Citation
Spatial Models for Analyzing the Effects of Land Use Patterns on Automobile Ownership and Usage

Material Information

Title:
Spatial Models for Analyzing the Effects of Land Use Patterns on Automobile Ownership and Usage
Creator:
Nowrouzian, Roosbeh
Place of Publication:
[Gainesville, Fla.]
Florida
Publisher:
University of Florida
Publication Date:
Language:
english
Physical Description:
1 online resource (119 p.)

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Civil Engineering
Civil and Coastal Engineering
Committee Chair:
SRINIVASAN,SIVARAMAKRISHNAN
Committee Co-Chair:
ELEFTERIADOU,AGELIKI
Committee Members:
WASHBURN,SCOTT STUART
YIN,YAFENG
KHARE,KSHITIJ
ZWICK,PAUL D
Graduation Date:
8/9/2014

Subjects

Subjects / Keywords:
Bandwidth ( jstor )
Land use ( jstor )
Predetermined motion time systems ( jstor )
Regression analysis ( jstor )
Socioeconomics ( jstor )
Spatial models ( jstor )
Statistical models ( jstor )
Transportation ( jstor )
Travel ( jstor )
Urban design ( jstor )
Civil and Coastal Engineering -- Dissertations, Academic -- UF
car-ownership -- geographically-weighted-regression -- mixed-geographically-weighted-poisson-regression -- person-mile-traveled -- quasi-geographically-weighted-poisson-regression -- vehicle-distance-traveled -- vehicle-time-traveled
Genre:
bibliography ( marcgt )
theses ( marcgt )
government publication (state, provincial, terriorial, dependent) ( marcgt )
born-digital ( sobekcm )
Electronic Thesis or Dissertation
Civil Engineering thesis, Ph.D.

Notes

Abstract:
Estimating location-sensitive or spatial models is gaining a lot of interest in transportation literature. In the conventional global (or aspatial) models, the marginal effect of every factor is assumed to be fixed over space and inherent similarities in behaviors of neighboring entities are largely ignored. However, the sensitivities to the same explanatory factor can change over space due to a variety of reasons such as differences in attitudes, preferences, and contextual effects. The need for spatial models is particularly relevant in the context of modeling the interactions between land-use patterns and travel behavior. This study contributes to the literature by developing a Geographically Weighted Regression (GWR), mixed-GWR, and Quasi Geographically-Weighted Poisson models to improve our understanding of impact of land use on person-miles traveled, vehicle-distance-traveled (VDT), vehicle-time-traveled (VTT), and car ownership at a disaggregate level. This research also demonstrates the benefits of estimated spatial models over simpler methods. The study area covers South East Florida (SE) metropolitan area which includes three counties: Miami-Dade, Broward, and Palm Beach County. In this dissertation both micro level, and meso scale urban design indicators were created. The National Household Travel Surveys of 2008/2009 and the 1999 South East Florida Household Travel Surveys constituted the primary source of travel and socio-economic data. Various micro level built environment measures were created to examine their impacts on PMT, VDT, VTT and car ownership. The 2010 statewide NAVTEQ road network file was used to determine the shortest-distance travel paths between each origin-destination pair. Overall this research contributes on understanding and quantifying relationship between PMT, VDT, VTT, car ownership and land use patterns using spatial non-stationary process in estimating impact of explanatory variables. ( en )
General Note:
In the series University of Florida Digital Collections.
General Note:
Includes vita.
Bibliography:
Includes bibliographical references.
Source of Description:
Description based on online resource; title from PDF title page.
Source of Description:
This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Thesis:
Thesis (Ph.D.)--University of Florida, 2014.
Local:
Adviser: SRINIVASAN,SIVARAMAKRISHNAN.
Local:
Co-adviser: ELEFTERIADOU,AGELIKI.
Statement of Responsibility:
by Roosbeh Nowrouzian.

Record Information

Source Institution:
UFRGP
Rights Management:
Copyright Nowrouzian, Roosbeh. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Resource Identifier:
969977002 ( OCLC )
Classification:
LD1780 2014 ( lcc )

Downloads

This item has the following downloads:


Full Text

PAGE 1

SPATIALMODELSFORANALYZINGTHEEFFECTSOFLANDUSEPATTERNSONAUTOMOBILEOWNERSHIPANDUSAGEByROOSBEHNOWROUZIANADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPHYUNIVERSITYOFFLORIDA2014

PAGE 2

c2014RoosbehNowrouzian 2

PAGE 3

Tomyparents,FarhadandMahnaz,andValehfortheirunconditionallove,support,andencouragement 3

PAGE 4

ACKNOWLEDGMENTS Iwouldliketothankmyadvisor,Dr.SivaSrinivasan,forhisvaluablesupport,guidance,andencouragement,throughoutthecourseofmystudiesatUniversityofFlorida.HecontinuallyconveyedaspiritofadventureduringmyresearchandmadethisjourneyanexcitingpathtowardbecomingtheresearcherIamtoday.Iamdelightedforhavingtheopportunitytoworkwithhim.IwouldalsoliketothankDr.LilyElefteriadou,Dr.ScottWashburn,Dr.YafengYin,Dr.PaulZwick,andDr.KshitijKhareforservingonmyPh.D.committeeandfortheirinsightfulcommentsandsuggestions.DuringmyyearsatUniversityofFlorida,Ihadthepleasureofmeetingmanyamazinggraduatestudentcolleagueswhobecamegreatfriendsofmine.Iwouldliketothankmyofcemates,BarbaraMartin,NagendraDhakar,EvangelosMintsis,FangHe,IreneSoria,andCoreyHillforallthehelpandsupporttheyhavegiventome.IwouldalsoliketothankmydearfriendsMahmoodZangui,NimaShirmohammadi,SeckinOzkul,XiaoyuZhu,MaLu,KwangkyuLim,MiguelLugo,XuRuoying,MdMamun,DimitraMichalaka,DonWatson,andZhuofeiLiforcreatingsomanywonderfulmemoriesforme.Lastbutcertainlynotleast,mydeepestgratitudegoestomydearestparents,FarhadandMahnaz,andmylovelysister,Valeh,whotaughtmehowtofollowmydreamsandencouragedmeineverydecisionImadeinmylife.FinallyIwanttothankmygirlfriend,Bita,forherloveandsupportandallthelaughtershegavemethroughouttheseyears. 4

PAGE 5

TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 7 LISTOFFIGURES ..................................... 8 ABSTRACT ......................................... 9 CHAPTER 1INTRODUCTION ................................... 11 1.1BackgroundandMotivation .......................... 11 1.2DissertationObjectivesandApproach .................... 13 1.3StructureoftheDissertation .......................... 14 2LITERATUREREVIEW ............................... 15 2.1Overview .................................... 15 2.2SpatialRegressionandApplicationinTransportationEngineering ..... 16 2.2.1SpatialHeterogeneity ......................... 17 2.2.1.1Geographicallyweightedregression ............ 17 2.2.1.2Mixedgeographicallyweightedregression ........ 23 2.2.1.3Geographicallyweightedquasi-Poissonregression .... 25 2.2.2SpatialClustering ............................ 28 2.2.2.1GlobalmoransI ....................... 29 2.2.2.2Localindicatorofspatialassociation ............ 30 2.2.3SpatialDependency .......................... 31 2.2.3.1Endogenousinteractionamongdependentvariables ... 31 2.2.3.2Interactioneffectsamongtheerrorterms ......... 35 2.3ApplicationContext ............................... 37 2.3.1DistanceTraveledModeling ...................... 37 2.3.2CarOwnershipModeling ........................ 38 2.4Summary .................................... 41 3PERSONMILESTRAVELEDMODELING ..................... 42 3.1Overview .................................... 42 3.2DataDescription ................................ 42 3.3ModelStructure ................................ 47 3.4EmpiricalResults ................................ 49 3.4.1StatisticalComparisons ........................ 49 3.4.2AverageMarginalEffects ........................ 50 3.4.3SpatialVariationintheMarginalEffects ............... 53 3.4.4VarianceDecompositionandImplications .............. 57 5

PAGE 6

3.5Summary .................................... 58 4VEHICLETIMEANDDISTANCETRAVELEDMODELING ............ 61 4.1Overview .................................... 61 4.2DataDescription ................................ 62 4.2.1MicroLevelUrbanDesignPattern ................... 67 4.2.1.1Density ............................ 67 4.2.1.2Design ............................ 69 4.2.1.3Diversity ........................... 69 4.2.1.4Transitaccessibility ..................... 70 4.2.2MesoLevelUrbanDesignPattern ................... 70 4.3ModelStructure ................................ 70 4.4EmpiricalResults ................................ 70 4.4.1AverageMarginalEffects ........................ 75 4.4.2StatisticalComparisons ........................ 83 4.4.3Summary ................................ 84 5CAROWNERSHIPMODELING .......................... 87 5.1Overview .................................... 87 5.2DataAssembly ................................. 87 5.3ModelStructure ................................ 90 5.4EmpiricalResults ................................ 94 5.4.1OverallFitMeasuresandSpatialAutoCorrelations ......... 94 5.4.2AverageMarginalEffects ........................ 95 5.4.3SpatialVariationintheMarginalEffects ............... 98 5.5Summary .................................... 99 6SUMMARYANDCONCLUSIONS ......................... 104 6.1Overview .................................... 104 6.2Contributions .................................. 105 6.3FutureWork ................................... 107 REFERENCES ....................................... 109 BIOGRAPHICALSKETCH ................................ 119 6

PAGE 7

LISTOFTABLES Table page 3-1SummaryofresidentialLand-Usecharacteristics ................. 44 3-2Summaryofpersonlevelandhouseholdlevelcharacteristics .......... 45 3-3Summaryofsocioeconomiccharacteristics .................... 46 3-4Summaryofempiricalmodelresults(meaneffects) ................ 51 3-5SummaryofGWRmodelresults .......................... 54 3-6Decompositionofvariance ............................. 59 4-1Socioeconomicexplanatoryanalysis ........................ 62 4-2Meanurbandesigncharacteristics ......................... 71 4-3Meanurbandensityexplanatoryvariables ..................... 71 4-4GWRresultsforVDT ................................. 73 4-5GWRresultsforVTT ................................. 74 4-6VTTMGWRresults ................................. 76 4-7VDTMGWRresults ................................. 77 5-1Socioeconomicexplanatoryanalysis ........................ 89 5-2Meansocioeconomicexplanatoryanalysis .................... 89 5-3Landusevariablesummary ............................. 91 5-4Comparisonsofmodelestimationresults ..................... 96 7

PAGE 8

LISTOFFIGURES Figure page 3-1PMTproleofSouthEast(SE)Floridaresidents ................. 47 3-2Spatialproleofmarginalimpactsforstatistically-signicanteffects ....... 56 4-1ViolinplotsofVDT(inmile)andVTT(inminute) .................. 64 4-2BagplotofVDTandVTT .............................. 65 4-3HouseholddailyVTTandVDTatdifferentdays .................. 66 4-4SpatialproleoftheimpactofEntropyonVDT .................. 75 4-5SpatialproleoftheimpactofEntropyonVTT .................. 78 4-6SpatialproleofdistancetonearestregionalactivitycenteronVDT ...... 79 4-7SpatialproleoffractionofdevelopedareathatisinstitutionalonVTT ..... 80 4-8SpatialproleoffractionofdevelopedareathatisinstitutionalonVDT ..... 81 4-9HouseholddailyVDTatdifferentresidentialdensity(tractlevel)percentiles .. 85 4-10HouseholddailyVTTatdifferentresidentialdensity(tractlevel)percentiles .. 86 5-1Spatialcorrelogramsofmodelresiduals ...................... 97 5-2Densityplotofcoefcientsonlandusevariables ................. 100 5-3Densityplotsofcoefcientsonsocioeconomicvariables ............. 101 5-4Spatialproleofmarginalimpactsoflandusevariables ............. 102 5-5Spatialproleofmarginalimpactsofsocioeconomicvariables .......... 103 8

