ContentslistsavailableatScienceDirect
Accident
Analysis
and
Prevention
jo u r n al hom e p a g e :w w w . e l s e v i e r . c o m / l o c a t e / a a p
A
dynamic
analysis
of
motorcycle
ownership
and
usage:
A
panel
data
modeling
approach
Chieh-Hua
Wen
a,∗,
Yu-Chiun
Chiou
b,
Wan-Ling
Huang
baDepartmentofTransportationTechnologyandManagement,FengChiaUniversity,100WenhuaRoad,Taichung40724,Taiwan,ROC
bInstituteofTrafficandTransportation,NationalChiaoTungUniversity,4F,118,Sec.1,Chung-HsiaoW.Road,Taipei100,Taiwan,ROC
a
r
t
i
c
l
e
i
n
f
o
Articlehistory:
Received31July2010
Receivedinrevisedform11January2011
Accepted7March2011 Keywords: Motorcycle Paneldata Discretechoice Regression Accident
a
b
s
t
r
a
c
t
Thisstudyaimstodevelopmotorcycleownershipandusagemodelswithconsiderationofthestate dependenceandheterogeneityeffectsbasedonalarge-scalequestionnairepanelsurveyonvehicle owners.Toaccountfortheindependenceamongalternativesandheterogeneityamongindividuals,the modelingstructureofmotorcycleownershipadoptsdisaggregatechoicemodelsconsideringthe multi-nomial,nested,andmixedlogitformulations.Threetypesofpaneldataregressionmodels– ordinary, fixed,andrandomeffects–aredevelopedandcomparedformotorcycleusage.Theestimationresults showthatmotorcycleownershipinthepreviousyeardoesexerciseasignificantlypositiveeffectonthe numberofmotorcyclesownedbyhouseholdsinthecurrentyear,suggestingthatthestatedependence effectdoesexistinmotorcycleownershipdecisions.Inaddition,thefixedeffectsmodelisthepreferred specificationformodelingmotorcycleusage,indicatingstrongevidenceforexistenceofheterogeneity. Amongvariousmanagementstrategiesevaluatedunderdifferentscenarios,increasinggaspricesand parkingfeeswillleadtolargerreductionsintotalkilometerstraveled.
© 2011 Elsevier Ltd. All rights reserved.
1. Introduction
In many Asian countriessuchasChina,Indonesia, Malaysia, Taiwan,Thailand,andVietnam,motorcyclesareaprimarymode ofurbantransportation.Duetolowpurchaseandrunningcosts andconvenientparking,demandformotorcycleshascontinuously riseninthesecountries.Trafficcongestion,accidents,parking dis-order,andairpollution aretheinevitableconsequences ofhigh ratesofmotorcycleownershipandusage.Especially,inthe con-textofaccidentanalysis,becauseoflimitedprotectiondesignof motorcyclesincomparisontocars,motorcyclesarethemost dan-gerousformofmotorizedtransport,withinjuryrateseighttimes, andfatalityrates35timesthatofcaroccupants(pervehiclemile traveled)(NHTSA,2007;Ranneyetal.,2010).Inmostdeveloped countries,motorcyclefatalitiestypicallycomprisearound5–18%of overalltrafficfatalities(Mohan,2002;Koornstraetal.,2003;WHO, 2006).Thisproportionreflectstherelativelylowownershipand usageofmotorcyclesinmanydevelopedcountries.However,the ownershipandusageofmotorcyclesandothertwo-wheelersis generallyrelativelyhighinmanydevelopingcountries.Reflecting thisdifference,thelevelsofmotorcyclistfatalitiesasaproportion ofthoseinjuredontheroadsaretypicallyhigher indeveloping countriesthanindevelopedcountries.Forexample,inIndia,69%
∗ Correspondingauthor.Tel.:+886424517250x4679;fax:+886424520678.
E-mailaddress:chwen@fcu.edu.tw(C.-H.Wen).
ofthetotalnumberofmotorvehiclesaremotorizedtwo-wheelers (Mohan,2002)and27%ofroaddeathsinIndiaareamongusers ofmotorizedtwo-wheelers,whilethisfigureisbetween70%and 90%inThailand,andabout60%inMalaysia(Mohan,2002;Umar, 2002;SuriyawongpaisalandKanchanusut,2003).InChina, motor-cyclesaccountedfor23.4%ofallregisteredmotorvehiclesin1987, increasing to63.2% in 2001.Motorcyclist fatalitiesand injuries increased5.5-foldand9.3-foldbetween1987and2001, respec-tively.Atotalof7.5%ofalltrafficfatalitiesand8.8%ofalltraffic injuries were sustained by motorcyclists, withthe correspond-ingproportionsincreasingto18.9%and22.8%in1987and2001, respectively.Thechangingproportionsofbothtrafficfatalitiesand injuriessustainedbymotorcyclistswerepositivelycorrelatedwith thechangeintheproportionofmotorcyclesamongallmotor vehi-cles(Zhangetal.,2004).InIndonesia,thepopulationofmotorcycles hasreached78.3%oftotalmotorvehiclesand75%fatalityvictimsof trafficaccidentsweremotorcyclists(IndriastutiandSulistio,2010). InTaiwan,themotorcycleownershipratereaches645per thou-sandpeople(thehighestmotorcycleownershiprateintheworld). In2009,motorcyclistfatalitiesinTaiwanaccountedfor56.69%of total trafficdeaths(MOTC,2010).Tocomparethefatalitiesand injuries of car and motorcycle crashes further, Tables 1 and 2, respectivelygivethestatisticsofcarandmotorcycleownership, usageandcrashvictimsof23counties/citiesinTaiwan.Notedfrom Table1,thereareatotalof5.7millionregisteredpassengercars with502fatalitiesand71,564injuriesin2009,whilethenumber ofregisteredmotorcyclesreaches14.6millionwith889fatalities
0001-4575/$–seefrontmatter © 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.aap.2011.03.006
194 C.-H.Wenetal./AccidentAnalysisandPrevention49 (2012) 193–202
Table1
Carownership,usageandcrashof23counties/citiesinTaiwan.
