ICT Implications and Trip Timing Choice
Maria Börjesson
Doctoral Thesis inInfrastructure
withspecialisation inTransport andLocation Analysis, August2006
DivisionofTransportandLocationAnalysis
DepartmentofTransportandEconomics
RoyalInstituteofTechnology
ISSN1653-4468
ISBN 13: 978-91-85539-04-8
ISBN 10: 91-85539-04-X
Akademisk avhandling som med tillstånd av Kungliga Tekniska Högskolani Stockholm
framläggestill oentliggranskningför avläggandeavteknologie doktorsexamenfredagen
den 29september2006,KungligaTekniskaHögskolan,Stockholm
c
cations and Trip Timing Choice. Department of Transport and Economics,
KTH, Stockholm. ISBN10: 91-85539-04-X.
Abstract
Travel demand forecasting is essential for many decisions, such as
infrastructure investments and policy measures. Traditionally travel
demand modelling has considered trip frequency, mode, destination
and route choice. This thesis considers two other choice dimensions,
hypothesised to have implicationsfor travel demand forecasting. The
rstpartinvestigateshowtheincreasedpossibilitiestoovercomespace
that ICT (information and communication technology) provides, can
beintegratedintraveldemandforecastingmodels. Wendthat
possi-bilitiesofmodellingsubstitutioneectsarelimited,irrespectiveofdata
source and modelling approach. Telecommuting explains, however, a
very smallpart ofvariation inworktrip frequency. Itis thereforenot
urgenttoincludeeectsfromtelecommutingintraveldemand
forecast-ing. Theresults indicate thattelecommuting isa privilege forcertain
groups of employees, and we thereforeexpectthat negative attitudes
from management, job suitability and lack of equipment are
impor-tant obstacles. We nd also that company benets can be obtained
from telecommuting. No evidences that telecommuting gives rise to
urbansprawlis,however,found. Hence,thereisgroundforpromoting
telecommuting froma societal, individualand company perspective.
Thesecondpartdevelopsadeparturetimechoicemodelinamixed
logit framework. This model explains how travellers trade-o travel
time, travel time variability, monetary and scheduling costs, when
choosing departure time. We explicitly account for correlationin
un-observed heterogeneity over repeated SP choices, which was
funda-mental for accurate estimation ofthe substitution pattern. Temporal
constraints at destination are found to mainly restrict late arrival.
Constraints at origin mainly restrict early departure. Sensitivity to
traveltimeuncertaintydependsontriptypeandintendedarrivaltime.
Given appropriate input data and a calibrated dynamic assignment
model, the modelcan beapplied to forecast peak-spreading eects in
congested networks. Combined stated preference (SP) and revealed
preference (RP) data is used, which has provided an opportunity to
compare observed andstated behaviour. Such analysishaspreviously
notbeencarriedoutandindicatesthattherearesystematicdierences
inRP andSP data.
Keywords: Travel Demand, ICT, Tele-substitution, Revealed
Pref-erence, Stated Preference, Mixed Logit, Unobserved Heterogeneity,
In research, as in most human activities, interdependence certainly is more
valuable than independence. This thesis is the result of ve years of work
through which I have been accompanied and supported by others who have
added to this work indierentways.
First,IwouldliketothankmysupervisorStaanAlgers,atrulygenerous
and considerate person with a refreshing sense of humour. He has very
unselshly shared his ideas, knowledge and experiences with me during the
past years. Thanks alsotoLars-Göran Mattsson,head of the department of
Transport and Economics, for always having kept an eye on my work and
supporting mewith valuable commentsand advice.
I would further like to thank all members of the SILVESTER project
team, Leonid Engelson, Andrew Daly, Staan Algers, Jonas Eliasson and
Fredrik Davidson. They have substantially contributed to the development
of the second part of this thesis. I am particularly grateful to the project
leader Leonid Engelson for careful reading and valuable comments on the
work. Leonidhasalsobeen veryencouragingand helped metostructure the
work. ExtensivecommentsfromanddiscussionswithAndrewDalyhavehad
adirectimpactonthenalformandquality ofthesecondpartof thethesis.
Jonas Eliasson has contributed to the SP design and to inspiring, amusing
and interesting discussions. I would also like to thank Pia Koskenoja for
valuablecomments inthe nal seminar.
Iamgratefultomyfriends andcollegesatTLAfor allgenerous support.
I will miss them all. I want to thank my friend and former college Muriel
Besser Hugosson in particular. Her wise, cheerful and generous coaching in
researchaswellasmorepersonalmattershavemadetheseyearseasier,more
ecient and a lotmore fun.
IfurtherwanttoacknowledgetheSwedishAgencyforInnovationSystems
and the Swedish Road Administration forhaving provided necessary funds.
ToLisa and Peter, thank youfor always helpingme out in critical
situa-tions; you have been wonderful during this lastphase. Finally,I would like
to thank my darlingsKalle, Malte and Johan for all moral support,
encour-agement, understanding and love.
Stockholm, August 2006
1 Introduction and Scope of the Thesis 1
2 Research Context 3
2.1 ICT and Travel
A Backward Glanceand WiderPerspectives . . . 3
2.2 DepartureTimeChoiceModellingand TravelTimeUncertainty 5
3 Dierent Approaches in Travel Demand Modelling 9
4 Methods and Data 11
4.1 Methods . . . 11
4.2 Data . . . 13
5 Results and Discussion 17
5.1 Paper I, 'A Communication Choice Model' . . . 17
5.2 Paper II, 'Telecommuting and work travel demand modelling
in Sweden' . . . 18
5.3 PaperIII,'CompanyIncentivesandToolsforPromotingT
elecom-muting' . . . 19
5.4 PaperIV,'DepartureTimeModelling: ApplicabilityandTravel
Time Uncertainty. ' . . . 20
5.5 Paper V, 'Joint RP-SP Data in Mixed LogitAnalysis of Trip
TimingDecisions' . . . 22
6 What Have the Findings Brought to Travel Demand
Mod-elling? 25
(I) Börjesson, M., 2003. A Communication Choice Model.
(II)Börjesson, M., 2003. Telecommutingandwork traveldemand modelling
in Sweden.
(III) Robèrt, M., Börjesson, M., 2006. Company Incentives and Tools for
PromotingTelecommuting. Environment and Behavior, 38(4).
(IV) Börjesson, M., 2006. Departure Time Modelling: Applicability and
Travel Time Uncertainty.
(V) Börjesson, M.,2006. JointRP-SP Datain MixedLogitAnalysisof Trip
Urban regionsthroughoutthe world arestrugglingwithfast growingcar use
and increasing congestion as a consequence. Foreconomical, environmental
and socialreasons, and considering the large amount of latent demand, new
physical infrastructurecannot alonesolvethese problems. Society is, onthe
otherhand,dependingonhighaccessibilityinallsegmentsofthepopulation,
particularly to labour markets. This development underscores the need for
policy innovations and eective utilization of new technology, which is an
importantchallenge for modern travel behavioural modelling.
Thisthesis addressestwoissues,suggested aspotentialsolutionstothese
problems. The objective of the rst two papers I - II is to investigate if
and howthe increasedpossibilitiestoovercomespace thatICT(information
andcommunicationtechnology)provides,canbeintegratedinexistingtravel
demand forecast models. Another important objective is to investigatehow
a new data source, the Communication Survey (SIKA 2002), can be used
to improve modelling with regard to ICT. KOM is a one-day travel and
communication diary survey, carried out on a regular basis, on commission
oftheSwedishInstituteforTransportandCommunicationsAnalysis(SIKA).
