User attitudes towards a corporate Mobility as a Service... · demand, and even attract current public transport users towards taxi and car‐ pool alternatives. In this study we

Author’sPre‐printversionSubmittedtoTransportation

UserattitudestowardsacorporateMobilityasaServiceJuan Manuel Lorenzo Varela (Corresponding author)

KTH Royal Institute of Technology / KTH – ITRL: Integrated Transport Research Lab Address: Teknikringen 10a, 100 44, Stockholm, Sweden. E‐mail address: [email protected] Tel.: +46 8 790 66 95. URL: https://www.linkedin.com/in/jmlorenzovarela/ OrcID: 0000‐0003‐4512‐9054

Yusak Susilo

KTH Royal Institute of Technology / KTH – ITRL: Integrated Transport Research Lab Address: Teknikringen 10a, 100 44, Stockholm, Sweden. OrcID: 0000‐0001‐7124‐7164

Daniel Jonsson KTH Royal Institute of Technology Address: Teknikringen 10a, 100 44, Stockholm, Sweden. OrcID: 0000‐0001‐8901‐5978

AbstractMobility as a service (MaaS) envisages enabling a co‐operative andinterconnected single transport market which provides users with hassle freemobility. Among MaaS postulated benefits, MaaS enthusiasts claim that MaaSsolutionscouldpersuadepeople togiveuptheircar.Conversely, there isa fearthatMaaScouldinfactinducelesssustainabletravel,bymeansofinducingextrademand, and even attract current public transport users towards taxi and car‐poolalternatives.InthisstudyweinvestigateuserattitudesandexpectationstowardsacorporateMaaSsolution,througha latentclassandlatentvariablemodel.Resultssupportthat there is a trend fromcarownership tousership.Wealso findnoevidencethatMaaS solutions could produce a shift frompublic transport users to otherless space‐efficient shared‐mobility solutions such as taxis or car‐poolalternatives under our experiment conditions. In connection with user’spreference to share a car journeywith strangers,we find the existence of twooppositetrends.Thisfindingsuggeststhattheremightbeappetiteforbothtypesof solutions,where users could choose between private or shared journeys bycar. Moreover, we find that normative beliefs impact usermobility styles, andthat theneedand feeling for flexibility is found tobeoneof thekey factors foruserstoembraceaMaaSsolution.Keywords: Mobility as a Service, MaaS, Travel behaviour, Attitudes, Norms,LatentClassandLatentVariableModel(LCLVM).

UserattitudestowardsacorporateMobilityasaService

2

1 INTRODUCTION

As urbanization trends continue to worsen gridlock plagues in a growingnumberofcitiesaroundtheworld,transportplannersareembracingnewwaysoftacklingtheoldproblemofcongestion;andamongthesolutions,MobilityasaService (MaaS) has caught international attention in recent years. MaaSenvisages enablinga co‐operative and interconnected single transportmarketwhich provides users with hassle free mobility, whilst accounting for therealitiesofspatialandtemporalefficiency(Wongetal.,2017).This potential of MaaS to induce significant changes in current transportpracticeshastriggereda lotofon‐goingdiscussionsabout thedevelopmentofthe technical and regulatory elements (backend / frontend systems, routeplanners, legislation, responsibilities, business models, etc.). Nevertheless, farlessattentionhasbeenpaid touserpreferences forMaaSproducts,aswellasthe impact thatMaaS could have onmode choice behaviour andmodal shift,despiteofthefactthatunderstandingtravellers’choicesandbehaviouriskeytodesigningsolutionsthatwilladdressuserdemandsandexpectations.MaaS enthusiasts claim that travel behaviour changes derived from theadoptionofMaaSsolutionscouldpersuadepeopletogiveuptheircar.Thisisaveryambitiousgoalandonethatdoesnot justrelyonthepresenceofaMaaSplatform in a city or region, but more importantly on the availability ofalternative transportmodes (public transport, taxi, bikes, etc.), their effectivecombination, and thewillingness to adopt such amodality styleby the users.Regarding the latter, (Sochor et al., 2016), presented evidence of travelbehaviourchanges froma6‐month field testofUbiGo ‐aMaaSbrokerserviceforeverydayurbantravel‐whereallusergroups’modechoiceshiftedinamoresustainable direction. Nevertheless, authors pointed out that field testparticipants may not represent the “average traveller”, as the project targeturban households, with a certain level of access to the existing transportsolutions, and large enough travel needs for the service to be financiallycompetitive.Hence,theseencouragingresultsmightnotextrapolatewellacrossthe larger group. This finding was also supported at a larger scale by (MaasGlobal,2017).Thisreportshowshowmodesharesshiftedamongusersofwhim‐ a MaaS solution implemented in Helsinki ‐. Here, private car trips werereduced to half, and the share of Public Transport (PT) trips increased by afactor of 1.5, when compared with levels before whim was available.Furthermore,amorerecentstudy,(Kamargiannietal.,2018),foundevidenceofacar‐ownershiptocar‐usershipamongLondoners.Conversely, there is the fear that MaaS could in fact induce less sustainabletravel,bymeansof inducingextrademand, andevenattract currentPTuserstowardstaxiandcar‐poolalternatives.Hence,quantifiedevidenceoftheimpactofMaaSontravelbehaviourisstillneeded.Moreover, (Haahtela and Viitamo, 2017), showed that that themost relevantunit of analysis is not an individual commuter but the family and household


3

whichdetermines theprerequisites for travellingof the familymembers.Thisfinding shows the importance of household dynamics in explaining userdecisions,whichmightseemirrationalifthesefactorsareneglected.Hence,theneedforflexibilityatahouseholdlevelispostulatedasakeyelementforMaaSsystemstosucceed.In this paper we study the interrelation of household dynamics, normativebeliefs1; modality styles2; and, user attitudes towards different transportsolutions.We use a latent class latent variablesmodel to quantify the impactthatacorporateMaaSsolutionmighthaveontravelbehaviourandmodalshiftamong a company´s employees. To our knowledge, no previous study hasexplored in this way user demands and expectations about MaaS solutions,althoughusersarethecornerstoneofMaaS.Fromamethodologicalpointofview,thispaper includesacomparisonoftwodifferentmodelling techniques for ordered variables.Where previous studiessimplified the modelling of ordered variables by assuming a logisticdistribution,whichhasaclosedcumulativeform,(Dalyetal.,2012;andKruegeretal.,2016),thisstudyestimatesthemodelsusingbothanormalandalogisticcumulativedistribution,andcomparestheresults.The rest of the paper is structured as follows: Section 2 presents themethodology.Section3providesanoverviewofthedata.Section4presentstheuserinsightstowardsacorporateMaaSsolutions,andSection5concludes.

