A Vehicle Use Forecasting Model Based on Revealed and ... - CiteSeerX

A Vehicle Use Forecasting Model Based on Revealed and Stated Vehicle Type Choice and Utilization Data Thomas F Golob David S Bunch David Brownstone

UCTC No 598

The University Transportation

of CMifornia Center

The Umvers2ty of Califorrda TransportaUon Center(UCTC) 2sericoftenregional urn-is mandated by Congress and established in Fall198gto support research, cducauon, and trazmng m su~ace ~nsponaUon.The UC Center servesfederal RegionIX and is supposed by matching gTan~ from the U.S. Department of Transpo~a~on, the California Department of Transportauon (Calt~ans), and ~c lJmvcrszzy Basedon theBerkeley Campus,UCTC draws upon existing capabflmcs and resources ofthe Instnntes of Transportation Studzes at Berkeley, Davzs, Irvme,and Los Angei~,the Insumte of Urbanand RegmnalDevelopment at Berkeley, and several acadcmzc departments at me Berkeley, Davis, Irvme, and Los Angeles campuses Faculty andstudents on other Umverszty of California campuses may partzcipate m

Centeraefiwnes. Researchers at otheruniversiues within the region also have opportunities to coIlaborate with UCfaculW on selected smches. UCTC’s educanonaI and research pro~"ams are focuses on strategic ptam’ung for lmprovmg metropohma accessibiIhy, with emphasis on thespot:el conditions m Region IX.Part,cularattention is d~rected to s:ratemes for using transportation as an instalment of economzc development, while also accommodat:ng to the regmn’s p~rs~stent expansmv and while mammmmgand enhancmg the quahty of life there The Center dzsr_nbutes ~pons on tr~ research m worhng papers, monogaphs, and :n reprints of pubhshed amctes It alsopubhshes Acces.% a magazine presenting summanes of selected smdmsFor a hst of pubhcatmns m pnnt, write to the address below

N$CLANER

Umvermtyof C~LIforn~ Transportation Center I08 Naval Axchttectum Bmld~g Bcmeley, Cahfotma94-720 Tat 5101643*7378 FAX.510/643-5456

The contents of thisreport reflecttheviews of theauthors, who are responsible forthefactsandtheaccuracy oftheinformation presented herein. This document is disseminated under thesponsorship ofthe Department of Transportation, University TmnspDrtatJon Centers Program, i. theinterest of informabon exchange. TheU.S.Government assumes no Jiabiiit7 forthecontents orusethereof.

Theconten~ofthis r~port reflect the wowsof the au~orwhom responslble forthe facts andaeeur~, of~edata presented hereto Thecent-ate,do not nct:essaniy reflect fl~officud views or pohczcsof~eState ofCahfomaa or the U S DepartmentofTnmspormnoa. 2"msr~pon d~snot consUtulz a standard, spedficauoa, or mgula~um

A Vehicle Use Forecasting Model Based on Revealed and Stated Vehicle Type Choice and Utilisation Data

Thomas F. Golob Institute of TransportationStudies Umversltyof Cahfomia,Irvme David S. Bunch Graduate School of Management University of Cahforma,Davis David Brownstone Departmentof Economics Umvers~tyof Cahfomla.Irvme

Reprinted from Journal of Transport Economtcsand Pohcy pp 69-92 (January 1997)

UCTCNo. 598 The University of California Transportation Center University of California at Berkeley

A Vehicle Use Forecasting Model Based on Revealed and Stated Vehicle Type Choice and Utilisation Data

By ThomasE Golob, David S. Bunch and David Brownstone*

1. Introduction ~as research describes a new model of household vehicle use behavlour by type of vehicle. Forecasts of future velucle emissions, including potential gains that might be attributed| omtroducUons ofalternaUve-fuel (clean-fuel) vehicles, cntmally dependupon the ability to forecast vehicle-miles travelled by the fuel type, bodystyle and raze, and wntageof the vehicle. Householdsacqmrevebacles to satasfy both the transport needs and the preferences of household membersConsequently, vehicle use by type of vehicle can be consldered to ee a function of three categories of variables" (1) householdcharacteristics, (2) principal dr~ver char actenstics, and (3) characteristics of the vehicle itself Examplesof household charactert ~ttcs are income,residential locatmn, numberof vehicles, numberof hcenceholders, numberof workers, and numberof household membersby age group. Useof a specific vehicle dependsheavily on winchhouseholdlicence-holder typically drives the vehicle, that is, the principal clnver. Importantprmcipal driver characteristics mclucle age, gender and employmentstatus Workers, young people and males are likely m drive more, as demonstrated m the modelsof Hensher (1985), Hensher et al. (!992), Mannering (1983), Mannering and Winston (1985) and Tram (1986) *

ThomasF Golob Institute of Transportauon Studies, Umvers]ty ofCahforma, Irvme David S Bunch Graduate School of Management,Umverslty of Calfforma, Dav~s DawdBrownstone Department of Economlcs, Umverslty of California, Irvme This research xs part of a project to develop a model systemto forecast demandfor clean-fuel vehicles m Calfforma, sponsored by Southern CahfommFAlson Companyand the Cahforma Energy CommissmnThe authors thank thelr colleagues whohave madethe researchreported here possible Jane Torous, WeipmgRen and SeyoungKan of the University of Cahforma, Irvme, RymchlKatamuraof the Umverslty of Cahforma, [)avis, and Kyoto Umverslty and Mark Bradtey of the Bradley Research and Consulting Group Ernest N orales, Richard Rice and Debbie Brodt of Southern CahformaE&sonCo., and Gary Occhluzzo and Mike J~ske of the CalfformaEnergyComrmsslon have also workeddiligently on behalf of the project. Theauthors are solely responsible for the views expressed here and for any remainingerrors

69

January1997

Journalof TransportEconomtcs andPohcy

Useis also affected by vehicle characteristics, such as vet’acle age (vintage). operating and capital costs, passenger and cargo capacity, and body style. Moreover,alternatavefuel vetacles are distingtushed by velucle attributes that potenUally have an important influence on patterns of use. Examplesinclude: linuted range betweenrefuelhng, coupled wath lim.tted fuel availability; or the necessity to refuel or recharge the vetncle at home overnight. D~fferences between convenUonal-fueIand alternaUve-fuel vehicles in terms off-ael costs, cargo capacaty, performanceand imageare also expectedto influence vehicle use (van Wissen and Golob, 1992) Applying a vehicle-type use model m travel demandforecasts requires obtaining or developing forecasts of all the model’s exogenous variables. The first category of varlables, householdcharaetertsttcs, can be readily forecast using census data or household sociodemograptuc models used in regaonal planmng For example, the model developedhere as part ofa nucrosimulataonforecasting system (Brownstoneet aL, 1994, Bunchet at., 1995) that as driven by a competang-nskhazard modelof changing household demograptucs (Kazirni, 1994, 1995; Kazlml and Brownstone, 1995) Forecasts of the second category of varmbles, charactensUcsof the principal drtver, are problematic for multi-vehicle households, and also for smgle-vel~clehouseholdswith more than one hcence-holder. For such households, modethng vehicle use behawour involves all~aUng vehtcles to licence-holders m order to satisfy their acUvxtyneeds (Golobet aL, 1995) In add~Uon,~t revolves d~stributmg total travel amongthe vehicleand licence-holders. While, m principle, only forecasts ofhouseholcl and vebacle charactensUcs are needed to forecast vebacle use for smgle-velucle households w~th only one hcence-holder, exogenousforecasts of the characterisUcs of the principal driver of each velucle mmulti-vebacle and multt-dnver householdsare needed. To address ttus assue, the modelsdescribed here samultaneouslyincorporate alIocataon of hcence-holders to vetncles along w~th vehicle utihsation Finally, exogenousforecasts of household vel~acle holdings by type of vehzcle are obtainable using vehtcle-type choice models, such as those developed by Laveand Train (1979), Manskaand Sherman(1980), Hensher and Manefield (1982), Hochermanet aL (1983), Berkovec (1985), Hensher and Le Plastner (1985), Marmermgand Winston (1985), Tram(1986), McCarthyand Tay (1989) and Hensher et al (1992). Such vehicletype choice models are based on vel-acle holdings and transacUons data (so-called "revealed-preference", or RP, mc~lels), Because consumersdo not have e×penencewith alternative-fuel vehicles that are likely to be available an 1998and beyond,a vebacle-type choice modelbased on stated preference (SP) data ~s reqmredto forecast demandfor these newvehicle types (see, for e×ample,Bunchet al., 1993; or Golobet al., 1993). Onesuch model (Brownstoneet al., 1995) is being coupled with the use modelsdescribed m this paper to forecast alternative-fuel vehicle use for the State of Cahforma. These modelsare similar to previous modelsof vehicle allocation and use in multivehicle households (Mannenng,1983; Hensher, 1985, Tram, 1986; and Hensher et aL, 1992) in that separate equaUonswith correlated error terms are developedfor each vehicle in the household. However,this research chffers from prewousstuches because there are additional equations describing the most ~mportantcharactenstacs of the principal driver 7O