Paper I, 'A Communication Choice Model', focuses on substitution
ef-fects of postal and banking activities. Paper II, 'Telecommuting and work
travel demand modelling inSweden', investigateshow telecommuting could
be implemented in conventional travel demand models. Specically, it aims
at forecasting future commuting frequency. Paper III, 'Company Incentives
and ToolsforPromotingTelecommuting',alsoconsiders telecommuting,but
is more policy oriented. It aims at identifying company incentives for
pro-moting telecommutingto reducecommuting.
The second part of the thesis, paper IV - V, develops a departure time
choice model. Thismodelexplainshowtravellerstradeotraveltime,travel
time variability, monetary and scheduling costs, when choosing departure
time. Thisworkwaspart ofalargerprojectcalledSILVESTER (Simulation
of choice between starting times and routes), which aims to develop an
ap-plication to forecast peak-spreading eects in Stockholm. SILVESTER was
largely triggered by the tollstrial inStockhom, taking place January- July
2006. SPand RPdata, usedto estimatedthe demand model, were collected
within theSILVESTERproject. The demand modelpresented inthis thesis
willnowbeimplementedwithacalibrateddynamicassignmentmodelwithin
the SILVESTER project. Eects frominfrastructureinvestments andpolicy
measures can then be calculated usingthis application.
InpaperIV,'DepartureTimeModelling: ApplicabilityandTravelTime
mation. However, there are certain downsides using stated preference data,
as opposed to revealed preference data (RP), describing travellers' actual
choices. A common feature of SP data is that response scale is distorted.
Paper V, 'Joint RP-SP Data in Mixed Logit Analysis of Trip Timing
Deci-sions', therefore extends the analysis in paper IV by estimating the model
2.1 ICT and Travel
A Backward Glance and Wider Perspectives
Thediscussionaboutwhethertelecommunicationssubstitutepassenger
trans-portationhasbeenappearingintransportliteraturesincethe1970s.
Accord-ingtoMokhtarian(1998),theoilcrisiswasthestartofitall,whichcoincided
with the startof theinformationera. One ofthe leadingideas wasthatICT
would create the 'death of distance' in a longer perspective, i.e. that the
spatial distance between home, work and stores etc., would play a less
im-portant role for activity participation. In this context redenitions of the
concept accessibilityhave frequently been discussed, see Shen (2000).
Thisrevolutionaryperspective wasverymuchanoutcome ofthe
fascina-tion ofthe tremendoustechnologicaldevelopment. This resultedina
simpli-ed and technologyfocusedway oflookingathuman activities. Mostly very
specic and direct eects from telecommunication services were discussed
and analyzed withthe prospects that thesecould be asolutiontothe
trans-portationproblems. Theinteractionwasoftencategorizedasby Mokhtarian
and Meenakshisundaram (1999):
Substitution: Telecommunicationsreplace travel.
Generation: Telecommunications stimulate or complement travel, for
in-stance by increasing access to information.
Modication: Telecommunicationsneithergeneratenorsubstitutetravel,but
change the travel pattern, for instance by time, mode orroute.
Neutrality: The use of telecommunications and travel is not inuenced by
each other.
Indirect(orrebound) eects: The eectsdescribedaboveinuencethe travel
demand and land use in a longer perspective. This may give rise to more
spread out residential locationsand subsequent eects onlocal and regional
travel patterns and mode choices.
Numerousstudieswere carriedoutwiththeobjectivetodeterminewhich
of these eects that were the most dominant. However, very little
clear-cut results came out of these. Mokhtarian and Meenakshisundaram also
note that many of these studies were short-term, and therefore tended to
overestimate substitution eect and underestimate the more long-term and
complex generation eects. When correlation between amount of transport
and telecommunications was found, which was the case in many studies,
chair studies', whichmainlyexposedideas forfutureresearch. Thesewere in
general verytechnology focused, and Salomonemphasises that these should
not beconfusedwithforecasts. OneexampleisGrübler (1989),whoshowed
howdierenttechnologieshaveemergedandvanishedinS-shapessinceabout
1800, with increasing frequency. He argued that the length of canals,
rail-roads and surface roads, relative to their saturation level, all followed a
S-shapedpattern. Grüblerthenarguedthatthisdevelopmentpatternis
gener-ally applicable whenforecasting penetration patterns of modern technology,
which is just what Salomon objected to. Mattsson and Höjer and
Matts-son (2000)alsocriticizedGrüblerforthisreason,andfurtherquestioned the
quality of the data.
It is now clear that ICT impacts on the transportation system reach far
beyond and are far more complex than the ones most frequently addressed
in literature. The progressive information and communication technology
development has changed, and will in the forthcoming decades continue to
change, organisation of work and rms (Castells 1996), education, shopping
and entrainment industries and land use patterns etc. This reformation
oc-curs in various spheres and levels of society and has thereby a fundamental
impact onconsumers' behaviour.
On the other hand, analyzing very specic behavioural adaptations like
the ones listedabove,inresponse tospecictechnologicalinnovations has in
general not been very successful. These studies have also become
substan-tially fewer. The limitations of them are primarily related to two factors.
First,itseemsthatICTprimarilyacts atthetravelpatternlevel,incontrast
tothenumberoftrips. Importantimpactsarepresumablysucheectsastrip
chaining,destinationchoice,departuretimechoiceandmeetingsarranged at
shortnoticeinuencingtravelplanning. In alongerperspectivealsolanduse
and mode choice may be aected. Deriving causal relationships and
calcu-lating the 'net'eect of allthese eectsis noteasily done. Second, there are
many factors that simultaneously act on travel patterns, inaddition to and
jointlywithinformationtechnology. Such factorsare, forinstance, economic
growth, environmentalawareness, work situationsand policymeasures, etc.
Inthisperspectiveitisnot onlydicultbut alsolessmeaningfultocalculate
very specic impactsfrom informationtechnology.
To exemplify, at rst glance the substitution eect from
teleconferenc-ing is obvious. On the other hand, this contact (trip or communication)
might never been taken if society had been less information and
technol-ogy dependent and globalized. These kindsof discussions become, however,
highlyhypotheticalandthusnot veryinteresting. Anotherexampleisurban
urban sprawl. However, also avariety of other factors rule this process, e.g.
decreasing travel times and economical growth, and it is very diculty to
assess the eects of these factors isolatedfrom eachother.
2.2 Departure Time Choice Modelling and Travel Time
Uncertainty
Trip timing choice has received increased interest as a consequence of the
time-of-day varying tolls trialin Stockholm (Jan-July 2006). It is expected
that driver's preferences for dierent departure times are an important
di-mension for policy evaluation in congested regions. In aninternational
per-spective there is quitesubstantialevidence of peak spreading eects (Porter
etal.1995). Tracowsinpeakhourhavegrownlessthanthoseinthe
over-allpeakperiodandpeakperiodtracvolumeshavegrownless thanthosein
the o-peak. Thereisalsothe reverseofpeakspreading, peakconcentration,
resulting fromcapacity expansion. In Stockholm, this was manifestedwhen
the capacity of the most essential motorway was increased, with subsequent
increase of peakiness of the trac ow.
Travel demand modelling and forecasting has traditionally been based
upon a four dimensional model, includingthe frequency, destination, mode
and route choice dimensions. The three former choice dimensions are
nor-mallyuniedunder theconcept demand modelling. Given level-of-service in
the trac system (travel times and cost, etc.), a demand model computes
trac volumes ineach originand destinationrelation. A routechoice model
assignsthexed demandtothe roadnetwork andcalculatestraveltimesand
cost, etc. Assignment models have traditionallybeen based on steady-state
equilibrium trac ows and thus lack the time dimension. Since demand
diers largely between on and o peak, trac is often modelled separately
for peak and o peak conditions, and the distribution of demand between
peakando-peakhours isusuallyxed. Peak-spreading eectscan thusnot
bestudied by this approach.