PAGE 9

AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosophySPATIALMODELSFORANALYZINGTHEEFFECTSOFLANDUSEPATTERNSONAUTOMOBILEOWNERSHIPANDUSAGEByRoosbehNowrouzianAugust2014Chair:SivaSrinivasanMajor:CivilandCoastalEngineeringEstimatinglocation-sensitiveorspatialmodelsisgainingalotofinterestintransportationliterature.Intheconventionalglobal(oraspatial)models,themarginaleffectofeveryfactorisassumedtobexedoverspaceandinherentsimilaritiesinbehaviorsofneighboringentitiesarelargelyignored.However,thesensitivitiestothesameexplanatoryfactorcanchangeoverspaceduetoavarietyofreasonssuchasdifferencesinattitudes,preferences,andcontextualeffects.Theneedforspatialmodelsisparticularlyrelevantinthecontextofmodelingtheinteractionsbetweenland-usepatternsandtravelbehavior.ThisstudycontributestotheliteraturebydevelopingaGeographicallyWeightedRegression(GWR),mixed-GWR,andQuasiGeographically-WeightedPoissonmodelstoimproveourunderstandingofimpactoflanduseonperson-milestraveled,vehicle-distance-traveled(VDT),vehicle-time-traveled(VTT),andcarownershipatadisaggregatelevel.Thisresearchalsodemonstratesthebenetsofestimatedspatialmodelsoversimplermethods.ThestudyareacoversSouthEastFlorida(SE)metropolitanareawhichincludesthreecounties:Miami-Dade,Broward,andPalmBeachCounty.Inthisdissertationbothmicrolevel,andmesoscaleurbandesignindicatorswerecreated.TheNationalHouseholdTravelSurveysof2008/2009andthe1999SouthEastFloridaHouseholdTravelSurveysconstitutedtheprimarysourceoftravelandsocio-economicdata. 9

PAGE 10

VariousmicrolevelbuiltenvironmentmeasureswerecreatedtoexaminetheirimpactsonPMT,VDT,VTTandcarownership.The2010statewideNAVTEQroadnetworklewasusedtodeterminetheshortest-distancetravelpathsbetweeneachorigin-destinationpair.OverallthisresearchcontributesonunderstandingandquantifyingrelationshipbetweenPMT,VDT,VTT,carownershipandlandusepatternsusingspatialnon-stationaryprocessinestimatingimpactofexplanatoryvariables. 10

PAGE 11

CHAPTER1INTRODUCTION 1.1BackgroundandMotivationEstimating“context-sensitive”or“spatial”modelsisgainingalotofinterestintransportationliterature.Intheconventionalglobal(or“aspatial”)models,themarginaleffectofeveryfactorisassumedtobexedoverspaceandinherentsimilaritiesinbehaviorsofneighboringentitiesarelargelyignored.However,thesensitivitiestothesameexplanatoryfactorcanchangeoverspaceduetoavarietyofreasonssuchasdifferencesinattitudes,preferences,andcontextualeffects.Further,therstlawofgeographywhichstatesthat“Everythingisrelatedtoeverythingelse,butnearthingsaremorerelatedthandistantthings”(Tobler1970[ 104 ])suggestingspatialclusteringinbehavioralpatterns.Spatialmodelshelpstoderivelocation-speciceffectsofexplanatoryfactorsratherthanglobaleffects.Iftherelationshipbetweenachoiceoutcomeandanexplanatoryfactordoesindeedvaryacrossthestudyarea,thenaglobalmodelisnotapplicableorrepresentativeofbehaviorinmanypartsofregions.Therefore,spatialmodelsaremoreusefulinidentifyingandassessinglocation-specicstrategiesandpolicies.Fromstatisticalpointofviewignoringspatialheterogeneitycancausebiasedparameterestimatesandmisleadingsignicancelevelsandhypothesistests(AnselinandGrifth1988)[ 11 ].Theneedforspatialmodelsisparticularlyrelevantinthecontextofmodelingtheinteractionsbetweenland-usepatternsandtravelbehavior.Traditionaldevelopmentalpatterns(typicallycharacterizedas“sprawling”giventhesuburbanresidentialdevelopmentshappensignicantlydistantformtheurbancore)havebeenassociatedwiththeneedforextensivetraveltosatisfyactivity-participationneedsandincreaseddependencyonthepersonalautomobileformakingthesetrips(Mooreetal.2010)[ 80 ].Inturn,thesepatternsoftravelcanbelinkedtoincreasedtrafccongestion(wastedtimes),greaterdemandsforgasoline,increasedpollutionfromtail-pipeemissions,anddetrimental 11

PAGE 12

publichealthimpactsbecauseofreducedphysicalactivity.Specically,75%ofcarbondioxideemissionsisfrommobilesourcesandhouseholdvehiclemilestraveledconsistsmorethan80%ofroadwaymilestraveled(Gomezetal.2009;Parryetal.2007)[ 50 , 89 ].Asthecurrentdevelopmentalparadigmisfoundtobeunsustainable,severalstrategieswiththebroadumbrellaof“coordinatedland-useandtransportationplanning”isbeingexaminedasanoptiontofacilitateeaseoftravelandactivity-participationwhileminimizingthenegativeimpactsonthesociety(theseareinadditiontotheuseofnewer/efcienttechnologiesincludingelectricvehiclesandpolicyinstrumentssuchaspricing).Inparticular,thereisinterestinthedevelopmentofcompact(highdensity)communities,creatinganefcientbalanceofcomplementarylanduses(includinghousing,educational,employment,recreational,retail,andserviceopportunities),publictransportation,andurbandesignsthatencouragewalkingandcycling.Theseareoftendescribedassmartgrowth,sustainablecommunities,andnewurbanism.Quantifyingtheimpactsofsuchland-usechangesintravelpatternsrequireintegratedland-useandtransportationmodels.Whileseveralpaststudieshaveprovidedvaluableinsightsintothestrengthofrelationshipbetweenland-usedescriptorsandtravelbehavior,oneofthecharacteristicfeaturesofthesestudiesistheuseofstationary(aspatial)analysistechniques.Irrespectiveofhowmanyvarietiesofland-usedescriptorsweincorporateinsuchmodelstherearealwayslikelyotherunobservedfactorsabouttheregionleadingtospatialcorrelations.However,aspatialmodelsignorespatialcorrelations/clusteringinbehavior.Further,factorssuchasadifferencesinattitudesandperceptionsacrosspeopleinthedifferentlocationsofthesameregion/citycouldmakethemdifferentiallysensitiveeffects(Fotheringhametal.2002[ 45 ])tothesameland-useattribute(suchasdensityorland-usediversity).Onceagain,aspatialmodelsignorethesespatialdifferencesinsensitivitiesbyassumingxedglobaleffects.Theuseofspatialmodels 12

PAGE 13

canalleviatetheseshortcomingsbycapturingthedifferencesinchangesintravelpatternsacrosstheregionevenifthesameland-usechangewereimplemented. 1.2DissertationObjectivesandApproachInthiscontext,theintentofthisstudyistoexpandthebodyofknowledgeonland-useandtravelinteractionsbydevelopingspatialmodels.Landuse/builtenvironmentimpactsavarietyoftravelbehavioraspectssuchaslonger-termchoicesofresidentiallocationandnumberofvehiclesownedandtheshorter-term(daily)choicesoftripfrequencies,triplengths,modeandtime-of-dayoftravel.Tosupportpolicystudies,itwouldalsobedesirabletounderstandtherelationshipsbetweenlanduseandaggregatetravelmeasuressuchastotaldailydistance/timetraveled.Thefocusofthisstudyisonunderstandingandquantifyingtheimpactoflanduseon(1)autoownershipand(2)aggregate(household-level)travelvolumemeasureslikethedailydistance/timetraveled.Amongthevarietyofspatial-analysismethodsthatareavailable,thegeographicallyweightedregression(GWR)isadoptedinthisstudy.Anselin(2010)[ 9 ]inhispaperreviewingthirtyyearsofspatialeconometriccalledadventofGeographicallyWeightedRegressionas“themostimportantdevelopmentasawaytomodelparametervariabilityacrossspace”.TheGWRtakesintoaccountcontinuousspatialprocessandusesthedatatodeterminethespatialboundaries.Theapproachcanbeusedtomodelbothcontinuous-andcount-variables(i.e.,bothdailydistancetravelledandthenumberofcarsowned).Further,theprocedurecanbeadaptedtoallowforcertainparameterstobexedoverspacewhileothersareallowedtovaryleadingtothemixedgeographicallyweightedregressionorMGWR.Thegeographicalscopeofthisworkisthethreecounty(PalmBeach,Broward,andMiami-Dade)regioninSouthEastFlorida.Thisregionhasmultiplecitiesandamixofurbanandrurallocationswithavarietyoflandusecharacteristicsand,hence,largerheterogeneitytoaccountforbythemodels.TheNationalHouseholdTravelSurveysof2008/2009andthe1999SouthEastFloridaHouseholdTravelSurveysconstitute 13

PAGE 14

theprimarysourceoftravelandsocio-economicdata.Theavailabilityofgeo-codedresidential-locationsdatasupportsspatialanalysisatanescale.Thesedataaresupplementedwithlandusedatawhichweregeneratedfrommultiplescales.Thisallowsfortoconstructavarietyofland-usemeasuresatvariousscalesandalsomapthesetothehouseholdlocation.Dataonthetransportationnetworkcharacteristicsoftheregionarealsoused. 1.3StructureoftheDissertationTherestofthisdocumentisorganizedasfollows.Chapter 2 presentsareviewofthestudiesrelatedtospatialregressionandautomobileownershipandusagemodels.Chapter 3 , 4 ,and 5 describesthePerson-Miles-Travel,Vehicle-Time-and-DistanceTravel,andcarownershipmodelingacrossSouth-EastFlorida.Eachchapterexplicitlyprovidesdatadescription,modelingframework,discussionsofempiricalmodels,andsummaryattheend.Intheend,Chapter 6 presentsthesummaryandconclusionsofthestudyandidentiesareasthatneedfurtherresearch. 14

PAGE 15

CHAPTER2LITERATUREREVIEW 2.1OverviewAnalyzingimpactoflanduseontravelbehaviorcanbeclassiedtoaggregate,disaggregate,andsimulationstudies(Handy2005)[ 55 ].Simulationstudiesanalyzetheimpactofchangeinurbandesignfactorsusingtraveldemandforecastingmodels.Butbeforesteppingtowardsimulationimpactanalysis,itisnecessarytohaveareliablepolicysensitivepredictivemodel.Aggregatelevelstudieslinkandmodelgeographicallyaggregateddeterminantsatmacroscalesuchascensustract,ormetropolitanleveldatatoobtaintravelbehaviordirectlyataggregatelevel.Aggregateanalysisissuitableinanalyzingoverallimpactofurbanformontravelbehaviorinspatiallyaggregateunits(e.g.,city,region).Inaggregatestudiescollectiveactionsofagentsinpredenedspatialresolutionsisevaluated.Whilethisecologicaltravelbehaviorapproachiscomputationallyandcostlyefcient(reduceddatacollectionrequirements),however,theaggregationindatareducestheexplanatorypower(limitedvariance)ofthevariables,thereforebothexplanatoryanddependentvariableshavelimitedvariancetoexplaintherelationship.Theoutcomesuffersfromcollinearityamongcovariates,largestandarderrorsofestimatedparameters,andlackofvariationsamonghouseholds.Thereforethenatureofdatalimitsthenumberandtypeofvariablesincluded.Ignoringvariationsamonghouseholdscanresultinaggregationbias(i.e.,ecologicalfallacy)andself-selectionbias.Theecologicalfallacyrefersthattherelationshipestablishedataggregatelevelmaynotbeappropriatetoapplyatindividualandhouseholdlevel.Inferringbehavioralpatternatdisaggregatelevelfromaggregatestudiescanbemisleading.Self-selectionandcausalrelationshipamongtravelbehaviorandurbandesignvariablescandistortandbiasedtheresults.Demographicandsocio-economiccharacteristicsofhouseholdsandindividualmightbethereasonoftheirtraveloutcome.Forinstancehighincome 15