A VchlctcUseFore~asungModd

T F Golober al

of each veincle. Althoughthese characteristics cannot be readily forecast for use in a rmcroslrnulaUonsystem, they can be "solved out" of the problem;reduced-formequatmns ,are developed for forecasting purposes through a structural specification of veincle allocation to drivers. Anotherunique feature ~s that the modelsuse both RPand SP data smmltaneously. In other words, the models are esUmated with a mix of RP and SP observations, mal~ngthe modelssensitive to attributes associated with future alternativefuel vebacles. The current version of the modelstakes the household’s vehicle holdings (both the immberof vehicles and the vehicle types) as Dven.Tins modelstructure is theoretically unappealing(as described in Golobetal, 1995) because a household’santicipated travel behawouris likely to influence its velucle choices If the error terms of the veincle choice modeland the vehicle use modelare correlated, the parameterestimates will be b~ased. Oneapproachis to apply a linear correctaon term mvolwng a transformaUonof predicted vehicle choice probabihties to the veincle use modelto account for self-select~wty bias ,(McFaddeaetal., 1985; Marmenngand Winston, 1985; Train, t986, Hensher et al., t 992). Emplncally, however,the selectlv~ty correcttons apphedm utilisataon modelsto ,~ccount for endogeneltybias havenot had substanUaIeffects on esumationresults (Tram, 1986, Hensher, 1992) Morecomplexmodels, including dtrectjo~m estimation of choice and use, are plannedfor future research 2. Data ’Ihe data are from a 1993 survey conducted using geographically straUfied pure random digit dialling° The survey, covenngmost of urbardsed Califorrda (excluding San D~ego County), contained three distract components,as described m Brownstoneetai. (1994) and Golob etal. (1995) An ~mUalcomputer-aided telephone mterwew(CA’H)coliected inforrnatio a on householdstructure, vehicIe inventory, housingcharacteristics, employmerit data, commuting for all workers and students, and talon’nation about the intended next veincIe transacUon. These CATIdata were then used to produce a customised postal quesUonna~rewinch asked detatled quesUonsabout each household member’scommuting and ve~acle use The questionnazrealso contained two SP (stated preference) veincletype choice experimentsfor each household, the responses to winchwere collected m the third part of the survey, a follow-on CATIsurvey. Eachof the SP experimentsdescribed three hypothetacalvehicles m terms of attributes such as body type, fuel type, refuelhng range, purchase price, and so on. These hypothetic~dvehicles included both alternalave-fuel and petrol vehicles. Attribute descriptions werevaned accordingto an experimentaldesign. For attributes related to body t~)e and purchase price, canchdate levels m the design were custonusedto be consistent with the types of vehicles that householdsmdacatedan interest m for their next intended ,~einclepurchase, mordertomakethechoicetaskmorerealisUc and relevant Households were asked to choose their preferred vehicle and indicate whether the chosen veincle wouldreplace an erasrJng householdvehicle (and, if so, winchone) or be added to the householdlleet. Thequestionnaire was custonusedso that the choices were characterised as transacUonsrelauve to the household’scurrent vetucle holdings. 71

January 1997

Journal of Transport Economms andPohcy

Vehicleuse SPquestions followedthe choice experiment.Thequestions on use asked the householdto assign princlpaldrivers to eachvetuclemthe newveNclefleet (including the SPchosenvehicle), and to mchcatehowmanymiles per year the chosenvebaclewas likely to be drlven. Theflowof the surveyensuredthat respondentsfirst reportedthe pnncipaldrivers and patterns of use for their current veNctesbefore performingthe SP task, allowingthemto makereformedjudgementsbased on tl~s mformaUon, as well as their ownperceptions of principal driver and vehicle characteristics. Tfus approach allowedeach completedsurvey to prowdeboth RPand SP measuresof annual vehiclemiles travelled that couldbe jointly analysedman appropriatemodelstructure. TheSP observaUons had the potential to providedata on the effect of alternative-fuel vehicle attributes on annualvehicle-nulestravelled. Ofthe 7,387householdswbachcompletedthe ~mtial CAT[survey,66 per cent, or 4,747 households,successfully completedthe postal pomonof the survey. Acomparisonwith census data reveals that the sample is shghtly biased towards home-owmng larger householdswith higher incomes,and weightsare being developedto balancethe sample to the knownpopulation(Brownstone et aI., 1994). Anunweightedsampleis used here. Thebreakdownof the 4,747 householdsby vehicle ownersl~plevel was: 1 per cent zerovetacles,34 per cent one vetncle, 47per cent twoveI~cles,13per cent three veh~cies; and 5 per cent four or morevehicles Of the one-vehiclehousehokts,75 per cent had one driver, w~le25 per cent hadtwoor moredrivers. Thus,approximately 73 per cent of the samplehouseholdsweree~ther multi-vehacleor single-vehicle/multi-driver,wheredriver allocation behaviourIs relevant The modelvaraables are chwdedinto three groups (1) behav~ouraivehicle use characteristics, capturingthe waysin wbachhouseholdsuse their velucles, (2) physical vehiclecharactenstics;and(3) household structural characteristics. Vebacieuse for each household’svehicles (RPuse data) is self-reported mterms of"How manyrmlesper year ~s this vehicle dnven’~"It wouldbe moreaccurate to calculate annualuse fromvehicle odometerreachngsone year apart, but such data are not avmlablem a cross-sectional survey. Vetucle use for the hypothetical future vetncles (SP use data) wascollected through a series of questaons aslang howmanymiles the vehicle selected m a choice experimentwouldbe driven each week,and whoin the householdwouldtypically use the chosenvebacle to commute to workor school. Becausethese modelsare being used mconjunction with a tbrecastmgsystem, the householdvanables selectedwereh matedto those producedby the avail able demographic forecastang model. Thevariable "meanage of householdheads" wascomputedas the meanof the ages of partners mhousehoIdsincluding spouses, or the age of the single parent or personwhocan be identified as the majorincome-earner.Thedummy variable "household headsare retired" is set to one if oneor both household headsare retired and neither householdheadis employed or temporarilyunemployed (it ~s poss~Nethat another person, perhapsa grownchild, ~s employedmsuch a household). Separate modelsare developedfor singte-vetncle householdsand multi-vehicle households. Thetwo samplesizes are: 2,260 smgle-vebacteobservations, madeup of householdscurrently hot&ngone vetucle (RP data) madhouseholdsen&ngup with one 72

A Vehicle Use Forec~sung Model

T F Oolob et al

Table I EndogenousVariables for EachVehtcle Vanable

Acronym

Natural log of vehicle-miles travelled per year Ageof pnn,zlpal driver myears Genderof principal driver (0 = male, 1 = female) Employment status of prmclpal driver (I = working)

Lnfvlvrr) Driver Age I)nver Gender Driver Empl. St

Table 2 ExogenousVariables for Each Vehicle Variable

Acronym

VehMeage (m years from 1993) Vehicle ela~s dumrraes(12), base = luxury class Mmlclass Subcompactcat class Compactcar class Mld-s~zeor full-size car class Full-size (standard) car class Sports car Compactp~ckup truck Full-size (standard) pickup truck Mmlvan (compact van) Full-size (standard) van Compactsport utility vehicle Full-size (standard) sport utlhty vehicle Operating cost per mile (¢) Elecmc veb lcle (dummy) Range betx~een refuelhng m wales

Vehicle Age Type Mm~ Type Subcompact Type Compact Type Mid-sine Type Full-slze Type Sports Car Type Small Truck Type Std Truck Type Mmlvan Type Van Type Compact SUV Type Full-stze SUV Operating Cost Electric Vehicle Range

velncle after the SP choice task, and 3,150 multi-vehicle observations, madeup of households currently holding two or more vetucles (RP data) and households enchngrap wlth two or morevehicles after the SP choice task (In adchtion, a modelof third-vehicle use wasdevelopedusing a sampleof 445 householdswahthree or more vetucles, but tins modelis not reported here.) Eachof the samplesconsists of householdswith knowntype and vintage of their current velncle (single-vehicle sample), or no missing data on the newesttwo vetnctes m their fleet (multi-vetncle sample). It ~s also required that there no missing data on the age, sex and employment status of the principal drivers of each of these vehicles. 73