County/city Registered cars Averagekilometers traveled(km) Totalkilometers traveled(million veh-km)
Victims Fatalities(millionveh-km)
Victims Fatalities(million veh-km)
Victims Injuries(million veh-km) Taipeicity 637,354 9842 6273 28 0.005 5908 0.942 Kaohsiungcity 367,265 9734 3575 17 0.005 8023 2.244 Taipeicounty 776,945 8694 6755 20 0.003 6093 0.902 Yilancounty 111,717 8987 1004 19 0.019 1988 1.980 Taoyuancounty 535,501 9636 5160 49 0.010 7437 1.441 Hsinchucounty 152,342 9157 1395 16 0.011 2007 1.438 Miaolicounty 159,682 7152 1142 26 0.023 1605 1.406 Taichungcounty 438,394 8908 3905 36 0.009 6668 1.708 Changhwacounty 343,522 8978 3084 26 0.009 4509 1.462 Nantoucounty 147,459 10,035 1480 19 0.013 1644 1.111 Yunlincounty 179,534 8901 1598 34 0.021 2006 1.255 Chiayicounty 134,520 8437 1135 29 0.026 1242 1.094 Tainancounty 287,513 8198 2357 31 0.013 3561 1.511 Kaohsiungcounty 298,871 7783 2326 41 0.018 4140 1.780 Pingtungcounty 194,016 7582 1471 25 0.017 2230 1.516 Taitungcounty 49,223 7354 362 13 0.035 710 1.963 Hualiencounty 84,012 9701 815 22 0.027 1379 1.692 Ponhucounty 18,047 7702 139 3 0.023 141 1.016 Keelungcity 77,898 8010 624 6 0.010 847 1.358 Hsinchucity 114,352 10,057 1150 11 0.009 1886 1.640 Taichungcity 316,596 8936 2829 12 0.004 3969 1.403 Chiayicity 69,624 7900 550 8 0.015 1769 3.217 Tainancity 190,162 9744 1853 11 0.006 1799 0.971 Average – 8758 – – 0.014 – 1.524 Total 5,684,549 – 50,982 502 – 71,564 –
and129,200injuriesinTable2.Evenintermsof accidentrates
(i.e.,numberofvictimspermillionkilometerstraveled), motorcy-cleusage(intermsoftotalkilometerstraveled)stillexhibitshigher ratesoffatalandinjuredvictimsthancarusage.Onanaverage, thereare0.017fatalitiesand2.038injuries permillion kilome-terstraveledbymotorcyclesincomparisonto0.014fatalitiesand 1.524injuriespermillionkilometerstraveledbycars.Tofurther
examinetherelationshipbetweenmotorcycleusageandnumbers ofvictimsoftwoseveritylevels,thePearsoncorrelationtestis per-formed.Resultsshowthatcorrelationcoefficientsofmotorcycle usagewithfatalitiesandinjuriesare0.586(p-value=0.003)and 0.886(p-value<0.0001),respectively,indicatingasignificantand positivecorrelationbetweenmotorcycleusageandcrashvictims. Comparedwiththerelationshipbetweenmotorcycle usageand
Table2
Motorcycleownership,usage,andcrashesof23counties/citiesinTaiwan.
County/city Registered motorcycles Averagekilometers traveled(km) Totalkilometers traveled(million veh-km) Fatalities Injuries
Victims Fatalities(million veh-km)
Victims Injuries(million veh-km) Taipeicity 1,092,788 4438 4850 24 0.005 9513 1.962 Kaohsiungcity 1,207,026 4778 5768 41 0.007 14,950 2.592 Taipeicounty 2,259,828 4912 11,099 66 0.006 13,600 1.225 Yilancounty 292,879 4730 1385 34 0.024 3064 2.212 Taoyuancounty 1,081,978 4665 5047 47 0.009 9919 1.965 Hsinchucounty 271,233 4769 1293 35 0.027 2844 2.200 Miaolicounty 350,202 3800 1331 23 0.018 2117 1.591 Taichungcounty 1,008,400 4199 4234 59 0.014 12,121 2.863 Changhwacounty 902,353 4220 3807 62 0.016 8245 2.166 Nantoucounty 349,862 3883 1359 34 0.025 2888 2.125 Yunlincounty 486,157 4308 2094 58 0.028 3384 1.616 Chiayicounty 365,747 4597 1681 43 0.025 2123 1.263 Tainancounty 805,813 4580 3691 57 0.015 7594 2.058 Kaohsiungcounty 1,014,396 5119 5192 79 0.015 9968 1.920 Pingtungcounty 697,431 4697 3276 86 0.026 5462 1.667 Taitungcounty 177,999 4006 713 30 0.042 1579 2.215 Hualiencounty 241,958 4472 1082 28 0.025 2435 2.251 Ponhucounty 68,426 4125 282 4 0.014 585 2.074 Keelungcity 190,771 4569 872 11 0.013 1653 1.895 Hsinchucity 262,338 4307 1130 16 0.014 3507 3.103 Taichungcity 646,739 3813 2466 23 0.010 6127 2.484 Chiayicity 202,586 4961 1005 11 0.011 2312 2.300 Tainancity 583,436 4885 2850 17 0.006 3210 1.126 Average – 4471 – – 0.017 – 2.038 Total 14,560,346 – 66,509 889 – 129,200 –
crashvictims,thecorrelationcoefficientsofcarusagewith
fatal-itiesandinjuriesarealsosignificantlytestedwithslightlylower
coefficientsof0.475(p-value=0.022)and0.876(p-value<0.0001),
respectively.Thus,effectivemanagement strategiesfor
decreas-ingtheownershipandusageofmotorcyclesandcarsareurgently
required.
Inthepast,substantialresearchhasbeendevotedtothe
devel-opmentofcarownershipandusagemodels,whichhascontributed
toenhanceourunderstandingofchoicebehaviors(DeJongetal.,
2004).Thesestudieshavebeenmostlyconductedinthedeveloped countriesbecausepeoplerelyheavilyonprivatecarsforurban travelintheseareas,andmotorcyclesplayasecondaryroleasa modeoftransportation.Asaresult,fewstudieshavebeendonein thecontextofmotorcycleownershipanduse.Inaddition, previ-ousworksonownershipandusageofmotorvehiclesoftenused eithercross-sectionalortime-seriesdata.Datasetsthatpoolcross sectionsandtimeseries(calledpanelorlongitudinaldata)have becomeincreasinglycommonintransportationandotherfields. Panel dataanalyseshave many advantages, suchascontrolling forindividualheterogeneity,lesscollinearityamongthevariables, moredegreesoffreedom,andmoreefficiency,overanalysesonly basedoneithercross-sectionalortime-seriesdataalone(Baltagi, 2005).Moreover,thepaneldatamodelsallowforanalyzingthe repeatedchoicesovertimeandcancapturethestatedependence effectbyincorporatingpastchoices.
Thecurrent literaturelackstheanalysisof motorcycle own-ershipand usageusingpaneldata.Toexamine dynamicchoice behaviorsassociatedwithmotorcycleownershipandusage,this studyconductedalarge-scale,nationwidetwo-wavepanelsurvey onownersofmotorcyclesinTaiwan.Thus,disaggregatemodelsof motorcycleownershipandusageusingpaneldataareproposed. Thediscretechoicemodelhasbeenwidelyusedasan appropri-atemethodologyforexaminingvehicleownership.Comparedwith theorderedchoicemodel,theunordered discretechoicemodel derivedfromrandomutilitytheoryprovidesatheoreticalbasefor modelingthenumberofvehiclesownedinthehousehold(Bhat and Pulugurta,1998).In addition totheuseof a standard dis-cretechoicemodel,i.e.,themultinomiallogitmodel(MNL),this studyalsoadoptsnestedlogit(NL)andmixedlogit(MXL) mod-elstoaccommodatethepossibleindependenceamongalternatives and parameter heterogeneity among individuals. For modeling theusageofmotorcycles,paneldataregressionmodelsinvolving fixedandrandomeffectsapproachesaccountingfor heterogene-ity aredeveloped and compared. These modelscan beusedto identifyfactorsinfluencingmotorcycleownershipandusage,such ashouseholdstructure,residentiallocation,transportationsystem performance,driver’stravelpatterns,andvehiclecharacteristics. Ourproposedpaneldatamodelsprovidereliableparameter esti-matestoevaluatetheeffectsofmanagementstrategiesinreducing motorcycleownershipandusage.Duetoasignificantrelationship betweenmotorcycleusageandcrashvictims,thesestrategieswill alsoleadtoreductionsintrafficaccidentsbymotorcycles.
Theremainderofthispaperisstructuredasfollows.Section2 providesabriefoverviewofpreviousliterature.Section3presents theframeworkof themotorcycle ownershipand usagemodels. Section 4 describes the data sources and estimation results of motorcycleownershipandusagemodelsaswellasscenariosand analysesofmanagementstrategies.Finally,thispaperconcludes withtheresearchfindings,andgivesdirectionsforfurtherresearch.