PAGE 16

householdsarelikelytoownandusemorecarcomparetolowincomehousehold.Asaresult,lowincomehouseholdslocatethemselvesinhigherdensityareaswherepublictransportisavailableandshoppingstore,andschoolexistinwalking/bikingdistance.Asaresulttheyarelesslikelyandsensitivetobuildenvironmentsuchasdensityandmissingitcanoverestimateurbandesignimpacts.Incorporatingsocioeconomicvariablesintravelrelateddecisionmakinghelpstodisentanglethespuriousandtruecasualeffectofurbandesign.Mostdisaggregatelevelstudiesthatcontrolfordecisionmakingandhouseholdcharacteristicssuchashouseholdincomeobtainrelativelytrueandlessaffectedfromresidentialself-selectionorresidentialsortingeffects.Eventhoughaggregatemodelstthedatagoodenoughforforecasting,thereinnotmuchtodowiththeoutcomeforanalyzingpolicestochangethecurrentpattern.Toovercomethedecienciesofaggregatedapproach,household(individual)isthedecisionmakingunitatdisaggregateapproach.DisaggregateapproachcapturestheunderlyingbehavioralmechanismsthatrevealthehouseholddecisionprocessandimprovederivingcausalrelationshipamongexplanatoryvariablesandtravelbehavioroutcomessuchasVDT,VTT,andcarownership.Consequently,disaggregatemodelsarepreferredapproachtomodeltravelbehaviorspeciallyforanalyzingpolicysensitiveanalysis. 2.2SpatialRegressionandApplicationinTransportationEngineeringThethreespatialanalyticissuesinthemodelestimationsinclude:spatialdependency,spatialheterogeneity,andspatialheteroscedasticity.Spatialheterogeneityreferstorelationshipbetweendependentandexplanatoryvariablesvaryacrossspatialcontext.Spatialdependencyoccurswhenthereisatleastoneofthethreedifferenttypesofinteractionamongdecisionmakingagents:Spatiallaggeddependentvariableinteraction(endogenousinteractioneffectsamongdependentvariableswherethedecisionofaspatialagentdependsonthedecisiontakenbyotherspatialagents),spatiallaggedexplanatoryvariable(exogenousinteractioneffectsamongthe 16

PAGE 17

explanatoryvariableswherethedecisionofaspatialagentdependsontheotherspatialagentindependentvariables),and/orspatiallaggederrorterm(interactioneffectsamongtheerrortermswhereunobservedindependentvariablesresultsincorrelationandsimilarbehaviorofspatialagents)isinmodelspecication.Spatialheteroscedasticityreferstoheterogeneityinthevarianceoftheunobservedcomponentacrossspatialunit. 2.2.1SpatialHeterogeneitySpatialheterogeneityreferstorelationshipbetweendependentandexplanatoryvariablesvaryacrossspatialcontext.Ignoringspatialheterogeneitywouldleadtomisspecicationandbiasedestimateofstandarderrors.Modelspecicationdealingwithspatialheterogeneitycanbeclassiedtodiscrete-heterogeneityandcontinuousheterogeneity.Indiscreteheterogeneity,studyareaisdividedtothepre-deneddisjoint(mutuallyexclusive)clusters(discretesub-regions,units).Whileallobservationsinaclusterarehomogenousandinterconnected,thereisnotanyconnection(links)amongtheclusters,norcanobservationsbelongtodifferentclusters.Modelspecicationsandcoefcientsareallowedtovarybetweenspatialregimes.Incontinuousheterogeneitytheregressioncoefcientchangesoverthespaceeitherbypre-speciedfunctionalform(SpatialExpansionMethod)orasdeterminedbythedatathroughalocalestimationprocess(GeographicallyWeightedRegression(GWR)). 2.2.1.1GeographicallyweightedregressionGWRissimilartoLocallyWeightedRegression(LWR),whichisalsonamedlazylearning,model-on-demandorjust-in-timelearning(Kimetal.2013)[ 62 ]introducedbyCleveland(1979)[ 31 ]andClevelandandDevlin(1988)[ 32 ].InLWRalocalmodelisestimatedbyprioritizingeachpointinaccordancewithitssimilaritytoattributespaceandkernelfunctionisdenedbasedondistanceamongexplanatoryvariables(McMillen2010;McMillenetal.2010)[ 76 , 77 ].Brundstonetal.(1998)[ 18 ]andFotheringhametal.(1999)[ 44 ]developedGWRthatweightseachpointbasedongeographicspace 17

PAGE 18

ratherthanattributespace.GeographicallyWeightedRegressionisspeciedas:Y(uj,vj)=0(uj,vj)+pXk=1(k(uj,vj)Xjk)+j (2)Intheequation( 2 ),(uj,vj)denotesthegeographiccoordinatesofthejthobservationinthespaceand0(uj,vj)andk(uj,vj)representthevaluesoftheparameters0andkatestimationlocationj.TheparametersintheGWRmodelarederivedlocallyatlocation(uj,vj)byminimizingtheweightedleast-squareproblem,equation( 2 )withrespectto0andk(k=1,2,...,p):nXi=1[Yi)]TJ /F9 11.955 Tf 11.96 0 Td[(0(uj,vj))]TJ /F6 7.97 Tf 18.27 15.21 Td[(pXk=1(k(uj,vj)Xik)]2wi(uj,vj) (2)wi(uj,vj)isagivenweightorkernelfunction(Wangetal.2008)[ 109 ].Theestimationofcoefcientsspecictoeverylocationisaccomplishedbyperforminglocalregressions,eachusingasub-sampleofdatainthevicinityofthelocationunderconsideration.Further,intheestimationofparametersatsaypointj,itisassumedthatdataassociatedwithobservationsspatiallyclosertopointjhavemoreinuence(orweight)intheestimationofthecorrespondinglocalcoefcientk(xj,yj)comparedtodatafrompointsfartheraway.Oncetheweight(kernel)functionisdened,thesolutionoftheweightedleastsquareis:^(uj,vj)=[^0(uj,vj),^1(uj,vj),...,^p(uj,vj)]T=[XTW(uj,vj)X])]TJ /F5 7.97 Tf 6.58 0 Td[(1XTW(uj,vj)Y (2)whereX=0BBBBBBB@1x11...x1p1x21...x2p............1xn1...xnp1CCCCCCCA,Y=0BBBBBBB@y1y2...yn1CCCCCCCA,andW(uj,vj)=diag[w1(uj,vj),w2(uj,vj),...,wn(uj,vj)]= (2) 18

PAGE 19

0BBBBBBB@w1(uj,vj)...000w2(uj,vj)...0............00...wn(uj,vj)1CCCCCCCAwherewi(uj,vj)representstheweightofpointiontheestimationofthemodelatpointj.W(uj,vj)variesforeachj.W(uj,vj)differentiatesGWRfromtraditionalweightedleastsquarewithxedweightingmatrix.Thereforethettedvaluesoftheresponseareobtainedby:^Y=(^y1,^y2,...,^yn)=HY (2)wherethehatmatrix,His:H=0BBBBBBB@xT1[XTW(u1,v1)X])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTW(u1,v1)xT2[XTW(u2,v2)X])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTW(u2,v2)...xTn[XTW(un,vn)X])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTW(un,vn)1CCCCCCCATheresidualvectorandresidualsumofsquarearerespectively,^=Y)]TJ /F8 11.955 Tf 13.89 2.66 Td[(^Y=(I)]TJ /F4 11.955 Tf 11.96 0 Td[(H)Y (2)andRSS=^T^=YT(I)]TJ /F4 11.955 Tf 11.95 0 Td[(H)T(I)]TJ /F4 11.955 Tf 11.95 0 Td[(H)Y (2)TheexpectedvalueofRSSis:E[RSS]=21 (2)where1=tr[(I)]TJ /F4 11.955 Tf 13.53 0 Td[(H)T(I)]TJ /F4 11.955 Tf 13.53 0 Td[(H)].Todervivelocalcoefcientstandarderrors,Var[^(uj,vj)]=[(XTW(uj,vj)X))]TJ /F5 7.97 Tf 6.58 0 Td[(1XTW(uj,vj)][(XTW(uj,vj)X))]TJ /F5 7.97 Tf 6.58 0 Td[(1XTW(uj,vj)]T2 19

PAGE 20

wheretheestimatedstandarderroris^2=RSS 1,andtr(H)isthetraceofhatmatrix,H(Leung2000)[ 69 ].Therearetwoaspectsofinterestwithrespecttodeningtheweight:thebandwidth(i.e.,thenumberofdatapointsconsideredtobenearthelocationunderconsideration)andtherateofdecayintheinuenceofadjacentdatapointswithincreasingdistancewithinthebandwidth(Thedatapointsoutsidethebandwidthhavenoinuence).Theweightingfunctionisbasedonconventionalwisdominheritedfromspatialautocorrelationwhichresultsinnon-stationarypatternsinestimatedcoefcient.ThereforeinGWRthelocalweightmatrixW(uj,vj)assignshigherweightstothelocationsthatarefartherfromcalibrationlocation.Onepossibleapproachistodeneabinaryweightfunction(box-car)thatexcludesobservationsfurtherfromhjthanaspeciedradiusdistance.wi(uj,vj)=8>><>>:1,ifdijhj0,otherwise,wheredijisEuclideandistancebetweenpointsiandj,andhjisthebandwidthforlocationj.Theweightforpointsfartherthanhjisdenedzeroforpointjastheseexertnoinuenceinthedeterminationofthelocalparameter.Butspatialprocessarecontinuousandclassifyingdiscretesub-regionsimpliesspatialprocessvaryabruptly.Toalleviatetheproblemofdiscontinuityoverthestudyarea,localweightmatrixW(uj,vj)usuallyisderivedfromakernelfunctionwhichassignsweightsthatmonotonicallydecreasewithlocationsthatarefartherfromcalibrationlocationsuchasGaussian,exponential,orbi-squarekernelfunction.Gaussianorbisquarearethemostcommondistancedecaykernelfunctions.Gaussiandistancedecayweightingfunctionisdenedas:wi(uj,vj)=8>><>>:1)]TJ /F8 11.955 Tf 11.96 0 Td[(exp((dij hj)2),ifdijhj0,otherwise, 20

PAGE 21

Notethattheinuencedecaysexponentiallyinthiscase.Bisquaredistancedecayfunctionisdenedas:wi(uj,vj)=8>><>>:[1)]TJ /F8 11.955 Tf 11.95 0 Td[((dij hj)2]2,ifdijhj0,otherwise,Theweightforpointsfartherthanhjisdenedzeroaswell.Therearetwoapproachestospecifyingthebandwidth:xedandadaptive.Thexed-bandwidthapproachassumesaconstantvalueofdistanceindeterminingthevicinityofanyestimationlocation.Whenthespatial-densityofdataisfairlyuniform,thismethodologyisattractive.However,whencertainregionshavemoredatapointscomparedtoothers,thesparserareasmaynothaveenoughsampleswithinthechosen,xedbandwidthdistancetosupporttheestimationoflocalparameterswithadequatecondence.Alternately,ifalargebandwidthdistanceischosentoensureadequacyofsamples,theestimatesforlocationswithhighdensityofsamplesmaybebiased.Theadaptive-bandwidthapproachvariesthisdistanceofinuencebasedonthesampledensity.Specically,regionsofhighdensityofobservationsusesmallerdistanceforbandwidthsandthoseoflowdensityofobservationsuselargerdistanceforbandwidths.Itisalsoensuredthatthetotalnumberofdatapoints(neighbors)inthebandwidthisthesameirrespectiveofthedistance.Asolutiontotheoptimalbandwidthselectionisacross-validation(CV)approachwhichaccountsforpredictionaccuracysuggestedforlocalregressionbyCleveland(1979)[ 31 ]andkerneldensityestimationbyBowman(1984)[ 15 ].Thedeterminationoftheoptimalbandwidthselectionisaccomplishedbyminimizingthecrossvalidation(CV)scorewhichisdenedas:CVscore=argminhXj[yj)]TJ /F8 11.955 Tf 12.24 0 Td[(^y6=j()]2 (2) 21