Journal of Transport Econommsand Pohcy

January 1997

Table3 ExogenousVariables. HouseholdCharacteristics Variable

Acronym

Household membership variables Total number of household members> 15 years Numberof chgdren m household aged 0 to 5 Numberof household members aged 16-20 Total numberof children (all ages) mhousehold Household is a couple (dummy) Household income less than $31,000 (dummy) Household income more than $60 000 Household head(s) are retired (dummy) Mean age of household heads Total Numberof workers in household Household has three or more vehicles (dummy)

No 16+ Yr Olds No less than 5 Yrs Old No 16-20 Yr Olds Total Children Couple HH Income < $31k Income > $60k Retired HH Ave Age of Heads No Heads Working 3+ Vehlcle HH

3. Specification 3.1 Partition of the variablesinto endogenous andexogenoussets A ~sUngms~ng feature ofttus research 1s the endogenous treatment of driver allocation behaviour.In order to avoidolmtted-var!ablesbias, velficIe-mites travelled (VMT) specified as a functaon of pnnclpal driver charactensUcs, in addmonto exogenous householdand vebacle-typecharactenstmsHowever,principal driver charactermticsare also specified as a functionof the exogenous variables Tlus allowsthe principal driver characteristics, for whichno exogenous forecasts are available, to be replacedby their pre&ctorsm the final forecasting equaUons Thereare four endogenous variables for each vehicle Theseare hsted mTable 1. In the mulU-velucle case, householdvehicles are arrangedso that the newestof the vehtcles is definedas "velucle1", describedby the first four endogenous variables andthe first group of velucle-type exogenousvariables Thesecond-newestvehicle is defined as "vebacle2", and It is describ~by the last four endogenous variables and the last group of veNcle-typevariables If twovebaclesare of the samevintage, the order of listing by the respondentIs preserved ° physical vebacle Theexogenousvanables meach modelare divided into two Nocks characteristicsandhousehold claaractensticsThefirst block,listed mTable2, ~s madeup of 16 physical vetucle charactensUcsfor each vebacle The second block of exogenousvariables ~s composedof up to eleven household characteristacs Tlus hst ~s reproducedw~thassociated acronyms for further reference m Table 3. Thedummy variable for three or morevelucles is used only mthe two-vehicle model.Thesevariables, togetherwith the driver characteristic variableshsted mTable1, were selected on the bas~s of publishedvebacle use modelresults (Mannering,1983; Hensher,1985; Manneringand Winston,1985; Hensh-erand Smith, 1986, Trmn,1986, Golob,1990; Hensheret at, 1992). 74

A Vehlcle UseForecasting Model

T F Golob etad

3.2Thestructural equation model form Thestandard structural equations model (without latent variables) Isgiven y-By+rx+~ (I) wherey is an m× 1 columnvectorof endogenous variables, andxis an n × 1 columnvector ofexogerousvariables. Thestructural parametersare the lementsof the matrices. B = matrix of causal links amongthe endogenousvariables, (mxm) and effects fromthe n exogenous variablesto the F= matrixof direct causal(regressmn) (mxn) m endogenousvariables Theerror term parametersare the elementsof the varianceco-matrix: = E(~’) = symmetricvanance-covariance matrix of the unexplained,or unique, (m x m)po~ons oftheendogenous variables Foridentificataon ofsystem (I), B must bechosen such that (I- 13)isnon-singular, where I d~notes theidentity matrix ofdimension m. Thetotal effects oftheendogenous variables oneach other areglven by H = (I - B)-I - I. (2) Theto~Ial effects of the exogenous variableson the endogenous variablesin a structural equationsmodelof this type are g~venby -IF, T = (I - B) (3) whmhare the parametersof the reduced-formequaUons. 3.3 Divisi, onof the problem into separatemodels Comparisonsof sample sizes with the numberof vanab|es and potential numberof parametersrevealed that separate modelscouldbe developedfor smgle-vehiclehouseholds and for two-veNclehouseholdsHowever,the numberof householdswith three or morevehtcles wasinsuffmlentfor the development of a dedicatedthree-vehicle model. TheaI~emat~vewasto expandthe two-vehiclemodelto cover householdswith two or moreveNctes, and to add a third-veNcle modelfor householdswith three or more vehiclesIt n~ghtbe preferableto modelutthsation of three velucless~multaneously, but the expansionof the present structural equationsystemto 12 endogenous variables and up to 58 exogenousvariables (16 for each vehicle plus 10 householdvariables) is not feasible wRhthe presentdata Theuse of a Third-VehicleModel(not reportedhere), with only 4 enciogenousand 26 exogenous variables, is a pragmatlcsolutaonto the problem. TheTwo-VehicleModel.covenngthe two newest vebacles in multi-veNclehouseholds, is the mostcomplex,andits specificataon~s describedherein detail. TheSingleVehicleModelis a simplification of the Two-Vehicle Model

3.4 Specification of the Two-Vehicle Model Eachmodelspec~ficatton can be subchvldedinto: endogenouseffects given by the B matrixin equationsystem(I); exogenous effects (the r matrix);and error-termvariance75

January 1997

Journal ofTransport Econonucs andPohcy

Table 4 Two-Vehtcle Model Postulated Direct Effects BetweenEndogenous Variables (posture effectsdenoted by+,negative effectsdenoted by-) InfluencingVariable Influenced variable

Ln (VMT~)

Driver Age 1

Driver IMver Ln Gender1 Empl St I ,(VMT2)

Driver Age 2

Driver Driver Gender2 Empl. St2

Ln (VMTO Driver

Age1

~3,2(-)

63.7(-)

=I~I 6~.6 265.7=f~i.3 65 8 ~ ~1,4 Drlver Age2

~6,4 ~ ~2,8

Driver

Gender 2

~7,3 = 133,7

~7,6 = ~3 2

~g,4 ---- 64,8

covariances(the ~Pmatrix)Ttns specificaUon is ba.gedonthe structure of the RP(revealed preference) utihsatlon modeldevelopedmGolobet al. (1995), but the present model extfiblts additional features p~-’ticularly related tojointSP-RPestimaUon. Thepostulated causal relatlonships among theendogenous variables areshown m Table4 Thereare twotypes of direct effects: wlthin-ve!ncleeffects andbetween-vehicle effects. Thewithin-vehicleeffects are those mthe upperleft-hand (first vehicle) and lower right-hand(secondvehicle) quadrantsof the B malrix Eachof these effects Is expected to be identical for the twovehicles, andequityrestriclaons are specifiedfor corresponding pairs of B-malri× parameters.Useis postulatedto be less for vehiclesprimarilydriven byolderpersons(131,2 = ~s.6) andwomen (~51,3 = [55,7), anduseis postulatedto be greater 76

