
(0.057) (0.125) (0.397) (1.781) (6.502) (1.766) (1.616) (2.249) (0.628) (0.799) (4.468)
Type prof 2−2.486∗
−3.909∗
−28.213∗∗∗
−19.123∗∗∗
−17.104∗∗∗
−6.408 1.713∗∗∗ 4.562∗∗∗
−38.436∗1.428 −8.943
(1.390) (2.068) (4.649) (1.294) (0.074) (14.159) (0.004) (1.413) (22.135) (6.598) (7.107)
Type prof 4−2.550∗∗
−3.709∗
−13.273 11.873∗∗∗
−10.352∗∗∗
−10.815∗∗∗ 6.467 1.361 22.814∗1.088 3.461
(1.293) (2.029) (13.339) (1.473) (0.197) (0.021) (17.554) (1.885) (13.098) (11.231) (4.272)
Type prof 5−2.225 −2.290 −38.297∗∗∗
−19.152∗∗∗ 4.677∗∗∗ 6.721∗∗∗
−10.456 −17.139∗∗∗ 4.344 −15.664∗∗∗
−4.980∗∗∗
(1.442) (2.379) (0.003) (1.866) (0.007) (0.079) (16.239) (4.819) (9.870) (0.061) (0.012)
Type prof 6 1.094 −1.285 −13.608∗∗∗ 2.344∗∗∗
−2.414∗∗∗
−1.505∗∗∗ 22.472∗∗∗ 4.065∗∗∗ 5.560∗∗∗ 35.564∗∗∗ 2.828∗∗∗
(2.491) (3.615) (0.00000) (0.00000) (0.018) (0.375) (5.636) (0.00001) (0.000) (10.485) (0.0002)
Type prof 7−5.203∗∗∗ 0.584 0.430∗∗∗ 0.378∗∗∗ 0.237∗∗∗
−0.544∗∗∗ 0.522∗∗∗ 0.476∗∗∗ 0.495∗∗∗ 0.401∗∗∗
−0.742∗∗∗
(0.001) (0.000) (0.000) (0.000) (0.00003) (0.000) (0.000) (0.000) (0.000) (0.000)
Type prof 8−3.334 −7.898∗∗
−4.637 1.409∗∗∗
−9.259∗∗∗ 0.333 −22.631∗∗∗
−0.385 0.904 25.381∗∗∗ 2.324
(2.298) (3.372) (4.167) (0.0001) (0.161) (0.281) (0.000) (0.700) (13.630) (8.482) (4.023)
Type prof 9−46.101 −0.448 −24.069∗∗∗ 1.858∗∗∗ 17.983∗∗∗ 33.334∗∗∗
−8.711∗∗∗ 2.094∗∗∗
−22.019∗∗∗
−5.069∗∗∗
−1.600
(5.474) (0.00000) (0.008) (0.144) (0.931) (0.905) (0.000) (0.053) (0.000) (12.408)
Type prof 10 10.587∗∗ 28.473∗∗∗
−7.671∗∗∗ 1.117∗∗∗ 4.131∗∗∗ 3.450∗∗∗
−1.034 1.965∗∗∗ 2.919∗∗∗ 2.728∗∗∗ 1.265∗∗∗
(4.524) (1.056) (0.001) (0.00000) (0.000) (0.000) (1.056) (0.00001) (0.00000) (0.000) (0.00000)
Transp subs 2 0.015 3.365∗∗
−11.910 −11.536∗∗∗
−5.213∗∗∗ 10.784∗11.768∗
−2.316∗∗∗
−18.269∗∗∗
−7.428∗∗
−12.054∗∗∗
(1.011) (1.689) (11.617) (0.0001) (0.639) (6.207) (6.855) (0.140) (0.069) (3.308) (0.266)
Transp subs 3 2.627∗∗ 4.336 6.859∗∗∗
−0.764∗∗∗
−4.558∗∗∗
−9.671∗∗∗ 9.274∗∗∗ 36.930∗∗∗ 27.424∗∗ 18.227∗∗
−2.107∗∗∗
(1.301) (4.258) (2.347) (0.024) (0.047) (0.835) (0.0001) (3.961) (11.667) (8.326) (0.003)
Gender 1.570∗2.446∗
−12.241∗∗∗
−0.793 1.007∗∗
−3.414 12.693 −16.093∗∗∗ 2.955 −21.541∗
−11.449∗∗∗
(0.913) (1.367) (0.081) (3.232) (0.406) (5.889) (18.412) (4.354) (9.066) (12.672) (1.438)
age 0.001 0.043 0.030 0.591 0.558 −0.903 0.726 1.282 0.637 0.536 0.592
(0.050) (0.073) (0.157) (1.122) (5.516) (1.538) (1.288) (1.389) (0.677) (1.144) (3.414)
Marital 2 0.161 −0.367 14.389∗∗
−1.805 9.694∗∗∗ 7.948∗∗∗
−6.677 14.716∗∗∗
−5.439 −18.631∗∗ 1.677
(1.301) (2.580) (6.099) (1.663) (0.144) (0.704) (7.122) (0.924) (9.946) (7.718) (5.831)
Marital 3−1.986 −4.283 2.591 −3.879∗∗∗
−13.526∗∗∗
−10.594∗∗∗
−9.060∗∗∗
−8.551∗∗∗ 22.011 −10.792∗∗∗
−7.101
(2.140) (2.864) (6.100) (0.024) (0.075) (0.00000) (1.175) (0.008) (15.222) (3.822) (7.001)
Marital 4−7.429∗∗∗ 0.549∗∗∗ 1.049∗∗∗ 0.383 −0.043∗∗∗ 0.742∗∗∗
−3.409∗∗∗ 0.145∗∗∗ 78.228∗∗∗
−0.049∗∗∗ 0.538∗∗∗
(0.00000) (0.000) (0.000) (0.000) (0.000) (0.00000) (0.000) (0.063) (0.000) (0.000)
Akaike Inf. Crit. 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 77
Table A.20: Activity type - Episode 8 (2/3)
Dependent variable:
1234567891011
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
Location type 2 20.082∗∗∗
−7.105 8.580∗1.264∗
−17.041∗∗∗ 9.886 −18.803 −12.163∗∗∗
−6.771 3.829 3.043
(1.514) (7.225) (4.977) (0.674) (0.202) (10.915) (15.411) (2.042) (9.578) (9.309) (10.475)
Location type 3 17.695∗∗∗
−14.508 12.157∗∗∗
−0.219∗
−27.031∗∗∗ 9.724∗∗∗ 0.125 −2.584∗∗∗ 21.068∗
−3.680 2.713∗∗∗