PAGE 22

Notethat^y6=j()istheestimateatpointjwithexcludingtheregressionpointitself(j)fromtheestimationandhiskernelbandwidth.Thispseudo-sum-of-square-erroriscalledtheCVscore.InfactCVisaniterativeprocessthatsearchforthekernelbandwidthtominimizerootmeansquarepredictionerror(RMSPE).ItcanbeinterpretedasleaveoneoutorN-FoldCrossValidationasonlyoneobservationisremovedfromtheestimation.AnotherapproachtoderivekernelbandwidthiscorrectedAkaikeInformationCriterion(AICC)whichprovidesatrade-offbetweengoodnessoftanddegreesoffreedom.AICCforGWRwithabandwidthhisdenedas:AICC=2nln^+nln2+nn+tr(H) n)]TJ /F8 11.955 Tf 11.96 0 Td[(2)]TJ /F4 11.955 Tf 11.96 0 Td[(tr(H) (2)TheGWRmodelscanalsobecomparedtotheOrdinaryLeastSquare-basedmodelsusingFtests.Leungetal.(2000)[ 68 ]developedtwostatisticaltestsfortestingthegoodnessoftoftheGWRmodel.Intherstapproach,theteststatisticF1usestheresidualsumofsquaresanditsapproximateddistributiontotestwhetheraGWRmodeldescribesagivendatasetsignicantlybetterthananOLS.TheteststatisticF2usethemethodofanalysisofvariance(ANOVA).BycomparingResidualSumofSquare(RSS)ofGWR(unconstrainedmodel)andOLS(constrainedmodel)anditsdistribution,thehypothesistestH0:k(u1,v1)=k(u2,v2)=...=k(un,vn)8k=0,1,...,pcanbetestedtoexamineifGWRmodelimprovementsindescribingdataisstatisticallysignicantcomparedtoOLSmodel.F1=RSS 1 RSS0 n)]TJ /F6 7.97 Tf 6.58 0 Td[(p)]TJ /F5 7.97 Tf 6.58 0 Td[(1 (2)Leungetal.(2000)[ 68 ]showedF1canbeapproximatedasF-distributionwith(21 2,n)]TJ /F4 11.955 Tf -444.08 -23.91 Td[(p)]TJ /F8 11.955 Tf 12.1 0 Td[(1)degreesoffreedomwhere2=tr[(I)]TJ /F4 11.955 Tf 12.11 0 Td[(H)T(I)]TJ /F4 11.955 Tf 12.11 0 Td[(H)]2.Healsodevelopedanotherteststatistics,F2whichisbasedonanalysisofvariance.F2=RSS0)]TJ /F6 7.97 Tf 6.59 0 Td[(RSS 1 RSS0 n)]TJ /F6 7.97 Tf 6.58 0 Td[(p)]TJ /F5 7.97 Tf 6.58 0 Td[(1 (2) 22

PAGE 23

F2canbeapproximatedasF-distributionwith(21 2,n)]TJ /F4 11.955 Tf 12.18 0 Td[(p)]TJ /F8 11.955 Tf 12.17 0 Td[(1)degreesoffreedomwhere1=n)]TJ /F4 11.955 Tf 11.73 0 Td[(p)]TJ /F8 11.955 Tf 11.72 0 Td[(1)]TJ /F9 11.955 Tf 11.73 0 Td[(1and2=n)]TJ /F4 11.955 Tf 11.73 0 Td[(p)]TJ /F8 11.955 Tf 11.72 0 Td[(1)]TJ /F8 11.955 Tf 11.73 0 Td[(21+2.Leungetal.(2000)[ 68 ]alsodevelopedastatisticaltestfortestingwhetherestimatedGWRparametersexhibitsignicantvariationoverthestudyregion.F3(k)Examinesspatialvariationofcoefcients,whetherH0:k(u1,v1)=k(u2,v2)=...=k(un,vn)foragivenk20,1,2,...,pornot.GWRhavebeenappliedinawidespectrumsuchasEcology(Zhangetal.2004)[ 117 ],PublicHealthandepidemics(Nakayaetal.2005)[ 84 ],Sociology/PublicPolicy(MalczewskiandRinner2005)[ 74 ].Anselin(2010)[ 9 ]inhispaperreviewingthirtyyearsofspatialeconometriccalledadventofGeographicallyWeightedRegressionas“themostimportantdevelopmentasawaytomodelparametervariabilityacrossspace”.IntransportationrelatedapplicationsGWRhasbeenappliedintrafccountandaccidentanalysisandprediction(Zhaoetal.2004[ 118 ];Hadayeghietal.2010[ 54 ]),publictransitridershipanduse(Cardozoetal.2012,2011[ 20 , 52 ];BlaineyandPreston2010[ 13 ]),individualsorcommutersdistancetraveled(e.g.,NowrouzianandSrinivasan2013[ 86 ]andLiyodandShuttleworth2005[ 72 ]).Alietal.(2007)[ 3 ]mentioneditspowerinderivinglocalrelationshipandinformationforpolicyapplicationisoneofthereasonsforitspopularityrecently. 2.2.1.2MixedgeographicallyweightedregressionGeographicallyWeightedRegression(GWR)extendstheconventionalregressionmodelbyallowingtheparameterstovaryoverspace.However,GWRmodelsrequiretheestimationofaverylargenumberofparametersandassuchnotefcient.WhiletherearealotofapplicationsforGWRbutthesemodelsarenotparsimonious.InGWRallcoefcientsareallowedtovaryacrossthestudyarea.Whileinapplicationsomeofthecoefcientsmayvaryspatiallywhileothersnot,sothereisneedforMixedGeographicallyWeightedRegression.TheMixedGeographicallyWeightedRegression 23

PAGE 24

hastheform,Y(uj,vj)=0(uj,vj)+qXk=1(kXjk)+pXk=q+1(k(uj,vj)Xjk)+j (2)Intheequation( 2 ),(uj,vj)denotesthecoordinatesofthejthobservationinthespaceand0(uj,vj)andk(uj,vj)representthevaluesoftheparameters0andkatestimationlocationj.Thereaderwillnotethattherstqparametersareassumedtobespatiallyinvariantwhereastheremaining(q+1thoughp)parametersvaryoverspace.Byspecifyingsubscriptscandvtotheconstantandvariableparts,letXc=0BBBBBBB@x10x11...x1qx20x21...x2q............xn0xn1...xnq1CCCCCCCA,Xv=0BBBBBBB@x1,q+1x1,q+2...x1px2,q+1x2,q+2...x2p............xn,q+1xn,q+2...xnp1CCCCCCCA,Y=0BBBBBBB@y1y2...yn1CCCCCCCA,c=0BBBBBBB@01...q1CCCCCCCA,v(ui,vi)=0BBBBBBB@q+1(ui,vi)q+2(ui,vi)...p(ui,vi)1CCCCCCCA,andHv=0BBBBBBB@xTv1[XTvW(u1,v1)Xv])]TJ /F5 7.97 Tf 6.58 0 Td[(1XTvW(u1,v1)xTv2[XTvW(u2,v2)Xv])]TJ /F5 7.97 Tf 6.58 0 Td[(1XTvW(u2,v2)...xTvn[XTvW(un,vn)Xv])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTvW(un,vn)1CCCCCCCAHc=Xc(XTcXc))]TJ /F5 7.97 Tf 6.58 0 Td[(1XTc (2)Thenthevectorofconstantcoefcients^cisestimatedby(Fotheringhametal2002;Leungetal.2000)[ 45 , 69 ],^c=(^0,^1,...,^q)T=[XTc(I)]TJ /F4 11.955 Tf 11.95 0 Td[(Hv)T(I)]TJ /F4 11.955 Tf 11.95 0 Td[(Hv)Xc])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTc(I)]TJ /F4 11.955 Tf 11.95 0 Td[(Hv)T(I)]TJ /F4 11.955 Tf 11.96 0 Td[(Hv)Y (2) 24

PAGE 25

andthevectorofspatiallyvaryingcoefcientsat(uj,vj),^v(uj,vj)isestimatedby,^v(uj,vj)=(^q+1(uj,vj),^q+2(uj,vj),...,^p(uj,vj))T= (2)[XTvW(uj,vj)Xv])]TJ /F5 7.97 Tf 6.58 0 Td[(1XTvW(uj,vj)(Y)]TJ /F4 11.955 Tf 11.95 0 Td[(Xc^c)Thereforethevectorofttedvaluesatnlocationsis:^Y=(^y1,^y2,...,^yn)T=Hv(Y)]TJ /F4 11.955 Tf 11.95 0 Td[(Xc^c)+Xc^c=HvY+(I)]TJ /F4 11.955 Tf 11.95 0 Td[(Hv)Xc^c=HY (2)whereH=Hv+(I)]TJ /F4 11.955 Tf 11.95 0 Td[(Hv)Xc[XTc(I)]TJ /F4 11.955 Tf 11.96 0 Td[(Hv)T(I)]TJ /F4 11.955 Tf 11.96 0 Td[(Hv)Xc])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTc(I)]TJ /F4 11.955 Tf 11.96 0 Td[(Hv)T(I)]TJ /F4 11.955 Tf 11.95 0 Td[(Hv) 2.2.1.3Geographicallyweightedquasi-PoissonregressionTheclassicalassumptionongenerallinearmodelsisthaterrorisnormallydistributed,theerrorvarianceisconstantandindependentofthemean.Whentheresponseiscountdata(integernonnegative),GWRandOLSregressionmaybeinappropriatebecausecountdataarenon-negativeanddiscrete,andtendtobehighlyskewedandnon-normallydistributed,andvariancewillbechangewiththemean(varianceismeanrelatedvariance).Atypicalapproachistotransformthedataforinstancereplacestheresponsewithanotherfunctioninthehopethattheassumptionsfortheclassicalgenerallinearmodelwillbesatised.TransformationofdependentvariableYbyY=f(Y)forsomemonotonic,non-afnefunctionf(.),mayworktononlinearityorheterogeneityofvariance.Howeverwearemodeling=E(Y)=E(f(Y))ratherthan=E(Y).Ingeneralfornon-afnef(.),=E(f(Y))6=f(E(Y))=f())6=f)]TJ /F5 7.97 Tf 6.59 0 Td[(1().Thereforewecannotuseestimatedf)]TJ /F5 7.97 Tf 6.59 0 Td[(1()todirectlyestimate.NelderandWedderburn(1972)[ 85 ]developedanewapproachandframeworkso-calledGeneralizedLinearModels(GLM).Insteadoftransformingtheobservations,formulatelinearmodelsforatransformationofthemeanvalue,thelinkfunctionandkeeptheobservationsuntransformed,therebypreservingthedistributionalpropertiesoftheobservations.SointheGeneralizedLinearModelsthelinearpartofthemodels 25

PAGE 26

Xdescribesafunctionofthemeanvalue(andnotthemeanvaluedirectlylikegenerallinearmodels).GeneralizedLinearModelsistiedtoaspecialclassofdistributions,theexponentialfamilyofdistributions.GLMmodelsrequirespecicationofthesystematicandrandom(stochastic)components.Thesystematiccomponentrelatesthevectorofexplanatoryvariablestothemeanresponse=E(Y).Itconsistsoftwocomponents:thelinearpredictor=Xandthelinkfunction(whichdescribesafunctionofthemeanvaluewhichcanbedescribedlinearlybytheexplanatoryvariables)g()=wheregisasmooth,monotonicfunction.TherandomorstochasticcomponentofaGLMspeciesthedistributionalformoftheresponses.OneimportantpropertyofGLMisexibilityinspecifyingresponsedatathatfollowsaverygeneraldistributioncalledexponentialfamilywhichincludesnormal,Poisson,geometric,negativebinomial,exponential,gamma,andinversenormaldistributions.Countordiscrete(integernonnegative)variables,oftenfollowaPoissondistribution.InPoissonregressionitisassumedthatVar(Yi)=i.PoissonregressionalleviatesOLSassumptionofthesymmetricerrordistributionandconstantvarianceandpredictingnegativecountvalues.Weareassumingforthespeciccombinationoftheregressorsordesignvariables,thedistributionofthemeasuredresponseisPoissonwithmean(andvariance)i.UnderGLMmethodthemeanandvarianceoftheresponseateachdatapointarerelated.Whenthevarianceexceeds/fallsbehindthemean,over/under-dispersionoccurs.Tocountfortheover/under-dispersion(varianceexceeds/fallsbehindthemean),Quasi-LikelihoodGLPMmodelisestimated(GLPMestimatedbyquasimaximumlikelihood).ForacompletediscussionseeMcCullaghandNelder(1989)[ 73 ].Thepoissonregressionisappropriateformodelingcountdataandfallswithintheframeworkofgeneralizedlinearmodels(NelderandWedderburn1972)[ 85 ].However,thisapproachassumesequidispersion(meanisequaltothevariance),whichhasbeenfoundtobeuntrueoften.Toaddressthisissue,negative-binomialregressionmodelsarepopularlyused.Thesemodelsassumethatthevarianceisaquadraticfunctionof 26