A Vcblcle UseForecasting Model

T F Golobet al

for vehicles primarily driven by employed persons (151,3 = 155,s). Maleprincipal drivers are = morehkelyto be employed (154,3 15s.7), as are youngerprincipal drivers (154,2 = 158,6), and older drivers are expected to be male (153,2 = 157,6)" All important feature of this specification is that, for each of the two householdvehicles, VMT is postulated to be a function of all three of the princ:pal driver variables. Thus, while driver allocataon is endogenous,VMT is specified as a function of dr~ver characteristics. Thepastulated cross-vehtcle effects are those m the lower left-hand and upper righthand quadrants of the B matrix of Table 4 Thesereciprocal effects capture the relationships betweenthe characteristics of the pnncipa] drivers of the two vehicles Weexpect strong negative relationships betweenprincipal-driver genders and employment status in the between-vehicleeffects, and tlus is operationahsed by specifying equated pairs of reciprocaleffects (~3,7= 157,3)and(154,8= 158,4).In addition,weexpectthat, if the driver either velucle is employed,the driver of the other vehicle is likely to be youngerthan otherwise e×pected (152.8 = 156.4) Two-vemcle households with no employeddrivers are likely to contain retired (and therefore older) members.Thepostulated modelis parsimomousm that it has only Pane free parametersin the B matrix, representang rune pairs of equatedcareer effects. The l~)Stulated structure of the vehicle-charactensUcexogenouseffects ~s shownin Table 5. The vehicle-type effects specified m the exogenousvariable structure were developedby considering vehicle-use stereotypes. For example,there are typically more male principal drivers of compactand full-size p~ckuptrucks, subcompactcars mighthave younger principal dr~vers, and minivans are hkely to be driven by females Logically, older vetttcles andhigher operatangcost vetncles shouldbe driven less, other thmgsbe, rig equal. The major resmcuonsapplied m specifying these exogenousvehicle-type influences are that the effects should be the same for the two vetacles It is a straightforward procedure subsequently to test whether the model can be significantly ,reproved by releasing these cross-vebacle parameterequaht3’ resmcuonsIt is also quite possiNethat the charactensracs of the first velucle can affect the VMTand pnnclpal driver characterlst~cs of the second veNcle, and conversely. The modelwas ~mttally specified by setting all such cross-vehicle effects to zero Tests were then conducted to ascertain whether cross-vehicle effects significantly ,mprovedmodelfit Examplesof d~rect household effects to be tested include, principal drivers m households wlth more workers and in lugh-mcomehouseholds are more likely to be employed;use is bagher in householdswith morechildren and in high-incomehouseholds; pnnc:paI drivers are younger in households with young children, drivers m retired households are older and less likely to be employed(although somedrivers m rettred households, such as adult children hying with their parents, could be employed), and finally, householdsw~ththree or morevehicles havelower levels of use on their first and secondvehicles, all else held constant Thedefault restriction on all of these postulated householdinfluences involves equaUngthe corresponchngeffects on the two vehicles, and then tesUng whether the relaxataon of each equoJity results in a significant model tmprovement 77

January 1997

Journal of Transport Econom~cs and Pohcy Table 5 Two- Vehwte Model Postulated Dtrect Effects from the Exogenous Variables EndogenousVartable

Exogenous Variable

Ln Driver I(VMTI) Age

Driver Drover Ln Drtver Gender1 Empl St I (VMT2) Age 2

Dr~ver Driver Gender2 Empl S~

Vetncte Age I Vehicle Classes 1 hOperating Cos Elec~c Vehicle 1 Range I

F~rstvehicleeffects of vehlc]e characteristics on VMT and driver allocaUon (densesub-matrtx,equated w~thsecond vehicleeffects)

Effects of charactensracsof the first vehicle on VMT and driver allocatmnof the second (sparse sub-matr~ mmally specifiedto be nu1i)

Vehlcle Age 2 Vehlcle Classe~ 2 OperaUng Cost 2 Electric Vehlclc 2 Range 2

Effects of charactenstacsof the secondvehxcle on VMT anddriver allocataonof the first vehxcle(sparsesub-matrix, lmtmliyspecifiedto be null)

Second vehicle effects of vehmle charactcnstlcs on VMTanddriver allocauon (dense sub-matrLx, equated with first vehlcle effects)

No 16-20 Yr Olds No 16+ YrOlds No 1-5 Yr Olds Toted NoChildren Income> $60k Ave Age of Heads 3+ Vehicle HH

Effects ofhousehold characterstlcs onVMT and prmclpaldnver allocatmnof first vehicle (equated across vehlcles)

Effects of household charactensUcs on VMT andprincipal driver allocat~onof secondvehlcte (equated across vehlcles)

Thefinal specificationstep mvoives the error-term vafiance-covanance matrixW.If the unique(error)component of anyoneof the fourendogenous variablesof the first velucle~s correlatedwiththe umque component of the corresponchng varmble for the second vehicle,thenweshould findstatisticallys~grdficant coefficientsfor the~ matrix termsV5,1,V6,2,V7.3or ~ts,7 That~s, if what~s notexplained abouta varmble for one vetncle~s correlated withwhatis notexplained aboutthe samevarzable for the other vetucle, these sub-d~agonal parameters shouldbe foundto be sigmficantThefreely estimatedmain-d~agonal variances of the ~Pmath× produce R2 values" R2 =(s~,,- ~,,,)/s,,, (4) where s,,, is thesarnple variance of endogenous variable t.

4. Estimation Method Structuralequauons systemsof ~s typecanbe generallyesUmated usingmethods of moments (also known as varianceanalys~smethods) Thesemethods proceedbydefimng 78

AVehicleUseForecastingModel

T F Golobetal.

the sample variance-covariance matrix of the combinedset of endogenousand exogenous variables,, partJUonedwith the endogenous variables first:

where S..~y denotes the vanance°covanancematrix of the endogenousvariables, Syx denotes the covanance matrix between the endogenousand exogenousvariables, and $~ denotes the varlance-covariance matrix of the exogenousv&"iables. In the Two-Vehicle Model, there are eight endogenousvarmbles and 38 exogenousvarmbles, so $ ~s a (46 46) symmetricmamx.It can be easily shownusing matrix algebra that the corresponding varlance-covariance mmrixreplicated by model system (1), denoted

’ + V)[(l Y_~= (I- B)-iCrSxxr - B)-~]I-I

(7)

= CI- S)-rs

(8)

andEr~= $x~is taken as g~ven Thestructural equataon system ~sestamatcd using the varianceanalysis normal-theory maxlmum hkehhood method (Bollen, 1989)Thefitting function forstructural equations maxamum hkehhood method (ML)estimauon FML= LogIE(0)I - Logt$t + tr[SY..(0)] - (p + q) whereZ(0) represents Z (equataons6-8) implied by the vector of modelparameters, 0. fitting fu~actionis (-2/n) tamesthe log of the hkehh~dfunctton that S Is observedffZ(0) Js the true multivariate normal vanance-covariance matrix M_tmmlsatJonof FML 1S equwaler~tto maximasauon ofthe hkehhc~lfuncUonUnder the assumpttonsof multivariate normahtyand the modelbeing correctly specified, nFML lS Z2 Chstnbuted, providanga test of modelrejection and criteria for teslmg hierarchical modelsFuncUon(9) is mimm~sed m the LISREL8program using a modafied Fletcher-Powell algorithm (Joreskog and Sorbom, 1993a). Note that if we fit the unrestricted reduced form by regressing each endogenous varmbleon all the exogenousvariables, then Z(0) would exactly equal S and FMLwould be zero. Alternatively, wewouldget exactly the same result if weonly ~mposedenough restrictions on the underlyingparameters,0, to identify the system. All our modelsampose morerestricuons than are necessaryto identify the systemIt turns out that nFML xS simply the hkelilaood rat~o test statistac for the null hypothemsthat these over-identifying parameterrestriclaons are consistent with our observeddata. Becausefour of the eight endogenousvariables are dichotomous,the structural error terms corresponding to these chchotomousvarmbles will have unequal variances. Although fl~e coefficient esttmates will still be conmstent, the estimates of parameter standard errors and the overall modelZz go3dness-of-fit will be mconmstent (Bentler and Bonett, 1980). Consistent esumates can be generated using the asymptotically dislribuUon-free weighted least-squares method(Browne,1982, 1984), but tlus requires a much larger sampleraze. (Therule-of-thumbis that the sampleraze mustbe at least three times 79