(3.814) (9.255) (0.000) (0.133) (0.0004) (0.062) (2.003) (0.119) (10.994) (8.661) (0.019)
Location type 4 21.977∗∗∗
−5.050 21.967∗∗∗
−9.488∗∗∗
−11.626∗∗∗ 17.528 10.153 19.910∗∗∗ 7.559 22.375∗∗ 2.549
(1.689) (7.075) (4.977) (0.873) (0.219) (10.782) (13.623) (2.728) (14.462) (10.490) (10.907)
Location type 5 21.721∗∗∗
−8.080 4.125∗∗∗
−1.520∗∗∗
−13.070∗∗∗ 1.562∗∗∗
−5.884∗∗∗ 3.506∗∗∗
−3.785∗∗∗ 41.626∗∗∗
−5.353∗∗∗
(2.070) (7.935) (0.003) (0.000) (0.049) (0.000) (0.015) (0.000) (0.00002) (7.511) (0.0004)
Home occ 2−6.524∗
−4.858 −14.445∗∗ 3.046∗∗∗
−6.187∗∗∗
−25.929∗∗∗
−22.843∗∗∗
−10.074∗∗∗ 2.238 2.193 2.521∗∗∗
(3.628) (3.310) (6.097) (0.133) (0.00000) (0.023) (1.553) (1.240) (14.630) (9.872) (0.009)
Home occ 3 2.830 3.863 −4.497∗∗∗ 1.486∗∗∗ 3.305∗∗∗
−3.736∗∗∗
−34.798∗∗∗ 5.324∗∗∗
−14.252∗∗∗
−14.991∗∗∗ 0.918∗∗∗
(1.929) (3.019) (0.002) (0.00000) (0.00000) (0.007) (0.645) (0.015) (0.00001) (0.545) (0.002)
Income 2 35.706∗∗∗
−3.665 −0.261 1.088∗∗∗
−1.781∗∗∗ 22.665∗∗∗
−1.429 13.688∗∗∗
−34.810∗∗ 16.451∗∗
−6.417∗∗∗
(1.555) (6.035) (6.097) (0.0001) (0.006) (0.775) (0.905) (4.359) (17.755) (7.431) (0.010)
Income 3 37.873∗∗∗
−5.222 6.086 −7.232∗∗∗
−3.210∗∗∗
−7.294∗∗∗ 18.063 9.262∗∗∗
−31.784∗18.062 1.039
(1.138) (6.326) (6.375) (0.026) (0.104) (0.012) (13.288) (0.933) (17.072) (11.944) (3.892)
Income 4 38.266∗∗∗
−4.391 −7.923 −6.787∗∗∗
−11.684∗∗∗
−8.836∗∗∗ 14.584 −1.575 4.481 42.780∗∗∗
−4.726∗∗∗
(0.959) (6.183) (6.345) (0.135) (0.048) (0.023) (10.500) (1.491) (13.400) (11.861) (0.035)
Income 5 37.531∗∗∗
−1.304 16.754∗∗∗ 13.219∗∗∗ 13.554∗∗∗ 6.362 −7.549 −16.077∗∗∗ 20.148 21.476∗∗∗ 14.337
(1.179) (6.811) (5.848) (1.526) (0.207) (9.525) (5.356) (1.133) (14.246) (4.407) (10.702)
Income 6 39.519∗∗∗
−0.466 0.055∗∗∗
−12.695∗∗∗ 1.351∗∗∗ 18.452∗
−12.018∗∗∗ 3.619∗∗∗ 3.580∗∗∗ 10.336∗∗∗
−18.340∗∗∗
(1.286) (5.594) (0.00001) (0.0002) (0.049) (10.835) (0.00000) (0.00000) (0.003) (0.001) (0.00000)
PrevAct type 2−9.989∗∗∗
−60.635∗∗∗
−11.671∗∗∗
−0.532 −39.278∗∗∗
−22.214∗∗∗
−19.059∗2.561 −16.974 −16.157 −11.833∗∗∗
(3.534) (0.001) (1.273) (22.814) (0.041) (0.728) (10.538) (4.057) (14.726) (10.775) (4.208)
PrevAct type 3−2.167 −3.573 38.782∗∗∗ 10.114∗∗∗
−18.083∗∗∗ 5.899 −3.953 8.632∗∗∗ 14.012∗∗∗ 3.049∗∗∗ 4.532
(3.402) (3.186) (9.365) (0.008) (0.140) (18.378) (13.416) (0.042) (2.187) (0.477) (12.619)
PrevAct type 4−54.312∗∗∗
−9.478∗∗
−8.512∗∗∗
−5.800∗∗∗
−10.646∗∗∗ 1.134∗∗∗
−14.814∗∗∗
−2.390∗∗∗
−6.462∗∗∗ 13.104 −5.682∗∗∗
(0.000) (3.842) (0.000) (0.000) (0.00000) (0.00000) (0.000) (0.00001) (0.00001) (11.637) (0.0001)
PrevAct type 5 16.379∗∗∗ 18.026∗∗∗ 5.408∗∗∗ 4.038∗∗∗
−1.478∗∗∗
−19.085∗∗∗ 53.183∗∗∗ 5.075∗∗∗ 5.032∗∗∗
−9.741∗∗∗ 3.543∗∗∗
(1.409) (1.414) (0.000) (0.000) (0.006) (0.040) (1.556) (0.000) (0.272) (0.545) (0.116)
Akaike Inf. Crit. 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 78
Table A.21: Activity type - Episode 8 (3/3)
Dependent variable:
1 2 3 4 5 6 7 8 9 10 11
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11)
PrevAct type 6−6.205 −11.254∗∗ 5.374∗∗∗ 0.780∗∗∗
−8.669∗∗∗
−23.479∗∗∗ 43.760∗∗∗ 3.823∗∗∗
−5.898∗∗∗
−0.008 2.000∗∗∗
(4.309) (5.258) (0.000) (0.000) (0.00000) (0.001) (6.638) (0.00000) (0.00001) (3.107) (0.00004)
PrevAct type 7−21.718∗∗∗ 38.601∗∗∗ 2.230 2.297∗∗∗ 1.597∗∗∗ 3.763∗∗∗ 2.701∗∗∗ 3.163∗∗∗ 3.761∗∗∗ 45.228∗∗∗ 3.040∗∗∗
(0.000) (4.389) (0.000) (0.000) (0.000) (0.000) (0.000) (0.000) (4.389) (0.000)
PrevAct type 8−52.329∗∗∗
−11.661∗∗∗
−6.962∗∗∗
−0.689∗∗∗
−62.693∗∗∗
−18.841∗∗∗
−25.242∗∗∗ 1.689∗∗∗
−5.692∗∗∗
−14.320∗∗∗
−5.007∗∗∗