2. Literaturereview
Studiesonvehicleownershipoftenuseeitheraggregatemodel (e.g.,Jansson, 1989; Buttonet al., 1993)or disaggregatemodel (e.g.,Train,1980;Mannering,1983).Theaggregatemodelsmay
suffertheshortcomingsofaggregationbiasandmulticollinearity betweenexplanatoryvariables(PotoglouandKanaroglou,2008). Onthecontrary,thedisaggregatemodelingapproachovercomes theweaknessesofaggregatemodelsbycapturingindividualchoice behavior and explanatory variables at anindividual level, with theabilitytoobtainmorereliableestimates,andtherefore,ithas receivedgreaterattentioninrecentvehicleownershipresearch. The number of vehicles in a household, which is a categor-ical dependent variable, is typically analyzed by using either the ordered- or unordered-response choice models (Bhat and Pulugurta,1998).Conditionalonthehousehold’sdecisiontoown vehicles,individual’svehicleusagemeasuredasmiles(or kilome-ters)peryearandvehicletypecanbemodeledsimultaneouslyto examinetherelationshipbetweenthesetwovariables(Mannering andWinston,1985).
The recent development of vehicle ownership models has movedfromstatictodynamicmodelingbycollectingpaneldata thatcontainhouseholds’vehicleinformationforconsecutiveyears. Studies of household vehicleownership with paneldata often adopted ordered-response models accounting for state depen-denceandheterogeneity(KitamuraandBunch,1990;Hanlyand Dargay,2000;GiulianoandDargay,2006;DargayandHanly,2007). The panel data analysis of vehicle usage commonly employed random-andfixed-effectsspecificationsallowingforunobserved heterogeneity(e.g.,Dargay,2007;Woldeamanueletal.,2009).A substantialbodyofresearchdocumentsmodelingandempirical findingsofcarownershipandusageworldwide.However,therehas beenrelativelylittleresearchintomotorcycleownershipandusage. Forexample,Senbiletal.(2007)appliedorderedprobitmodelsand tobit/probitregressionmodelswithsampleselectiontoexplore motorcycleusageinthemetropolitanareaofJabotabek,Indonesia; thesemodelshaveidentifiedkeyexplanatoryvariables(e.g., resi-dentiallocation,landuse,transportationsystemperformance,and socioeconomic/demographiccharacteristics). Burgeet al.(2007) presentedatwo-levelNLmodelcontainingthehousehold’s deci-siontoownmotorcyclesattheupperlevelandthechoiceofengine sizesatthelowerlevel.LaiandLu(2007)appliedathree-levelNL modeltoanalyzehouseholdjointchoicesofthenumberofcars, thenumberofmotorcycles,andworkmodeoftransportation.More recently,Chiouetal.(2009)proposedanintegratedmodelthat ana-lyzeschoicebehaviorsassociatedwithownership,type,andusage ofcarsandmotorcyclesinTaiwan.Theexistingliteraturestilllacks dynamicanalysisofmotorcycleownershipandusage,whichallows forincorporatinghouseholdpreferencesovertimeandidentifying factorsthatinfluenceownershipandusagebehavior.
3. Methodologicalframework 3.1. Motorcycleownershipmodel
Motorcycleownershipmodelexaminesthenumberof motor-cycles(i.e.,0,1,...)ownedbyhouseholdsineachyear.Households mayalterormaintainthenumberofmotorcyclesnextyear.The existing methods for modeling this type of dependentvariable typically used discrete choice model, ordered logit or probit model,andcountdatamodel(KarlaftisandGolias,2002),butthe unordereddiscretechoicemodelispreferredbecauseitprovides atheoreticalframeworkonabasisofrandomutilitytheorythat iswidelyusedtoexplainhumanbehavior.Undertheframework of discrete choiceanalysis, a decision-maker (i.e.,a household) isassumedtochoosethealternative(i.e.,numberofmotorcycles ownedbyahousehold)withthehighestutilityundertheprinciple of utility maximization. The attractiveness (interms of utility) ofeachalternativecanberepresentedbythesumofsystematic
196 C.-H.Wenetal./AccidentAnalysisandPrevention49 (2012) 193–202
(observable)andrandom-error(unobservable)components.The totalutilityofanalternativeiforhouseholdnatyeartisspecifiedas Uitn=Vitn+εitn=˛i+ˇStn+Ltn+ıCitn+Ditn+εitn (1)
whereVitnistheobservedcomponentofutility;Stn(household’s
characteristics)andLtn(locationcharacteristicsandtransportation
systemperformance)arevectorsofalternativespecificvariables thatdonotvaryoveralternatives;Citnisavectorofgeneric
vari-ablesincludingfixedandvariablecostsoftheirownmotorcycles; thesearevariedbyalternativesandallowthesamemarginaleffect oneachalternative’sutility;˛i(alternativespecificconstant),ˇ,
,ı,andaretheunknownparameterstobeestimated;εitnisthe
randomerrorterm.
InEq.(1),laggeddummyvariables wereusedtoaccountfor statedependence.Foreachalternative,ifthenumberof motorcy-clesownedbythehouseholdninyeartisthesameasthenumberof motorcyclesownedbythehouseholdninthepreviousyear(t− 1), Ditn=1,and0otherwise.Ifthestatedependenceparameteris
sta-tisticallysignificant,itindicatesthatmotorcycleownershipinthe currentyearisaffectedbymotorcycleownershipintheprevious year.
Thediscretechoicemodelcanbederivedunderdistributional assumptions with respect to the error term. The MNL is the mostusedmodelduetoitssimplifiedprobabilityformulationand tractablecomputation.TheprobabilityformoftheMNLisgivenby: Pitn=
exp(Vitn)j
exp(Vjtn)
(2)
ThepanelMNLmodelassumesthattheerrortermhasan inde-pendent and identical Gumbel distribution across individuals, alternatives,andyears,andthusitexhibitstheindependenceof irrelevantalternatives(IIA)propertythatimposesrestrictionson therelativeprobabilitiesofeachpairofalternatives.TheNLmodel, whichhasbeenwidelyusedtorelaxtheIIAproperty,allowsfor cor-relationbetweentheutilitiesofpairsofalternativesinacommon nestandaccommodatessimilaritiesamongalternativesinthesame nest.Theoretically,anyalternativecanbegroupedintoanest. How-ever,thenumberofmotorcycleshasthecharacteristicsofordered nature.Thisstudyonlyconsidersbehavioralinterpretablenested structure.Forexample,theadjacentalternatives(e.g.,households owning1and2motorcycles)canbegroupedintoanest,and alter-nativesof3and4motorcycleshavesinglealternativenests.For theNLmodel,theprobabilitythatanalternativeiwillbechosen byhouseholdnatyeartcanbewrittenas:
Pitn=
exp(Vitn/m)×
j∈Nmexp(Vjtn/m)m−1m
j∈Nmexp(Vjtn/m)
m (3)where thelogsum parameterfor thenest m,m,indicatesthe
degreeofcorrelationbetweenalternatives’utilitiesinthenest.Nm
isasetofalternativesinthenestm.
Thevalueoflogsumparametermustliewithintherangeof0–1 forconsistencywithutilitymaximization.IftheNLmodelcannot outperformtheMNLmodel,theMNListheappropriate specifica-tion.ThelikelihoodratiotestcanbeusedtocomparetheMNLand variousNLmodelstoverifywhethertheIIAassumptiondoesor doesnothold(HausmanandMcFadden,1984).