Fewexistinglarge-scale modellingapplicationsintheworlddeal withthe
trip timingdimensionoftravellers'behaviour. Suchanapplicationhasto
in-tegrateadynamicassignmentmodelwithabehaviouralmodelthatdescribes
individuals'responsestonewtravelconditions. Dynamicassignmentmodels
dier from steady-state assignment models in the possibility of modelling
temporalvariationsinows and traveltimeswithin apeakperiod,resulting
from queues buildingup and dissolving.
(1982). However, the work by Vickrey (1969), who originated much of the
conceptually frameworkof subsequent studies, consideredboth demand and
supply inacontinuousanddeterministicframework. This,rathertheoretical
contribution, appliedasinglelinkbottleneckmodel. Focusingmainlyon
op-timaltolling schemes, this work was further developed ina series of articles
by Arnott, de Palma and Lindsey (1998). Hyman (1997) and van Vuren et
al. (1999) also further developed the work by Vickery, and named this
ap-proach "equilibriumschedulingtheory". Buildingonthis theory, a software
product was developed by Hague Consulting Group (1999), in which the
demand model was integrated with the dynamic assignment models
CON-TRAM, SATURN or TRIPS.
Theutilityfunctiondeveloped byVickeryand Smallisbasedontradeo
between traveltime and shiftfrom preferredarrival(or departure)time. As
notedby Smallandmanyotherauthors, the MNLapproachisinappropriate
for departure time modelling (section 4.1 discusses dierent model types
in more detail). This relates to the obvious ordering of consecutive time
intervalsinatriptimingmodel,whichinducesaspeciccorrelationstructure.
The MNL model cannot accommodate a correlation structure among the
alternatives.
Several studies on departure time choice have therefore used more
com-plex model structures. de Jong et al. (2003) and Hess (2005) estimated
error component (mixed) logit models, designed to induce correlation and
heteroskedasticity between departure time intervals. Alternatives with
de-parture times closer to each other are assumed to share more unobserved
attributes than more distant alternatives. The models also include a mode
choice dimension. Estimationthusindicatesthe relativesensitivityof
depar-ture time andmode choice shift,resultingfromchangesingeneralised travel
costs. They nd that, unless the time shifts considered are very large, the
departure time choice is normally more sensitive to travel cost than modal
choice.
Thereis alsoanother class of choice models,GEV models, which can
ac-commodate a exible correlation structure, but still has closed form choice
probabilities. In particular Ordered GEV (OGEV) modelshould beapplied
in cases where there is a natural ordering of the alternatives, and is thus
well suited fordeparture time choice. This modelwas rst appliedby Small
(1987), to an arrival time choice model. Bhat (1998a) jointly modelled
de-parture time and modelchoice applying acombined OGEV-MNL model.
Triptimingmodels inthe literature haveused eitherSP orRP data. No
previous published study has, however, used joint RP and SP data. There
These have, as virtually all choice models including travel time variability,
used SPdata. The exception is Lamand Small (2001), who have used data
from a HOT lane 1
project in California,in order to estimatewillingness to
payforreductionintraveltimevariability. Thisstudy doesnot, ontheother
hand, consider departure time choice.
There are two dierent approaches in previous studies on travel time
uncertainty modelling,the mean-variance approach and the endogenous
ex-pected scheduling approach. The former argues that the disutility of travel
time variation arises from the diculty of planning the day because of not
knowing exactly when to arrive. This disutility could be captured directly
by aseparate variablesuchasthe standard deviationoftraveltimeor
some-thing equivalent. The latter approach assumes that the cost of travel time
uncertaintyarisefromtheriskofarrivingatatimedierentfromthedesired
one and should be represented by the expected value of the schedulingcosts
variables(equallingthe arrivaltime shiftearlier/later fromthe preferred
ar-rival). The theory of this approachwasrst proposed by Garver(1968) and
Polak (1987). Noland and Small (1995) then further developed the theory,
combining itwith the workof Small (1982).
AnexampleinSmalletal.(2000)illustrateshowtheexpectedscheduling
delayvariablescapturethedisutilityarisingfromtherisk ofarrivingatatime
dierent from the preferred arrival time (
P AT
). They describe a situation wherevepossiblearrivaltimes(AT
)aregiven, intermsof shiftfromP AT
. These time shifts are -7, -4, -1, 5 and 9, where early arrival is coded as anegativenumber. Assumingthesame probabilityforthese arrivaltimes, the
expected schedulingcosts are:
E(SDE) = E(max[P AT − AT, 0])) = (7 + 4 + 1 + 0 + 0)/5 = 2.4
(1)E(SDL) = E(max[AT − P AT, 0])) = (0 + 0 + 0 + 5 + 9)/5 = 2.8
(2) This shows that the averages values of bothE(SDE)
andE(SDL)
are positive, although the arrival time is either late or early. This is aconse-quence of the uncertain travel time and ofthe fact that expected scheduling
costs arenon-linear functionsoftraveltime. This explainswhythe expected
scheduling variablescapture some of the disutilityarising fromit.
Sincepreferredarrivaltimeisrequired forcomputingthe schedulingcost
variables,thisapproachhasmostlybeenappliedinassociationwithtrip
tim-ingchoicemodelling. Therearerelativelyfewexamplesofstudiescomparing
1
A set of express lanes on an otherwise free and congested road oers high-quality
provide alinkbetween theseapproaches by theoreticallyshowing that if the
mean travel time is independent on time-of-day (which, on the other hand,
inunlikelyinacongested network)andnodummyvariableforlate arrivalis
included (a discrete arrivinglate penalty), both approaches are equivalent.
Thediscussionabout ofhowtraveltimevariabilityshouldberepresented
in choice models, partly originates from the fact that perception and
be-havioural response to uncertain travel times is still relatively unexplored.
Brownstone and Small (2005) review studies on value of time and value of
reliability using RP and SP data from two HOT lane projects in southern
California. These studies found that standard deviation captures
individ-ual's preferences only imperfectly according to the RP data, and that the
dierence between the 90th and the 50th percentile gave a better model t
(see for instance Lam and Small (2001)). Brownstone and Small suggest
that a likely reason for this is that driver's preferences are not symmetric.
That is, the chance of a shorter travel time does not inuence the driver's
preferences as much as the chance of a longer travel time. The dierence
between the 90th and the 50th percentile is abetter indicator of the chance
of being considerablydelayed, ascompared tothe standard deviation,which
might better reect what travellers actually care about. Brownstone and
Small also suggest that the poor results obtained from using the standard
deviation may arise fromincorrect measurements.
For an extensive review of travel time uncertainty and departure time
choice, see Bates et al. (2001), Noland and Polak (2002) and de Jong et al.
elling
The trend of transportation modelling is obvious. It is becoming
increas-ingly disaggregate in various directions, e.g. in the analysis of individual
behavioural responses and its underlying mechanisms, as well as in spatial
and temporal resolution. During recent years, much attention has been put
on activity-based research in the eld of travel demand, as a way to better
understand the underlying mechanisms of travel behaviour.