(0.000) (3.926) (0.00000) (0.000) (0.000) (0.023) (0.00005) (0.000) (0.0004) (0.000) (0.011)
PrevAct type 9−24.036∗∗∗ 18.426∗∗ 5.442∗∗∗ 1.920∗∗∗ 2.358∗∗∗ 3.667∗∗∗ 3.123∗∗∗ 3.030∗∗∗ 4.814∗∗∗ 30.213∗∗∗ 3.637∗∗∗
(0.000) (7.489) (0.000) (0.00000) (0.001) (0.000) (0.00000) (0.000) (0.0004) (7.487) (0.019)
PrevAct type 10 −6.486∗
−10.360∗∗∗
−8.504∗∗∗
−19.450∗∗∗
−30.798∗∗∗
−29.228∗∗∗
−37.999∗∗∗
−27.060∗∗∗
−8.673 −27.791∗∗∗
−8.615∗∗∗
(3.524) (3.676) (0.002) (0.000) (0.00000) (0.00000) (0.305) (0.0001) (15.934) (0.00000) (0.095)
PrevAct type 11 −8.531∗∗∗ 9.336∗0.879 0.285∗∗∗ 0.518∗∗∗ 0.150∗∗∗ 0.509∗∗∗ 0.419 0.791∗∗∗ 0.762∗∗∗ 0.609∗∗∗
(0.000) (4.989) (0.000) (0.000) (0.002) (0.000) (4.989) (0.000) (0.000) (0.000)
Children −1.567∗∗
−2.588∗∗
−2.266 −11.906∗∗∗ 5.833∗∗∗
−4.984 6.360 0.233 −0.437 20.663 −2.372
(0.796) (1.218) (24.158) (1.144) (0.648) (10.746) (4.467) (3.385) (14.221) (12.789) (10.375)
Constant −48.853∗∗∗ 25.234∗0.730∗∗∗
−12.847∗∗∗
−26.642∗∗∗ 8.158∗∗∗
−14.293∗∗∗ 13.231∗∗∗
−78.706∗∗∗ 30.516∗∗∗
−14.893∗∗∗
(3.713) (13.569) (0.024) (1.680) (0.064) (0.212) (5.356) (0.905) (4.685) (3.670) (0.360)
Akaike Inf. Crit. 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742 986.742
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 79
Table A.22: Activity duration - Episode 1
Dependent variable:
Duration
Work hours 0.022∗∗∗
(0.004)
Type prof 2 0.150∗∗
(0.069)
Type prof 3 0.287
(0.341)
Type prof 4 0.237∗∗∗
(0.062)
Type prof 5 0.311∗∗∗
(0.067)
Type prof 6−0.172∗
(0.092)
Type prof 7 1.301∗∗∗
(0.326)
Type prof 8 0.247∗∗∗
(0.083)
Type prof 9 0.145
(0.117)
Type prof 10 0.114
(0.122)
Transp subs 2−0.042
(0.049)
Transp subs 3 0.107∗∗
(0.053)
Gender −0.105∗∗∗
(0.039)
age 0.001
(0.002)
Marital 2 0.245∗∗∗
(0.057)
Marital 3 0.051
(0.079)
Marital 4 0.111
(0.226)
Education 2a 0.199
(0.170)
Education 2b 0.182
(0.151)
Education 3a 0.115
(0.155)
Education 3b 0.217
(0.159)
Education 4 0.093
(0.155)
Education 5 0.093
(0.156)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Duration (cont.)
Location type 2−0.067
(0.072)
Location type 3−0.174∗
(0.094)
Location type 4−0.188∗∗
(0.075)
Location type 5−0.104
(0.098)
Home occ 2−0.048
(0.045)
Home occ 3 0.152∗
(0.082)
N cars 2−0.118∗∗
(0.051)
N cars 3−0.043
(0.068)
N cars 4−0.002
(0.130)
Ncars 5 0.328
(0.283)
N cars 6 0.083
(0.185)
Income 2 0.062
(0.095)
Income 3−0.032
(0.097)
Income 4−0.036
(0.101)
Income 5−0.061
(0.110)
Income 6−0.210∗
(0.126)
Children 0.021
(0.034)
Observations 3,954
R20.053
Max. Possible R21.000
Log Likelihood −28,692.010
Wald Test 212.420∗∗∗ (df = 40)
LR Test 215.987∗∗∗ (df = 40)
Score (Logrank) Test 216.075∗∗∗ (df = 40)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 80
Table A.23: Activity duration - Episode 2
Dependent variable:
Duration
Work hours −0.007∗
(0.004)
Type prof 2 0.217∗∗∗
(0.071)
Type prof 3 0.408
(0.341)
Type prof 4 0.255∗∗∗
(0.065)
Type prof 5 0.170∗∗
(0.070)
Type prof 6 0.081
(0.095)
Type prof 7−0.109
(0.326)
Type prof 8 0.403∗∗∗
(0.086)
Type prof 9 0.154
(0.121)
Type prof 10 0.387∗∗∗
(0.128)
Transp subs 2 0.161∗∗∗
(0.050)
Transp subs 3−0.108∗
(0.056)
Gender 0.200∗∗∗
(0.041)
age 0.003
(0.002)
Marital 2 0.029
(0.058)
Marital 3−0.027
(0.082)
Marital 4 0.172
(0.236)
Education 2a 0.032
(0.178)
Education 2b −0.085
(0.157)
Education 3a −0.075
(0.163)
Education 3b 0.003
(0.167)
Education 4 0.166
(0.162)
Education 5 0.014
(0.164)
Location type 2−0.051
(0.073)
Location type 3 0.041
(0.096)
Location type 4−0.001
(0.077)
Location type 5−0.011
(0.100)
Home occ 2 0.090∗
(0.048)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Duration (cont.)