PAGE 27

themeanandarethereforeappropriateforoverdisperseddata(variance>mean).Analternativeapproachisthequasi-Poissonmodelthatisappropriateforbothover-andunder-dispersedcases.TheQuasi-Poissonmodelassumesalinearrelationshipbetweenmeanandvariance:Var(YjjXj)=j (2)If0<<1underdispersionoccurs,andwhen>1overdispersionhappens.Incase=1thespecicationwouldbethesameasstandardPoissonregression.Foracompletediscussionofquasi-PoissonmodelsrefertoMcCullaghandNelder(1989)[ 82 ].IntransportationcontextGiuffraetal.(2011)[ 46 ]andYannisetal.(2007)[ 114 ]haveusedquasi-poissonmodelinthecontextofsafetyanalysis.Thequasi-GeographicallyWeightedPoissonRegressionmodelrelaxesthespatialstationaryassumptionbyallowinggreaterexibilityandensuringsmoothnessinthevariationofthemodelcoefcientsoverspace.Specically,Inquasi-Poissonmodel,thecoefcientsareconstantacrossthestudyarea(spatialstationaryassumption).Specically,ln(E(YjjXj))=0+pXk=1(kXjk)+j (2)Thequasi-GeographicallyWeightedPoissonRegressionmodelrelaxesthespatialstationaryassumptionbyallowinggreaterexibilityandensuringsmoothnessinthevariationofthemodelcoefcientsoverspace.Specically,ln(E(Y(uj,vj)))=0(uj,vj)+pXk=1(k(uj,vj)Xjk)+j (2)Theestimationofcoefcientsspecictoeverylocationisaccomplishedbyperforminglocalquasi-GWPM,eachusingasub-sampleofdatainthevicinityofthelocationunderconsideration.Further,intheestimationofparametersatsaypointj,itisassumedthatdataassociatedwithobservationsspatiallyclosertopointjhavemoreinuence(orweight)intheestimationofthecorrespondinglocalcoefcientk(uj,vj)comparedto 27

PAGE 28

datafrompointsfartheraway.Oncetheweightfunctionisdened,theparametersareestimatedinapointwiseway.Avariantofthelocallikelihoodprinciple(GeographicallyWeightedlikelihoodprinciple)isemployed.ThemaximizationproblemcanbesolvedbyamodiedlocalFisherscoringprocedure,aformofiterativelyre-weightedleastsquares(IRLS).AfterrunningIRLSatallthepoints,theparametersareconvergedsuchas,^(uj,vj)=[(XTW(uj,vj)A(uj,vj)X)])]TJ /F5 7.97 Tf 6.59 0 Td[(1XTW(uj,vj)A(uj,vj)ln(E(Y(uj,vj))) (2)WhereW(uj,vj)isanbyndiagonalmatrixwithdiagonalelementsrepresentinggeographicalweightingofeachofnobserveddataforregressionpointj.A(uj,vj)denotesthevarianceweightsmatrixassociatedwithFisherscoring.Theestimationprocedureissimilartoconventionalkernelregression(localregression)modelstomaximizeaweightedlikelihoodfunction.ThedifferenceisthatinGWPMtheweightsarebasedonkernelfunctioningeographicalspacewhileinstandardkernelregressionthekernelfunctiongeneratesweightsinattributespace.ForacompletediscussionseeFotheringhametal.(2002)[ 45 ]andNakayaetal.(2005)[ 84 ]. 2.2.2SpatialClusteringToinvestigatespatialheterogeneityofmodelresiduals,variationofmean(variance)ofresidualsfromonecluster(subregion)toanotherclustershouldbeexamined.Existenceofspatialautocorrelationamongglobalregressionscancauseinefcientleastsquareestimatorsandfalsiedstatisticalinferenceresults(Leungetal.2000[ 69 ]).Assessingspatialautocorrelationofresidualsisinvestigatedattwoscalesofanalysis,globalandlocal.Globallevelapproachassumesthatthesamepatternexistsintheentiregeographicaldataset.Itisusedtogureoutifthereisanyspatialclusteringofresidualvalues.Forexploringwherethesepatternsoccurlocalmeasuresarecalculated.Inthelocallevelanalysisaspecicvalueisassignedforeachpartofregion.Differentpatternthatmightexistindifferentpartsofregioncanbecapturedbylocallevelapproaches. 28

PAGE 29

GlobalMoransIisthemostcommonglobalmeasureofspatialautocorrelationwhichisapplicabletothepointsorpolygonscontainingcontinuousdata.Itvariesbetween-1to+1.Itisanindexofdispersion,randomness,andclusterpatterns.LocalIndicatorofSpatialAssociation(LISA)isusedtoevaluatelocalspatialheterogeneityofmodelresiduals.Foramodelestimatedinalargespatialregion,theremightbesomesub-regionsthattheexistingpatterninthoseisdifferentfromnormalpattern.LISAcanbeusedtorevealHotSpots(positiveautocorrelation)andColdSpots(negativeautocorrelation). 2.2.2.1GlobalmoransIMoran(1950)[ 81 ]developedateststatisticsforanalyzingspatialautocorrelationamongneighbors.SimilartoDurbin-Watsontestforserialcorrelationintimeseriesanalysis(DurbinandWatson1950[ 40 ]),CliffandOrd(1972,1981)[ 33 , 34 ]extendedtheteststatisticsforOLSresidualswhenthedatahaveageographicalordering.Theyprovedundersomeassumptions,thedistributionofMoransIisasymptoticallynormalunderthenullhypothesisofnospatialautocorrelationamongthenormallydistributederrors.However,thatteststatistics(MoransI)sometimesdoesnotfollownormaldistribution.ThereforeexactdistributionofMoransIshouldbecalculated.SimilartoderivingexactdistributionofDurbin-Watsonteststatisticsforserialautocorrelation,onthebasisoftheoreticalgroundsdevelopedbyImhof(1961)[ 59 ]andKoertsandAbrahamse(1968)[ 64 ],TiefelsdorfandBoots(1995)[ 103 ],andHepple(1998)[ 56 ]derivedtheexactdistributionofMoransIstatisticsforOLSresidualsunderthenullhypothesisofnospatialautocorrelationamongthenormallydistributederrors.ExactdistributionofMoransIiscomputationallyintenseevenformediumsizedatasets.Tiefelsdorf(2002)[ 102 ]showedsaddlepointapproximationofexactdistributionresultsincomputationallyefcientandreasonablepreciseapproximationofthesamplingdistributionofMoransI.AnotheralternativeisMonteCarlo(random)permutationforMoransI.Randomlyarrangethevaluesacrossstudyareaandcalculateteststatistics 29

PAGE 30

eachtime.IftheactualMoransIliesbeyond(less)95%(5%)ofthesimulatedMoransIvalues,itisconcludedthatitispseudopositive(negative)signicantat5%level.GlobalMoranforregressionresidualsiscalculatedbydividingcrossproductoftheerroranditsspatiallagtotheerrorcrossproductadjustedforspatialweights,I=nnPi=1NPj=1W(uj,vj)(ei)]TJ ET q .478 w 276.66 -110.06 m 283.01 -110.06 l S Q BT /F4 11.955 Tf 276.66 -117.38 Td[(e)(ej)]TJ ET q .478 w 316.38 -110.06 m 322.74 -110.06 l S Q BT /F4 11.955 Tf 316.38 -117.38 Td[(e) (nPi=1nPj=1W(uj,vj))nPi=1(ei)]TJ ET q .478 w 302.54 -142.76 m 308.9 -142.76 l S Q BT /F4 11.955 Tf 302.54 -150.08 Td[(e)2 (2)WhereIiscalledGlobalMoranIndex,ejandeiaremodelerrorataparticularlocationjandi, eisthemeanoftheerrortermovernlocations,andW(uj,vj)istheweightindexinglocationofirelativetoj.TheGlobalMoranIdeterminesthecrossproductofdeviationfromthemeanofallerrorsintheneighboringdistance.Iftheerroratjejandtheerrorsataneighborofjarebothgreaterthan(orlessthan)thantheaverageerror,thecrossproductproducesapositivenumberandthisimpliesspatialclusteringoferrors.WhenGlobalMoranturnsouttobestatisticallyinsignicantitimpliesresidualsarerandomlydistributedoverspace.Itiscanbeinterpretedasthecorrelationbetweenej,andthespatiallagofejformedbyweightedaveragingallthevaluesofpointjneighbors.GlobalMoransIcanbegivenaniceintuitivegeometricinterpretation.TheexpectedvalueisE(I)=)]TJ /F8 11.955 Tf 9.3 0 Td[(1=n)]TJ /F8 11.955 Tf 11.96 0 Td[(1andvarianceisderivedunderassumptionofrandomnessornormality.Theseassumptionsdeterminethegenerationprocessoferrorvaluesweregeneratedunderthehypothesisofrandomlyplacederrorvalues. 2.2.2.2LocalindicatorofspatialassociationToexploreexistenceoflocalpositivespatialautocorrelation(HotSpot)andnegativespatialautocorrelation(ColdSpot)ofmodelresiduals,Anselin(1995)[ 6 ]developedLocalIndicatorsofSpatialAssociation(LISA)todecomposeMoranI(disaggregateversionofGlobalMoran).Thislocalstatisticsdetermineswheretheclustersareandwhatistheirspatialextent.Ij=(ej)]TJ ET q .478 w 198.18 -608.53 m 204.54 -608.53 l S Q BT /F4 11.955 Tf 198.18 -615.85 Td[(e)Pni=1W(uj,vj)Pni=1(ei)]TJ ET q .478 w 345.22 -608.53 m 351.57 -608.53 l S Q BT /F4 11.955 Tf 345.22 -615.85 Td[(e)Positive,negative, 30

PAGE 31

andstatisticallyinsignicantvalueofLISAindicatesclustered,dispersed,andrandomdistributionofresidualsvaluesaroundlocationjinthecorrespondingneighborhood. 2.2.3SpatialDependencySpatialdependenceimpliesthereisdependencyamongnearbylocationswhereastraditional(aspatial)crosssectionmodelsassumethatobservationsareindependentfromeachother.Thisassumptionmightbeviolatedforinstanceinspatialcontext,theclosertwoobservationaretoeachother,themoresimilarthandistantthingsarewhichisconsistentwithTheToblerrstlawingeography“Everythingisrelatedtoeverythingelse,butnearthingsaremorerelatedthandistantthings”(Tobler1970[ 104 ]).Toalleviateandremovespatialdependency,itisnecessarytoinvestigatethereasonsthatcausedthisdependency.Thesereasonscanbeclassiedtothreegroups:Modelspecicationdealingwithspatialdependencetakestheformofincludingeitheroneofthespatiallagofdependent(e.g.,spatialautoregressiveorspatiallagmodels),independent(e.g.,spatialcross-regressivemodels),orerrorterms(e.g.,spatialerrormodels)oracombinationofthose(e.g.,spatialdurbinmodel).Thesethreespatialreasonsthatviolatetheindependencyassumptionwillbediscussedindetailhere. 2.2.3.1EndogenousinteractionamongdependentvariablesAgents,households,individuals,anddecisionmakersbehaviorisinuencedwithbehavioroftheotheragents,households,etc.Agents/decisionmakersinteractwitheachotherandinmostofthetimestheoutcomeormotivationformakingadecisiondonotdependonlyonthecharacteristicsofparticularagent/decisionmakerbutalsothroughtheirinteractionwithotheragents/decisionmakers.Theconcepttakesdifferentlabelsandjargonsineacheldincluding:neighborhoodeffects,peerinuence,copy-catting,yardstickcompetition,racetothebottom,contagion,epidemics,socialcapital/norm,andstrategiccompetition/interaction(Ansein2002[ 7 ]).Ineconomics,theissueisreectedintheoryofinteractingagentsandsocialinteraction.Theycapturehowtheinteractionsamongagentsresultsinemergentcollectivebehaviorandaggregation 31