2 METHOD

We estimate a latent class and latent variable model (LCLVM) to classifyindividuals based on their responses to a variety of survey questions. Thismodel was first used by Krueger et al. (2016), and it combines elements oflifestyle oriented and socio‐psychological approaches to travel behaviouranalysistoexplaintheformationoflatentmodalitystyles,whichareidentifiedthroughlatentclasssegmentation.Under this approach, the likelihood of an individual to belong to a particularmodalitystyleisafunctionoftheindividualsocio‐demographiccharacteristics,aswellasnormativebelievestowardstheuseofdifferenttransportmodes.The indirectmeasurement of latent variables, normative beliefs andmobility‐related attitudes, is done by collecting various indicators. These indicatorsconsist on a list of statementswhere each respondent is asked to state theiragreement using a Likert scale (Likert, 1932). We start the analysis by

1 Normative beliefs are defined as an individual’s perception of the beliefs of others regarding a specific behaviour. 2 Modality styles represent the part of an individual’s lifestyle that is characterised by the use of a certain set of modes


4

performing a Confirmatory Factor Analysis (CFA) to understand if theformulated latentattributescanbe identified.Although thesestatementshavebeendesigned tocapturesomepre‐determinedaspects, it isuseful to identifywhat are the indicators that reveal most of the information about the latentvariables.Therefore,wecombinetheCFAwithanExploratoryFactorAnalysis(EFA)ontheindicators.

2.1 Modelling framework

ThemodelinthispaperusetheframeworkproposedbyKruegeretal.(2016).Theframeworkconsistsofthreedifferentparts:first,alatentnormativebeliefsubmodel with structural and measurement components; second, a latentmodalitystylessubmodel;andthird,class‐specificsubmodelsforbinomialandordered variables. Figure 1 below shows the conceptual framework. (SeeappendixAfordetails)Figure 1. Conceptual model framework. Figure adapted from Krueger et al., 2016)

Forconsistency,wehaveretainedthesameformulationandterminologyinthisstudy. We also complement the approach taken by Krueger et al. (2016) byimplementing an alternative modelling assumption regarding the stochasticcomponent of the ordered response variables; hence, ordered variables aremodelledusingorderedlogitandorderedprobitmodels. .(SeeappendixAfordetails).Asexplained inKruegeret al. (2016), the largermodel includes several class‐specific submodels, which are constants‐only models, conditional onmembership in a latent class . Depending on the format of the dependentvariable these models are either binomial or ordered logit models. Table 1providesanoverviewofthedifferentclass‐specificsubmodels.


5

Table 1. Overview of class‐specific ordered and categorical data models

Dependentvariable Class‐specificsubmodelcategory

Typeofmodel

Modefrequentuser(car,publictransport,walk,bicycle,etc.)

Mode‐usefrequencies Binomial

Carownership(Yes,no) Mobilityattributes BinomialBicycleavailability(Yes,no) Mobilityattributes BinomialWalktimetoaccessPTlessthan5min Mobilityattributes BinomialTeleworkingfrequency(Morethanoncepermonth,lessthanoncepermonth) Mobilityattributes Binomial

Mode‐specificattitudesStrongly agree, agree, neutral, disagree, stronglydisagree

Attitudes Ordered

Thelikelihoodfunctionisformedbycombiningthedifferentcomponentsofthemodel across individuals to obtain the unconditional probability of observingthe data. In order to estimate the model, maximum simulated likelihoodmethodsarerequired,astheobjectivefunctionlosesitsclosed‐formbecauseofthelatentvariablerandomerrorterms.(SeeAppendixAfordetails)

3 DATA

The data used for the analysiswas collected among employees of a companyimplementing a corporate MaaS solution. Note that this setup has certainparticularitieswhencomparingittootherMaaSstudies.Forinstance;theMaaSsystem only has one provider, which owns and operates all modes (buses,bicycles, taxis); and there is only one group of users, employees, which havestabletravelpatters(workcommuteandintra‐campustrips).Havingclarifiedthat,thedatawascollectedbymeansofanonlinequestionnairewhich provided 433 observations. The questionnaire was designed to collectsocio‐economic variables, transport related information, and answers to 50attitudinalquestionsonaLikertscalewith5levels.Socio‐economicvariablesinclude:age;gender;numberofchildrenlivinginthehousehold;homepostalcode;typeofjob;carandbicycleaccessibility;aswellastelework habits. Regarding the transport related information, the surveyprovidesinformationon:accesstimestopublictransportfromhome;frequentusageofdifferenttransportmodestocommute,travelinsidetheworkarea,andin their free time; average number of work trips inside the work area on anormalday,aswellashowmuchtimedotheynormallyspendtravellinginsidetheworkarea.Thefinalpartofthesurveywasdesignedtotargetspecificlatentnormative beliefs that were hypothesised to influencemodality styles. Thesefactors included car affective and symbolic attitudes (Steg, 2005); socialinfluence and status; having an environmental mindset; expectations aboutMaaS;publictransportinstrumentalutility;benefitsofactivetravel;anduseoftravelinformation.


6

Table 2 reports the marginaldistributions of selected mobilityattributes and socio‐economicvariables.Aftercomparingthis informationwiththeemployerdatabaserecordsforthewholepopulation,wecansaythat thesample used for the analysis isrepresentativeofthelargerpopulationwithinthecompany.Moreover, it was hypothesised thatresidentiallocationplaysacriticalroleon the traveller adopted modalitystyle; hence, three large areas weredefinedforfurtheranalysis.

Table 2. Sample descriptive statistics (N=433)

VariablevalueSamplemargin

%

Age 18‐24 425‐44 5645‐59 3660‐60 4Gender Male 75CarOwnership Yes 85Bikeaccessibility Yes 83Driverlicense Yes 97Childrenlivinginthehouseldyes

56

Managerialposition Yes 12PTaccesstime 5minorless 525to10min 2910to15mi 11Morethan15min 8

These areas are the greater Stockholm area, the Södertälje campussurroundings, and the Stockholm/Södertälje commuter railway line. Thereasonstodefinetheseareaswere:

‐ Greater Stockholm area (10km buffer around Stockholm city centre).Peopleleavinginthisareahavemorealternativesavailabletocommute,makingeasiertofindacompetitivealternativetotheuseoftheprivatecar.Thisincludesthecommutertrains,buses,andtheemployersexpressbusshuttle.

‐ Södertälje area. (10 km buffer around Scania campus). These users donothavesomanymotorizedalternativesaspeoplelivinginthepreviousarea; nevertheless, they have a better situation to use non‐motorizedmodes(walkandbicycle).

‐ 1kmbufferaroundthecommuterrail linebetweenStockholmandSodSödertälje. People leaving along the commuter rail linemight bemoreprone to commute by railway, and even theymight have chosen theirresidentiallocationbasedonthisfact.

Using a geographic information system (QGIS, 2018) a dummy variable wasdefined for each of the areas described above. These variables take the valueoneifanypartoftheuserhome‐postcodeintersectswiththebuffers.Finally,histogramsofresponsestothe50attitudinalindicatorsarereportedinAppendixB.