Home occ 3 0.235∗∗∗
(0.085)
N cars 2 0.033
(0.052)
N cars 3 0.052
(0.069)
N cars 4−0.041
(0.138)
N cars 5 0.381
(0.284)
N cars 6−0.707∗∗∗
(0.207)
Income 2 0.290∗∗∗
(0.103)
Income 3 0.347∗∗∗
(0.105)
Income 4 0.274∗∗
(0.109)
Income 5 0.283∗∗
(0.118)
Income 6 0.212
(0.134)
prev travel −0.003∗∗∗
(0.001)
Act type 2−4.330∗∗∗
(0.139)
Act type 3−2.897∗∗∗
(0.209)
Act type 4−0.333
(0.716)
Act type 5−2.054∗∗∗
(0.207)
Act type 6−0.530∗∗
(0.258)
Act type 7−1.211∗∗∗
(0.388)
Act type 8−0.270
(0.322)
Act type 9−1.172
(0.713)
Act type 10 −3.108∗∗∗
(0.064)
Act type 11 −0.075
(0.264)
tb −0.001∗∗∗
(0.0002)
Children 0.012
(0.035)
Observations 3,758
R20.501
Max. Possible R21.000
Log Likelihood −25,873.920
Wald Test 3,267.730∗∗∗ (df = 52)
LR Test 2,615.254∗∗∗ (df = 52)
Score (Logrank) Test 5,349.875∗∗∗ (df = 52)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 81
Table A.24: Activity duration - Episode 3
Dependent variable:
Duration
Work hours −0.018∗∗∗
(0.005)
Transp subs 2 0.065
(0.071)
Transp subs 3−0.311∗∗∗
(0.087)
Gender −0.037
(0.058)
age −0.006∗
(0.003)
Marital 2−0.113
(0.090)
Marital 3−0.033
(0.127)
Marital 4−0.170
(0.391)
Location type 2 0.055
(0.114)
Location type 3 0.121
(0.148)
Location type 4−0.035
(0.118)
Location type 5 0.162
(0.155)
Home occ 2−0.136∗
(0.075)
Home occ 3−0.217
(0.150)
N cars 2−0.106
(0.081)
N cars 3−0.108
(0.113)
N cars 4 0.033
(0.223)
N cars 5 0.359
(0.458)
N cars 6 0.205
(0.328)
Income 2 0.165
(0.165)
Income 3 0.392∗∗
(0.164)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Duration (cont.)
Income 4 0.343∗∗
(0.168)
Income 5 0.287
(0.176)
Income 6 0.244
(0.196)
prev travel −0.003∗∗∗
(0.001)
Act type 2−1.870∗∗∗
(0.118)
Act type 3−1.664∗∗∗
(0.124)
Act type 4−1.396∗∗∗
(0.115)
Act type 5−1.210∗∗∗
(0.117)
Act type 6−0.727∗∗∗
(0.163)
Act type 7−1.739∗∗∗
(0.252)
Act type 8−2.126∗∗∗
(0.240)
Act type 9−2.471∗∗∗
(0.235)
Act type 10 −3.218∗∗∗
(0.115)
Act type 11 −1.300∗∗∗
(0.213)
tb −0.002∗∗∗
(0.0002)
Children −0.079
(0.052)
Observations 1,694
R20.595
Max. Possible R21.000
Log Likelihood −10,138.700
Wald Test 1,440.050∗∗∗ (df = 37)
LR Test 1,533.142∗∗∗ (df = 37)
Score (Logrank) Test 1,912.286∗∗∗ (df = 37)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 82
Table A.25: Activity duration - Episode 4
Dependent variable:
Duration
Work hours −0.015∗∗
(0.006)
Transp subs 2 0.173∗∗
(0.082)
Transp subs 3 0.125
(0.105)
Gender −0.174∗∗
(0.070)
age −0.002
(0.004)
Marital 2 0.002
(0.108)
Marital 3 0.100
(0.153)
Marital 4−0.494
(0.432)
Location type 2 0.082
(0.137)
Location type 3 0.129
(0.178)
Location type 4 0.049
(0.143)
Location type 5 0.0005
(0.187)
N cars 2−0.020
(0.094)
N cars 3−0.060
(0.128)
N cars 4 0.158
(0.262)
N cars 5 0.644
(0.463)
N cars 6 0.498
(0.395)
Income 2 0.159
(0.191)
Income 3 0.139
(0.188)
Income 4 0.175
(0.192)
Income 5−0.003
(0.200)
Income 6−0.045
(0.227)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Duration
prev travel −0.0002
(0.001)
Act type 2−2.266∗∗∗
(0.147)
Act type 3−2.215∗∗∗
(0.152)
Act type 4−1.433∗∗∗
(0.156)
Act type 5−1.265∗∗∗
(0.128)
Act type 6−1.685∗∗∗
(0.189)
Act type 7−1.945∗∗∗
(0.257)
Act type 8−2.534∗∗∗
(0.375)
Act type 9−2.551∗∗∗
(0.194)
Act type 10 −2.918∗∗∗
(0.129)
Act type 11 −1.752∗∗∗
(0.266)
tb −0.003∗∗∗
(0.0002)
Children 0.027
(0.062)
Observations 1,187
R20.643
Max. Possible R21.000
Log Likelihood −6,609.624
Wald Test 1,189.350∗∗∗ (df = 35)
LR Test 1,221.652∗∗∗ (df = 35)
Score (Logrank) Test 1,650.629∗∗∗ (df = 35)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 83
Table A.26: Activity duration - Episode 5
Dependent variable:
Duration
Work hours −0.012
(0.009)
Transp subs 2 0.128
(0.125)
Transp subs 3 0.237
(0.176)
Gender −0.082
(0.108)
age −0.008
(0.006)
Home occ 2−0.068
(0.136)
Home occ 3−0.208
(0.305)
Income 2 0.200
(0.299)
Income 3 0.276
(0.288)
Income 4 0.253
(0.280)
Income 5 0.134
(0.292)
Income 6 0.278
(0.335)
prev travel 0.004∗∗
(0.002)
Act type 2−1.665∗∗∗
(0.168)
Act type 3−1.829∗∗∗
(0.200)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Duration (cont.)