January 1997

Journal of Transport Economms andPohcy

greater thanthe number of free entries in the asymptoticvariance-covariance matrixof the correlation mamx, that is, the fotu’th order moments; wlth 36 variables~tins requires approximately3,250 observations.) However,Amen~ya (I981) showsthat for moderate samplesizes, the weightedleast squares methodmayproduceworseresults than the unweightedmaximum likelihood esUmatesused here. In an), case, the coefficmnt estimatesare still consistent, and they havebeenshownto be fatrly robust (Boomsma, 1983). 5. ResLdts: Two-Vehicle Modal 5.1 Modelfit and structure TlueTwo-Vehicle Modelfits well accordingto the standardgoodness-of-fitcriteria. The likelihoodratio test statisUc associatedwith the nunhypothesisthat the esUmated model is consistent with the observedsamplevar tance-covariancematrix is 210.5 with 237 degreesof freedom,correspondingto aprobabihtyvalue of 0.892 Thus,the modelcannot be rejected at the p = 0 10 level Theestimatedg2 value for VMT of the first (newest) vehicleis 0.115,andthat of the second(oldest) velucleis 0.131.Asexpected,significant positive error-term covariartce.s werefoundbetweenthe VMTsof the two vetncles (tstaUstic = 11.3), betweenpnnc~paIdriver ages (t-staustic = 8 4) and betweenprincipal driver genders(t-statistic = 15.2). Theestamateddirect effects betweenendogenousvarmblesare hsted with their tstaUstlcs in Table6 This endogenous variable structure modelis basically maccordance with the hypothesisdepletedmTable4. All six of the witinn-velaicteendogenous variable effects postulatedfor eachvehiclewerefoundto be statisucallysignificant,andfive of the six effects are equalacross the twovel-dcles Thethree postulatedcross-vehicleeffects were also foundto be significant and symmemc. ThreeaddationaIcross-vehicleeffects werefoundto be necessaryfor goodmodelfit. Theseare identified bythe itahcisedcells in Table6: (t) if the driveroftbe first veincle as older, use of the secondvetucleis less thanotherwiseexpected(effect [55.2), an effect winchfurther links the influenceof driver age on VMT for the twoveincles, (2) If the driver of the secondveincle~s female,use of the first vehicle is greater than expected (effect ~1,7);and(3) if the driverof the first vetucleis employed, the driverof the second vehicleis morelikely to be female(effect ~7,4) The last two ef-fecIs al"e consistentw~th the travel and acUvitypatterns mmanyhouseholdsmwinchthere are worlongmaleheads and non-worldngfemale heads whobear the primary cl-dldcare and homemanagement responsib~htaes (Robinson, t977; Pas, 1984; Townsend,1987; Golob and McNaUy, I995). 5.2 Totaleffects Thetotal effects of the endogenous variables on the two-vehicleuse variables are hsted in Table7. For s~mplic~ty, only the total effects on the twoVMT variables, elements~l ~j and rlsj (] = 1 to 8) of matrixHdefinedin equationsystem(2), are shown.Resultsshow 8O

A Vehmle Use Forecastang Model

T F Golob et al

Table 6 Two- Vehwle Model Esttmated D~rect Effects Between EndogenousVariables (t-statmt~cs m parentheses) Influencing Variable Influenced vartable Ln (VMT,)

Ln (VM’rl)

Driver Age I

Driver Driver Gender I Empl S h

Ln (VMT2)

Driver Age 2

-0.0043 -0.131 0.179 (.-.4.as) (--6.s2) (454)

Drwer Gender 2 00~7

Driver Age1

--2.81 (-S.85)

Dr~ver Gender 1

-0 0051 (-7 78)

I~rlver

-0.0065 (-10.3)

Empl S h Ln (VM’T2)

-0.693 (-21.6) -0.103 (-11.3) --0.0043 (4.05)

--0.131 (--6.52)

0.179 (4.54)

-2.81 (-S.SS)

Age2 Gender 2

--6.140 (-1S.9)

--0 OO28 (-2 ]2)

I~IV~I"

Driver

Driver Empl S~

--9.693 (-21.6)

0 506 (3 72)

-0 0036 (-6 46) -0.6065 (-10.3)

--0.103 (-11.3)

Note Coefl~cmntsthat are restrmted to be equal for the two vehicles are shownmbold type, cross-vehwle effects are 1clentafmdby ltahc type

that drive,, agehas a slgrdficant effect on vebacleuse that is uniformfor the twoveNcles; if either driver is younger,both the first and secondvehicles are hkely to be used more. In contrast, the genderand employment status effects are consistent and reciprocal across the two velucles If the prinmpaldriver of either vehicle ~s female, that vehicle ~s driven less and the other vehicle is driven more, andff either driver ~s employed,that vehicle ~s driven more, and the other vel~cle is driven less, other ttungs being equal Thesereclprocal pzars of effects are generally strongest for the driver’s ownvebacle. Thetol al effects of the endogenous variables on the vehicle-use endogenousvariables ~e hsted Jn Table 8. These are the coefficients of the reduced-formequations for two of lhe eight endogenousvariables, whichare given by matrix equation (2). 81

January 1997

Journal of Transport Economics and Pohcy

Table 7 Two-Vehtcle Model Total Effects of the Other Endogenous Variableson the TwoUseVariables lnfluencedVa~ab~ EndogenousVariable

Ln (VMT2)

Ln (VMTI) Totaleffect

t-stat~sttc

Totaleffect

t-statistic

DroverAge~ Driver Gender 1 IDriver EmplSt

-4300358 --040013 0 20385

-3 36 -713 5 03

--000363 0 19975 --002565

-2 79 5.33 -310

Driver Age 2 Driver Gender 2 zDriver EmplSt

-000116 0 35896 -002098

-472 5 60 -3 26

-000362 -028460 0 15278

-291 --624 384

Thetotal effects of vehtcle age on VMTare strongest for the secondvebacle,but the effectsare consistentfor bothvebacles" the oldera vetucle~s, the less it is used,othertNngs beingequalAlso,the olderthe first vehicle~s, the less the othervei~c!eis usedas welt Theforecasting ~mphcanon of tlus is reduceduse of the householdfleet over nine ff no velucIe transacnonsoccur Ifhouseholdstructure, incomeand employment donot change, the reduction~n the fleet VMT will be furhheraccentuatedthroughthe negativetotal effect of 6r~ver age on use TNsimplies that householdsw~sbangto accommodate newtravel demandare morelikely to replace a vehicle w~tha newerone; while householdswith dechmng travel demandare morehkelyto hold on to their existing vetncles. Thetotal effects of operatingcost are ~mpreciseiyestimated, but the s~gnsof the wltban-vetu cIe effects are as expected.Also,a h~gheroperanngcost for the secondvehacle lmphesa shift of use fromthe secondvetacle to the first vehicle, but the coefficientsin the reduced-formequanonshave relanvely 1ughstandard errors. Theavaflabihtyof the SP use data y~eldeclreformationaboutthe effect of a lirmted rangevehicle on annualVMTthatwouldnot otherwisebe avmlablefromthe RPresponses alone The effects of the elecmc vehicle (EV)dummy variable on VMTare potentially importantfor pollution and energypohc~es.If e~ther of the first twoveNclesin multivehiclehouseholds is a future EV,the modelresults implythat the EVwilt be drivenless, otherttungs beingequal. Moreover, if the EVIS the newest(first) vet~cIemthe household, the secondvehiclewill be drivenmorethan otherwiseexpectedThus,this modelcaptures a shift in use fromEVsto convenUonal-fuel vehicles, somewhat n’nt~gatingthe emissions gmnsof the electricay versus conventionalfuels Themagmtude of ttus cross-vehicle substatutton effect can be assessed by using tl~s utfl~satmnmodelfor forecasting m comNnationwlth demograpt~c,vehicle transaction and vehicle-type choice models (Brownstone et al., 1994)o 82

A Vehicle Use Forecasting Model

T F Golob et al

Table 8 Two-Vehicle Model Total Effects of the ExogenousVariables on the TwoVehicle Use Variables Influenced Variable

Ln (VMT~)

Exogenous Variable

t-smusuc

Total effect

Typel Subcompact Type a- Compactcar Type1 Mut-slze ear TypeI Full°sine car Type~ Sports car Typ%Small truck Type~ Std truck Type 1 Muavan Type1 Std van Type 1 Small SUV Type~ Std SUV Operating Cosq Elecmc Vehacle I Range 1

-4) 01301 -0 23091 0 01675 0 08289 -4) 01416 -0 08733 0 03025 0 07210 0 08037 0 12686 0 02095 0 23267 0 07242 .0 00057 -0 25025 0 00153

-2 98 -5 33 3 33 2 41 -2 69 -1 57 3 97 8 35 8 37 2 44 2 05 4 33 6 74 -I 36 -2 51 3 30

-0 00095 .0 00043 0 01983 0 00710 .0 06500 -0 00872 0 01494 --43 03599 -0 04012 0 01668 -0 01046 .0 02145 -0 02177 .0 00058 0 09420 0

-2 48 .0 10 3 49 2 39 -1 57 -2 32 244 -6 52 -653 4 12 -201 -3 25 -2 80 -1 31 1 25

Vehicle Age 2 Type2 Mira car Type2 Sub:ompact Type2 Compact car Typ%Mid-size car Type2 Full-s~ze car Type2 Sports car Type2 Small ~uck Type2 Std lruck Typ% Mmwan Type2 Stcl van Type2 SmalI SUV Type2 Std SUV Operaung Cos h Electric Velucl% Range2

0 00443 0 00323 -4) 000290 0 00140 0 01319 0 .0 09826 -0 06246 -0 08374 0 -0 01879 -0 03829 -0 05326 0 00427 0 -0 00096