4 RESULTS

Inthissectionwepresenttheparameterestimates,log‐likelihoodandgoodnessoffitresultsforthemodelsestimatedwithpythonbiogeme(Bierlaire,2016).Toidentifytheoptimumnumberoflatentclasses,modelswithone,twoandthree


7

classeswereestimated.Themodel specification terminatedwhennoneof thespecificationswiththreelatentclassesconverged.The estimation results for the model specifications with one and two latentclasses are meaningful in a statistical and behavioural sense. Yet, both theAkaike Information Criterion (AIC) and the Bayesian Information Criterion(BIC)indicatethatthemodelspecificationwithtwolatentclassesoutperformsthemodelspecificationwithasinglelatentclass,asshowninTable3.Table 3. Summary statistics for different model specifications

Latentclasses

Latentvariables

ObservationsDraws

Estimatedparameter

Log‐likelihood

BIC AIC

1 5 433 500 182 ‐23148.7 47402.3 46661.42 5 433 500 225 ‐22737.8 46841.5 45925.63 5 433 500 301 Failuretoconverge

Table 4. Comparison of model fit between models with two latent classes, but different assumptions regarding how ordered variables were modelled.

Model Logistic Normal

Assumptiondifference , , ~ 0, 6 , , ~ 0,1

Iterations 491 361Runtime(minutes) 35.32 29.43Finalloglikelihood: ‐22740.1 ‐22737.8

Also, results from comparing the two different modelling assumptionsregarding the stochastic component of the ordered response variables showthatusinganormaldistributionprovidesabetterfitinlessiterations(seeTable4).Hence,intherestofthepaperwepresentresultsfromthemodelwithtwolatent classes, and ordered responses modelled by ordered probit models,whichgoodnessoffitstatisticsarereportedinTable3.

Latent Classes description

The class membership function calculates the probability of individual nbelongingtoeachclass.Inthisstudy,theselatentclassesmodellatentmobilitystyles, which are hypothesised to be a function of socio‐demographiccharacteristics; mobility‐related attitudes; and normative beliefs towards theuseofdifferentmodesoftransport.Belowwesummarisethecharacteristicsofeachclass.

Class1:Car‐orientedclass(~75%)Inthisclass,carownership,regularcommutingbycar,andtheuseofthecar for travelling inside theworkarea ishigher than for theother class.Hence,wecallthisclasscar‐oriented.Coefficient estimates for observed attributes show that the presence ofchildren in the household, andhaving amanagerial positionmakeusersmorelikelytofallwithinthiscarorientedclass.Thesefindingsmakesensefromabehaviouralpointofview,asmanagersandparentswithchildreninthehouseholdareusuallytimeconstrained,makingthemmorelikelyto


8

want to reduce its travel time and have the flexibility to adapt to thehousehold/jobneeds;hence,optingfrequentlyforthecaralternative.Amongthelatentfactorsconsideredintheclassmembershipfunction,wefindapositive influenceof the latent construct “CarAffective”and “HighExpectationsforMaaS”withtheusersfallinginthisclass.Class2:Shared‐mobilityorientedclass(~25%)Userswithinthisclassaremore likelytocommutebyPT;theemployersexpress shuttle; and use non‐motorized transport modes (walk andbicycle).WithinthisclasswefindpeopleleavinginthegreaterStockholmarea,andwithgoodaccesstoPT.Thisresultindicatestheimportanceofcompetitivealternatives in the mode choice. Travellers from the greater Stockholmarea have the ability to choose between a larger set of PT alternatives,including the employers express shuttle which is restricted to a fewroutes.Also young people (18 to 24 years old) are more likely to be shared‐mobilityoriented.Wefindtwocomplementaryexplanationsforthis.First,youngpeopledonothavethepurchasepowertoaffordacar;andsecond,they want to live in the city and they adopted more sustainable travelbehaviours, due to the ability to choose between a larger set ofalternatives.Among the latent factors, we find that “Having an environmentalmind‐set”,andhaving“HighexpectationsforMaaSinsidetheworkcampus”helptoexplainbelongingtothisclass.Thisresultisinteresting,ashavinghighexpectations for MaaS also had explanatory power for the car‐orientedclass.Webelievethatthedifferenceresidesonthefactthatusersfromtheshared‐mobilityclasshavealreadyasuitablecommutingexperience,butitis very difficult to travel inside the campus and they are in need ofsolutions.

Othervariablesincludedintheclassmembershipfunctionare:gender;dummy variables for leaving close to the working place, or the commuterrailway line; and the latent factor “social influence”.Unfortunately,we cannotfind statistical significant (at a 95% confidence level) of these variables inexplaining the classes. Detail parameter estimates of the class membershipfunctionvariablesand themobilityattributesareprovidedonTables5 and6.Furthermore, detail parameter estimates of the latent attributes entering theclassmembershipfunctioncanbefoundinTablesC1andC2inAppendixC.


9

Table 5. Coefficient estimates of the class membership model

Parameter Class1 Class2

Value t‐val Value t‐val

Constant ‐1.80 ‐1.32*

Referenceclass

Gender(reference=female) Male ‐1.23 ‐1.92*Age(reference=45to59yearsold) 18to24 ‐3.21 ‐2.3225to44 ‐0.19 ‐0.37*60ormore 3.18 1.62*Typeofemployment(reference=employee) Manager 1.79 2.07Typeofworker(reference=blue‐collar) White‐collar ‐1.00 ‐1.48*Frequenttravellerinsidecampus(reference=lessthan3trips/day) 3ormoretrips/day 0.77 0.88*Childrenlivinginthehousehold(reference=none) 1ormore 3.94 4.04Homelocation Sodertalje10kmbuffer 1.25 1.88*Stockholm10kmbuffer ‐0.94 ‐1.75*CommutinglineStockholm‐Sodertalje1kmbuffer 0.169 0.34*Latentattributes Caraffective 2.48 3.05Environmentalmindset ‐0.40 ‐2.29HighexpectationsaboutMaaS 0.80 2.10HighexpectationsaboutMaaS(onlyinsideworkarea) ‐0.90 ‐2.19Socialinfluence ‐0.60 ‐1.49*

*Parameternotstatisticallydifferentfromzeroat95%confidence

Table 6. Estimate of coefficients on mobility attributes



Bicycleaccess(reference=no)

Yes 1.77 11.3 1.19 5.05

Caraccess(reference=no)

Yes 4.42 8.56 ‐0.45 ‐2.08

Regularlycommutebybicycle(reference=no)

Yes ‐2.38 ‐11.9 ‐1.62 ‐6.01

Regularlycommutebycar(reference=no)

Yes 4.26 8.12 ‐2.97 ‐4.22

Regularlycommutebypublictransport(reference=no)

Yes ‐2.24 ‐11.1 1.04 4.53

Regularlycommutebyemployer’shuttle(reference=no)

Yes ‐4.41 ‐8.69 ‐0.94 ‐4.23


10

Table 6.(cont). Estimate of coefficients on mobility attributes



Regularlycommutebywalking(reference=no)

Yes ‐2.56 ‐11.9 ‐1.67 ‐6.10

Regularlytravelinsidecampusbybicycle(reference=no)

Yes ‐3.70 ‐10.3 ‐2.97 ‐6.47

Regularlytravelinsidecampusbycar(reference=no)