Act type 4−1.099∗∗∗
(0.271)
Act type 5−0.940∗∗∗
(0.171)
Act type 6−1.410∗∗∗
(0.250)
Act type 7−2.056∗∗∗
(0.305)
Act type 8−1.617∗∗∗
(0.525)
Act type 9−2.369∗∗∗
(0.246)
Act type 10 −3.093∗∗∗
(0.202)
Act type 11 −0.566
(0.519)
tb −0.002∗∗∗
(0.0004)
Children 0.090
(0.092)
Observations 533
R20.599
Max. Possible R21.000
Log Likelihood −2,573.865
Wald Test 437.270∗∗∗ (df = 25)
LR Test 487.291∗∗∗ (df = 25)
Score (Logrank) Test 550.512∗∗∗ (df = 25)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 84
Table A.27: Activity duration - Episode 6
Dependent variable:
Duration
Work hours 0.009
(0.013)
Transp subs 2 0.078
(0.167)
Transp subs 3 0.349
(0.240)
Gender 0.287∗
(0.163)
age −0.002
(0.010)
Income 2−0.124
(0.376)
Income 3 0.117
(0.358)
Income 4−0.005
(0.352)
Income 5−0.274
(0.364)
Income 6−0.267
(0.442)
prev travel 0.004∗
(0.002)
Act type 2−2.064∗∗∗
(0.259)
Act type 3−1.805∗∗∗
(0.264)
Act type 4−1.493∗∗∗
(0.333)
Act type 5−1.275∗∗∗
(0.253)
Act type 6−1.976∗∗∗
(0.404)
Act type 7−2.160∗∗∗
(0.390)
Act type 8−2.142∗∗∗
(0.567)
Act type 9−2.191∗∗∗
(0.285)
Act type 10 −3.166∗∗∗
(0.307)
Act type 11 −1.184∗
(0.658)
tb −0.001∗∗
(0.001)
Children −0.050
(0.132)
Observations 273
R20.543
Max. Possible R21.000
Log Likelihood −1,155.365
Wald Test 200.050∗∗∗ (df = 23)
LR Test 213.489∗∗∗ (df = 23)
Score (Logrank) Test 253.953∗∗∗ (df = 23)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 85
Table A.28: Activity duration - Episode 7
Dependent variable:
Duration
Work hours −0.028
(0.022)
Transp subs 2 0.370
(0.315)
Transp subs 3 0.387
(0.433)
Gender −0.293
(0.314)
age 0.049∗∗∗
(0.018)
Income 2−0.522
(0.833)
Income 3−0.609
(0.813)
Income 4−1.354∗
(0.780)
Income 5−1.043
(0.836)
Income 6−0.528
(0.804)
prev travel −0.007
(0.007)
Act type 2−3.484∗∗∗
(0.495)
Act type 3−4.122∗∗∗
(0.581)
Act type 4−3.934∗∗∗
(0.839)
Act type 5−2.377∗∗∗
(0.503)
Act type 6−2.948∗∗∗
(0.777)
Act type 7−2.961∗∗∗
(0.641)
Act type 8−1.791
(1.114)
Act type 9−3.336∗∗∗
(0.599)
Act type 10 −4.868∗∗∗
(0.617)
Act type 11 −0.208
(1.247)
tb 0.00000
(0.001)
Children 0.068
(0.230)
Observations 115
R20.640
Max. Possible R20.999
Log Likelihood −375.149
Wald Test 95.160∗∗∗ (df = 23)
LR Test 117.621∗∗∗ (df = 23)
Score (Logrank) Test 146.050∗∗∗ (df = 23)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 86
Table A.29: Travel time - Episode 1 to Episode 2
Dependent variable:
Travel time
Work hours −0.001
(0.004)
Type prof 2 0.044
(0.070)
Type prof 3−0.076
(0.341)
Type prof 4−0.065
(0.063)
Type prof 5−0.142∗∗
(0.069)
Type prof 6−0.049
(0.096)
Type prof 7 0.048
(0.328)
Type prof 8 0.153∗
(0.084)
Type prof 9 0.266∗∗
(0.119)
Type prof 10 0.077
(0.124)
Transp subs 2−0.063
(0.050)
Transp subs 3−0.347∗∗∗
(0.055)
Gender 0.033
(0.039)
age −0.003
(0.002)
Marital 2−0.041
(0.057)
Marital 3−0.164∗∗
(0.082)
Marital 4−0.057
(0.231)
Education 2a 0.137
(0.174)
Education 2b 0.042
(0.155)
Education 3a −0.117
(0.161)
Education 3b −0.137
(0.165)
Education 4−0.208
(0.161)
Education 5−0.312∗
(0.162)
Location type 2 0.129∗
(0.073)
Location type 3 0.099
(0.096)
Location type 4 0.013
(0.077)
Location type 5−0.015
(0.101)
Home occ 2 0.093∗∗
(0.047)
Home occ 3−0.074
(0.085)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Travel Time (cont.)