Althoughmorerecentdevelopeddiscretechoicemodels,such asthegeneralizednestedlogitmodel(WenandKoppelman,2001), havemoreflexiblecorrelationstructuresthantheNLmodel,the NLmodelremainspopular.However,eitherthestandardMNLor NLmodeldoesnotallowforindividualpreferenceheterogeneity andmaynotadequatelydescribethebehaviorofallusers.Onthe contrary,theMXLmodelallowstheestimationofeachparameter byassumingadistributionoverthepopulationinsteadofasingle
fixedvalueimposedinthestandardMNLorNLmodel.TheMXL modelhasbeenappliedinanalyzinghouseholds’preferencesfor automobilestocapturethesystematicheterogeneityindecision makers’preferences(Brownstoneetal.,2000).Thepresentstudy appliestheMNL,NL,andMXLmodelsformotorcycleownership. 3.2. Motorcycleusagemodel
Thecontinuousdependentvariableusedforrepresentationof motorcycleusageisannualmileage traveledacrossthree years. Mostpaneldataapplicationsusefixedandrandomeffectsmodels tocontrolindividualeffects(heterogeneity).Duetotheunobserved individualeffectsthatmaybecorrelatedwiththeincluded vari-ables,thefixedeffectsmodelallowseachobservationtohaveits ownintercept,bycreatingasetofdummy(either0or1)variables includedasregressors.Thismodelisalsocalledtheleastsquares dummyvariableandcanbeestimatedbystandardleastsquares. Theformulationoffixedeffectsmodelinthisstudyis
lnytn= N
n=1
nDtn+Rtn+Htn+Ctn+ωWtn+εtn (4)
wherelnytnisannualkilometerstraveledbyprincipaldrivernat
yeart,takingthenatural logarithm.Rtn (principaldriver’s
char-acteristics),Htn(principaldriver’shouseholdcharacteristics),Ctn
(fixedandvariablecostsatyeart),andWtn(principaldriver’s
vehi-clecharacteristics)arevectorsofexplanatoryvariables.Dtnisaset
ofinterceptdummies;eachobservation(principaldriver)hasits ownintercept.n,,,,andωaretheunknownparameterstobe
estimated.εtnisthenormallydistributederrorterm.
Underthenullhypothesisthatthepooledregressionmodelwith asingleconstanttermusingordinaryleastsquaresestimationis theproperspecificationforthedata,theFtestisappliedtotestthe significanceofthefixedeffectsparameters.Rejectionofthenull hypothesisindicatesthatthefixedeffectsmodelispreferred.
Alternatively, the unobserved effects can be captured by a randomlydistributedparameter,iftheindividualeffectsare uncor-relatedwiththeregressors.Thistypeofpaneldatamodelisreferred totherandomeffectsmodelandcanbeestimatedbythe general-izedleastsquaresprocedure.Therandomeffectmodeliswritten as
lnytn=(+
v
n)+Rtn+Htn+Ctn+ωWtn+εtn (5)whereisthesingleconstantterm;nistheunobservablerandom
componentonlyspecifictothenthobservation,andεtntheelement
oftheerrorthatvariesoverobservationandyear.
Therandomeffectsmodelreducesthenumberofparameters tobeestimated. However,if theassumptionof random effects specification is not held, therandom effectestimator is incon-sistent.TheLagrangemultiplier(LM)test, proposedby Breusch andPagan(1980),is usedtojustify appropriatenessofthe ran-domeffectsmodelversusthepooledregressionmodel.Underthe null hypothesis thatthe pooledregression model hasa correct modelspecification,rejectionofthenullhypothesisindicatesthat therandomeffectsmodelisfavored.Theselectionbetweenfixed andrandomeffects specificationsisbased ontheHausmantest (Hausman,1978).Thenullhypothesisisthattherandomeffects modelisthepreferredspecification.Rejectingnullhypothesis indi-catesthefixedeffectsspecificationisthepreferredone.
4. Empiricalanalysis 4.1. Thedata
Thesurveyquestionnairecontainsthreemaincomponents.The firstpart includeshouseholdcharacteristics, suchashousehold
C.-H. Wen et al. / Accident Analysis and Prevention 49 (2012) 193– 202 197
Distributionofhouseholdcharacteristicsundervariousnumbersofmotorcyclesownedinthesecond-wavesurvey.
Item Level Motorcycleownership(innumber) Motorcycleownership(inpercentage)
1 2 3 4 Subtotal 1 2 3 4 Householdsize 1 12 4 1 1 18 65.94 22.05 6.03 5.99 2 39 77 16 11 143 27.17 53.97 11.32 7.54 3 65 89 45 22 221 29.54 40.14 20.52 9.80 4 77 133 100 52 362 21.18 36.84 27.58 14.40 5 29 77 64 43 213 13.40 36.16 30.08 20.36 ≥6 22 73 52 29 177 12.69 41.27 29.47 16.56
Numberofmotorcyclelicenses 1 108 27 2 2 140 77.39 19.43 1.59 1.58
2 91 255 34 5 385 23.60 66.23 8.77 1.40
3 26 106 107 28 267 9.59 39.70 40.16 10.55
4 14 49 103 87 253 5.68 19.33 40.72 34.26
≥5 5 16 32 36 89 5.46 18.27 36.20 40.07
Householdmonthlyincome 0–5 93 176 88 52 409 22.74 43.03 21.48 12.74
5–10 99 202 123 69 493 20.10 40.94 24.89 14.07 10–15 28 45 43 28 145 19.07 31.36 30.08 19.50 15–20 10 11 9 5 35 29.15 30.95 24.58 15.31 20–25 6 5 2 1 14 43.66 32.45 15.96 7.93 25–30 0 6 4 1 12 0.00 54.54 35.93 9.53 ≥30 8 9 10 1 28 29.25 31.29 35.44 4.03 Carownership 0 78 91 51 31 251 30.89 36.27 20.30 12.54 1 132 266 157 87 641 20.55 41.43 24.52 13.51 2 27 77 56 34 193 13.74 39.82 29.09 17.34 3 8 17 13 2 41 19.84 42.43 32.27 5.46 ≥4 0 2 1 4 8 0.00 29.00 14.27 56.73
Yearlykilometerstraveled 0–2500 106 159 74 38 377 28.08 42.07 19.66 10.20
2500–5000 63 141 101 58 363 17.25 38.90 27.83 16.03
5000–7500 30 58 44 14 146 20.68 39.66 29.91 9.74
7500–10,000 24 60 37 29 150 15.81 40.28 24.76 19.15
≥10,000 21 35 23 19 98 21.99 35.75 23.12 19.14
198 C.-H.Wenetal./AccidentAnalysisandPrevention49 (2012) 193–202 Table4
Estimationresultofmotorcycleownershipmodel.