Yes 1.04 8.07 ‐3.15 ‐5.97

Regularlytravelinsidecampusbytaxi(reference=no)

Yes ‐2.03 ‐11.7 ‐1.16 ‐4.96

Regularlytravelinsidecampusbywalking(reference=no)

Yes ‐0.44 ‐3.86 0.12 0.59*Walkingtimetoaccesspublictransport(reference=morethan5min)

Goodaccess(5minutesorless) 0.004 0.04* 0.32 1.56*

Telework(reference=lessthanonepermonth)

Morethanonepermonth ‐1.11 ‐8.67 ‐1.07 ‐4.68


User attitudes towards MaaS trends

Below we discuss user´s attitudes regarding the opportunities that MaaSsystems are envisaged to bring, andwe point out key differences/similaritiesbetween the two classes. Table 7 shows the estimated distributions of therelatedattitudinal indicatoracross latent classes,onwhich these in‐sightsarebased. Detailed parameter estimates from which the distributions of Table 7werecalculatedcanbefoundinTableC3inAppendixC.

CarownershiptousershiptrendWeobservethatindividualsinbothlatentsegmentsexpresstheirpreferencetonotownacar,butstilltheywantthebenefitsofcartrips.Thisfindingsupportsthe existence of a car ownership to usership trend, as suggested by(Kamargianni et al., 2018). Furthermore, over67%ofusers in anyof the twoclasses agree or strongly agree with the fact that owning a car is a bigexpenditure for thehousehold, suggesting that the costofowninga car is thekeyintheirdecisiontoabandoncarownership.

PTtocarshiftOur results show that the introduction of a MaaS solution is not likely toproduceashift intransportmodesfromPTtocarundertheconditionsofthisexperiment.64%ofusersintheshared‐mobilityclass,and57%ofusersinthecarorientedclass,statedthattheywillnotreplacePTtripswithcariftheMaaSapp will allow them to book taxis. We believe that this result must beinterpretedwithcaution,as inthesetupofthisMaaSexperiment,taxiscanbebookedbut theircost isbilledtotheemployee’sdepartment;hencebookingataximustbejustifiedwhilstusingtheothermeansoftraveldonot.


11

VehiclesharingRegardingthewilltosharecarjourneyswithstrangersdifferentopinionsarise.Around50%ofusersinbothcar‐orientedandshared‐mobilityclasses,seemtobeinfavourofsharingcarjourneys,whilstaround20%ofusers,inanyofthetwoclasses,reportednottofeelcomfortablewithsharingacartripwithpeoplethat they do not know. These results point out that it might be a feasiblebusinesscaseforbothtypesofuserneeds,whereusersmightchoosebetween“private”3orsharedjourneysbycar.

Uptodateandreal‐timeinformationResultsalsoshowtheimportanceofaccesstoinformation.97%ofusersintheshared‐mobility class reported that they check real‐time information to plantheirtripsorcheckfordisruptions.Butthisneedforinformationisnotlimitedtotheshared‐mobilityclass,as60%ofthecarorientedclassreportedthesamebehaviour.Inaddition,39%oftheshared‐mobilityclassand70%ofthecar‐orientedusersreported that theyalwayschoose the fastestalternative.Thisresulthighlightsthe need for adequate information so users can make informed choices,especiallywhenMaaSwillincreasethenumberofalternatives.

FlexibilityFrom the estimated distributions we can observe that the majority of users,49%on theshared‐mobility classand71%on thecarorientedclass, agreeofstronglyagreewiththeneedforflexibilityduetoirregularschedules.Also,weobserve that 64% on the shared‐mobility class and 88% on the car orientedclass reported to agree or strongly agreewith the fact that non‐work relatedactivitiesconditionthecommutingmodechoice.Furthermore,weobservethat42%on theshared‐mobilityclassand51%on thecarorientedclassreportedthat they travel with their current mode because they do not have otheralternative.

3 Private in this sense mean that the car journey is not shared with strangers, but does not imply that the car needs to be privately owned.


12

Table 7. Estimated distributions of attitudinal indicators across latent classes

Statement/ResponseValue Class1(%) Class2(%)Iwouldlovetohaveaccesstoacarwithoutthehassleofowningone. Stronglydisagree 3 3Disagree 11 10Neutral 23 22Agree 28 27Stronglyagree 35 38Owningacarisabigexpenditureformyhousehold Stronglydisagree 1 1Disagree 6 6Neutral 24 26Agree 41 42Stronglyagree 28 25Ifthenewsystemallowedtobookcabs,IwillusethatfeatureinsteadofthePT Stronglydisagree 5 7Disagree 17 21Neutral 35 36Agree 28 25Stronglyagree 15 11Ialwayschoosethefastesttravelalternative Stronglydisagree 0 2Disagree 4 15Neutral 26 44Agree 43 31Stronglyagree 27 8Iuseinformationontheinternettochecktimetablesanddelays Stronglydisagree 6 0Disagree 13 1Neutral 21 2Agree 32 11Stronglyagree 28 86IdonotfeelcomfortablewhensharingacartripwithpeoplethatIdonotknow Stronglydisagree 15 13Disagree 34 32Neutral 33 34Agree 14 16Stronglyagree 4 5IneedflexibilitybecausenormallyIhaveanirregularschedule Stronglydisagree 2 6Disagree 11 23Neutral 16 22Agree 39 34Stronglyagree 32 15ItravelthewayIdobecauseIdonothaveanyotheralternatives Stronglydisagree 4 6Disagree 16 21Neutral 29 31Agree 35 31Stronglyagree 16 11Inormallytravelinthesamewayanddonotplanmytrips Stronglydisagree 1 3Disagree 7 12Neutral 18 25Agree 35 35Stronglyagree 39 25Iplanmycommutetripbasedonother(non‐work)tripsaswell Stronglydisagree 0 3Disagree 2 9Neutral 10 24Agree 26 33Stronglyagree 62 31Ioftenthinkaboutmovingclosertoworkinordertoreducemytraveltime Stronglydisagree 85 66Disagree 10 18Neutral 4 10Agree 1 4Stronglyagree 0 2Acarprovidesstatusandprestige Stronglydisagree 25 24Disagree 35 35Neutral 30 31Agree 10 10Stronglyagree 0 0Iliketodrive Stronglydisagree 0 1Disagree 2 10Neutral 17 38Agree 38 36Stronglyagree 43 15