Activity-based approaches are characterized by two key ideas. First,
travel is a derived demand, arising from the need or wish to participate in
activities withinatime andspace continuum. This standsincontrast tothe
conventional model approach, in which the trip itself is the focus. Second,
activity-basedresearchemphasises thetemporalandspatialpossibilitiesand
constraints as well as the context in which individuals or households make
theirdecisions. Travelisassumedtocausedisutilityintermsofthemonetary
and time resources it requires. It occurs therefore only when the net utility
ofanactivityperformedatadistantlocation,andthetripthatisrequiredto
gothere,exceedsthe utilityoftheperformance oftheactivityatthe original
location.
The family of activity-based models includes a set of approaches and
methods rather than a model per se. These have been developed with a
varying degree of detail. Many of the eorts to implement activity-based
modelshaveappliedmicrosimulation,whichbynatureishighlydisaggregated
in that eachagent ismodelled individually. (Millerand Salvini2002)
Since the hypothesised implications of ICT-based services concern the
scheduling possibilities and relaxed constraints in time and space, it has
often been argued that activity based modelling isbettersuited for analysis
of this process. The work in this thesis starts, however, out from the more
conventionalmethodswherethetripisinfocus. Thisismotivatedforseveral
reasons. IncludingimplicationsfromICTinanykindofactivity-basedmodel
is complicated, and would require a considerable research eort, in terms
of time and data resources. It is therefore important to, at the rst stage,
enlargeknowledgeandunderstandingabouthowtomodelbehaviouraleects
from ICT, as much as possible. At this stage, it is less important what
model approach that is used. Many of the results from this work apply to
the modelling of implications from ICT use per se, and not to a specic
model approach. The conventional modelling approach is further very well
established and accumulated experiences from data collection, estimation
Ithasalsobeenarguedthattheactivitybasedframeworkisbettersuited
tomodeltriptimingchoice,sincethisiscloselyrelatedtoactivityscheduling.
Some activity-based modelsalsooperateinatime-space continuum. Insuch
models, it is feasible to model trip timing choice as a continuous choice
and which reduces aggregation errors. The more conventional, trip based,
modellingapproachis,however, motivatedforthe samereasons asdiscussed
above. By developing activity and trip based applications parallel, sound
4.1 Methods
Discrete choice methods are used in all papers in this thesis. The rst two
papersapplythemultinomiallogitmodel(MNL).Thiswidelyusedmodel
re-liesontheassumptionthattheerrortermsofallalternativesareindependent
and identically Gumbel distributed. From this condition, the irrelevance of
independent alternatives(IIA) property follows. This property implies that
the cross-elasticity isconstant, i.e. thatif anattribute intheutilityfunction
of one alternative changes, the choice probabilities of all other alternatives
change by the same percentage. The MNL model also maintains the
as-sumptions of response homogeneity. All individuals are hence assumed to
be equallysensitive to the attributes included in the utility function.
(Ben-Akivaand Lerman1985)
However, there are also more general forms of discrete choice models
that relax these assumptions to various degrees. The most general of these
is the mixed logit model (Train 2003). The utility functions may include
several error components. The utility function of each alternative includes
oneindependentandidenticallydistributederrorterm,asintheMNLmodel.
Othererrorcomponentscanbespeciedsoastoachieveadesirable
variance-covariance structure of the model. The modeller species the distributions
of these. Several namesareusedforthis model,dependingonwhatproperty
that is emphasized. 'Mixed logit' isoften used for itsgenerality. It refers to
the fact that it is a mixture of logit models with a specied mixing
distri-bution, i.e. the distribution of the additionalerror components (Revelt and
Train 1998). The name 'Probit with a logit kernel' was used by Ben-Akiva
and Bolduc (1996). 'Probit' indicates that the mixing distribution in this
modelis normal.
Otherfrequentlyused termsare 'randomtaste logit'or'errorcomponent
logit'. Theformernameindicatesthatthismodelcanaccommodateresponse
heterogeneity. Thelatterreferstothefactthatthemodelallowsspecication
of an arbitraryerror-component structure. The dierence between the
'ran-domtastelogit'andthe'errorcomponentlogit'is,however, entirelyamatter
of interpretation (Train 2003). The substitution pattern of this model can
beveryexibleanddeterminedempirically. Thecross-elasticityoftwo
alter-nativesislargerthe smallerthestandard deviationofthe utilitydierenceis
between them. That is, the cross-elasticity of the choice between two
alter-natives is larger the more common unobserved attributes these alternatives
share, relativeto other alternatives.
two papers, assuming response heterogeneity also gives rise to a favourable
error-components structure.
In paper III, the normal distribution is applied to the cost parameter.
This distribution is by far the most common mixing distribution, because
it is easy to implement and works well technically. The limited number of
distributionsavailableinexitingsoftwarepackages isanotherreasonforwide
use of the normal distribution. However, more attention has recently been
paid tothe factthat thenormaldistribution isinappropriateinmany
appli-cations, due tothe fact that it isunbounded onboth sides and symmetrical
(Hess 2005). Itistypicallyunsuitablefortravelcostparameters, becausewe
knowapriori thatindividualsdonothaveapositivevaluationoftravelcost.
In paper III, the parameter value for the cost attribute is actually negative
for a non-negligiblepart of the population. Now, in this case it is still
pos-sible that some individualswould be willing to accept a salary reduction in
order to receive a exible oce. Since the normal distribution is used this
conclusion can, however, not be inferred fromthe data.
To avoid problems with wrong-signed random parameters we have
ap-plied the bounded and non-symmetrical
Johnson
0
s S
B
distribution toesti-mateschedulingandmonetarysensitivityinpapers IVand V.
Johnson
0
s S
B
distribution is highly exible and a nonlinear transformation of the normal
distribution. It can approximate the normal and log-normal distributions
but can also be specied to have a plateau or to be bi-modal shape (Train
and Sonnier 2005).
An important advantage of the MNL model is that ishas a closed form,
i.e. the choice probability can be expressed analytically. In contrast, the
(unconditional) choiceprobabilityof themixed logitmodelcannotbesolved
analytically. Thisisinsteadobtainedbyintegration oftheconditionalchoice
probability,overallpossiblevaluesoftherandomparameters. Thedimension
ofthisintegralequalsthenumberofrandomparametersorerrorcomponents
inthe model. Forestimation the simulatedmaximumlikelihoodapproachis
most widely used. The simulatedmaximum log-likelihoodfunction is
devel-oped fromthe simulated choice probabilities.
The downside of the mixed logit model is that the estimation times are
relativelylong. Analternativetothe mixedlogitisthe 'Probit'model. This
model also requires approximation of a multi-dimensional integral of each
iterationstep. The run time isstillshorter becausethe more ecient
GHK-estimator(see Train (2003))canbeapplied. However, this modelisalsoless
general since only normally distributed error components can be used, and
has not been used in the work of this thesis.
implemented in large-scale applications. As mentioned, there is an
alterna-tive class of closed form models, GEV models that also can accommodate
exible correlationstructures. McFadden (1978)rst derived the GEV-class
of models, and ve dierent model structures in this class have so far been
developed. Foranoverview of these see Bhat (2002).
Fortrip timing choice the Ordered GEV (OGEV) has received much
at-tentionbecausethismodelshouldbeappliedincaseswherethereisanatural
ordering of the alternatives. The disadvantage of GEV models is, however,
that they, incontrast to the simulation-basedmodels, cannotaccommodate
correlationintheunobserved utility(theerror term)across dierentchoices.
This kind of correlation often arises across choices within the same
indi-vidual in data sets that contains more than one choice per individual (this
often is the case in SP studies). The OGEV model structure was tried in
the departuretime choicemodellingprocess,but wasrejectedbecauseof the
importanceof correlation in the unobserved utilityacross dierentchoices.