Variables MNLModel1 MNLModel2
Coefficient t-Statistic Coefficient t-Statistic
Alternativespecificconstants
Onevehicle 1.408 3.869*** −0.750 −1.611 Twovehicles 0.674 1.315*** −0.883 −1.348 Threevehicles −3.542 −5.511 −3.284 −4.044*** Fourvehicles −6.232 −8.121*** −5.706 −5.754*** Laggeddummies Zerovehicle – – 4.237 5.883*** Onevehicle – – 3.239 17.899*** Twovehicles – – 3.002 19.200*** Threevehicles – – 2.659 15.594*** Fourvehicles – – 4.460 14.082***
Numberofhouseholdmembers
Twovehicles 0.216 6.375*** 0.149 2.505*
Threevehicles 1.480 22.844*** 0.923 9.452***
Fourvehicles 2.022 23.150*** 1.463 11.248***
Memberswithoutmotorcyclelicense
Threevehicles −1.256 −22.608*** −0.785 −9.602***
Fourvehicles −1.806 −22.051*** −1.410 −11.308***
Numberofcarsinthehousehold
Threevehicles −0.372 −5.299*** −0.228 −2.059** Fourvehicles −0.527 −5.724*** −0.574 −3.750*** Transitdensity Twovehicles −0.006 −4.349*** −0.008 −2.919*** Threevehicles −0.004 −2.215** −0.007 −2.272** Fourvehicles −0.008 −2.951*** −0.007 −1.412
(Annualinsurancefee)0.5 −0.008 −3.612*** −0.006 −1.789*
(Annuallicensetaxandfuelfee)0.5 −0.043 −2.657*** −0.044 −2.235**
(Maintenancecostperkm)2 −0.001 −2.795*** −0.001 −1.686*
Finallog-likelihoodvalue −3731.604 −1542.354
Likelihoodratio2 0.282 0.703
Adjustedlikelihoodratio¯2 0.278 0.699
Samplesize 3228 3228
* Significanceoft-statisticsat10%. ** Significanceoft-statisticsat5%. ***Significanceoft-statisticsat1%.
location,age,andgenderofhouseholdheads,familysizeand struc-ture,household income, number of workers in the household, numberofvehiclesinthehousehold,distancefromhometo near-estpublictransitstop,andpurchasesorsalesofmotorcycles.The secondpartincludesprincipaldriver/riderdemographics,suchas gender,age,occupation,educationallevel,income,driving expe-rience,commutingmode,andtraveltimetowork.Thethirdpart includesvehiclecharacteristics,suchasyearofproduction,year ofacquisition,brand-neworsecond-hand,brandname,purchase price,enginesize,gasmileage,annualkilometerstraveled, cumu-latedkilometerstraveled,majorareasinuse,weeklycommuting days,weeklyrecreationaldays,totalannualusagecosts(e.g.,fuel, maintenance,parking,roadtoll,andinsurance).
Thequestionnairesweredisseminatedbyposttoownersin23 cities/countiesinTaiwanproportionallytothenumbersof motor-cyclesregisteredinthosejurisdictionsaccordingtothestratified randomizationsamplingmethod.Thefirst-wavesurveyrandomly selectedatotalof45,000driversfromTaiwan’sVehicleRegistration (VR)DatabasefromOctober1toNovember30,2007.Atotalof2536 validquestionnaireswerereturned,withaneffectiveresponserate of5.6%.In2008,thesecond-wavesurveysentthequestionnaireto therespondeddriversinthefirstwaveandreturned1134valid questionnaires,withaneffectiveresponserateof44.7%.
Table3presentsthedistributionofhouseholdcharacteristics undervariousnumbersofmotorcyclesowned.NotedfromTable3, thehouseholdswithalargersizetendtoownmoremotorcycles. Thesamedistributionalsoexistsfor thenumberof motorcycle licenses.However, thenumber of motorcycles willdecrease as thehouseholdincomeincreases,suggestingthatmotorcyclesare consideredasinferiorgoods.Inaddition,carusageisusually con-sideredasa substituteof motorcycleusage.However,fromthe
cross-tabulationofcarandmotorcycleownership,thesubstitution effectbetweentheownershipofthesetwomodelsisnot signif-icant.Oneofthereasonsfortheinsignificantsubstitutioneffect maybebecausepeopleinTaiwantreatmotorcyclesasashort-range transportmodeevenforthosewhohavealreadyownedcars.Asto theeffectofmotorcycleusage,thereisstillnoobvious distribu-tionthatcanbeidentifiedfortherelationshipbetweenmotorcycle usageandownership.Furthermore,astothedistributionof house-holdswithvariousnumbersofmotorcycles,thelargestpercentage of sampled households (39.96% and 453 households) own two motorcycles,followedbythehouseholdsthatownthree motorcy-cles(24.57%and279households).Thedemographicbreakdowns of motorcycle owners and characteristics of sampled motorcy-clesof thefirst-wavesurvey canrefer toChiouetal.(2009)for moredetails.
4.2. Estimationresultofmotorcycleownershipmodel
Formotorcycleownership,theestimationresultsofMNL mod-elswithandwithoutstatedependencevariablesarereportedin Table4.Thequestionnairesurveyrevealedfewhouseholds pos-sessingmorethanfourmotorcyclesorreplacingmorethanone motorcycleannually,andthusthemotorcycleownershipmodel considersfivealternatives(zerotofour-vehiclehouseholds).The alternative of zero motorcycles in the household was selected asthereferencethat itsconstantis zero.Theexplanatory vari-ables usedinthemodelsaremainlyobtainedfromoursurvey, exceptforonevariable:transitdensity(definedastransit vehicle-kilometerpercapita),whichwascalculatedforeachcity/county. Thisvariablereflectsdifferencesinpublictransportation environ-mentamongcounties/cities,andrespondentslivinginthesame
residentialareassharethesamevalue.ComparingtwoMNL mod-elsintermsoflikelihoodratioindexindicatesthattheMNLmodel 2performsbestwith2=0.703.Thecoefficientsofstate
depen-dence variables (lagged dummies) are positive and statistically significant,suggestingthatthestatedependence effectplaysan importantroleinhouseholds’decisiontoownamotorcycle. Ignor-ingstatedependenceastheMNLmodel1specificationmaylead tobasedestimatesandinferiormodelfit.Our findingis consis-tentwiththeresultsreportedinKitamuraandBunch(1990)and Hanlyand Dargay (2000)and further indicatestheimportance ofstate dependencein modeling vehicleownership withpanel data.
The parameter coefficients of explanatoryvariables are sig-nificantlydifferentfrom zeroatthe10% level,except for some alternative specific constants. The generic variables, namely, annuallicensetaxplusfuelfee,annualmaintenanceandinsurance costsperkilometerinowningandusingmotorcycleshavenegative parametercoefficients,indicatingthatincreasingthevalueofeach variablewillreducetheutilityofthealternative,andthusthe prob-abilityofthealternativebeingselected,providedallelseremains unchanged. Negative coefficients, as expected, were related to thesecosts,suggestingcarownershipreduceswithincreasingfixed orvariablecosts.Judgingfromarelativemagnitudeofcoefficient estimates,licensetax plusfuel feeshows thelargest andmost significanteffectonmotorcycleownership,followedbyinsurance cost.Maintenancecosthasthesmallestcoefficient.
Theremainingexplanatoryvariablesarealternativespecificto thenumberof motorcycles.Thetransitdensityvariable reveals threenegativecoefficientsrelatedtothreealternatives(two,three, and four motorcycles in the household), indicating that more convenient transit service reduces motorcycle ownership. The householdsizeestimatesspecifictothreealternatives(two,three, andfourmotorcyclesinthehousehold)areallpositiveand sig-nificant;thevalueofcoefficientbecomeslargeasthehousehold sizeincreases,suggestingthatastherearemoremembersinthe household,more motorcycles areneeded tosatisfy theirtravel requirements.Numbersofcarsinthehouseholdshowsignificant andnegativeeffectsontwoalternatives(threeandfour motorcy-clehouseholds),indicatinghighsubstitutionrelationshipsbetween carsandmotorcycles.Asthenumberofhouseholdmembers with-outmotorcyclelicensesincreases,thehouseholdsareinclinednot toowntoomanymotorcycles,particularlythreeandfour motor-cycles.Householdincomeisanimportantexplanatoryvariablein thecar ownershipstudy.Itisa surprisethathouseholdincome isnotassociatedwithmotorcycleownershipinthecurrentstudy. Thismaybeduetolowpurchasecostofmotorcycles.Most low-incomehouseholdscanaffordtobuyinexpensivemotorcyclesas theyneed, and high-income households prefer more cars than motorcycles.