1 14 3 19 -0 59 2 26 2 46

--0 03372 -0 16784 0 02421 0 00436 -0 01784 0 0 02713 0 05819 0 05863 0 0 01490 0 12405 0 05463 -0 00860 -0.22579 0 00072

-9 34 -3 29 3 99 1 93 -3 21

No 16-20 Yr Olds No less thma 5 Yrs Old Total no Cl~fldren lncome> $60k Rettred HH Ave Age of Heads No Heads Working 3+ Vehicle HH

0 -0 0 0 -0 -0 0 0

4 80 -3 41 3.30 4 i3 -4 84 -4 24 5 5I

0 0 0 0 -0 -0 0 -0

1 55 195 3 88 3 10 -419 -5 88 5 20 -I 59

Total effect Vehicle A ge 1

Type I l~nl car

00956 00917 03060 11339 05129 00350 11234

-2 12 --6 43 -6 86 -2 01 -4 68 -5 30 0 626 -1 53

00822 04667 03662 08506 04452 00545 10618 04580

3 63 6 99 6 70 2 2 5 -t -1 0

02 17 84 19 20 81

83

January1997

Journalof Transport Economxcs andPohcy

The range vanable also captures a reduced VMTeffect for all hmited-range vetucles (potentially mclu&ng dechcated compressednatural gas vebacles in addation to EVs). For hmited-rangesecond vehicles, there ~s also a shift in use from the second vehicle to the first vehicle The number of household members between 16 and 20 years old has a positive influence on VMTof both the first and second vetncle. However,the numberofdrtvers in the householdhas negative effects on VMTofbothvetucles, possibly m&catinga shift of use towards third and fourth vehicles in the household. The numberof children ] to 5 years old positively mfiuencesVMT,mostly of the second vehicle, while the total number of children posltavely influences VMTof both the first and second vehicles. The income effect has the expectedsign, but, as rathe case of averageage of the heads, the effects are ~mprec~selyestimated. Finally, as expected, the presence of three or more household vehtctes reduces VMTof both the first and second vehicles. 5.3 Scenarios of changes in VMTimplied by the total effects The endogenousvariables are expressed m terms of the natural logarithms of VMT,so the natural exponent of each 1educed-formequaUoncoefficlent represen~ a mult~phcattve factor apphed to the endogenous VMTvariable m question That Is, exp(Sfqj) and exp(6fcsj) express mulUphersof VMTfor vehicles 1 and 2, respectively, where the T matrix of total exogenouseffects is defined m equauonsystem (3), and 5j is the level change in thejth exogenousvariable Someselected VMTmultipher effects are hsted in Tables 9 and 10. Each scenario depicted m Tables 9 and 10 assumes that all factors not defined m the scenmoremmnconstant ] n the case of vehicle replacements, this mc!udesthe vehicle type class and operatmg cost However, to provide reahsm, whenveNcles are assumed to be replaced with ~dent~calvelucles w~thdifferent rangesor fuels, ~t ~s assumedthat the replacement vehicle ~s one ;year newer. Of alI the modelpred~ctaonscomputedm Table 9, the most substantial effects are those attributable to vehicle range and the electric vehicle (EV)designator In the case of the first (newer) vetucle, a reduction m range of 150 wales reduces VMT by a factor of 0 81, but there ~s no effect on VMTof the second vetucle In the case of the second vehicle, a s~milar reduction in range of 150 miles reduces VMTbya factor of only 0 93, but firstvehicle VMT~spredictedtoincreasebyafactorof1.15 The weaker second-vehicle range effect ~s pamally caused by an offsemng stronger second-vehicle age effect Comb~mng reducedrange w~ththe EVeffect, the modelpre&ctsthat ffthe first vehicle is an EVw~th I00 miles range, W¢Twill reduce by a factor of 0.58, and second*vebacle VMTwill mereaseby a factor of 1.10. If the second vetacle ~s an EVwith 100 males range, VMT on this vehicie will reduceby a factor of 0 70, but there will be moreof a shift to use of the first vehicle, w~thfirst-veNcle VMTincreasing by a factor of 1.24 Given these results, weregard the SP data as prowgunguseful ~mprovementsto the quahtyof our VMT forecasts for future alternata ve-fuel vel~cles, andespecially for electric vetdctes. However,there were also someposmblehm~tatmnsinherent in the data from the 84

A Vehicle Use Forecasting Model

T F Gotob et al

Table 9 Two-Vehicle Model Exponent~ated Total Effects on VMT of Selected Changes in Vehtcle Charactemstics Exogenous Change

Mulnpher Effect on Vehicle I VMT

Multtpher Effect on Vehicle 2 VMT

Vehicle AgeI (m years) vehlcle ages one year replace with same type vehicle 1 year newer replace with sametype vehicle 5 years newer

098 I 01 1 07

099 1 00 1 01

Electric Vehlcl%(EV1), Range~(m miles) replac~ 300 mite vehicle with 200 m~le non-EV,i year newer replace 300 mile vehicle with 150 mile non-EV,1 year newer replace 300 mile vehicle with 100 mile EV1 year newer replace 300 mile vehicle with 75 rmle EV, 1 year newer

0.87 081 0 58 0 56

1 00 100 1 10 1 10

Vehlcle Age2 (in years) vehMeages one year replace vnth sametype vehicle 1 year newer replace with sametype vehicle 5 years newer

1 O0 0 99 098

0.97 I 03 1 I8

EtecmcVehicle2 (EV2), Range2 (in miles) replace: 300 mile vehicle v, ath 200 mile non-EV,I year newer replace, 300 m~evehicle with 150 male non-EV,1 year newer replace 300 mile vehicle with 100 mile EV, 1 year newer replace 300 mile vehlcte with 75 male EV, I year newer

1 10 1 15 1 21 1 24

0 96 0 93 0 72 0 70

SPexperiment,wtachwasmainlyfocusedon the issue of vehicle choice. Respondents wereapparently ableto reflect the generaleffect of limited-range electric vehiclesonuse patternsthroughboththe allocationof the vehicle andsomeadjustments to VMT. More subtleeffects onutihsaUon atmbutable to otherfactors, suchas hmitedfuel availabihty (for example,away-from-home rechargingfor electric vehicles, or smallernumbers of stattonsfl~r naturalgasvetucles),or chfferences in fuel operating costs, may" nothavebeen as easily capturedusingthis experimental formatTl’as couldhaveresultedin an overestimaUo a of rangeandEVeffects, andanunderestimation of the effects of, for example, improved operatingcosts In fact, the coefficientonoperatingcost mourcurrentmodel is modest,implyingthat the rangeandEVscenarioresults wouldnot be substantially changed byimposingaccompanying reahsticchangesmoperatingcosts. Invalidatingthis result wouldrequireaddtUonal research. 85

January !997

Journal of Transport Economiesand Pohey

Table 10 Two-VehzcleModel ExponenttatedTotal Effects on VMT of Selected ChangesmHousehold Characteristtcs Exogenous Change

MuIt~pherEffecz on Vehicle ) VMT

Multtpher Effect on Vehicle 2 VMT

Numberof i 6-20 year-olds chzJdpasses 16th birthday, no other changes

i 0!