13

5 CONCLUSIONS

MaaS envisages enabling a co‐operative and interconnected single transportmarketwhichprovidesuserswithhasslefreemobility.ThissituationhasmadeMaaSahottopicinthelastfive‐yearperiod;andeventhoughplentyofresearchisbeingcarriedoutonMaaS,andpilotMaaSsystemsarebeingdeployedallovertheworld,verylittle isknownabouttheconsequencesthatMaaSsystemswillhave on travel behaviour, especially when MaaS systems become the normratherthantheexception.Thisstudyprovidesanswers toopenquestions thatpreviousstudiesonMaaSsystems rose about travel behaviour and user attitudes using a Latent ClassLatentVariablesModel. This analysis is expected tobeof interestnot only tothisparticularemployer,butwouldalsobeinstrumentalinsupportingrelevantstakeholdersallovertheworldwhichdiscusstheintroductionofsimilarMaaSsolutions. By doing so, we hope to contribute to a more evidence andknowledge‐baseddecisionmaking.Thelatentsegmentationanalysisidentifiestwoclasses,whichaccordingtotheirmanifestedbehaviourcanbecharacterisedas:(1)car‐orientedand(2)shared‐mobility oriented. We find statistically significant influence on classmembership from socio‐demographic and normative beliefs for age; childrenliving in the household; having a managerial position; being environmentallyconscious;havingcaraffectiveattitudes;havinghighexpectationsaboutMaaSsolutions;andchoiceofresidentiallocation.Othersvariablesforwhichwecouldnot find statistically significant influence on classmembership, e.g. gender orsocialinfluenceshouldnotbediscardedimmediatelyinotheranalysis,asvaluesandnormsdifferwidelyamongsocieties;hence, it is important tore‐evaluatethe importance of these factors in each new environment. Furthermore, theoutcomeofthistypeofanalysiswillinformpolicymakersonhowtodesignhighimpactpolicies.Regarding the existence of a car ownership to car usership trend, we findevidence to support this hypothesis as the majority of respondents in bothclassesexpressedtheir interest ingettingaccess toacarwithoutowningone,and pointing out that owning a car is rather costly for their households. Thistrend can lead to a reduction of congestion and driven kilometres due to asmallernumberofprivatelyowncars.We also find evidence to support thatMaaS is not likely toproduce a shift intransport modes from PT to car under the conditions of this experiment.Nevertheless,webelievethatthisresultmustbeinterpretedwithcaution,asinthesetupofthisMaaSexperiment,taxiscanbebookedbuttheircostisbilledtothedepartmentof theemployee;hencebookingataximustbe justifiedwhilstusingtheothermeansoftransportdonot.Webelievethatthissituationmightinfluencetheresponsesofthesurveyparticipantsandnotberepresentativeoftheactualbehaviouriftheusersdidnotneedtojustifythetripcost.


14

Regardinguser’spreferencetoshareacar journeywithstrangers,weobservetwoopposite trends suggesting that theremightbe appetite forboth typesofsolutions,whereuserscouldchoosebetweenprivateorsharedjourneysbycar.Resultsalsoshowthatuserswantaccesstoadequateinformationtoplantheirtripsandcheckfordisruptions.Itisforeseenthatthisneedforinformationwillbe evenmore avid, as MaaS systems increase the number of available modealternativesforeachtrip,anduserstrytomaximizetheutilityoftheirchoices.Inthebrightside,MaaSplatformswillbeinanexcellentpositiontosatisfythisavid demand for information from a single location, reducing the burdenassociatedwiththechoice.We also find evidence that users value very highly the possibility toaccommodateirregularschedules,wheretherelevantunitofanalysisisnotanindividualbutthefamilyandhousehold.Hence,successfulMaaSsystemsshoulddemonstrate how they can fulfil these demands, especially if they want toattractuserswithamorecar‐orientedmind‐set.

Finally, these are empirical findings for trips within a particular company inSweden,andwhethertheycanbegeneralisedtootheremployers/citiesshouldbefurtherexplored.Observablelong‐termeffectsofMaaSsolutionsstillremaintobe assessedandwill allowdeterminingwhether the results attained in theanalysisaresustained.Inthemeantime,wehopetheseuserinsightshelpdesigntheMaaSsolutionsthatuserswant;thatenthusiastsclaimwillmakebetterofftheirusers;butalsothatbenefitsocietyasawhole.

6 ACKNOWLEDGEMENTS

This work is part of the project “Sustainable Mobility Services Södertälje”carried out by the Integrated Transport Research Lab (ITRL) at KTH RoyalInstituteofTechnology.WewouldliketothankScaniaAB,inparticularScaniaCity Solutions for their support. The projectwas “funded byVinnova throughtheStrategicInnovationProgramDriveSweden(grantnumber2017‐01976)”.

7 COMPLIANCE WITH ETHICAL STANDARDS

Conflictof interest: On behalf of all authors, the corresponding author statesthatthereisnoconflictofinterest.

8 REFERENCES

Bierlaire, M. 2016. PythonBiogeme: a short introduction. Report TRANSP‐OR160706, Series on Biogeme. Transport and Mobility Laboratory, School ofArchitecture, Civil and Environmental Engineering, Ecole PolytechniqueFédéraledeLausanne,Switzerland.

Daly, A., Hess, S., Patruni, B., Potoglou, D., Rohr, C., 2012. Using orderedattitudinalindicatorsinalatentvariablechoicemodel:Astudyoftheimpactof


15

securityonrailtravelbehaviour.Transportation,39(2)(2012),pp.267‐297.https://link.springer.com/article/10.1007%2Fs11116‐011‐9351‐z

Greene, D.L., Hensher, D, 2010.: Modelling Ordered Choices: A Primer.CambridgeUniversityPress,Cambridge

Haahtela, T., and Viitamo, E., 2017. Searching for the potential of maas incommuting‐comparison of survey and focus group methods and results.ConferenceProceedings1stInternationalConferenceofMobilityasaService.ICoMaaS.http://www.tut.fi/verne/aineisto/ICoMaaS_Proceedings_FULL_webres.pdf

Kamargianni,M.,Matyas,M., Li,W., andMuscat, J., 2018. Londoners’ attitudestowards car‐ownership and Mobility‐as‐a‐Service: Impact assessment andopportunitiesthatlieahead.MaaSLab‐UCLEnergyInstituteReport,PreparedforTransportforLondon.https://www.maaslab.org/reports

Krueger, R., Akshay V., Rashidi H. T., 2016. Normative beliefs and modalitystyles: a latent class and latent variable model of travel behaviour.Transportation.10.1007/s11116‐016‐9751‐1.

Likert,R. (1932).A technique for themeasurementofattitudes.Arch.Psychol22(140),55.

MaaSGlobal,2017.SERAConference.https://www.era.europa.eu/Document‐Register/Documents/Impulse%20speech%20Jonna%20P%C3%B6ll%C3%A4nen.pdf

QGIS, 2018. “QGIS Development Team. QGIS Geographic Information System.OpenSourceGeospatialFoundationProject.”. http://qgis.osgeo.org

Sanko, N., Hess, S., Dumont, J., Daly, A., 2014. Contrasting imputation with alatent variable approach to dealingwithmissing income in choicemodels. J.ChoiceModel.12,47–57.