PAGE 32

pattern.Location,space,andspatialinteractionisakeypartofthat(Ansein2002[ 7 ]).Modelsforneighborhoodspill-overeffects,macroeconomicmodelswithmeaneldinteractions,interactingparticlesystemsandrandomeldsmodels,pathdependenceandimperfectcompetitionunderlyingtheneweconomicgeographyareexamplesoftheoreticalframeworkineconomics(Anselin2002[ 7 ]).Sociologistshavearguedthat,individualsarenotisolatedagentsandentitiesbutarepartofnetworkoffamily,friends,neighbors,andcolleaguesthatexchangeinformation,inuenceandaffecteachother,makeculturalnorms,andeconomicopportunity.“One'sneighborsmatterindeningone'sopportunitiesandconstraints”(Burt1997[ 19 ]).SAR(SpatialAuto-Regressive)regressionusesspatialautoregressiveprocesstoextendconventionalregressionmodels.Itincorporatesspatialdependencebyaddingspatiallagofdependentvariabley.AsitisshowninthereducedformofSAR,modelspecicationisasimultaneousmodelwithfeedbackbetweenobservations.Inotherwords,ashockinthesystemthataffectsindependentvalueatapoint,willbetransmitted(propagated)throughtheconnected(wij6=0)observations.SARregressiontakesthestructuralformof[ 5 , 10 , 66 , 67 ],Y=WY+X+ (2)orreducedformof,Y=[(In)]TJ /F9 11.955 Tf 11.96 0 Td[(W)])]TJ /F5 7.97 Tf 6.59 0 Td[(1X+[(In)]TJ /F9 11.955 Tf 11.96 0 Td[(W)])]TJ /F5 7.97 Tf 6.59 0 Td[(1 (2)whereYisn1vectorofdependentvariables,Xnpamatrixofindependentvariables,andisparameterofspatialassociationfordependentvariables,andWisannmatrixconsistingofwijelementsrepresentingthedegreeofrelatednessamongobservations.InSAR,E(Y)=(In)]TJ /F9 11.955 Tf 12.24 0 Td[(W))]TJ /F5 7.97 Tf 6.59 0 Td[(1X.Byassumingj1jandwij1(row-standardizedspatialweights),LeontiefExpansionoftheinversematrix(SpatialMultiplier)resultsin,(In)]TJ /F9 11.955 Tf 11.95 0 Td[(W))]TJ /F5 7.97 Tf 6.58 0 Td[(1=I+W+2W2+3W3+... (2) 32

PAGE 33

PowerofthespatialweightmatricesW,W2,W3,...representsorderofneighbors.Wistherst-orderneighbors(theobservationsthataredenesasneighbor),andW2reectsecond-orderneighbors(thosethatareneighborstotherstorderneighborsforeachobservations).Sowecanseethatspatialdependencyisglobalanditrelatesallobservationstoeachother.Inotherwordsachangeinexplanatoryvariablewillbetransmittedanddiffusedtothewholesystemandstudyareathroughtheinversematrix.(In)]TJ /F9 11.955 Tf 12.95 0 Td[(W))]TJ /F5 7.97 Tf 6.59 0 Td[(1,theLeontiefinverse,connectseachytoallxsthroughaspatialmultiplier.Itshouldbetakentoconsiderationthatchangesinaparticularexplanatoryvariablenotonlychangesthecorrespondingdependentvariablebutalsoallthedependentvariablesinthesystem.Thelatterislabeleddirecteffectandtheformerindirectorspillovereffect.Thetotalimpactofachangeinanexplanatoryvariablecanbedecomposedintodirectandindirect(spatialspillover).ThetraditionalOrdinaryLeastSquareregressiondoesnotallowforspillovereffectssinceitassumesthatthedependentvariablesareindependent,whichmightbeviolatedwithspatialdata.Oneoftheapplicationsofsimultaneousspatialdependenceanddiffusionthroughoutthewholesystemiscommutingtime.KirbyandLesage(2009)[ 63 ]andParentandLesage(2010)[ 88 ]foundaglobalspilloverincommutingtimeforresidentsofcensustract.Sincecommutersshareacommonroadway,congestiononroadinoneregion(censustract)willproducespilloverexternalityoncommutingtimeofotherregions(censustract)throughoutmetropolitanareawithmoreimpactonnearerregionsversuslongerregions(spatialspillover).Transportationinfrastructureplaysanimportantroleinboostingtheregionaleconomic.Thetransportationinfrastructureandit'sspatialspilloverimpactonvariousaspectsoftheeconomyhasbeeninvestigatedinmanystudies.Bothpositiveandnegativespatialspilloverofinfrastructure-economyrelationshipispossible[ 105 ].Forinstanceimprovingtransportationnetworkinastatecanhelptheeconomicsofadjacentstatesthroughtheintegratedtransportationnetwork(positivespatialspillover).Ontheotherhand,economicactivitiescanrelocatefrom 33

PAGE 34

theadjacentstatestothestatewithbetterconnectivity(negativespatialspillover).Thereforetransportationinfrastructureofonestatecanaffecttheproductivityofprivatesectorsintheneighborstates.Theassessmentofspatialspillovereffectsisrelevantintransportationinfrastructurefrompointofviewofprioritizing,nancingandallocatingfundstothetransportationinfrastructureprogramssuchashighways,ports,airports,etc.withconsideringwhich/where/when,toboosttheeconomy,andemployment.ForinstanceTrans-EuropeanTransportNetworks(TEN-T)policyseekandsupportprojectsofcommoninterest.EuropeanValueAddedisdenedasthemagnitudethatapolicyimprovestransportationefciencyandboostnewdevelopment.Thereforeinspillover(trans-national)impactoftheprojectsareimportantcomponentofTENprojectevaluations.Projectswithhighestspatialspillovereffects(animprovementinonepartofthenetworkproducepositiveeffectsinmanyotherneighboringnetworks)meetthegoalofEuropeanintegrationanddevelopingmethodologiesforappraisalofthespillovereffectshasbeenstudied[ 53 , 90 ].Tongetal.(2013)[ 105 ]analyzeddirecteffectandspatialspilloverimpactoftransportationinfrastructureonagriculturaloutputacross44statesintheUS.Theyconcludedthatroadinfrastructureexpansionstimulatesagriculturaleconomicsinbothwithinastateandotherstates(spatialspillover).Theresultswouldleadtoprioritizinginvestmentinconstructionandmaintenanceofroadwaystogeneratethegreatestagriculturalproductivitygiventhebudgetdecitandlimitedavailableresources.TheyrecommendedinvestinginroadstructureinthecentralUSstatesthatproducethemostspatialspillovereffects.Cohen(2010)[ 36 ]arguesthatignoringspatialspillovereffect(beyondthegeographicboundarieswithinwhichtheinfrastructureinvestmentsareundertaken)canresultinmisstatementsofaccuratelyassesstheimpactsofinfrastructure.Itisnecessarytoincludethespatialspillover(indirect)effectsoftransportationinfrastructuretoavoidderivingbiasedinfrastructureelasticityimpact.Itwouldguidepolicymakerstodecidewhereandwhattypeofinfrastructurehasthemostbenet.Italsohelpsinprioritizingamongprojects 34

PAGE 35

tofund.Heexaminedspatialaspectsofpublicinfrastructureinpreviousstudiesandconcludedthatthereisnotaconsensusindirectionofinfrastructureimpactanditdependsongeographiclocationsunderconsideration.CohenandMorrisonPaul(2003)[ 37 ]analyzedairportinfrastructurespilloversinanetworksystem.DelayatonenodeintheUSairtransportationnetworkwillexacerbatedelaysthroughoutthesystem.Theyfoundevidenceofpositivespatialexternalitiesacrossstatesthatshouldbetakenintoaccountwhendevelopingpolicytoboostairtrafcnetworkefciency.CohenandMonaco(2008)[ 35 ]used48contiguousstateleveldatatomodelmanufacturingproductionandcostincorporatinginvestmentinportandhighwayinfrastructureasexplanatoryvariable.Theyfoundincreasedportinfrastructurelowersmanufacturingcostsincorrespondingstatesbutincreasesthemanufacturingcostsintheneighboringstates(negativespatialspillover).Thisnegativespatialspilloverisevidenceofexternaldiseconomyofscale(Agglomerationeconomies).Insumweconcludethatoneoftheapplicationsofspatialdependencyintransportationinfrastructureanalysisisassessingandappraisingtheprotabilityoftransportationinfrastructureinvestmentsandprioritizingamongthembasedonbothdirectandindirect(spatialspillover)impacts. 2.2.3.2InteractioneffectsamongtheerrortermsInpreviousdiscussion,thespatialdependencydescribedwasinherentpropertyofdependentvariablewhichiscalledInherentspatialdependency.OneoftheviablereasonsforInducedSpatialAutocorrelationistheerrorterm.Thisismostlikelythecaseifomitted(un-modeled,not-captured)variablesthataresubsumedintheerrorterm,arejointlyspatiallydependentbythemselvesandleadtospatialautocorrelation.Inotherwords,decisionmakerssuchashouseholds,individualsandrmsclosetoeachotherinspatialunitarelikelytobesimilarbysomeunobservedfactors(spatiallyclustered)evenaftercontrollingwithexistingexogenousvariables.Examplesofthesignicantspatiallystructuredindependentvariablesthatmaynotbecapturedare:agentslivingin 35

PAGE 36

proximitycanshareunobservedcharacteristicssuchasattitudethatarenotavailabletoincludeasexogenousvariables.Anotherexamplewouldbetheimpactofaccessibilityandpublictransportationqualityandlocationamenitiesinhedonicmodelswhentherearenoprecisemeasuresavailableforthesevariableorzone-relatedattributessuchaspedestrianortransitfriendlinessofazoneintransportationplanning.Basedonthescopeofspatialimpactwecanderivetwospecicationsfortheerrorterm.Inonespecication,SpatialErrorModel(SEM)denesglobalspilloverinerrorterm.SEMregressiontakesthestructuralformof,Y=X+u (2)u=Wu+ (2)SEMtakesreducedformof[ 5 , 10 , 66 , 67 ],Y=X+[(In)]TJ /F9 11.955 Tf 11.95 0 Td[(W)])]TJ /F5 7.97 Tf 6.59 0 Td[(1 (2)Theerrortermisrelatedtoalltheobservationsinthesystemorintheotherwordsashockintheerrortermatanypointwillbetransmittedtoallotherpoints.Indeedforsmallvalueofafterarelativelysmallnumberofpowers(orderofneighbors)mayapproachzero.SpatialErrorModel(SEM)specicationisappropriatewhennoiseandothernuisances(un-modeledeffect)haveregionalimpacts.Inotherspecicationlocalspilloverinerrortermisdened.Inadditiontothespatialautoregressiveerrorprocess,wecanincorporatespatialdependencyinerrortermusingmovingaveragespatialerror.SpatialMovingAverage(SMA)speciesLocalspilloverinerrortermas[ 5 , 10 , 66 , 67 ]:Y=X+u (2)u=W+ (2) 36

PAGE 37

SMAtakesthereducedformof:Y=X+(In+W) (2)SpatialMovingAverage(SMA)isusedtomodellocalizedeffectsandconsiderlocalspatialspillover.Byitsdenitionspatialeffectswillaffectrst-orderneighborsasdenedbyweightmatrix.Thisspecicationisusedwhennoiseandothernuisances(un-modeledeffect)areconstrainedtotheimmediateneighborsandtheydonotaffectthedependentvariableafterimmediateneighbors.InthereducedformofSMAthereisnotaninversetermwhichresultsinlocalrangefortheinducedspatialcovariates.Thetwodescribederrorspecicationsaresimilartoautoregressiveandmoving-averagemodelintimeseriesanalysis.Theerrorspatialinterdependencycanbespeciedinaconditionalprobabilisticframeworkaswell(besidessimultaneousviewpoint).Conditionalperspectiveismoreusedinhierarchicalmodelsinspatialstatisticswhilestudiesinspatialeconometricsandappliedeconomicfocusonmodelingthecompletesimultaneityofthespatialinteraction[ 8 ].TheConditionalAutoregressiveModels(CAR)reliesonconditionaldistributionofspatialerrorterm.Cressie(1993)[ 38 ]showedthatthedistributionofeiconditioningonfej,i6=jgdependsonlyonneighborsofithobservation.ThepairwisedifferentspecicationofspatialdependencyisdenedbyaGaussianconditionalautoregressive[ 12 ]. 2.3ApplicationContext 2.3.1DistanceTraveledModelingEwingandCervero(2001,2010)[ 41 , 42 ],Handy(2005)[ 55 ],TRB(2009)[ 50 ],Boarnet(2010)[ 14 ],Mooreetal.(2010)[ 80 ]andSalonetal.(2012)[ 95 ]reviewedanddiscussedtheexistingempiricalstudiesmodelingrelationshipbetweenlanduseandvehicleuse.Differentmodelingtechniqueshavebeenappliedtoestimatevehicleuse(distanceortime)suchasordinaryleastsquareregression[ 26 , 83 , 115 ],tobitregression[ 24 ],seeminglyunrelatedregressionmodels[ 2 ],thetwopartmodelwith 37