1 02

Children bn’th of chtld, no other changes 19 year-old chxld movesout of home,no other changes

1 02 0 96

1 09 0 96

Income and EmploymentStatus incomer~ses above $60k, no other changes +1 head workingand income rises above $60k 1 head workingretares, income drops below $60k 2 headsretare at sametame, incomestays above$60k

1 12 1.25 0 85 0 90

109 1 21 0 88 092

Ownershipof a Third Vehlcle householdadds third vehlcle household&sposesof third vehxcle

1 O0 1 O0

096 t 05

In contrastto the rangeeffects, the vehtcIeageingeffects are weaker for the first (newer)vehiclethanfor the second(older)vehicle. If the newestvetuclemthe household as replaced witha vebaclethatisident~calantype,operating cost, rangeandfuel, butis five yearsnewer,the modelpredictsthat VMT for that vebaclewill increaseby approximately 7 per cent, withveryhtt|e effect on VMT of the secondvehicle However, ff the second vehicleis replacedwatha vehiclethatts ~denticalin type, operating cost, rangeandfuel, but is five yearsnewer,the modelpre&ctsthat VMT for that vehicle will increaseby approximately 18 per cent, andVMT of the first vehicle will shghtlydecrease(by about 2 percent). Predictedchangesin VMT associatedw~ththe scenariosrelatedto household characteristics are shownin Table10. Theinfluencesrelatedto the number of childrenmthe householdare smaller m magnitudethan expected, but the use behawour appears consistentwathconvenUonat notionsof first andsecondvehacles. Forexample,a new child placesmorepressureonthe useof the secondvehicle, the onethat ~s less likely to be used for commuUng 86

A Vehicle Use Forecasting Model

T F Golob et aL

In contrast, the predicted effects of incomeand the numberof householdheads working are relaavely strong, especially in combmaUon The joint impact of an ad&taonaIworker and a hJ gher householdincomeis a prechctedincrease of 25 per cent in use for the first householdvebacle, approximatelyhalf of whichas attributable to an incomeeffect; use of the second vehicle increases by a shghtly lower 21 per cent. If one workinghead retires andincomedrops belowthe bagh-incomecut-off, the modelpredicts that VMTofthe first and second vetncles will be reducedby the factors 0.85 and 0.88, respectively. If both household heads retire, the prechcted change in VMT is only 10 per cent for the first vehicle, provided that household incomeremzansabove (or below) the hagh-incomecutoff. Fm;tlly, the presence of a third householdvehicle has a modestinfluence on VMT of the second veI’ucle

6. Results:

The Single-Vehicle

Model

6.1 Moclelfit andfinal structure The structure of the Smgle-VebacleModelis also basically in accordance with the structured hypotheses Tlus model fits extremely well accorchng to all goodness-of-fit cratena, the Z2 statastic being 41o82 with 49 degrees of freedom, corresponding to a probablhty value of 0.757. The model cannot be rejected at the p = 0 10 level. The estamated R2 value for VMTis0 173. No sigmficant error-term covanances were found betweenany pairs of the four endogenousvariables. The endogenousvariable structure determined to be optamal m the Smgle-Vetucle ModelJ s~ similar to the w~ttnn-vebaclestructure found for the Two-VetucleModel(the structure depicted m the upper-left-hand and lower-nght-handquadrants of the 13 matrix shownm Table 6) The only difference ~s that an adchUonal d~rect effect was found betweenprincipal driver genderand age: female pnnc~paldrivers of a vehicle ma singlevehicle householdare youngerthan otherwise expected, other things being equal. 6.2 Total effects The total effects of the endogenousprmc~pal-dr~vervariables on VMTfor the SmgleVebacle Modelare shownm Table 11. As m the multi-velucle case, VMTis higher for younger~ male, employed drivers, but the gender and employmentstatus effects are relatively weaker for smgle-ve~cle households. Finally, the total exogenouseffects on VMT for the Singte-Vebacle Modelare shown m Table 12 Once again, these effects are similar to those found for multi-vehicle households, wath some excepUonsPatterns of use are consastent for eaght types of vel~cles, but sports cars, nnmvans,standard sport utihty vetucles, and full-slze cars exhibit &fferent use patterns in single-vebacle, as comparedwath multa-vetucle, households. Regarchngalternative fuel vehicles, the negaUveEVeffect and the positive effect of range on VMTare consistent between smgle-velucle and multi-vehicle households. 87

January 1997

Jourual of Transport Economiesand Pohcy

Table 11 Single- Vehicle Model Total Effects of the Other EndogenousVariables on Vehtcle Use Total Effect on Ln(VMT)

Driver Age Driver Gender Driver EmploymentStatus

Total Effect

t-statistic

-0 00396 -4? 08037 0 11671

-1 00 -2 30 2 46

Table 12 Smgle- Vehicle Model Total Effects of the ExogenousVartables on Vehicle Use Vanable

Total Effect on Ln(VMT) TataI Effect

t-$lallsttc

Vehicle Age Type Mini car Type Subcompact Type" Compact car Type Mld-slze car Type Full-size car Type Sports car Type Small Truck Type Standard Truck Type Mmlvan Type Standard Van Type Small SUV Type Standard SUV Operatang Cost Electr~c Vehicle Range

-0 01574 .0 27808 0 09798 0 12140 -4) 00259 0 00639 -0 00706 0 26612 0 52883 045711 0 34705 0 31306 0 00000 .0 01223 -0 15136 0 00138

-4 I4 -5 97 1 82 2 26 -0 91 1 38 -1 46 264 3 26 3 86 164 2 98 0 00 -1 31 -1 58 362

No 16-20 Yr Olds No 16+ Yr Old s No less than 5 Yrs Old Total no Children Income < $31k Income > $60k Couple HH Rettred HI{ Average Age of Heads No Heads Working

004246 003455 012448 -011225 -019112 010970 000833 -002178 --001071 005588

258 0 93 231 -3 78 -5 19 1 90 1 69 -1 67 -846 221

88

AVehlcteUseForecasting Model

T F Golobe~al

7. A Forecasting Method That Preserves Heterogeneity This m~Jelis being applied in a dynarmcmicrosimulationforecastmgsystem(Bunchet alo, 1996), wherea sociodemograptuc translUonmodeland vehicle transactions models are usedGoforecast changesmhouseholds’sociodemograpIuc structure and composition of the ve:tucle fleets Theuse modelis thenexercisedto forecast VMT for boththe before andafter situations for the householdThecalculatedchangein forecasts is then applied as a percentagechangeto the actual base level of use for the householdin the before situation. ]Even if the dynamicsociodemograpMc modelpredicts no change in household charactexisracs(householdcomposition,employment status or income),and the velucle transactumsmodelpredicts no vehicle transacttons for the householdfor the period in question, the present use modelwill in general predict changesin VMT Thls will be attributable to ageingof the household heads,ageingof the vehicles, andpossiblechanges mthe age categories of householdmembers, particularly cIuldren. Themost effective application of the use modelsma mtcrosimulaUon forecasting systemuses a"plvot"approach(describedbelow),rather than the traditional approach using the expectedvalue froma hnear modelTheplvot approachpreservesheterogeneity across householdsHeterogeneityresulting fromspatial and lifestyle factors is to be expected,somehouseholdsdrive moremiles per year than the modelwouldpredict, while others dr3 ve fewermilesper year than the modelwouldpredict. Byusing the residual difference betweenobserved and predicted VMT for each householdvehicle mthe modelestimation data set, wecan develophousehold/vehiclespecific raultlpliers that can be usedduringforecastmg.Suchmulttpherstake the form:

=

( l O)

wheret denotesthe tth vehicle, and 0 denotesthe "base year" of the forecasLwhich correspondsto the original esUmaUon sampleAnewpredicted VMTIs computedfor each forecasting period, and then "pivoted"by using the multlpherThesemulUphers capture effects fromheterogeneitythat m~gh~ be wassmg in the model,and preservethemin the forecasts. Onedifficulty with this approachis that vebactetransacUons will occurduring the course;of a forecast. Appropriate rules havebeendeveloped for reassigningmultlphers after a velucletransactton,andthese are usedin our imcrosimulataon forecastingsystem.

8. Conclusions and Directions For Further Research The struclural elegance of the models and their stalastlcal fit to the sample data prowde support to our modelling approach. Moreover, the correspondence between pure RP results (C,olob et al., 1995) and the present SP-RP results ~s encouraging. Weare also encouraged by the advantages associated with a jointly estimated RP-SP model that simultaneously captures the endogenous effects of vetucle reallocatmn along with perceived changes in utahsation associated w~th elecmc vePacle character~stms. These

89

January 1997

Journal of Transport Economlcs andPohcy

effects are not available from RP data alone The approach automatically produces estimates that are consistently scaled, and yields reduced-form equaUonsthat are convenient for forecasting utihsation of alterna~ve-fuel vehicles. However,the SP questions m the 1993 household survey from winch these data were extracted are pnmariIyfocused on the ~ssue of vehxcle choice, and are potentmlly limited in capturing the full range of effects on use attributable to fuel availability, peakandoffpeak recharge costs for EVs, cargo capacity, performance, and other vehicle and fuelsystem characteristics that might chstmgtush future vetucles These ~ssues are being pursued through a second householdsurvey, conductedin 1994, that contmneda ch fferent vebacle-use SP protocol. Whenthe 1994 data are avmlable, the robusmessof the present modelresults can be assessed~ and hopefully the model can be extended Potential selectivity bias can be accountedfor m ttns use modelby tinlang the model to a chscrete type-choice model(see, for example,Brownstoneet al., 1996). and adding into the structural equation system a correction term variable involwnga transformatmn of the household’s predicted vehtcle-type choice probabihties (McFaddenet al., 1985; Mannenng and Winston, 1985, Train, I986; Hensher et al., 1992). It ~s doubtful if such a correction term wouldhave a pronouncedeffect on the results. The known biases m the normal-theory maximumhkehhood estimation method apphed to dichotomousendogenousvariables are concentrated on coefficient standard errors and overall goodness-of-fit cnterm The fit of the model~s not in question, and hypothesis tesUng ~s subordinate to forecasUngcapability m ttus research However,~t would be possible to use unbiased generally weighted least-squares esUmat~on(Browne, 1982, 1984), as implemented m LISREL8with PRE-LIS2(Jbreskog and Sorbom, 1993b), w~th a slgmficanfly increased sample s~ze