Sochor, J., Karlsson, I. M., & Strömberg, H., 2016. Trying Out Mobility as aService: Experiences from a Field Trial and Implications for UnderstandingDemand. Transportation Research Record: Journal of the TransportationResearchBoard,(2542),57–64

Steg, L., 2005. Car use: lust and must. Instrumental, symbolic and affectivemotivesforcaruse.TransportationResearchA,39(2005),pp.147‐162

WongY.Z.,HensherD.A.,MulleyC.M., 2017,EmergingTransportTechnologiesandtheModalEfficiencyFramework:acaseformobilityasaservice(MaaS).Presentedatthe15thInternationalConferenceonCompetitionandOwnershipofLandPassengerTransport(Thredbo15),Sweden,Stockholm(2017),pp.13‐17.


16

9 APPENDIX A

The content in this appendix presents the framework of the model that wasintroducedbyKruegeretal. (2016)–Normativebeliefsandmodalitystyles:alatentclassandlatentvariablemodeloftravelbehaviour.‐.Forconsistency,wehavekeptequation1‐15,andtheirexplanationidenticalasdescribedinKruegeret al. (2016), the only difference in the methodology is the addition of acomparison of two different modelling techniques for ordered variables. Thereasonwhywe copied this section from Krueger et al. (2016) is to help ourreaderstounderstandtheframeworkandthesourceoftheestimation,eveniftheydonothaveaccesstoKruegeretal.(2016).Theframeworkconsistsofthreedifferentparts:first,alatentnormativebeliefsubmodel with structural and measurement components; second, a latentmodalitystylessubmodel;andthird,class‐specificsubmodelsforbinomialandorderedvariables.

Latent normative belief submodel

As explained inKrueger et al. (2016), the latent variablemodel estimates thevalues of the latent normative beliefs which enter the class membershipfunction as predictors. The structural component of the latent variablemodelrelatesthevalue , oflatentvariablem ∈ 1, … ,M tothevectorofobservedcharacteristicsX ofindividualn ∈ 1,… , N :

, ; , ; , , , (1)

, is a random disturbance ~ 0, i.i.d Normal across individuals withvariance , where is a parameter to be estimated. ; , is thedeterministic part specified to be linear in parameters including a constant

,andavectorofparameters .Consequently,thestructuralmodelof , is,

, ; , ∙ , ,(2)

where is the transpose of . This formulation leads to the probabilitydistributionfunctionofobservingavector of ∈ 1, … , latentvariablesforindividual :

| ; , , ∏ , ∙, (3)

where . denotes the probability density function of the standard normaldistribution.Since the modeller cannot directly observe the latent variables, a set of ∈ 1,… , Likert‐typeindicatorsprovidesapsychometricmeasurementofthelatentvariableinquestion.Forconvenience,anindexvalue ∈ 1,… , isassigned to the ordered response options , ,whereby a greater index valueindicatesgreateragreementwiththeprovidedstatement.SeeFigureA1below.


17

Figure A1. Example of Likert‐type indicator (i), for latent variable m, with 5 response categories (

Multiple indicators are related to the same latent variable by establishing alinear factormodel. Consequently, the item response function for indicator ,pertainingtolatentvariable ,forindividualnis, , , , , , ∙ , , , , (4)

where , denotes the factor loading, and , , is a stochastic component.Because the dependent variable of the item response function is ordinal, wemodel the responses to the psychometric indicators with an ordered model(Dalyetal.2012).AtthispointwecomplementtheapproachtakenbyKruegeretal.(2016),andimplement two different modelling assumptions for these ordered variables.First, we reproduce the modelling assumptions of Krueger et al. (2016) and

assume that , , ~ 0, i.i.d. By setting up the model in this way, the

probabilityofobservingresponse,μ , , , ismodelledbya logisticdistribution,which has closed form solution. The scale parameter of the logistic cdf isdenoted by Λ, and for convenience is set to one. Consequently, the responseprobabilitycanbeexpressedasfollows:

, , , ; , ; ,

, , , ∙ , 1,

, , , ∙ , , , , ∙ , 1 1,

1 , , , ∙ ,

(5)

where , is a set of 1 thresholdparameters for indicator of latentvariable . For identification, we set , , 0 and , 1 (see Daly et al.2012).Second, we assume thatν , , ~N 0, σ . Under this second approach, theprobability that the given response, μ , , , lies within a particular range of adistribution is no longer modelled by a logistic distribution, but rather by anormal distribution, which has no closed form. We decided to model theresiduals, , , ,withanormaldistribution,despiteoftheincreaseddifficultyin


18

estimation,asitisdeemedtheoreticallymoreappropriate.Foridentification,inthis case we set σ 1, , , 0 and , 1. This approach was used byGreeneandHensher(2010).Now, independently of the assumed distribution for ν , , , the probabilitydistribution function of observing the vector , of responses to the indicatorspertainingtolatentvariable ,isdefinedas:

, , , , ∏ , , | , , , , , , (6)AsalsodescribedinKruegeretal,(2016),notall indicatorswereobtainedforall of the latent variables. Hence, it ismore precise to introduce the notationM n to label the maximum number of latent variables as a function of theobservations. In addition, we introduce the dummy variable d , to indicatewhetherindicatorsofthe latentvariablewereobtainedforanobservation.Moreover,wemodifytheprobabilitydistribution functiontoreflect thatsomeindicatorsaremissingforsomeobservations:

, , , , , , , ∙ , , , , 1 , , (7)where , 0 if the indicators were not obtained. Then the value of theprobability distribution function is set to one and the contribution of thisobservationtothelog‐likelihoodiszero(Sankoetal.2014).To summarise, the probability distribution function of observing the matrix ofindicatorresponsesfor latentvariablesis:

| , , ∏ , | , , , ., (8)

Latent modality styles submodels

AsexplainedinKruegeretal.(2016),individualsareassumedtobedistributedacrossK latent classes representingmodality styles. Latent classmembershipprobabilities are predicted from the individual´s observed characteristics andavector oflatentvariables.This,theclassmembershipfunctionforclass∈ 1,… , ,isexpressedasfollows:

, , , ; , , , (9)

where , isthedeterministicpartoftheclassmembershipfunction,whichwespecifytobelinearinparameters.Hence,wehave:

, ∙ ∙ , , (10)where and arevectorsofparameterstobeestimated,and , isarandom

disturbance assumed to be , ~ 0, i.i.d. For observations, for which

indicatorspertainingtoalatentnormativebeliefwerenotobtained,thevalueofthe latent normative belief in question is imputed through the structural


19

equationofthelatentvariablesubmodelsothatthecoefficientonthelatentcanbeestimatedonthebasisofallobservations(Sankoetal.,2014).The class membership model is a multinomial logit model with the scaleparameterfixedtoone,asshownbyequation(11).