For the estimations, the software program ALOGIT was used in paper
I and II and partly in paper III. In paper III the estimation procedure of
the mixed logit model was individually programmed in the programming
languageOx. InpaperIV andVthe software programBIOGEME(Bierlaire
2005) was applied.
4.2 Data
In this thesis several data sources have been used, both SP (stated
prefer-ence) and RP (revealed preference) data. The stated preference method is
based on hypothetical choice situations presented to the respondent and is
widely used in the marketing eld toanalyze consumer behaviour(Louviere
et al. 2000). Revealed preference data exposes the respondent's actual
be-haviour inareal situation. SP dataand RP data both haveadvantages and
limitations. Combining RP and SP data in discrete choice modelling, have
therefore a long tradition (Ben-Akiva and Morikawa 1990). Methodologies
to use the two types of data sources simultaneously in model estimations
are well established, and the widely used logit (see section 4.1) model has
so far been the predominant methodology (Bhat and Castelar 2002). Joint
SP-RP estimation is carried out in the nal paper in this thesis. However,
as discussed more advanced model structures than the MNL model is used
to overcomecertain limitationsof the logitmodelstructure.
The data primarily used in the rst part of the thesis (paper I and II)
was not collected for the present studies in particular, but is a repeated
collectedKOMonayearlybasis(thecollectionisnowcarriedoutwithlonger
intermissions). The collection was rst designed in 1996, and the design is
very typical for that time, when there was a very general interest in the
explicitlinkbetweentravelandothermodesofcommunicationbehaviour. In
thisdataalltransmittedcontactsareincluded,andarecategorizedintomail,
e-mail, phone, cell phone, Internet, fax or video and teleconferences. KOM
further includes traditional socio-economic and geographic data, but also
dataconcerningaccesstotelecommunicationsequipmentandtelecommuting
habits. Thisdatawaspartlyformodellingpurposesandpartlyfordescriptive
purposes. The modelling studies in this thesis are, however, so far the only
ones thathave been carried out using this data.
Thedataused inpaperIIIwascollected throughaSPsurvey atthe
tele-com company Ericsson in one oce district in Sweden. The SP questions
were supplemented with questions concerning the respondents' preferences
for telecommuting and actual experiencewith this work form. The
employ-eesreceivedthequestionnairetroughthecompanyinternalmail-system. The
aim of this survey was to discover company incentives to promote
telecom-muting and identify tools that the company can use to promote this work
form. The company incentive addressed in the SP survey was the potential
savingsfromreducedrentalcosts thatacompany maygainfromintroducing
exible oces simultaneously with telecommuting.
The survey conducted to obtain data for the estimation of the
depar-ture time model(paper IV and V) was alsospecically designed within this
project. An SP survey, designed to explore the trade o that commuters
make between shifts from the preferred departure time, travel cost, travel
time and travel time uncertainty, was carried out inthe spring 2005.
The population from which the respondents were recruited consisted of
car drivers travelling toward the city centre during the extended morning
peak period (06.00-10.00). The drivers were rst registered by road side
number plate registration, and a survey agency then called them the same
evening. Information about the observed trip (i.e. purpose, departure time,
travel time and preferred departure time) as well as some socioeconomic
information (including income) was collected. An SP survey was mailed to
them the followingday.
The RP data included the same set of respondents. Information
con-cerning actual mean travel times was obtained from the dynamic
assign-ment model, CONTRAM (Leonard et al. 1989), calibrated for a Stockholm
network. From this application we obtained travel times for the extended
morningpeak (6.30- 9.30a.m) with15minutes intervalresolution. Data on
5.1 Paper I, 'A Communication Choice Model'
This paper investigates several models with dierent structures, assuming
substitution between contact- and travel-based postal and bank activities.
The nal modelshows interesting relationshipsbetween socio-economic
sta-tusandcommunicationandtraveldemandforpostandbankactivities.
Well-educated people perform more contacts bytelephone and Internetthan
oth-ers. Women use the telephone more, while men are more inclined to use
the othermodes of communication, includingtravel. Older individuals have
lower accessibility to Internet. The age eect for Internet use is, however,
rather weak. This is surprising, since younger people in general are more
inclined toadopt new technology. Itis possible that older people can better
aord Internet use.
The result from the communication model estimation indicates some
problems with the model specication and model approach, as a means to
simultaneously evaluate contact and travel behaviour. Some of the
prob-lems thatarose were relatedtolimitationsofthe data. Otherproblems were
related to the model approach. The dierent kinds of problems are partly
connected,sincethesizeandthetypeofthedatasamplerestrictedthemodel
approach toa large extent.
Since information about the actual activities is so limited in the data
that was used, only possibilities of modelling substitution at a one-to-one
levelcouldbe investigated, for agiven activity. Othereects (se section2.1)
are even more complex and modelling these requires that it is possible to
link tripsand contactsconcerning the sameactivity. More complex ICT
im-plicationstend alsotobe more dynamicand thereforelong-term. Modelling
these would hencerequire datathat ismore processoriented andfocused on
habits, stretching overa longerperiod.
Scenario-based stated preference (SP) data, or in-dept interviews could
bea feasible way of collecting more process or activity oriented data taking
attitudes and perception better into account. SP data collectionwould also
beless dependentofthetimelagofthebehaviouralresponsetonew
technol-ogy. That is, the adoption rate of the services is likely to be higher than in
RP data,enablingmore sophisticatedmodellinganalysisthaninthe present
analysis. However, a disadvantage is that SP data may cause behavioural
overreactions to new technology based onfuturisticideas.
Theproblemswiththemodelapproachisprimarilylinkedtodenitionof
activities. Whenassumingsubstitutionataone-to-onelevel,asinthismodel,
are in such cases not exactly the same. In reality, substitutable activities
may be very dierent. For instance, downloading a movie orplaying games
via Internetathome maysometimes substitute recreationactivitiesatother
places, such as shopping and visiting friends. Even in cases where the goal
of the activities is similar, like for postal and bank activities, are the actual
activities are dierent, depending on the mode of communication (physical
or non-physical). The broader the activities are dened, the more dicult
it is to identify and specify when substitution actually takes place. This
means that the observed substitution eect will in generalbedependent on
the denition of the activities inthis kind of analysis.
Manyofthe ICTbased activitiesdidfurtherpreviouslynot exist. Hence,
ICT implies in most cases that activities are substituted, rather than that
the mode of communication is substituted. As the technology develops,
ac-tivities will continuously be substituted. This process is very dicult to
forecast, since it is not only driven by technology development. Attitudes
and perceptions play alsoa criticalrole for new activity adoption.
Now, all transport modelling approaches is dependent on a pre-dened
set of activities. Thefactthat the actualactivities changeasaresultofICT
development, implies that the possibilities of modelling substitution eects
are limited,irrespective of data source and modelling approach.
5.2 Paper II, 'Telecommuting and work travel demand
modelling in Sweden'
The most commonly discussed substitution eect is telecommuting. At
present, telecommuting explains a very small part of the variation in work
trip frequency between dierent industrialsectors. Only 124 employees out
of 7576 actually telecommute full days athome. Inthis perspective itis not
urgenttoincludetheeectoftelecommutinginlarge-scaletracforecasting.