References Amem~yaT (1981) "Quahtauve Response Models A Survey" Journal ofEconomtc Ltterature. vol 19, pp 481-536 Bentler, P Mand D G Bonett (I980) *’Slgmficance Tests and Goodnessof Fit m the Analys~s of Covanance Structures" PsychologtcalBulletm, vol 88, pp 558-606 Berkovec, J (1985) "ForecasungAutomoblleDemandUsmgD1saggregateCholceModels" Transportatmn Research, vol 19B, pp 315-29 BoIlen. K A (1989) StructuralEquatwnswlthLatent Variables NcwYork Wiley Boomsma, A (1983) OntheRobustnessqfLISREL(MaxtmumLtkefihoodEstmzatton)agamstSmallSample StzesandNon-normaltty Amsterdam SocmmetncResearchFoundauon Browne, M W(1982) "Covanance Structures" In D M Hawkins (ed) Topics m Multwanate Anatys~s, pp 72-141 Cambridge Cambridge Umverslty Press Browne, M W(1984) "Asymptouc Distribution Free Methods m Analysls of Covarmnce Structures" Bnttsh Journal of Mathemattcaland Stattsttcal Psychology, vol 37, pp 62-83 Brownstone, D, D S Bunch and T F Golob (i994) "A Demand Forecasting System for Clean-Fuel Vehmles"Presented atthe OECDtntemat~onal Conference TowardsClean Tran.~port Fuel Effictent and Clean Motor Vehtcles, Mexico City, 28-30 March

9O

A Vehicle Use Forecasting Model

T F Golob et at

Brownstone, D, D S Bunch, T F Golob and WRen(1996) "A Personal Vehicle Transaclaons Choice Mode[ for Use m Forecasting Demandfor Future Alternative-Fuel Vehicles" ResearchmTransporiatton Econc~mcs,vol 4, pp 87-129 Bunch, D S, M BradIey, T F Golob, R Katamura and G P Occhmzzo (1993) "Demandfor Clean-Fuel PersonalVehlclesmCalfforma ADlscrete-CholceStatedPrefereneeStudy" TransportattonResearch, vo127A,pp 237-53 Bunch, D S, D Brownstone and T F Golob (1996) "A DynamacForecasting System for Vehicle Markets with Clean-Fuel Vehacles" Proceedings of the Seventh WormConference on Transport Research, Elseva,= Golob. T F (1990) "The Dynamics of Travel Tune Expendatures and Car Ownershxp Decaslons" Transportation Research, vo124A, pp 443-63 GoIob, T F.S KamandWRen(1995) "AStructuralModelofVehwleUsemMultl-VehlcleHouseholds" Paper presented at the Annual Meetang of the Transportation Research Board, WashingtonDC, 23-28 January’ Gelob, T F, R. Katamura~MBradley and D S Bunch(1993) "Pred:ctmg the Market Penetrat:on of Electric and Ct~an-Fuel Vehlcles" The Science of the TotalEnwronment,vol 134, pp 371-81 Golob, T F and MG MeNally (1995) "A Model of Household InteracUons m Acuvaty Panaeapatlon and the Deraved Demandfor Travel" Paper presented at the Conference on Acttvtty Based Approaches Acttvtr¢ Scheduhngand the Analysts of Acttvtty Patterns at EmdhovenUmverslty of Technology, The Netherlands, 25-28 May Hensher, D A. (1985) "An Econometric Model of Vehicle Use m the Household Sector" Transportation Research~vol 1913, pp 303-14 Hensher, D A and V Le Plastner (1985) ’q’owards a DynamicDiscrete-Choice Model of Household Automob:le Fleet Saze and Composltton" Transportat~onResearch, vo! 19B, pp 481-95 Henshel, D A and T Manefield (t982) "A Structured-Loglt Model of Automobile Acqulsatlon and Type Chome"Proceedtngs of the 7th Australian Transport Research Forum, Hobart, Tasmama,pp 625-54 Hensher, D A, and N C Snuth (1986) "A Structural Model of the Use of Automobiles by Households" TransportRev~ews, vol 6, pp 87-111 Hensher, 13 A, N C Smith, F WMflthorpe and P O Bernard (1992) D~mens~onsof Automobde Use A Longttudmal Stud) of Automobtle Ownershtp and Use AmsterdamNorth Holland t-locherman, I, J Prashker and MBen-Akwa(1983) "Esumatxonand Use of DynamacTransactaon Models of Automobile Ownershlp"Transportatton Research Record, no 944, pp 1.34-41 Joreskog, K G and D Sorbom (1993a) LISREL 8 User sReference Grade Ch:cago Scmntafic Software Jore~kog, K G and D Sorbom(I993b) PRELIS2 User’s Reference Gutde Chicago, Scientific Software Kazum, C (1994) "HazardModelsofChangmgHouseholdDemographacs" WorkangPaperUCI-ITS-WP94-10, ]~nsatute of Transportat:on Stu&es, Umversxtyof Caldomm at Irvme Kazmu~C~1995) "AM~crosnnulataonModelforEvaluatmgtheEnvn:onmentalImpactofAlternatlve-Fuel Vehlcks" Unpubhshed PhDchssertatlon, DepartrnentofEcouomles, Umvers:ty ofCalffomm, Irvme Kazmu, C and D Brownstone (1995) "Competing Rusk Hazard Models for Demographac Transmons" Paper presented at the Econometric Soclety WorldCongress, Tokyo, 27-29 August Lave, C /k and K Tram (1979) "A Dasaggregate Model of Auto-Type Choace" TransportattonResearch, vol 13A, pp 1-9 Mannenng,F (1983) "An Econometric Analysis of Vehacle Use an Multa-Vehacle Households" TransportatJont~esearch, vol 17A, pp 183-89 Mannenng,F and C Winston (1985) "Dynamic Ernpmcal Analysas of Household Vehacle Ownersh:p and UtlhzatLon’" RandJournal of Economics,vol 16, pp 215-36 Mansla~C F andL Sherman(1980) "AnEmpuucalAnalys:sofHouseholdChoaceAmongMotorVehaeles" Transportatton Research, vol 14A, pp 349-66 McFadden,D L, C Winston and A Borsch-Supan (1985) "Joint Estm:ataon of Freight Transportataon Dec~saoas Under Non-random Samphng" In A F Daugherty (ed) Analyttcal Studws tn Transport Economics, pp 137-60 Cambridge Cambridge Umvers:ty Press

91

January 1997

Journal of Transport Economicsand Pohcy

McCarthy,P S and R Tay (1989) "Consumer Valuataonof NewCar Attributes an EconometricAnalysis of the Demand for Domestaeand Japanese and WesternEuropeanImports" TransportationResearch, vol 23A,pp367-75 Pas, E I (1984) "The Effects of Selected Soexodemographlc Charaetermtlcson Daily Travel-Aclavlty Behavior" Environment andPlanmngA,vol 16, pp 571-81 Robmson,J P. (1977) HowAmencansUseT~me. ASoc~al-’PsychologtcalAna#,s~ofEverydayBehawor NewYork Praeger Townsend, T. A (1987) "The Effects of HouseholdCharacterastlcs on the Multx-dayTimeAUocataons and Travel-Act~vltyPatterns of Householdsand Their Members" Unpubhshed PhDdlssertaUon, NorthwestemUmverslty,Evanston, IL Tram,K. (1986) QuahgattveChoiceAnalysts" Theory, Econometrics,and an Apphcattonto Automobde DemandCambridge, MAMIT Press van Wmsen,L J and T F GoIob (1992) "A DynanucModelof Car Fuel Type Choice and Moblhty" TransportanonResearch, vo126B,pp 77-96

Date of recetpt

92

of final manuscrtpt: December 1995