Κ| , , , ; ,

∑ , , ; ,, (11)

Class‐specific submodels

The model includes several class‐specific submodels, conditional onmembershipinalatentclass .AsexplainedinKruegeretal.(2016),dependingon the format of the dependent variable thesemodels are either binomial orordered logit models. Table 1 (in Section 2) provides an overview of thedifferent class‐specific submodels. In our case study the class‐specificmodelsareconstants‐onlymodels.First,wespecifytheprobabilityoftheclass‐specificbinomiallogitmodels:

|Κ ; ∙

, (12)

wherez isadummyvariableindicating,whichobservationwasmade.Κ isaconstant, which is to be estimated. Since multiple binomial logit models areestimatedforeachclass,weusethesamesimplifyingnotationasKruegeretal.(2016), where ∈ 1,… , is an index denoting the binomial choice beingmodelled,

, , |Κ ; ∏ , , , |Κ , ; , (13)Similarly,therearemultipleorderedmodelsforeachclass.Inthesemodels,theprobability of observing response, , , is modelled in an identical way aspreviously seen for the observed latent variable indicators, , , . Then, theprobabilitydistributionfunctionoftheclass‐specificorderedmodelsisdefinedas:

, , |τ ; ∏ , , , | , ; , (14)where ∈ 1, … , isanindexthatdenotestheorderedchoicebeingmodelled,and , denotesthecodedordinaloutcomevariableforindividualn.Also as explained by Krueger et al. (2016), , is a vector of thresholdparameterscommontoalllatentclasses.Thissimplificationisjustifiedbecauseifthethresholdparametersareestimatedbasedontheentiresample,thenthedifference between the ordered categorical dependent variables is the sameacross all classes. The additive scale of the outcome variables is adjusted foreachclassbysubtractingtheclass‐specificconstantfromthesamplethreshold.


20

The class‐specific submodels for mode‐specific attitudes are specified asorderedlogitmodels.TheresponsesreportedinTableB1(appendixB)areusedas indicators and the corresponding threshold parameters are estimated asbeingspecifictotheentiresample.Moreover,itisassumedthattheinfluenceofindividual characteristics is sufficiently reflected in the latent segmentation.Hence, the structural model for each class only comprises a constant that isspecifictotheclass.

Likelihood function

As explained in Krueger et al. (2016), equations 3, 8, 11, 13, and 14 areiterativelycombinedacrossindividualstoobtaintheunconditionalprobabilityofobservingthecombinationofobservedchoices.Where , , arevectorsofobserved choices; being binomial choices; the coded ordinal outcomevariables;and observedindicatorsofthelatentvariables.

, , ; , , , , , , , Κ

| , ∙ , , | ; ∙ , , | ;

∙ | , , ∙ , ; , .

(15)

Because of the random error term that enters the structural equation of thelatentvariablemodel,thelikelihoodisintegratedoverthedensitiesofthelatentvariables , hence the dimension of the integral becomes the same as thenumberof latentvariables ( ).Asa consequence, theobjective functiondoesnotpossessaclosed‐formsolutionunderanyofthetwomodellingassumptionsusedtomodeltheorderedvariables;hence,modelsmustbeestimatedthroughmaximumsimulatedlikelihoodmethods.


21

10 APPENDIX B

This appendix presents the survey responses to the 50 attitudinal questionspresentedtotheusers.Histogramshavelevelsfrom0to5 lefttoright ,wherezeroshowsthenumberofnotrespondents,andlevels1to5showthelevelofagreementwiththestatementfrom1‐Stronglydisagree;2‐Disagree;3‐Neutral;4‐Agree;and5‐Stronglyagree.Statement(referencecode) HistogramofresponsesIamregularlystressedinmyeverydaylife.(AQ1)

Cyclingpreventsmefromhavingaprofessionallook.(AQ2)

Ialwayschoosethefastesttravelalternative.(AQ3)

Iuseinformationontheinternettochecktimetablesanddelays.

(AQ4)

IdonotfeelcomfortablewhensharingacartripwithpeoplethatIdonotknow.

(AQ5)

IneedflexibilitybecausenormallyIhaveanirregularschedule.

(AQ6)

ItravelthewayIdobecauseIdonothaveanyotheralternatives.

(AQ7)

Iplanmytripsthedaybefore.(AQ8)

Iplanmytripsearlyonthesameday.(AQ9)

Inormallytravelinthesamewayanddonotplanmytrips.

(AQ10)

Iplanmycommutetripbasedonother(non‐work)tripaswell.

(AQ11)

IoftenthinkaboutmovingclosertoScaniainordertoreducemytraveltime.

(AQ12)

IconsidertheenvironmentwhenIplanmytrips.(AQ13)

Socialandenvironmental


22

Statement(referencecode) HistogramofresponsesIcanmakeanimpactontheenvironmentbytravellingsustainably.

(AQ14)

Imakeactivechoicesonadailybasistoreducemycarbonfootprint.

(AQ15)

Fuelpricesshouldbeincreasedtoreducecongestionandairpollution.

(AQ16)

ItisimportantforScaniathatwetravelsustainable.(AQ17)

Peoplewhoareimportanttomehaveopinionsaboutmywayofcommuting.

(AQ18)

PeoplewhoareimportanttomesayIshoulduse(orbuy)acar.

(AQ19)

PeoplewhoareimportanttomesayIshouldusethepublictransport.

(AQ20)

PeoplewhoareimportanttomesayIshouldusethebicycle.

(AQ21)

PeoplewhoareimportanttomesayIshouldwalkmore.

(AQ22)

Transportmodes ItdoesnotmattertomewhichtypeofcarIdrive.(AQ23)

Acarprovidesstatusandprestige.(AQ24)

Youcantellwhoapersonisbylookingathisorhercar.

(AQ25)

Iliketodrive.(AQ26)

IknowofadreamcarthatIwouldlovetopossess.(AQ27)

IfeelfreeandindependentifIhaveaccesstoacar.(AQ28)

IliketousepublictransitbecauseIcanmakebetteruseofmytime.

(AQ29)

WhenIusepublictransport,otherpeoplecomeclosetomeinanunpleasantway.

(AQ30)


23

Statement(referencecode) HistogramofresponsesItisdifficulttousepublictransportwhenItravelwithmychildren.

(AQ31)

Itisdangeroustousethebicycle.(AQ32)

Icanreachmanyofmynormaldestinationsbybike.(AQ33)

Ienjoywalking.(AQ34)

Towalkandcyclehelpstokeepmehealthy.(AQ35)

MaaSExpectations IwasfamiliarwithSCANIAGObeforethissurvey(AQ36–binaryquestionyes/no)

IhavehighexpectationsforSCANIAGO.

(AQ37)

IthinkthenewSCANIAGOsystemwillaffecthowIcommutetowork.

(AQ38)

IthinkthenewSCANIAGOsystemwillaffecthowItravelinsideScania.

(AQ39)

Ifthenewsystemallowedtobookcabs,IwillusethatfeatureinsteadofthePT

(AQ40)

SCANIAGOisgoingtobeeasytouse.(AQ41)

SCANIAGOisgoingtobeareliableservice.

(AQ42)

SCANIAGOappisgoingtobeuserfriendly.

(AQ43)

Questionsonlyforcarowners IfIdidnotneedacar,Iwoulddisposeofitimmediately.

(AQ44)

IthinkthatthenewSCANIAGOsystemwillhelpmereducemycarusageintheshortterm.

(AQ45)

NewsystemslikeSCANIAGOwouldhelpmedependlessonmycar.