N cars 2 0.091∗
(0.051)
N cars 3 0.165∗∗
(0.068)
N cars 4 0.137
(0.134)
N cars 5 0.313
(0.287)
N cars 6−0.935∗∗∗
(0.193)
Income 2−0.038
(0.097)
Income 3 0.061
(0.100)
Income 4 0.009
(0.104)
Income 5 0.056
(0.113)
Income 6 0.059
(0.129)
prev dur 0.001∗∗∗
(0.0001)
Act type 2−1.628∗∗∗
(0.322)
Act type 3−2.514∗∗∗
(0.205)
Act type 4 1.844∗∗∗
(0.715)
Act type 5−1.942∗∗∗
(0.202)
Act type 6−0.385
(0.250)
Act type 7−1.844∗∗∗
(0.366)
Act type 8−1.730∗∗∗
(0.313)
Act type 9−2.227∗∗∗
(0.712)
Act type 10 −2.322∗∗∗
(0.052)
Act type 11 −0.274
(0.265)
Act type 12 −1.491∗∗∗
(0.509)
Children 0.028
(0.034)
Observations 3,883
R20.367
Max. Possible R21.000
Log Likelihood −27,324.870
Wald Test 2,433.410∗∗∗ (df = 52)
LR Test 1,775.419∗∗∗ (df = 52)
Score (Logrank) Test 3,320.393∗∗∗ (df = 52)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 87
Table A.30: Travel time - Epidose 2 to Episode 3
Dependent variable:
Travel time
Work hours −0.007∗
(0.004)
Type prof 2−0.101
(0.071)
Type prof 3 0.002
(0.341)
Type prof 4−0.246∗∗∗
(0.063)
Type prof 5−0.344∗∗∗
(0.069)
Type prof 6−0.010
(0.096)
Type prof 7−0.472
(0.331)
Type prof 8−0.076
(0.086)
Type prof 9−0.043
(0.125)
Type prof 10 −0.131
(0.126)
Transp subs 2−0.039
(0.052)
Transp subs 3−0.078
(0.055)
Gender −0.022
(0.041)
age −0.003
(0.002)
Marital 2−0.091
(0.059)
Marital 3−0.165∗∗
(0.084)
Marital 4−0.216
(0.238)
Education 2a 0.278
(0.183)
Education 2b 0.130
(0.164)
Education 3a 0.159
(0.170)
Education 3b 0.191
(0.172)
Education 4−0.092
(0.169)
Education 5−0.032
(0.170)
Location type 2 0.013
(0.073)
Location type 3 0.074
(0.096)
Location type 4−0.023
(0.077)
Location type 5−0.070
(0.101)
Home occ 2 0.068
(0.048)
Home occ 3−0.042
(0.086)
N cars 2 0.056
(0.052)
N cars 3−0.043
(0.071)
N cars 4−0.046
(0.138)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Travel time (cont.)
N cars 5−0.181
(0.288)
N cars 6−0.292
(0.196)
Income 2−0.134
(0.100)
Income 3−0.041
(0.103)
Income 4−0.017
(0.108)
Income 5 0.053
(0.117)
Income 6 0.142
(0.132)
prev dur 0.001∗∗∗
(0.0003)
Act type 2−0.103
(0.093)
Act type 3−0.076
(0.140)
Act type 4 1.855∗∗∗
(0.118)
Act type 5 1.090∗∗∗
(0.116)
Act type 6 1.173∗∗∗
(0.159)
Act type 7 0.391
(0.248)
Act type 8 0.884∗∗∗
(0.240)
Act type 9 0.957∗∗∗
(0.223)
Act type 10 −0.819∗∗∗
(0.116)
Act type 11 0.960∗∗∗
(0.212)
PrevAct type 2−0.218
(0.176)
PrevAct type 3−0.382∗
(0.232)
PrevAct type 4 1.687∗∗
(0.712)
PrevAct type 5−0.113
(0.221)
PrevAct type 6 0.731∗∗∗
(0.267)
PrevAct type 7 1.434∗∗∗
(0.420)
PrevAct type 8 0.263
(0.322)
PrevAct type 9 1.295∗
(0.725)
PrevAct type 10 −0.434∗∗∗
(0.134)
PrevAct type 11 0.348
(0.266)
tb 0.002∗∗∗
(0.0002)
Children 0.002
(0.035)
Observations 3,705
R20.313
Max. Possible R21.000
Log Likelihood −26,049.270
Wald Test 1,785.310∗∗∗ (df = 62)
LR Test 1,392.741∗∗∗ (df = 62)
Score (Logrank) Test 2,241.525∗∗∗ (df = 62)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 88
Table A.31: Travel time - Episode 3 to Episode 4
Dependent variable:
Travel time
Work hours −0.0001
(0.005)
Transp subs 2−0.090
(0.070)
Transp subs 3 0.033
(0.090)
Gender −0.004
(0.058)
age −0.006∗
(0.004)
Marital 2−0.230∗∗
(0.097)
Marital 3−0.233∗
(0.131)
Marital 4 0.182
(0.395)
Location type 2−0.032
(0.114)
Location type 3−0.124
(0.148)
Location type 4 0.012
(0.119)
Location type 5−0.191
(0.157)
Home occ 2 0.037
(0.074)
Home occ 3 0.226
(0.147)
N cars 2 0.113
(0.083)
N cars 3 0.054
(0.113)
N cars 4−0.024
(0.232)
N cars 5 0.962∗∗
(0.463)
N cars 6−0.448
(0.333)
Income 2 0.203
(0.166)
Income 3 0.192
(0.165)
Income 4 0.302∗
(0.168)
Income 5 0.434∗∗
(0.175)
Income 6 0.263
(0.196)
prev dur −0.002∗∗∗
(0.0003)
Act type 2−0.468∗∗∗
(0.085)
Act type 3−0.391∗∗∗
(0.144)
Act type 4 0.779∗∗∗
(0.148)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Travel time (cont.)
Act type 5 0.389∗∗∗
(0.125)
Act type 6 0.243
(0.187)
Act type 7 0.122
(0.239)
Act type 8−0.818∗∗∗
(0.275)
Act type 9 0.534∗∗∗
(0.195)
Act type 10 −0.198∗
(0.116)
Act type 11 0.307
(0.237)
PrevAct type 2 0.314∗∗
(0.122)
PrevAct type 3 0.338∗∗
(0.144)
PrevAct type 4 0.885∗∗∗
(0.111)
PrevAct type 5 0.524∗∗∗
(0.121)
PrevAct type 6 0.244
(0.162)
PrevAct type 7 0.634∗∗
(0.256)
PrevAct type 8−0.092
(0.244)
PrevAct type 9 0.803∗∗∗
(0.243)
PrevAct type 10 0.435∗∗∗
(0.141)
PrevAct type 11 0.301
(0.217)
tb 0.001∗∗
(0.0002)
Children 0.012
(0.053)
Observations 1,656
R20.273
Max. Possible R21.000
Log Likelihood −10,359.410
Wald Test 518.900∗∗∗ (df = 47)
LR Test 527.501∗∗∗ (df = 47)
Score (Logrank) Test 558.541∗∗∗ (df = 47)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 89
Table A.32: Travel time - Episode 4 to Episode 5
Dependent variable:
Travel time
Work hours −0.010
(0.006)
Transp subs 2 0.141∗
(0.085)
Transp subs 3−0.215∗∗
(0.106)
Gender 0.014
(0.071)
age −0.002
(0.004)
Marital 2−0.201∗
(0.111)
Marital 3−0.376∗∗
(0.158)
Marital 4 0.502
(0.467)
Location type 2−0.381∗∗∗
(0.140)
Location type 3−0.428∗∗
(0.183)
Location type 4−0.459∗∗∗
(0.146)
Location type 5−0.728∗∗∗
(0.194)
Home occ 2 0.091
(0.091)
Home occ 3−0.026
(0.169)
N cars 2 0.163∗
(0.093)
N cars 3 0.076
(0.128)
N cars 4−0.028
(0.271)
N cars 5−0.116
(0.466)
N cars 6 0.104
(0.401)
Income 2 0.099
(0.202)
Income 3 0.053
(0.197)
Income 4 0.134
(0.201)
Income 5 0.221
(0.210)
Income 6 0.074
(0.237)
prev dur −0.001∗
(0.0003)
Act type 2−0.034
(0.115)
Act type 3 0.512∗∗∗
(0.182)
Act type 4 0.987∗∗∗
(0.256)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Travel time (cont.)