Only 38 % of the working force works in industrial sectors in which
telecommutingisobservedatallinKOMandRES(SIKA2001)(RES is
sim-ilar to KOM, but includes only trips and no non-physical contacts). These
industrialsectors are concerned with computers, nance or media and
com-munication, authoritiesdealingwithissues concerninginfrastructureand
en-vironment and universities. Employees with high income and self-employed
persons have a larger propensity to telecommute. These groups have
pre-sumably high freedom in their work situation. The variables explaining the
private situation, i.e., children in the household, living with an employed
groups that have a possibility to actually choose. We can therefore expect
that anegativeattitudefromlabourmanagement,jobsuitabilityand lackof
neededequipmentexplainthelowtelecommutingfrequencytoalargeextent.
Telecommutingdoesnot seem tobeinuenced by orinuence low
acces-sibilitytothe labourmarket. Surprisingly,individualswithlowcar
competi-tion actually have larger propensity totelecommute. This alsopoints tothe
fact that accessibility tothe labour market was not animportant factor for
adopting telecommuting atthe time whenthe data was collected.
In summary, KOM data is useful to get an overview of the spread of
telecommuting. However, the data is too general and broad with respect to
the work force and also with respect to the information on facilitators and
constraints, in order to allow for deeper forecasting studies. The number of
telecommuters in the sample is very small, which makes model estimations
uncertain.
5.3 Paper III, 'Company Incentives and Tools for
Pro-moting Telecommuting'
As mentionedinthe introduction,this paperismoredirectlypolicyoriented
and seeks to nd ways to increase the amount of telecommuting. Since the
ndings in paper II indicated that constraints in the work situation and a
negativeattitudefromlabourmanagementhindertelecommuting,thispaper
aims at identifyingmeans for the company topromote telecommuting.
Thebasicideaisthattelecommutingimpliesthatalargeproportionofthe
company's oce space isunoccupied, providinga potentialto reduce rental
cost. However, to utilize this eciently, exible oces must be introduced
in which the employees donot have their own oce, but use any desk in an
open oce space. Employees' monetaryvaluationof thepresent oceplace,
in comparison to a exible oce is therefore tested, in order to estimate
the potential rental cost savings. The results indicate that employees are
in fact sensitive to the monetary compensation and that company benets
could be obtained from introducing exible oces. It is further indicated
that employees perceive an increase in work eciency from telecommuting,
whichis ultimately protable tothe company.
A majority of the employees in the study further want to telecommute
more that they actually do. This nding is coherent with the indication
that telecommuting is a privilege for certain groups that have a possibility
to actuallychoose, found inpaperII. Consequently we believe thatit would
computer supportand equipment to the employees.
In summary, the result suggests that package solutions, where the
com-panyallowstelecommunicationconditionallyontheemployees'acceptanceof
exible ocescouldprovetobeaneective companypolicy. In reverse, this
kind of package solution might reduce the levels of the monetary
compen-sations for introducing a exible oce. Analysis shows that telecommuters
actually demand less compensation to accept a exible oce than others.
The compensation to telecommuters could also be designed as company
-nanced technologicalequipmentand relatedservices.
Asmentionednoevidences that telecommutingshould giverise tourban
sprawl, was found in paper II. A joint conclusions from paper II and III
is thus that there is ground for promoting telecommuting from a societal,
individualand companyperspective. Itis,ontheotherhand, notlikelythat
telecommutingwillbealargefactor forthe overallworktripfrequency. This
is, however, dependingnot only onhowtelecommutingispromoted but also
on otherfactors, such as transportcosts.
5.4 Paper IV, 'Departure Time Modelling:
Applicabil-ity and Travel Time Uncertainty. '
Paper IV and V develops a departure time and mode which choice model.
Thismodelexplainshowcardriverstradeotraveltime,traveltime
variabil-ity,monetaryandschedulingcostswhen choosingdeparturetime. Given
ap-propriate inputdata andacalibrated dynamicassignment model,the model
can be applied tocalculate benets fromsuch as infrastructure investments
and policymeasures.
The modelshouldprimarily beused for analysesin relativelyshort-term
perspectives. Typical examples would be evaluation of congestion pricing
schemesandinfrastructureprojectsthatarerealizedwithinarelativelyshort
time. In longer-term perspective, alterations in trac generation and
desti-nation choice are more important dimensions. These are not considered in
themodel. Mode choiceisfurtheronlyconsideredpartially,bymodellingthe
propensity to switch from driving. It does not, however, take into account
that public transport travellers may switch todriving.
Previousstudiesanalysingtraveltimevariabilityusingtheexpected
schedul-ing approach (see 2.2) have started out from the assumption that travellers
intend to arrive at the most preferred arrival time,
P AT
. Travellers are thusassumed tochoose departure timesoastominimizetherisk ofarrivingtravel conditions. The departure time shifts considered are therefore
con-siderably larger than those considered in previous studies. Since we assume
thattravellersshiftintendedarrivaltime,expectedschedulingcost shouldbe
computed with respect to the new intended arrivaltime, and not to
P AT
. Now, the intended arrival time is usually not directly observable, whichmeans thatthe researcher has tomakesome assumptionabout it. A
reason-able assumption seems, however, to be that it is equal to expected arrival
time. For instance, assume that, for a given departure time and traveller,
arrival time usually vary between 8.00 a.m. and 8.30 a.m., and expected
arrival time is 8.10 a.m. It is then reasonable to assume that the traveller
use 8.10a.m. as areference. Arrivalearlierthan 8.10 a.m. isthusperceived
as early arrival, and arrival later than 8.10 a.m. as late arrival. This
as-sumption leads to equivalence of the expected scheduling approach and the
mean-variance approach, if the travel time distribution has certain
proper-ties. The lognormal distribution, given that the standard deviation of the
underlying normal distribution is xed, and the exponential distribution,
both have these properties.
In the present paper the model is estimated using SP data collected
for this particular purpose. Estimated schedulingdisutility, with respect to
travel time, is in linewith earlier studies. We found signicant unobserved
heterogeneity in scheduling costs sensitivity, but no observed heterogeneity
except between the three populationsegments. This points to the fact that
there are variousconditions thatdetermine the drivers schedulingexibility,
inadditiontoworkschedulingandtravelpurpose. Theexplanatoryvariables
availablein this study could,however, not capture these conditions.
Itwas alsofound that temporalconstraints are important both atorigin
and destination, but that they workin dierent directions. Hence, temporal
constraintsatoriginareprimarilyrestrictingearlydepartureand constraints
at destination primarily restricting late arrival. If not taking this into
ac-count,schedulingcostswasconsiderablyunderestimated. Specically,alarge
proportion oftravellers then appear tohave avery lowvaluation scheduling
costs. ThiscouldeasilybeinterpretedasSPartefacts,causedbyrespondents
not taken their restrictions properly into accounts, whereas it is actually an
eect of the less well specied
SDL
variable.Sensitivityfortraveltimeuncertainty,orreliability,isnormallycomputed
asa ratiobetween sensitivityfor standard deviationand sensitivity tomean
travel time and this ratio is often called reliability ratio. This study shows
also that reliability ratio is dependent on intended arrival time. If intended
arrival time is earlier than
P AT
, travel time uncertainty is not signicant in the two largest segments. Travel time uncertainty also proves to be leastThis resultwas expectedsince manyof thesetravellers donot havetoarrive
punctually.
In order to avoid travellers with wrong-signed scheduling cost and cost
sensitivity the highlyexible
Johnson
0
s S
B
distribution, bounded in the
in-terval[-1,0]was appliedtoallrandomparameters. Theparameter estimates
proved to be insensitive to the assumption about these bounds. Analysis
further showed that the pseudo random draws perform remarkably well in
comparison totheModiedLatin HypercubeSample(MLHS)draws as
pro-posed by Hess etal. (2006),incoherence with theirndings.