Tocapturethepossiblecorrelationstructurebetween alterna-tives,thisstudyestimatedavarietyofNLmodels.Theestimation resultsofallNLmodelsshowedthatthelogsumparametersfell out-sidethereasonablerangeorwerenotsignificantlydifferentfrom one.Inaddition,theNLmodelsdonotoutperformtheMNLmodel usinglikelihoodratiotest,indicatingthatthestandardMNLmodel, amoreparsimoniousmodelthantheNL,wouldbeappropriate.The MXLmodelwasfurtherusedtodetermineifindividual heterogene-ityexistsinthestandardMNLspecification.Randomcoefficients wereappliedtoallexplanatoryvariableswiththeexceptionofthe alternativespecificconstants,butnoneofthestandarddeviation estimateswerestatisticallysignificant.Additionally,theresultof likelihoodratiotestdoesnotsupporttheuseoftheMXL specifica-tion.TheIIAassumptionoftheMNLmodelholds,indicatingthat theuseofeitherNLorMXLmodelisnotrequired.Consequently,the standardMNLspecificationischosenasthepreferredmotorcycle ownershipmodel.
Table5
Resultsofgoodness-of-fitassessment forfixed-andrandom-effectsmodelsof
motorcycleusage.
Tests Statistics
Ftesta F(1133,1088)=4.04,p-value<0.001
LMtestb 2=73.58,p-value<0.001
Hausmantestc 2=28.18,p-value<0.001
aThepooledregressionmodelisrejectedbythefixedeffectsmodel. bThepooledregressionmodelisrejectedbytherandomeffectsmodel. c Therandomeffectsmodelisrejectedbythefixedeffectsmodel.
4.3. Estimationresultofmotorcycleusagemodel
Thisstudyestimatedtheparametersofmotorcycleusagemodel inastepwisemanner.Thedependentvariableisln(annual kilo-meterstraveled),andexplanatoryvariablesconsistofhousehold andprincipaldriver’scharacteristics,vehiclecharacteristics, own-ership and usage costs,and travel patterns of principal driver. Pooledregressionmodelusingordinaryleastsquaresestimation wasinitiallyestimatedandusedtoidentifyimportantexplanatory variables.Subsequently,therandomandfixedeffectsmodelswere estimated.TheF,LM,andHausmantestswereusedtodetermine thepreferredmotorcycleusagemodel.Testresultsofpooled,fixed, andrandomeffectsmodelsarepresentedinTable5.
TheFtestconcludesrejectionofthenullhypothesis,indicating thatthefixedeffectsmodelisthepreferredspecification.Moreover, theLMtestsuggeststherandomeffectsmodelisanappropriate specificationoverthepooledregressionmodel.Finally,the Haus-mantestwasappliedtoselectbetweenfixedandrandomeffects specifications.Thetestresultrejectstherandomeffects specifica-tion,whichindicatesthefixedeffectsspecificationisthepreferred one.
Table6liststheestimationresultofthepreferredfixedeffects model.ThecoefficientofdeterminationR2=0.809indicatesgood
modelfit.Theprincipaldriverincomenegativelyinfluences motor-cyclevehiclemilestraveledbecausehighincomedriverspreferto usecarsmorefrequentlythanmotorcycles;thecoefficientestimate alsoindicateslowincomeelasticityof−0.236.Thenegative coeffi-cientestimateofthesquarednumberofmotorcyclesinhouseholds indicatesthat theexistence of high substitution effects among motorcycleswithinhouseholdswouldleadtoareductionofeach motorcycle’susage.The positivecoefficientof principaldriver’s gendervariableprovidesevidencethatmalesusemotorcyclesfor travelmorethanfemales.Morecommutingandrecreationaldays Table6
Estimationresultofmotorcycleusagemodel.
Variables Coefficient t-Statistic
Genderofprincipaluser(male=1;female=0) 1.068 1.788*
ln(monthlyincomeofprincipaluser) −0.236 −1.729*
Dailycommutingtime 0.002 2.387**
Enginesize(c.c.) 2.347 2.808***
Ageofvehicle −0.011 −2.178**
Numberofhouseholdmembers 0.001 2.188**
Numberofworkersinhousehold 0.028 2.159**
(Numberofmotorcycles)2 −0.014 −2.043**
ln(maintenancecostperkm) −0.116 −4.256***
ln(fuelcostperkm) −0.232 −24.627***
ln(parkingcostperkm) −0.212 −7.813***
Weeklycommutingdays 0.031 3.071***
Weeklyrecreationaldays 0.001 2.343**
R2 0.809
Samplesize 2196
Theconstantsarenotshownduetoalargenumberofestimatesproducedbythe
fixedeffectsmodel.
*Significanceoft-statisticsat10%. **Significanceoft-statisticsat5%. ***Significanceoft-statisticsat1%.
200 C.-H.Wenetal./AccidentAnalysisandPrevention49 (2012) 193–202
Table7
Changesinnumberofmotorcyclesformanagementstrategiesundervariousscenarios.
Strategies Numberofmotorcycles Basevalue 10% 30% 50%
Increasinglicensetaxandfuelfee Total 14,281,206 14,251,215 14,171,241 14,072,700
Numberchange – −29,991 −109,965 −208,506
Percentchange – −0.21 −0.77 −1.46
Increasinginsurancecost Total 14,281,206 14,275,494 14,249,787 14,225,509
Numberchange – −5712 −31,419 −55,697
Percentchange – −0.04 −0.21 −0.39
perweektendtoincreaseannualkilometerstraveled.Theengine
sizevariableispositivelyassociatedwithannualkilometers
trav-eled,indicatingthatlargermotorcyclesaremoreintensivelyused
bythedriver.However,ageofmotorcyclehasanegativeimpacton
annualkilometerstraveled.Thecoefficientsofgasprice,
mainte-nancecost,andparkingcostareallnegative,indicatingthatthe
lowerusage costof motorcyclesincreases usageintensitywith
elasticitiesof0.232,0.116,and0.212,respectively.
4.4. Strategyanalysisresults
Thepredictedchoiceprobabilitiesforthenumberof
motorcy-clesownedbyhouseholdswerecalculatedusingtheownership
model.ThetotalnumberofmotorcyclesinTaiwanwasestimated
as14,281,206bymultiplyingtheaveragenumberofmotorcycles
ownedbyahousehold(thepredictedchoiceprobabilities×
corre-spondingnumberofmotorcycles)andtotalnumberofhouseholds.
Usingtheaverageannualkilometerstraveledobtainedfromthe
usagemodel, the estimated total kilometers traveled (in
thou-sands)is72,505,683.Basedontheparameterestimatesobtainedin
motorcycleownershipandusagemodels,changesinthenumber
ofmotorcyclesandtotalkilometerstraveledundervarious
man-agementstrategiescanbecalculatedtoevaluatetheeffectiveness
ofmanagementstrategiesinreducingmotorcycleownershipand
usage.Twomanagementstrategies(increasinglicensetax/fuelfee
andincreasinginsurancecost)influencebothmotorcycle
owner-shipandusage,whilegaspricesandparkingfeesmainlyaffectthe
usageofmotorcycles.