PAGE 38

instrumentalvariables[ 57 , 107 ],three-stageleastsquares[ 17 , 26 ],latentindexmodel[ 119 ],StructuralEquationModel[ 39 , 47 , 106 ],jointmodelsofvehicletypeandusesuchasmultiplediscretecontinuousextremevaluemodel[ 108 ],copulabasedjointmultinomialdiscrete-continuousmodel[ 99 ],jointmodelsofvehicleownershipandusesuchasmultivariateorderedprobitandtobitmodel[ 16 , 43 ],mixedmultiplediscretecontinuousextremevaluemodel[ 1 ].Despiteinterestincarownershipmodeling,relativelylittleattentionhasbeenpaidtospatialdependency(thecloserthelocation,themoresimilartoeachother)andspatialnon-stationary(thespatialvariationofmarginalimpacts).Bothspatialdependencyandspatialnon-stationaryarelikelytobecommonpropertiesofcarownership.ForinstanceZhangetal.(2012)[ 116 ]foundthateffectivenessofdifferentlanduseplansonvehicle-distance-traveledandpoliciesvariesbothacrossandwithinthecasestudyareas(Seattle,Virginia,Baltimore,WashingtonDC.).Zegras(2010)[ 115 ]suggestedaccountingforspatialvariationincarownershipandusagetoavoidbiasedforecastsandpolicyanalysis. 2.3.2CarOwnershipModelingCarownershipmodelsareappliedtoderivetraveldemand,oilandenergyconsumption,andemissions.InbothActivity-Basedandfour-steptravel-demandmodels,carownershipisconsideredasamedium-termhouseholddecisionthatimpactsdirectlyshort-termday-to-daytraveldecisionsofindividualssuchastripgeneration,andmodechoice.Thecarownershiplevels(obtainedfromautoownershipmodels)togetherwiththeusagepatterns(obtainedfromtraveldemandmodels)determinetheextentoffueluseandemissions.Thus,understandingcar-ownershiplevelsiskeytoeffectivelyassessingpoliciesaimedataddressingbothtrafccongestionandairqualityissues.Inthiscontext,thebroadfocusofthisresearchistounderstandtheeffectsoflandusepatternsonautomobileownership.Tobesure,thereisanextensiveliteratureinthisarea.Auto-ownershipmodelscanbeclassiedtoaggregate-anddisaggregate-modelsbasedonscaleofanalysis.Aggregatelevelstudiesmodelgeographicallyaggregated 38

PAGE 39

ownershiplevelsatmacroscalessuchasward[ 29 ],censustract[ 75 ],city[ 70 ].Incontrast,disaggregatemodelsoperateatthelevelofhouseholdsandcapturetheeffectofhouseholdcharacteristicsontheircar-ownershiplevels.Thedisaggregatemodelscancapturetheunderlyingbehavioralmechanismsthatrevealthehouseholddecisionprocessand,thus,canhelpcapturecausalrelationshipsbetweenexplanatoryvariablesandcarownership.Auto-ownershipmodelscanalsobeclassiedbasedonthetreatmentofthedependentvariable.Somestudiesanalyzedcarownershipasnominalquantity.ThesestudieshaveusedmodelsofRandomUtilityMaximization(RUM)suchasmultinomiallogit[ 21 , 92 , 115 ],multinomialprobit[ 111 ],nestedlogit[ 51 ]andthelatentclassmultinomiallogit[ 4 ].Anotherstructureusedincarownershipanalysisistheordered-response(OR)modelstructure,includingorderedlogit[ 61 ],orderedprobit[ 27 , 48 , 75 ],andthelatentclassorderedlogit[ 4 ].Thesestudiesexplicitlyaccountfortheinherently-orderednatureofthechoicesaboutcarownershiplevels.Somestudieshavecomparedtheunordered(MNL)specicationwithorderedlogitspecication[ 22 , 91 , 101 ].Athirdapproachisonethatexplicitlyrecognizesthecountnatureofthedependentvariable(automobileownershiplevelsofhouseholdarenon-negativeintegers).Ratherthanusingnominalororderedmodelingstructure,theyusecountregressionmodelssuchaspoissonregression[ 98 ],negativebinomialregression[ 96 , 97 ],andpoisson-lognormalmodel[ 60 ].Irrespectiveofthemodelingtypeused,allthemodelshavecapturedtheinuenceofvariousexplanatoryfactorsoncarownership.Thesefactorscanbeclassiedassocioeconomicvariablesandurbandesignindicators.Income,householdsize,householdcomposition(numberofadults,children,andworkers),lifecycle,andhouseholdraceareamongthemostimportantandconsistentlyquotedfactorsinuencingcarownership.Amongurbandesignvariables,effectsofresidentialdensity 39

PAGE 40

[ 101 ],employmentdensity[ 94 ],populationdensity[ 29 , 48 , 65 ],retaildensity[ 94 ],landusediversity[ 91 ],transitaccessibility[ 48 ],andhouseholdlocationtype(urbanversusrural)[ 22 , 61 ]haveallbeenexamined.However,practicallyallexistingmodelsemployaspatialmodelingtechniques.Evenstudiesatnationwideassumetherelationamongsocioeconomicandlandusefactorsonhouseholdcarownershipisconstantacrossallhouseholdsoverthestudyarea[ 48 , 61 , 65 ].Consequently,issuesofspatialdependencyandspatialnonstationarityhavenotbeenadequatelyaddressedinthecontextofcarownershipmodeling.Spatialdependencyimpliesthathouseholdslocatedclosertoeachotherarelikelytohavesimilarcarownershippatternsbecauseoffactorsthatarepotentiallyunobservedtothemodel.Theassumptionofaspatialstationaryprocessimpliesthatthemarginalsensitivityofcarownershiptosocioeconomicandland-useisconstantacrosstheregion.However,thismaynotbethecase.Forinstancevariationincostoflivingmakesobservedincomeandrealincomedifferent.Whiledisposableincomeisnotlikelytobederivedfromexistingdata,itisnecessarytoestimateregionalvariabilityinelasticityofcarownershiptoincometocapturedifferencesinspendingpoweracrosstheregioneveninthesameincomecategory.Thus,almosteverythingweknowaboutsocio-economicandlanduseimpactsoncarownershipareaverageeffects.Atthesametime,itisrecognizedthatspatialnon-stationarityisnot“amerestatisticalnuisance”[ 65 ].Notaccountingforspatialheterogeneitycausesbiasedparameterestimates,misleadingsignicancetests,andsuboptimalprediction[ 11 ].Totheauthorknowledgetherearethreestudiesthathavedevelopedspatialmodelsforcarownership[ 28 – 30 ]andalltheseareaggregatemodelsthatanalyzeimpactofincomeandpopulationdensityataggregatelevel(wards).ThesestudieshaveappliedtheGeographicallyWeightedRegression(GWR),andSpatialErrorModel(SEM)Models;bothofwhicharebettersuitedformodelingcontinuousdependentvariables. 40

PAGE 41

2.4SummaryUnderstandingtherelationshipbetweenland-useandtravelbehaviorisessentialtowardsquantifyingtheimpactsofland-usepolicies.Inparticular,examiningtheeffectsoflandusepatternsonvehicleusageandownershipiscriticalfromthestandpointofassessinggoalsaimedatreducingenergyconsumptionsandemissions.Whilethereisasignicantbodyofliteratureonthissubject,almostallofthestudiesemploystationarymethodsthatdonotallowforvariationsofmarginaleffectsoverspace.Thisstudycontributestotheliteraturebydevelopingageographically-weightedregressionmodelforpersonmiletraveledpresentedinChapter 3 ;mixedgeographicallyweightedregressionforvehicleusage(timeanddistance)presentedinChapter 4 ;andquasigeographicallyweightedPoissonregressionforcarownershippresentedinChapter 5 . 41

PAGE 42

CHAPTER3PERSONMILESTRAVELEDMODELING 3.1OverviewThisstudycontributestotheliteraturebydevelopingageographically-weightedregression(GWR)modelforcapturingtheimpactsoflanduseonperson-milestraveledanddemonstratingitsbenetsoversimplermethods.Travelsurveyandland-usedatafromSouth-EastFloridawereusedinthisanalysis.Theempiricalresultsreconrmthestrongimpactsofregionalaccessibility,land-usemixing,andconnectivityonPMT.Theseland-useeffectswereestimatedtobesignicantaftercontrollingforsocio-economicvariables.Further,theGWRmodeldemonstratesthatthemarginalsensitivitiesofPMTtovariousland-useattributesdovaryoverspace.Thisspatialvariationwasparticularlystronginthecaseoftheeffectofregionalaccessibility.TheempiricalresultsshowthatallowingforexibletrendsintheparametereffectsdoesimprovethemodelsandexplainagreaterproportionofthevarianceinPMTacrosstheregion.ThestudyalsohighlightsthestatisticalsuperiorityoftheGWRmodelovertheglobalregressionmodels. 3.2DataDescriptionThe1999SoutheastFloridaRegionalTravelCharacteristicsStudywhichhasone-day86travelinformationforabout5000householdsistheprimarysourceoftraveldata.Aftercleaningandconsistencychecks,datafrom3504adults(age18years)areusedinthisanalysis.Thissurveyincludesdetailedinformationonalltripsmadebytherespondentsincludingthetrip-endlocations(latitudeandlongitude).Theselocationswereoverlaidonaroadway-networkdatabasetodeterminetheshortest-distancetravelpathsbetweeneachorigin-destinationpair(thereaderwillnotethatconventionaltravelsurveysdonotdirectlycollectdataontravelroutes).Thedistancesofalltripsmadebyapersonwereaggregatedtodeterminethemotorizedperson-miles-traveled(PMT).TheaveragePMTinthesampleis24.87mile(SD=21.123).TheaveragePMTby 42

PAGE 43

countyare25.04miles(SD=23.11)forPalmBeach,24.96miles(SD=19.91)forBrevardand9624.59miles(SD=20.25)forMiamiDade.AspatialproleofthePMTisprovidedinFigure 3-1 .ThemapshowsindividualPMTsmoothedbyinterpolationbetweenindividuallocationsusingtheinversedistanceweighting(IDW)techniquewithincity-limitboundaries.AdarkershadeindicateshigherPMT.Aspatialpatternisclearlyevident.Specically,thePMTislowerforresidentsclosertothecoast(moreurbanizedareas)andhigherforresidentswhoarein-land(moresuburban/rural).Themapalsoshowsthelocationofthemajorcitiesinthethreecountyregioncoveredbythesurvey(Miami,Ft.Lauderdale,PalmBeach,andBocaRaton)andthePMTofhouseholdslocatedwithinthesecitiesaregenerallylower.Inthisstudyseveralmeasuresofland-usearoundtherespondentsresidencesweredetermined.Tables 3-1 , 3-2 ,and 3-3 presentstatisticalsummaryofsomeofthesemeasuresbycountyandintheoverall.Parcel-levellandusedatafromtheFloridaDepartmentofRevenuewereusedforthispurpose.Neighborhoodsweredelineatedbasedonafour-squaremilegridanditscharacteristicsweredeterminedbyaggregatingtheparcel-levelland-usedataandroadway-networkcharacteristicswithintheneighborhood.Measuresofresidentialdensity(units/acre)werecreated.Otherneighborhood-levelmeasuressuchastotalroadmiles,numberofintersectionspermileofroadwayandnumberofcul-de-sacspermileofroadwaywerecalculatedtoexamineimpactoftransportationinfrastructureonPMT.Inaddition,thenetworkdistanceofeachneighborhoodtoeachoffourregionalactivitycenters(oneineachofthefourmajorcities)wascalculatedasameasureofregionalaccessibility.Theregionalactivitycentersweredenedasneighborhoodswiththehighestcommercialsquarefootage(includes,retail,ofce,andentertainment).ThereaderisreferredSrinivasanetal.(2013)[ 100 ]forfurtherdetailsontheassemblyoflandusedata.Thetravelsurveyalsoprovidedsomeimportanttravelercharacteristicssuchasage,employmentstatus,householdcomposition,income,andcarownership.Table 3-2 alsopresentsastatisticalsummaryofthesemeasuresbycountyandintheoverall. 43