Takingcorrelation inthe unobserved heterogeneity acrossthe same
indi-vidual into account proved to be crucial for correctestimation of the
distri-bution of the random parameters. This implies that valuationof scheduling
disutilityisrandomlydistributedinthepopulation,but relativelyconsistent
across allSP choices. Ifneglecting the correlation of the unobserved
hetero-geneity we thus fail to estimate the substitution pattern correctly and the
modelevencollapsestoanestedMNLmodel. Theconsequenceofthisisthat
wecannotapplytheclosed-formOGEVmodeltothisdata,whichwouldhave
beenfasterintheapplication,sinceGEVmodelscannotaccommodate
corre-lation in unobserved heterogeneity. However, model implementation proves
that the run times of the simulation based mixed logit model are relatively
small compared to the run times of the assignment model, and the mixed
logitmodelmay thusbe used for implementation.
Finally,assumingthat departure timespresented inthe SPchoices dier
relatively much from
P DT
, involves a data collection problem. Including travel time uncertainty and departure time shifts jointly as a attributes,will in general imply that the disutility arising from departure time
shift-ing is larger than the disutility arising from travel time uncertainty. The
consequence is that it more dicult toproduce reasonable trade os, which
increases the risk of gettingpooraccuracy inthe estimates.
5.5 Paper V, 'Joint RP-SP Data in Mixed Logit
Anal-ysis of Trip Timing Decisions'
This paperextends theanalysis ofpaperIV,by usingjointRPandSPdata.
It has provided an opportunity to compare observed and stated behaviour,
whichis very valuable. Noother such comparisonsondeparture time choice
have been published. The present comparison indicates that there are
sys-tematicdierences inRPand SPdata. Thesedierences aremanifestedina
ual. Thesedierences seem,however, tobesmallerforcommuterswith xed
schedule. In order to assess the validity of the model, itis importantto try
to analyse howthese dierences arise.
67%oftherespondentsreportedthattheyactuallydepartedatpreferred
departuretime(
P DT
)intheRPdata,whichexplainsthehighresponsescale in the RP data. This proportion, and the response scale dierence, wouldprobably have been smaller if the cost attribute had been included in the
RP data, as well as in SP data. It might also have been smaller if we had
asked directlyabout preferredarrivaltime, insteadofdeparture time, inthe
interview.
Inthe RPdatawe have furtheraggregationerrors,since wehavedivided
the extended morning peak into15 minutes intervals, which might have
af-fected estimated trade-os. This would be one reason for the dierence in
scheduling disutility between RP and SP choices within individuals. The
temporal dierences in RP and SP choice situations might also be an
im-portant factor. In RP data observed behaviour is a result of a long-term
adaptation to actual travel conditions. The SP choice is short-term in the
sense that onlyone particular trip is considered. This has probably implied
higher temporalexibility for some travellers and smaller exibility for
oth-ers, in the SP choice. If this explanation is valid, SP data is generally less
trust worthy than RP data, given that long-termbehaviourand preferences
are what we wishto analyseand forecast. In this caselong-termrefers to at
least a couple of months, as opposed to one or a few days. For commuters
with xed schedule we expect that short-term and longer term exibility is
more equal. This would explain why the scheduling sensitivity is more
con-stant across RPand SPchoices within the same individualin this segment.
Itispossiblethattheproblemwithdierenttemporalperspectivesinthe
RPand SPchoicewould havebeenavoidedif wehadasked explicitlyforthe
long-termchoice. Thereare, however, two reasonsfor presenting the choices
as concerning only the observed trip. First, along-term choice relies on the
fact that the trip is made on a regular basis. Asking about the long-term
choice thus excludes trips that are not made regularly. Second, we tried to
make the choice situationas concrete and realisticas possible. We assumed
that this would help the respondents totake their restrictions properlyinto
account. Also, ten respondents taking part in a pilot survey were called
afterreturning the questionnaire. We asked if their choices wouldhavebeen
dierentif wehad asked about theirlonger-termchoiceand not particularly
about the observed trip. All ten respondents answered that their choices
would not dier. This was also a reason for our decision about the survey
Demand Modelling?
•
One-daycross-sectionaldataingeneralinunsuitableformodellingICT implications relatedto travel.•
The fact that the actualactivities changes asa result of ICT develop-ment,impliesthat thepossibilitiesofmodellingsubstitutioneects arelimited, irrespective of data source and modellingapproach.
•
Telecommutingexplainsatpresentaverysmallpartofthe variationin work trip frequency. There is no evidence that telecommuting shouldgiverise tourban sprawl.
•
Telecommuting seems to be a privilege for certain groups that have a possibility to actually choose and a majority of the employees atthe telecom company Ericsson want to telecommute more that they
actually do.
•
Employeesareinfactsensitivetomonetarycompensationandcompany benetscould beobtainedfromintroducingexible ocesincombina-tion with telecommuting.
•
A departure time and mode switch choice modelis estimated on joint RP and SP data. It explains how car drivers trade o travel time,travel time variability, monetary and scheduling costs when choosing
departuretime. Givenappropriateinputdataandacalibrateddynamic
assignmentmodel, the modelcanbeappliedtocalculatebenetsfrom
such asinfrastructure investments and policy measures.
•
Weassumethatintendedarrivalisequaltoexpectedarrivaltime. This implies that the expected schedulingapproach and the mean-varianceapproachareequivalent,ifthetraveltimedistributionhascertain
prop-erties.
•
Temporal constraints are important both at origin and destination, but constraints at the origin are primarily restricting early departureand constraintsatthe destinationare primarilyrestrictinglate arrival.
Sensitivity for travel time uncertainty isonly signicant of if intended
arrivaltime is later than
P AT
.•
Takingcorrelationinunobservedheterogeneityacrossthesame individ-ualintoaccountwascrucialforcorrectestimationofthedistributionofityis randomlydistributed inthe population, but relativelyconsistent
across allSP choices.
•
There are systematic dierences in RP and SP data, and these are manifested in a larger response scale in the RP data, and the factthat scheduling disutility is not constant across RP and SP choices
within each individual. The scale dierence is largely due to the fact
that 67 % of the respondents reported that they actually departed at
preferred departure time. The response dierences accros RP and SP
choices might be due to the temporaldierences in RP and SP choice
This thesis shows that conventional travel demand models are not suitable
for modelling implications from ICT. Activity-based approaches might, at
least fromsometheoreticalpointsofview, bebettersuitedforincorporating
ICT related eects, because of the focus on activity scheduling in time and
space. However, since denition of activities is crucial and ICT seems to
imply that these constantly change, it will be very dicult to include ICT
implications inany explicitmodellingapproach.
Even if it proved to be possible to model implications of ICT in travel
demand models, the problem of forecasting still remains. Perceptions and
attitudes, which are critical as explanatory variables for future ICT-based
serviceadoption,aremuchmorediculttoforecast,thanconventional
socio-economic data. The problem is even more severe if we add the diculty of
forecastingwhattypesoftelecommunication-basedservicesthat willemerge.
ICTandtravelsubstitutionwillpresumablyalsobeverydependent onother
factors in society, such as travel costs and economical growth.
Considering these issues, in-dept interviews and stated preference-based
data, where scenarios are presented to the respondents, are the only
possi-bility lefttoforecastingICTservice adoptionandtravelrelatedimplication.
This method was used in the European Commission 6th framework POET
project (de Jong et al. 2006). However, this kind of data collection is
de-pendent on a realistic description of the scenarios. It is further dependent
on the fact that respondents actually know howthey would respond to new
technology, withoutactually havingexperience of it fromtheir dailylife.