Table7 reportsthesimulationresultsforthechangesinthe numberofmotorcycleswithrespecttotwomanagement strate-gies(licensetax/fuelfeeandinsurancecost)underthebasecase (unchanged)andthree scenariosof10%,30%,and 50%increase. Totalnumberofmotorcyclesforthebasescenariois14,281,206
inTable7.Underscenarioof10%increaseinlicensefax/fuelfee andinsurancecost,thetotalnumberofmotorcyclesdecreasesby 0.21%and0.04%,respectively.Despiteasmallpercentageofchange inbothstrategies,increasinglicensefax/fuelfeeismoreeffective inreductionofmotorcycleownership.
Table8reportsthesimulationresultsforthechangesintotal kilometers traveled withrespectto fourstrategies under three scenarios.Totalannualkilometerstraveled(inthousands)byall motorcycles for thebase scenario is 72,505,683. Total kilome-terstraveledbymotorcyclesdecreaseby2.28%,6.15%,and9.34%, respectively, under scenarios of 10%, 30%, and 50% gas price increases.Totalkilometerstraveled bymotorcyclesdecreaseby 2.01%,5.42%,and8.25%,respectively,underscenariosof10%,30%, and 50% parking fee increases. These two management strate-gieswillleadtogreaterreductionsintotalkilometerstraveledby motorcycles.Inaddition,thesestrategiesareeffectiveinalleviating motorcyclefatalitiesandinjuries.Forexample,fatalitiesdecrease by28,76,and115,respectively,underscenariosof10%,30%,and 50%gaspriceincreases.
5. Discussionandconclusions
Anationwidepanelsurveyofmotorcycleownerswasconducted toexaminethedynamicchoicesofthenumberofmotorcyclesas wellasusage.Formodelingownership,thisstudyusestheMNL,NL, andMXLmodelstoaccommodatethepossibleexistenceof inde-pendenceamongalternativesandindividualheterogeneity.Panel dataregressionmodelsconsideringfixedandrandomeffectswere developedfortheusage.Todemonstratetheapplicabilityofthe ownershipandusagemodels,theeffectsofvariousmanagement strategiesindecreasingnumberofmotorcyclesandtotal kilome-terstraveledwerethenevaluated.
Table8
Changesinkilometerstraveledandnumberofvictimsformanagementstrategiesundervariousscenarios.
Strategies Kilometerstraveled(in
thousands)
Basevalue 10% 30% 50%
Increasinglicensetaxandfuelfee Total 72,505,683 72,475,692 72,395,718 72,297,177
Numberchange – −29,991 −109,965 −208,506
Percentchange – −0.21 −0.77 −1.46
Fatalitieschange −1 −2 −4
Injurieschange −61 −224 −425
Increasinginsurancecost Total 72,505,683 72,476,681 72,346,170 72,222,911
Numberchange – −29,002 −159,513 −282,772
Percentchange – −0.04 −0.22 −0.39
Fatalitieschange −1 −3 −5
Injurieschange −59 −325 −576
Increasinggasprice Total 72,505,683 70,849,063 68,049,947 65,736,391
Numberchange – −1,656,620 −4,455,736 −6,769,292
Percentchange – −2.28 −6.15 −9.34
Fatalitieschange −28 −76 −115
Injurieschange −3376 −9081 −13,796
Increasingparkingfee Total 72,505,683 71,049,000 68,578,351 66,521,858
Numberchange – −1,456,683 −3,927,332 −5,983,825
Percentchange – −2.01 −5.42 −8.25
Fatalitieschange −25 −67 −102
Motorcycleownershipandusagemodelsidentifyfactors
influ-encing motorcycle ownership and usage, such as household
structure,residentiallocation,transportationsystemperformance,
driver’stravelpatterns,andvehiclecharacteristics.Previous
stud-iesoncarownershipshowedthathigh-incomehouseholdsprefer
more vehicles. On thecontrary,household income is not
asso-ciatedwithmotorcycleownershipin thepresent studybecause
low-incomehouseholdscanafford tobuyinexpensive
motorcy-clesastheyneed,andhigh-incomehouseholdsprefermorecars
thanmotorcycles.Inaddition,ourresultindicatesthatthestate
dependenceeffect(previouschoices)playsanimportantrolein
households’decisiontoownmotorcycles.
Thesimulationresultsindicatethatthechangesintotalnumber
ofmotorcyclesareinsensitivetomanagementstrategiesofboth
increasinglicensefax/fuelfeeandinsurancecost.However,
rais-inggaspricesand parkingfeesshoweffectivenessin reduction
oftotalkilometerstraveled.InTaiwan,exceptforametropolitan
arealikeTaipei,whichhasametrosystemandanextensive
bus-transitnetwork, peopleinothercitiesandcounties relyheavily
onprivatemodesfor mobility.Thus, controllingfor motorcycle
ownershipisunlikelytobesuccessful.Incontrast,curtailingthe
usageofmotorcyclescan beachieved usingsomemanagement
strategies.Ourfindingsuggeststhatinadditiontotheuseofhigh
gaspricepolicy,effective parkingdemand managementshould
beexercised.InTaiwan,motorcyclesareubiquitousinurbanand
suburban areas where parking is convenientand free in many
regions.Forsustainabletransportationanddevelopment,
motorcy-cleparkingchargescheme,suchasanhourlyorflatratechargeto
on-streetparking,canbeimplementedinurbanareas,particularly
wherepublictransportationsystemhasbeensufficientlyprovided.
Becauseasignificantandpositiverelationshipbetweenmotorcycle
usageand crashvictimsisidentified,thesemanagement
strate-giescouldbeeffectiveinalleviatingmotorcycleaccidentsandcrash
victims.
Theresultofmotorcycleownershipmodelindicatesthatgood
transitservicesdecreasethetendencyofhouseholdstoownmore
motorcycles.Thefindingsuggeststhatimprovingtransitservices
canbeeffectiveinreductionof thenumber ofmotorcycles.For
example,TaipeiCitywiththebestpublictransitservicesamong23
counties/citieshasthelowestrateofmotorcyclesperhousehold
(InstituteofTransportation,2010).Asaresult,TaipeiCityhasthe lowestfatalcrashrateformotorcycles.Currently,onlytwocities inTaiwanhave rapidtransitsystems,andbus modesharesare relativelylowfortherest.Toreducemotorcycleownershipand usage,othercitiescouldimprovethelevelofbustransitservices byextendingthebusnetworkorincreasingbusfrequency.
Therewereseverallimitations tothisstudy, manyof which highlightdirectionsthatcouldbeconsideredinfuture research. Thisstudyonlydevelopedmotorcycleownershipandusage mod-els.Duetothehighsubstitutionbetweenmotorcyclesandcars, anintegratedmodelwithpaneldataincludingownersof motor-cyclesandcarscanbefurtherdeveloped.Thisstudyconducteda large-scalepanelsurveyonownersofmotorcyclesinTaiwan,and therefore,thefindingscannotbegeneralized toothercountries withdiversecharacteristics.Consequently,itmaybeworthwhile to validate our proposed models and results using data from different countries. The data sets in this study comprise two waves of surveys. The third or more waves of follow-up sur-veysshouldbe continuouslyconducted toenrich thedata sets for a comprehensive analysisof the dynamic choicebehaviors. Themotorcycle ownershipand usagemodelsdeveloped in this study consistof two separatecomponents. It is likely that the choicesofownershipandusageareinterrelated.Futureresearch couldexamineajointchoiceofmotorcycleownership(discrete) and usage (continuous) by using discrete/continuous modeling structures.