(AQ46)


24

Statement(referencecode) HistogramofresponsesNewsystemslikeSCANIAGOwouldhelpmesubstitutecartripswithothermodesinsidescania’sarea.

(AQ47)

IttakesalotoftimetofindaparkingspacewhenIusemycar.

(AQ48)

Owningacarisabigexpenditureformyhousehold.(AQ49)

Iwouldlovetohaveaccesstoacarwithoutthehassleofowningone.

(AQ50)


25

11 APPENDIX C

Table C1. Estimates of the structural coefficients of the latent normative belief models

Variable Caraffective Environmentalmindset

HighexpectationsforMaaSeverywhere

HighexpectationsforMaaS on‐campus

Socialinfluence

Est t‐val Est t‐val Est t‐val Est t‐val Est t‐valConstant 2.23 9.90 2.73 8.27 1.57 6.76 1.47 9.49 0.21 0.23*Gender reference female Male 0.271 1.96 ‐0.33 ‐1.23* ‐0.46 ‐2.86 ‐0.40 ‐3.59 0.23 1.50*Age reference 45‐59 18‐24 0.78 2.39 0.21 1.68* 0.36 1.01* ‐0.11 ‐0.76* 1.06 2.9525‐44 ‐0.29 ‐2.32 0.11 0.54* 0.19 1.32* 0.32 3.42 0.22 1.57*60ormore ‐0.27 ‐0.87* ‐0.82 ‐1.56* ‐0.31 ‐0.94* ‐0.67 ‐3.64 ‐0.24 ‐0.74*Childreninthehh reference no Yes ‐0.05 ‐0.43* 0.54 2.37 ‐0.12 ‐0.85* 0.15 1.80* 0.12 0.93Manager reference no Yes 0.13 0.71* 0.43 1.25* ‐0.14 ‐0.54* 0.09 0.68* 0.53 2.47Standarddeviation 1.28 10.0 2.91 10.1 1.52 8.80 1.18 12.47 1.29 6.69



26

Table C2. Estimates of measurement coefficients of the latent normative belief models

Variable Loading Threshold∗ Disagree Neutral Agree StronglyAgree Est t‐val Est t‐val Est t‐val Est t‐val Est t‐valCaraffectiveAQ26 1# ‐ 0# ‐ 0.81 6.42 1.73 9.48 2.61 10.6AQ27 0.43 9.46 0# ‐ 0.56 8.78 0.90 7.35 1.39 9.03AQ28 0.99 8.27 0# ‐ 0.47 4.38 1.06 6.42 2.15 11.1Environmental.mindsetAQ13 1# ‐ 0# ‐ 2.51 11.3 4.02 9.83 5.67 9.21AQ14 0.40 8.55 0# ‐ 0.26 5.42 0.79 8.63 1.47 11.1AQ15 0.41 11.0 0# ‐ 0.76 9.91 1.77 12.3 3.34 10.2AQ17 0.24 7.76 0# ‐ 0.20 5.24 0.79 9.81 2.06 12.5HighexpectationsforMaaS(everywhere)AQ37 1# ‐ 0# ‐ 0.59 6.88 1.56 9.49 2.46 9.01AQ38 0.32 5.62 0# ‐ 0.95 10.4 1.40 6.26 1.99 5.82AQ45 0.47 5.62 0# ‐ 1.07 9.19 1.76 6.78 2.46 5.17AQ46 0.90 6.53 0# ‐ 1.12 8.86 2.10 8.53 3.15 6.88AQ47 0.95 6.51 0# ‐ 0.37 5.21 1.01 7.54 2.06 10.1AQ48 0.57 7.18 0# ‐ 0.51 8.55 1.00 8.54 1.55 8.54HighexpectationsforMaaS(on‐campus)AQ37 1# ‐ 0# ‐ 0.68 7.87 1.65 11.2 2.50 10.0AQ39 1.02 10.8 0# ‐ 0.61 7.41 1.39 9.87 2.54 12.0AQ41 2.89 9.36 0# ‐ 1.04 4.71 3.63 9.98 6.55 10.4AQ42 2.91 8.79 0# ‐ 1.50 5.99 3.72 10.1 6.48 10.1AQ43 3.68 6.88 0# ‐ 1.47 4.25 4.59 8.16 7.75 7.67SocialinfluenceAQ24 1# ‐ 0# ‐ 0.85 8.55 1.74 8.69 2.80 7.75AQ25 0.81 4.97 0# ‐ 0.58 8.57 1.48 10.2 2.97 9.13#Constrainedforidentification*Categoricalvaluetotherightofthethresholdvalue;valuetotheleftofthethresholdvalueisomittedintheinterestofbrevity


27

Table C3. Estimates of coefficients on attitudinal indicators

AttitudinalQuestion AQ3 AQ4 AQ5 AQ6 AQ7Threshold* Est t‐val Est t‐val Est t‐val Est t‐val Est t‐val1 0# ‐ 0# ‐ 0# ‐ 0# ‐ 0# ‐2 1.12 7.02 0.68 6.67 1.02 10.66 0.98 8.36 0.91 8.693 2.34 10.90 1.31 7.56 1.95 10.29 1.56 7.67 1.72 9.814 3.48 11.88 2.15 9.53 2.81 8.05 2.59 11.30 2.73 10.88Class‐specificmodifiers Class1 2.88 13.91 1.57 10.81 1.04 8.45 2.10 13.10 1.73 12.13Class2 2.07 8.72 3.24 12.16 1.15 5.77 1.53 7.48 1.52 7.44

Table C3 (Cont). Estimates of coefficients on attitudinal indicators

AttitudinalQuestion AQ10 AQ11 AQ12 AQ24 AQ26Threshold* Est t‐val Est t‐val Est t‐val Est t‐val Est t‐val1 0# ‐ 0# ‐ 0# ‐ 0# ‐ 0# ‐2 0.82 6.91 0.68 5.70 0.58 6.92 0.92 10.62 1.29 6.253 1.58 8.49 1.49 8.08 1.11 5.53 1.95 10.45 2.50 9.884 2.51 10.68 2.34 9.63 1.67 4.64 3.41 8.18 3.56 11.41Class‐specificmodifiers Class1 2.23 13.20 2.63 14.13 ‐1.03 ‐8.16 0.68 5.95 3.37 13.37Class2 1.84 8.48 1.53 7.17 ‐0.40 ‐2.04 0.69 3.47 2.52 8.97Table C3 (Cont). Estimates of coefficients on attitudinal indicators

AttitudinalQuestion AQ40 AQ49 AQ50Threshold* Est t‐val Est t‐val Est t‐val1 0# ‐ 0# ‐ 0# ‐2 0.90 7.76 1.04 6.32 0.79 6.813 1.85 9.63 2.06 9.08 1.54 8.214 2.73 8.70 3.16 10.95 2.24 8.49Class‐specificmodifiers Class1 1.67 10.52 2.57 12.40 1.85 11.68Class2 1.49 6.46 2.50 7.51 1.94 6.21

#Constrainedforidentification*Categoricalvaluetotheleftofthethresholdvalue