Act type 5 0.621∗∗∗
(0.170)
Act type 6 0.239
(0.235)
Act type 7 0.610∗∗
(0.291)
Act type 8−0.697∗
(0.364)
Act type 9 0.142
(0.236)
Act type 10 0.152
(0.177)
Act type 11 0.435
(0.483)
PrevAct type 2−0.686∗∗∗
(0.154)
PrevAct type 3−1.490∗∗∗
(0.149)
PrevAct type 4−0.364∗∗
(0.156)
PrevAct type 5−0.590∗∗∗
(0.122)
PrevAct type 6−0.524∗∗∗
(0.185)
PrevAct type 7 0.172
(0.270)
PrevAct type 8−2.073∗∗∗
(0.378)
PrevAct type 9−0.172
(0.195)
PrevAct type 10 −1.745∗∗∗
(0.126)
PrevAct type 11 −1.134∗∗∗
(0.278)
tb 0.0004
(0.0003)
Children 0.127∗∗
(0.064)
Observations 1,171
R20.402
Max. Possible R21.000
Log Likelihood −6,806.270
Wald Test 583.090∗∗∗ (df = 47)
LR Test 602.029∗∗∗ (df = 47)
Score (Logrank) Test 653.958∗∗∗ (df = 47)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 90
Table A.33: Travel time - Episode 5 to Episode 6
Dependent variable:
Travel time
Work hours −0.011
(0.009)
Transp subs 2−0.008
(0.124)
Transp subs 3−0.105
(0.174)
Gender 0.285∗∗
(0.112)
age −0.0004
(0.006)
Home occ 2−0.064
(0.141)
Home occ 3−0.372
(0.350)
Income 2 0.396
(0.316)
Income 3 0.517∗
(0.300)
Income 4 0.477
(0.302)
Income 5 0.516∗
(0.310)
Income 6 1.207∗∗∗
(0.352)
prev dur 0.002∗∗∗
(0.001)
Act type 2−0.172
(0.142)
Act type 3 0.348
(0.229)
Act type 4 0.282
(0.295)
Act type 5 0.108
(0.239)
Act type 6−0.218
(0.383)
Act type 7−0.070
(0.355)
Act type 8−0.609∗
(0.351)
Act type 9 0.696∗∗
(0.279)
Act type 10 −0.483∗
(0.272)
Act type 11 −0.186
(0.753)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Travel time (cont.)
PrevAct type 2−0.222
(0.191)
PrevAct type 3−1.328∗∗∗
(0.201)
PrevAct type 4−0.449
(0.307)
PrevAct type 5 0.074
(0.177)
PrevAct type 6 0.196
(0.247)
PrevAct type 7 0.138
(0.308)
PrevAct type 8−1.322∗∗
(0.532)
PrevAct type 9 0.546∗∗
(0.248)
PrevAct type 10 −1.750∗∗∗
(0.219)
PrevAct type 11 −0.701
(0.520)
tb 0.003∗∗∗
(0.0005)
Children 0.070
(0.095)
Observations 527
R20.352
Max. Possible R21.000
Log Likelihood −2,665.351
Wald Test 217.940∗∗∗ (df = 35)
LR Test 229.032∗∗∗ (df = 35)
Score (Logrank) Test 234.628∗∗∗ (df = 35)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 91
Table A.34: Travel time - Episode 6 to Episode 7
Dependent variable:
Travel time
Work hours −0.008
(0.013)
Transp subs 2−0.359∗
(0.185)
Transp subs 3−0.374
(0.268)
Gender 0.089
(0.174)
age −0.011
(0.009)
Home occ 2 0.245
(0.199)
Home occ 3 0.403
(0.411)
Income 2 0.401
(0.385)
Income 3 0.496
(0.369)
Income 4 0.406
(0.359)
Income 5 0.540
(0.390)
Income 6 0.101
(0.429)
prev dur −0.0003
(0.001)
Act type 2−0.178
(0.240)
Act type 3 0.373
(0.362)
Act type 4 1.586
(1.107)
Act type 5−0.409
(0.409)
Act type 6−0.463
(0.572)
Act type 7 0.337
(0.527)
Act type 8−1.231∗∗
(0.556)
Act type 9 0.547
(0.541)
Act type 10 −0.047
(0.385)
Act type 11 1.870
(1.252)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Dependent variable:
Travel time (cont.)