Acknowledgements
The authors appreciate two anonymous reviewers for their insightfulcommentsandsuggestions,whichhelpedimprovethe qualityofthispaper.ThisstudywassponsoredbytheInstituteof Transportation,MinistryofTransportationandCommunicationsof theRepublicofChina,undercontractMOTC-IOT-98-SDB004.
References
Baltagi,B.J.,2005.EconometricAnalysisofPanelData.JohnWiley&SonsLtd., Eng-land.
Bhat,R.C.,Pulugurta,V.,1998.Acomparisonoftwoalternativebehavioralchoice
mechanismsforhouseholdautoownershipdecision.TransportationResearch
PartB32,61–75.
Breusch,T.S.,Pagan,A.R.,1980.TheLagrangemultipliertestanditsapplicationsto
modelspecificationineconometrics.ReviewofEconomicStudies47,239–253.
Brownstone,D.,Bunch,D.S.,Train,K.,2000.Jointmixedlogitmodelsofstatedand revealedpreferencesforalternative-fuelvehicles.TransportationResearchPart
B34,315–338.
Burge,P.,Fox,J.,Kouwenhoven,M.,Rohr,C.,Wigan,M.R.,2007.Modelingof
motorcy-cleownershipandcommuterusage:aUKstudy.TransportationResearchRecord
2031,59–68.
Button,K.,Ngoe,N.,Hine,J.,1993.Modelingvehicleownershipanduseinlow
incomecountries.JournalofTransportEconomicsandPolicy27,51–67.
Chiou,Y.-C.,Wen,C.-H.,Tsai,S.-H.,Wang,W.-Y.,2009.Integratedmodelingof
car/motorcycleownership,typeandusageforestimatingenergyconsumption
andemissions.TransportationResearchPartA43,665–684.
Dargay,J.,2007.TheeffectofpricesandincomeoncartravelintheUS. Transporta-tionResearchPartA41,949–960.
Dargay,J.,Hanly,M.,2007.Volatilityofcarownership,commutingmodeandtime intheUK.TransportationResearchPartA41,934–948.
DeJong,G.C.,Fox,J.,Daly,A.,Pieters,M.,Smit,R.,2004.Comparisonofcarownership
models.TransportReviews24,379–408.
Giuliano,G.,Dargay,J.,2006.Carownership,travelandlanduse:acomparisonof theUSandGreatBritain.TransportationResearchPartA40,106–124. Hanly,M.,Dargay,J.,2000.CarownershipinGreatBritain:apaneldataanalysis.
TransportationResearchRecord1718,83–89.
Hausman, J.A., 1978. Specification tests in econometrics. Econometrica 46,
1251–1271.
Hausman,J.A.,McFadden,D.,1984.Specificationtestsforthemultinomiallogit
model.Econometrica52,1219–1240.
Indriastuti,A.K.,Sulistio,H.,2010.Influencingfactorsonmotorcycleaccidentin urbanareaofMalang,Indonesia.InternationalJournalofAcademicResearch2, 252–255.
InstituteofTransportation,2010.AnnualStatisticsforPopulationandCar Owner-shipinTaiwanArea.InstituteofTransportation,MinistryofTransportationand
Communications,Taiwan(inChinese).
Jansson,J.O.,1989.Cardemandmodelingandforecasting:anewapproach.Journal
ofTransportEconomicsandPolicy23,125–138.
Karlaftis,M.,Golias,J.,2002.Automobileownership,householdswithautomobiles, andurbantrafficparameters:aretheyrelated?TransportationResearchRecord:
JournaloftheTransportationResearchBoard1792,29–35.
Kitamura,R.,Bunch,D.S.,1990.Heterogeneityandstatedependenceinhousehold
carownership:apanelanalysisusingordered-responsemodelswitherror
com-ponents.In:Koshi,M.(Ed.),TransportationandTrafficTheory.ElsevierScience
PublishingCo.,Amsterdam.
Koornstra,M.,Lynam,D.,Nilsson,G.,Noordzij,P.,Petterson,H.E.,Wegman,F.,
Wouters,P.,2003.SUNflower:AComparativeStudyoftheDevelopmentofRoad
SafetyinSweden,TheUnitedKingdomandTheNetherlands.InstituteforRoad
SafetyResearch,Leidschendam,SWOV.
Lai,W.-T.,Lu,J.-L.,2007.Modelingtheworkingmodechoice,ownershipandusage ofcarandmotorcycleinTaiwan.JournaloftheEasternAsiaSocietyfor Trans-portationStudies17,869–885.
Mannering,F.,1983.Aneconometricanalysisofvehicleuseinmultivehicle
house-holds.TransportationResearchPartA17,183–189.
Mannering,F.,Winston,C.,1985.Adynamicempiricalanalysisofhouseholdvehicle
ownershipandutilization.RandJournalofEconomics16,215–236.
Mohan,D.,2002.TrafficsafetyandhealthinIndiancities.JournalofTransportand Infrastructure9,79–94.
MOTC,2010.AnnualReportofRoadTrafficSafety.MinistryofTransportationand
Communications,Taiwan(inChinese).
NHTSA,2007.TrafficSafetyFacts:Motorcycles.U.S.DepartmentofTransportation.
ReportNo.:DOTHS810990.
Potoglou,D.,Kanaroglou,P.S.,2008.Modellingcarownershipinurbanareas:acase
studyofHamilton,Canada.JournalofTransportGeography16,42–54.
Ranney,M.L.,Mello,M.J.,Baird,J.B.,Chai,P.R.,Clark,M.A.,2010.Correlatesof
motor-cyclehelmetuseamongrecentgraduatesofamotorcycle trainingcourse.
AccidentAnalysisandPrevention42,2057–2062.
Senbil,M.,Zhang,J.,Fujiwara,A.,2007.Motorcycleownershipandusein metropoli-tanareaofJabotabekIndonesia.In:Presentedatthe86thAnnualMeetingofthe
202 C.-H.Wenetal./AccidentAnalysisandPrevention49 (2012) 193–202
Suriyawongpaisal,P.,Kanchanusut,S.,2003.RoadtrafficinjuriesinThailand:trends, selectedunderlyingdeterminantsandstatusofintervention.InjuryControland
SafetyPromotion10,95–104.
Train,K.,1980.Astructuredlogitmodelofautoownershipandmodechoice.The
ReviewofEconomicStudies47,357–370.
Umar,R.,2002.HelmetinitiativesinMalaysia.In:Proceedingsofthe2ndWorld
EngineeringCongress,Kuching,Sarawak,Malaysia,InstitutionofEngineers. Wen,C.-H.,Koppelman,F.S.,2001.Thegeneralizednestedlogitmodel.
Transporta-tionResearchPartB35,627–641.
WHO,2006.Helmets:ARoadSafetyManualforDecision-makersandPractitioners.
WorldHealthOrganization,Geneva.
Woldeamanuel, M.G., Cyganski, R., Schulz, A., Justen, A., 2009. Variation of
households’carownershipacrosstime:applicationofapaneldatamodel.
Trans-portation36,371–387.
Zhang,J.,Norton,R.,Tang,K.C.,Lo,S.K.,Jiatong,Z.,Wenkui,G.,2004.Motorcycle
ownershipandinjuryinChina.InjuryControlandSafetyPromotion11,159–