On the other hand, this work shows, inline with other studies, that the
expectation and hope that ICT would reduce travel was exaggerated. It is
clear that individualsand commercial business in general have not used the
opportunities given by telecommunications to reduce physical travel. The
POET study also showed that the eects on travel pattern implied by new
ICTservicearegenerallysmall. Asdiscussedinsection2.1,ICTimplications
arealsopresumablyinuencingtravelpatternsinaverycomplexmannerand
are dicult to distinguish from other factors acting on travel patterns. In
this perspective, and given the diculties of modelling the eects, it does
not seem veryurgenttospend a largeresearch eorttoincludeimplications
from ICT in large-scale trac forecasting.
Theprimaryresearcheort, fromatransportperspective,shouldinstead
focus on the challenge to make use of new service innovations to meet the
goals set up by society. In this perspective, ICT has the potential to
re-duce negativeeects from transportation. The focus shouldprimarily be on
improvedinformationservicesinthepublictransportandcarpoolingsystems
wouldincreasetheattractionofthese,lessresourceconsuming,travelmodes.
Along-termobjectiveofSILVESTER (seesection1),istolaterintegrate
the departure time model with the mode choice model of the Swedish
na-tionaltravelforecastingmodelSAMPERS(BeserandAlgers2001). However,
the demand models of SAMPERS are estimated in another choice context,
as compared to the departure time choice model. In the former frequency,
mode and destinationchoiceare modelledsimultaneouslywhereas the latter
assumes xed frequency and destination choice. Integrating the departure
time choice model directly in SAMPERS, would therefore cause
inconsis-tency. The easiesway toovercomethis problemwould betore-estimate the
mode and destination choice modelof SAMPERS. The inclusive value from
the departure time modelwould then replace time and cost variables in the
car alternative. This is possible since we explicitly have estimated the RP
scale of the utilityfunction of the departure time model.
Anobstaclewhenmodellingeects oftraveltimeuncertainty,andforthe
SP designinparticular, wasthe limitedknowledgeabout the distributionof
travel times, and drivers' perception about this. Little was further known
about travellers' preferences and behavioural response to this attribute.
Perceptions and behavioural response to uncertain travel times is very
diculttoanalyse fromdata collectionperspective. In theRP situationthe
problemisthataverylargenumberofobservationsisneededtogetacorrect
measurement of the travel time variability. Excluding observations in the
cleaning processing must therefore be made extremely careful. The obvious
risk is otherwise that observations of unusually high travel times, which are
the most important ones, are assumed because of expected measurement
errors. Theproblemof measuringvariabilityinthe SPdesignarises fromits
complexity,andthe factthatthis attributeingeneralisdiculttointerpret.
A key issue for further research is thus how travellers perceive travel time
variability and howit inuences their preferences.
ThepresentworkalsoindicatesdierencesinSPandRPdataconcerning
trip timingchoice. Fortracforecastingit wouldtherefore beveryvaluable
to analyse RP data includingthe cost attribute. Such data is presently not
available,but can be obtained fromthe tolls trialin Stockholm. Ideally the
respondents taking part in the present survey would be asked again about
their departure time during this trial. New respondents could also be
re-cruited and asked about their departure time before, during and after the
trial. This approachrelies, however, onrespondent's abilitytorecollectpast
Arnott, R., A. de Palma and R. Lindsey 1998. Recent developments in the
bottleneckmodel. In: K.J. Button and E.T.Verhoefeds. Road Pricing,
TracCongestionandtheEnvironment: Issuesof EciencyandSocial
FeasibilityCheltenham: Edward Elgar, 79-110.
Axhausen, K., Pendyla, R., 2002. Microsimulation: workshop report.
em-phIn: H.S. Mahmassani, ed. In Perpetual Motion: Travel Behavior
Re-search Opportunities and Application Challenged.. Oxford: Pergamon,
Elsevier Science, .
Bates,J.,Polak,J., Jones,P.andCook,A.J,2001.Thevaluationofreliability
for personaltravel.Transportation Research Part E, 37(2-3), 191-229.
Ben-Akiva, M. and Lerman, S.R., 1985. Discrete Choice Analysis: Theory
and Application to travel Demand. Cambridge: The MIT Press.
Ben-Akiva,M.andMorikawa,T., 1990.EstimationsofTravelDemand
Mod-elsfromMultiple DataSources. Proceedingsfrom the11
th
International
Symposium on Transportation and Trac Theory, Yokohama, Japan,
Elsevier New York,461-476.
Ben-Akiva, M. and Bolduc, D., 1996. Multinomial Probit with a Logit
Ker-nel and a General Parametric Specication of the Covariance
Struc-ture.[PDF]workingpaper,Departmentd'Economique,UniversiteLaval,
Quebec, Canada.
Ben-Akiva, M., Bowman, J. and Gopinath, D., 1996. Travel demand model
system forthe informationera, Transportation, Vol 23,241-266.
Beser, M. and Algers, S., 2001. SAMPERS - The new Swedish National
Travel Demand Forecasting Tool.emphIn: Lundqvist, L.and Mattsson
L-G.ed.Nationaltransport models,RecentDevelopments andProspects,
Springer.
Bhat,C.,1998a.AnAnalysisofTravelModeandDepartureTimeChoicefor
UrbanShoppingTrips.TransportationResearchPart B,32(6),361-371.
Bhat,C., 1998b.Accomodatingexiblesubstitutionpatternsin
multidimen-sionalchoice modelling: formulationandapplicationtotravelmodeand
revealed and stated preferences: formulation and application to
con-gestion pricing analysis in the San Francisco Bay area. Transportation
Research Part B, 36(7), 593-616.
Bhat,C.R., 2002.RecentMethodologicalAdvancesRelevanttoActivityand
Travel Behavior Analysis. emphIn: H.S. Mahmassani, ed. In Perpetual
Motion: Travel Behavior Research Opportunities and Application
Chal-lenged.. Oxford: Pergamon, Elsevier Science, 381-414.
Bhat, C.R., Sivakumar, A. and Axhausen, K.W., 2003. An analysis of
the impact of information and communication technologies on
non-maintenance shopping activities. Transportation Research B, 37 (10),
857-881.
Brewer, A.M. and Hensher, D.A., 2000. Distributed work and travel
be-haviour: The dynamicsofinteractiveagency choicesbetween employers
and employees, Transportation, 27,117-148.
Bierlaire, M., 2005. BIOGEME Version 1.4. Available from:
http://roso.ep.ch/biogeme
Black, I.G., Towriss, J.G., 1993. Demand Eects of Travel TimeReliability.
Craneld Institute of Technology: Centre for Logisticsand
Transporta-tion.
Brownstone,D.,Bunch,D,Train,K.,2000.Jointmixedlogitofstatedand
re-vealed preferences for alternative-fuelvehicles.Transportation Research
Part B, 34, 315-338.
Brownstone, D., Small, K, 2005. Valuing time and reliability: assessing
theevidencefromroadpricingdemonstrations.TransportationResearch
Part A, 39, 279-293.
Cao, X., Mokhtarian, P.L., 2005. The Intended and Actual Adoption of
Online Purchasing: A Brief Review of Recent Literature Institute of
Transportation Studies, University of California: Davis, Research
Re-portUCD-ITS-RR-05-07.
Castells, M., 1996. Rise of the Network Society . Cambridge, MA, USA.
Blackwell Publishers, Inc.
Cirillo, C., Torino, P. D., Daly, A., and Lindveld, K., 1996. Estimating bias
due to the repeated measurements probleminSPdata. The24th