PrevAct type 2−0.158
(0.295)
PrevAct type 3−1.227∗∗∗
(0.250)
PrevAct type 4−0.671∗
(0.365)
PrevAct type 5−0.244
(0.254)
PrevAct type 6 0.166
(0.421)
PrevAct type 7−0.024
(0.407)
PrevAct type 8−0.593
(0.669)
PrevAct type 9 0.517∗
(0.292)
PrevAct type 10 −1.844∗∗∗
(0.346)
PrevAct type 11 −0.123
(0.631)
tb 0.001
(0.001)
Children −0.036
(0.142)
Observations 266
R20.406
Max. Possible R21.000
Log Likelihood −1,153.733
Wald Test 123.230∗∗∗ (df = 35)
LR Test 138.376∗∗∗ (df = 35)
Score (Logrank) Test 137.799∗∗∗ (df = 35)
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 92
Table A.35: Mode choice - Commuting
Dependent variable:
CAR PT WALK
(1) (2) (3)
Travel Time 0.004∗∗∗
(0.001)
Home parking 2−0.176 −0.221 −3.404∗∗∗
(0.117) (0.149) (0.878)
Home parking 3−0.759∗∗∗ −0.514∗−16.248
(0.233) (0.270) (10266.180)
Children 2−0.038 0.064 −1.953
(0.085) (0.108) (0.816)
Gender 2−0.202∗∗ 0.282∗∗∗ 19.421
(0.082) (0.103) (2382.464)
Transp subs 2 2.075∗∗∗ 0.016 3.686∗∗∗
(0.216) (0.358) (0.925)
Transp subs 3−0.603∗∗ 0.160 −17.255
(0.118) (0.130) (4538.238)
Location type 1−0.745∗∗∗ −0.786∗∗∗ −0.096
(0.084) (0.107) (0.785)
N cars 2 0.439∗∗ −0.271∗∗∗ 1.614
(0.092) (0.112) (1.176)
N cars 3 0.913∗∗∗ −0.288 4.495∗∗∗
(0.154) (0.210) (1.382)
Departue time 2−0.247∗∗∗ −0.191∗−2.449∗∗
(0.081) (0.103) (1.126)
Constant 1.196∗∗∗ 0.349∗∗ -26.278∗
(0.146) (0.168) (2382,465)
McFadden’s R20.096
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 93
Table A.36: Mode choice - Discretionary activities
Dependent variable:
CAR PT WALK
(1) (2) (3)
Act type 2−0.101 −0.950 0.590
(0.520) (1.170) (0.594)
Act type 3−0.015 1.418∗1.524∗∗
(0.587) (0.751) (0.627)
Act type 4 0.879 2.232∗4.575∗∗∗
(1.057) (1.180) (1.072)
Act type 5−0.661 0.602 1.281∗∗∗
(0.410) (0.601) (0.453)
Act type 6 0.729 0.556 2.615∗∗
(1.042) (1.477) (1.069)
Act type 7−0.141 −10.726∗∗∗ 1.054
(1.047) (0.00003) (1.124)
Act type 8 8.739∗∗∗ 11.880∗∗∗ 14.299∗∗∗
(0.456) (0.721) (0.407)
Act type 9 0.012 0.791 1.355∗
(0.761) (1.083) (0.812)
Act type 10 −0.740∗∗ 0.691 1.258∗∗∗
(0.326) (0.493) (0.369)
Home parking 2−0.082 0.098 −0.317
(0.398) (0.552) (0.426)
Home parking 3−0.606 0.631 −0.276
(0.624) (0.775) (0.665)
Children 2 0.416 0.966∗∗ 0.264
(0.256) (0.377) (0.270)
Gender −0.472∗0.109 −0.140
(0.262) (0.360) (0.276)
Transp subs 2 1.817∗∗ 0.253 1.244∗
(0.735) (0.908) (0.747)
Transp subs 3−0.828∗∗
−0.324 −0.429
(0.326) (0.472) (0.349)
Location type 1−0.747∗∗∗
−0.378 −0.411
(0.264) (0.359) (0.278)
N cars 2 0.596∗∗ 0.131 0.035
(0.279) (0.382) (0.295)
N cars 3 1.112∗
−0.104 0.554
(0.628) (0.880) (0.648)
Departue time 2−0.011 0.456 −0.286
(0.299) (0.394) (0.315)
Constant 4.087∗∗∗
−1.272∗0.884∗
(0.443) (0.670) (0.487)
McFadden’s R20.168
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
APPENDIX A. PARAMETER ESTIMATES FOR THE FINAL MODELS 94
Table A.37: Location Choice
Dependent variable:
External Meal Shopping Serv. and Leisure Others
(1) (2) (3) (4)
TravelTime −0.34∗∗∗ −0.058∗∗∗ −0.070∗∗∗ −0.266∗∗∗
(0.040) (0.020) (0.014) (0.020)
Std Dev (TravelTime) - - - 0.286∗∗∗
- - - (0.016)
TimeNextFixed 0.075∗∗∗ −0.074∗∗∗ −0.194∗∗∗ −0.181∗∗∗
(0.028) (0.020) (0.014) (0.016)
AvailTime - 0.106∗∗ 0.026∗∗∗ 0.088∗∗∗
- (0.027) (0.007) (0.019)
Restaurants 0.013∗∗∗ ---
(0.001) - - -
Supermarkets - 0.072∗∗∗ - -
- (0.005) - -
Service - - 0.019∗∗∗ -
- - (0.002) -
Density - - - 0.001∗∗∗
- - - (<0.001)
McFadden’s R20.721 0.459 0.387 0.396
Correctly predicted 49.21% 28.89% 32.26% 31.32%
∗p<0.1; ∗∗p<0.05; ∗∗∗ p<0.01
Appendix B
List of Geographical Zones
In this Appendix, we present the list of geographical zones defined to generate the graphs
used to visualize the results in Figures 4.3 and 4.4. The regions are listed in Table B.1,
with the number of each zone being used in the graphs in Chapter 4.
Table B.1: Geographical zones in the Luxembourg-France border region
Number Name Country Number Name Country
13 Thionville France 104 Bettembourg Luxembourg
14 Longwy France 105 Clervaux Luxembourg
15 Meuse & Ouest MM France 106 Differdange Luxembourg
16 Longuyon France 107 Dudelange Luxembourg
17 Meuse & Ouest MM France 108 Echternach Luxembourg
18 Moselle Est France 109 Esch-sur-Alzette Luxembourg
19 Moselle Est France 110 Grevenmacher Luxembourg
20 Agglo Metz France 111 Hesperange Luxembourg
21 Agglo Thionville France 112 Kayl Luxembourg
22 Ouest Thionville France 113 Larochette Luxembourg
23 Agglo Metz France 114 Luxembourg City Luxembourg
24 P´eriph´erie Metz France 115 Mamer Luxembourg
25 Nord Moselle France 116 Nordstad Luxembourg
26 Ouest Thionville France 117 Petange Luxembourg
27 Thionville Est France 118 Redange Luxembourg
28 Longwy France 119 Remch Luxembourg
101 Alzette Luxembourg 120 Sandweiler Luxembourg
102 Bascharage Luxembourg 121 Sanem Luxembourg
103 Bert-Strass Luxembourg 122 Wiltz Luxembourg
95
Leuven Statistics Research Center
Celestijnenlaan 200B bus 5307
3001 LEUVEN, BELGI¨
E
tel. + 32 16 32 88 75
www.kuleuven.be