PAGE 44

Table3-1. SummaryofresidentialLand-Usecharacteristics VariablesOverallPalmBeachBrowardMiami-Dade MeanStd.Dev.MeanStd.Dev.MeanStd.Dev.MeanStd.Dev.NeighborhoodLandUseCharacteristicsLandUseMixFractionoflandareathatisdeveloped0.520.110.510.120.540.100.510.11Fractionofdevelopedareathatisresidential0.680.150.670.150.660.130.710.16Fractionofdevelopedareathatiscommercial0.060.050.050.040.060.050.060.05FractionofdevelopedareathatisOfce0.030.040.020.030.030.040.030.05FractionofdevelopedareathatisIndustrial0.040.060.030.050.040.060.040.09FractionofdevelopedareathatisInstitutional0.070.110.060.110.140.100.080.12Fractionofdevelopedareathatisother0.120.120.160.140.060.090.070.11DensityGrossresidentialdensity(unitsperacre)2.871.902.501.373.442.422.541.40Netresidentialdensity(unitsperacre)8.245.617.403.519.336.377.785.79DesignNumberofCul-de-Sacspermileofroadway1.520.751.990.781.270.531.290.71Numberofintersectionspermileofroadway8.531.747.791.908.461.569.501.22RegionalaccessibilityDistancetonearestregionalactivitycenter10.935.7910.896.9610.174.5611.955.57 44

PAGE 45

Table3-2. Summaryofpersonlevelandhouseholdlevelcharacteristics TravelerCharacteristicsOverallPalmBeachBrowardMiami-Dade Freq%Freq%Freq%Freq%Nocarsharing(carsadults)269676.994279.6105979.469570.3Carsharing(cars
PAGE 46

Table3-3. Summaryofsocioeconomiccharacteristics TravelerCharacteristicsOverallPalmBeachBrowardMiami-Dade MeanStd.Dev.MeanStd.Dev.MeanStd.Dev.MeanStd.Dev.Numberofchildreninhousehold0.801.050.711.040.781.040.951.08Numberofemployedadultinthehousehold1.680.941.480.971.740.891.840.92Numberofadultinthehousehold2.200.862.080.852.150.782.350.93Age62.98131.6469.37140.5762.16132.9256.45117.94Numberofobservations(Persons)350411831333988 46

PAGE 47

Figure3-1. PMTproleofSouthEast(SE)Floridaresidents 3.3ModelStructureTheperson-milestraveledislinkedtothetraveler-andtheland-use-characteristicsusingthreemodels:(1)Aglobalregressionmodel,(2)Anexogenous-segmentationmodel,and(3)Ageographically-weightedregressionmodel.Theglobalregressionmodelcanbeexpressedas:Yj=0+XKkXjk+j (3) 47

PAGE 48

WhereYjislogarithmofPMTforpersonj(theuseofthelogarithmofPMTasthedependentvariableensuresthattheestimatedPMTisalwayspositive),Xjkarelanduse,transportationandtravelerattributes,andjiserrortermcapturingtheeffectsofunobservedfactorsonlogarithmofPMT.Intheexogenous-segmentation(ES)model,theparametersareallowedtovaryacrossexogenouslydenedspatialboundaries.Forinstance,ifthemodelparametersareallowedtobedifferentacrossthethreecounties,thespecicationmaybestatedas:Yj=0+XKkXjk+j (3)WhereYjislogarithmofPMTforpersonj(theuseofthelogarithmofPMTasthedependentvariableensuresthattheestimatedPMTisalwayspositive),Xjkarelanduse,transportationandtravelerattributes,andjiserrortermcapturingtheeffectsofunobservedfactorsonlogarithmofPMT.Intheexogenous-segmentation(ES)model,theparametersareallowedtovaryacrossexogenouslydenedspatialboundaries.Forinstance,ifthemodelparametersareallowedtobedifferentacrossthethreecounties,thespecicationmaybestatedas:Yj=0+(0,2Cj,2)+(0,3Cj,3)+XKk,1(XjkCj,1)+XKk,2(XjkCj,2)+XKk,3(XjkCj,3)+j (3)IntheaboveequationCj,1,Cj,2andCj,3areindicatorvariablesrepresentingwhetherpersonjislocatedinPalmBeach,Broward,orMiami-Dadecountiesrespectively.Byinteractingeveryparameter()withthesespatialindicatorvariables,theeffectsofeachvariableisallowedtovaryacrossthethreecounties.However,ifthedifferencesintheeffectswerefoundtobestatisticallyinsignicant,theparameterscanbesettobeequaltherebysimplifyingthemodel.Theexogenous-segmentationmodelextendsthesimpleglobalregressionmodelbyallowingtheparameterstovaryacrossthepre-denedspatialboundaries.Further, 48

PAGE 49

boththeglobal-andtheexogenous-segmentation-modelscanbeestimatedusingtheclassicalordinary-least-squarestechniqueusingcommercialsoftware.However,theESmodelcouldresultintwohouseholdssituatedclosetoeachotherbutoneithersideofthepre-denedspatialboundarytohaveverydifferentsensitivitiesofPMTtothesameland-useattribute.Suchabruptspatialchangesinsensitivitiesmaynotbereasonable.Thegeographically-weightedregression(GWR)modelfurtherextendstheexogenous-segmentationmodelbyallowinggreaterexibilityandensuringsmoothnessinthevariationofthemodelcoefcientsoverspace.Theestimationofcoefcientsspecictoeverylocationisaccomplishedbyperforminglocalregressions,eachusingasub-sampleofdatainthevicinityofthelocationunderconsideration.Further,intheestimationofparameterssayatpointj,itisassumedthatdataassociatedwithobservationsspatiallyclosertopointjhavemoreinuence(weight)intheestimationof^k(uj,vj).Theparametersareestimatedusingweightedleastsquares. 3.4EmpiricalResultsThreemodelsforPMTwereestimatedusingeachofthemethodsdiscussedabove.IntheestimationoftheGWR,theadaptiveweightingschemewasusedgiventhatthedataweredenselyclusteredinsomeareasandsparselyinothers.Forinstance,thebandwidthtoinclude100neighborsvariedfrom3,634metersto54,831metersacrossthesample.Anoptimalbandwidthof1223neighborswasdetermined.AGaussianweightfunctionwasused.TheseanalysesweredoneusingRwhichisafreesoftwareenvironmentforstatisticalcomputingandgraphics. 3.4.1StatisticalComparisonsPriortothediscussionoftheempiricalresults,itisusefultopresentastatisticalcomparisonofthethreemodels.Thesumsofsquarederrorsfortheglobalregression,theexogenouslysegmentedmodel,andtheGWRmodelarerespectively2931.73,2884.49and2877.75.TheAkaikeInformationCriterion(AIC)measuresforthesamemodelsarerespectively9332.81,9291.98,and9268.62.Thesemeasuresgivea 49

PAGE 50

generalindicationofthesuperiorityoftheexogenously-segmentedmodeloverglobalregressionmodelandthesuperiorityoftheGWRovertheexogenously-segmentedmodel.Statistically,theglobalOLSandtheexogenously-segmentedmodelscanbecomparedusingthetraditionalFtest.Theexogenously-segmentedmodelsstatisticallytsdatabetterwith.05signicancelevel(F=11.428>2.21).TheGWRmodelscanalsobecomparedtotheOLS-basedmodelsusingFtests.Inoneapproach,theteststatisticF1usestheresidualsumofsquaresanditsapproximateddistributiontotestwhetheraGWRmodeldescribesagivendatasetsignicantlybetterthananOLS.Thesecondteststatistic(F2)usesthemethodofanalysisofvariance.IntheestimatedGWR,F1=0.98andF2=3.29andboththeseresultsverifythatthedatattedbyGWRisimprovedcomparedtoglobal-OLSstatisticallywith0.05signicancelevel.TheGlobalMoranIndex(I)valuesforglobalregressionandexogenouslysegmentedregressionarerespectively0.009and0.003andstatisticallysignicantindicatingspatialauto-correlationintheerrors.TheGlobalMoranfortheGWRmodelresidualsturnsouttobe0.001andstatisticallyinsignicantwhichindicatesmodelresidualsarerandomlydistributedacrossspace.Insummary,thestatisticalcomparisonofthemodelsestablishesthattheGWRdoesimprovethemodeltandreducesspatialauto-correlationintheerrorterms. 3.4.2AverageMarginalEffectsTable 3-4 presentsasummaryofempiricalresultsforthethreemodels.Intheexogenous-segmentationmodel,eachcoefcientwasallowedtovaryacrossthethreecounties;howeverwhentheinter-countydifferenceswereinsignicant,theparameterswerexedtobeequal.InthecaseoftheGWRmodels,onlythemeanvaluesofthecoefcientsarepresentedinTable 3-4 .Thespatialvariationofthesecoefcientsisdiscussedlateron.TheresultsshowthatPMTisaffectedbyseveralland-useandsocio-economicvariablesand,ingeneral,thedirectionalityoftheeffectsisconsistentacrossthemodelsandisintuitivelyreasonable.Theimpactofthefraction 50

PAGE 51

Table3-4. Summaryofempiricalmodelresults(meaneffects) GlobalRegressionExogenousSpatialSegmentationGWR VariablesPalmBeachBrowardMiami-DadeCoefcienttCoefcienttCoefcienttCoefcienttCoefcientFractionofdevelopedareathatiscommercial-1.09-3.27-1.18-3.52-1.18-3.52-1.18-3.52-1.01Distancetonearestregionalactivitycenter0.026.26-(*)-0.045.260.023.540.03Numberofintersectionspermileofroadway-0.05-5.01-0.10-6.38-0.02-3.12-0.02-2.48-0.04Numberofcul-de-sacspermileofroadway0.052.200.093.700.093.700.093.700.09Nocarsharing(carsadults)0.112.890.092.670.092.670.092.670.11Fulltimeemployed0.315.420.315.450.315.450.315.450.28Parttimeemployed0.182.580.192.570.192.570.192.570.22Retired-0.16-2.38-0.15-2.21-0.15-2.21-0.15-2.21-0.20Highincome(income80K)0.143.69--0.202.310.191.950.18Constant2.7122.373.2119.682.2715.962.3517.542.52Numberofobservation350435043499SumofSquareError2931.732884.492877.75AIC9332.819291.989268.62GlobalMoranCoefcient0.0090.0030.001 (*)-Indicatesinsignicantvariable 51

PAGE 52

ofdevelopedareaaroundtherespondentsresidencethatiscommercialisnegativeandsignicantinallmodels.Themoretheneighborhoodiscommercial,thegreateraretheopportunitiestoparticipateinactivitieslocallyleadingtoareductionofoverallPMT.Theexogenously-segmentedmodelindicatesthattheeffectofthisvariableisnotstatisticallydifferentacrossthethreecounties.Thedistancetothenearestmajoractivitycenter(i.e.,ameasureofregionalaccessibility)wasfoundtobestrongdeterminantofPMT.Specically,householdsthatarelocatedfatherfrommajoractivitycenterstravelmore.Theexogenously-segmentedmodelindicatessignicantvariationintheeffectofthisvariableacrossthethreecounties.TheeffectismorepronouncedinBrewardCountyandstatisticallyinsignicantinPalmBeachCounty.Numberofintersectionsandcul-de-sacspermileofroadwayhasastatisticallysignicantimpactonPMT.Thegreaterthenumberofintersectionspermileofroadwayimpliesgreaterconnectivity(i.e.,shorterdistancestotravelbetweenanytwolocations)and,hence,alowerPMTmaybeexpected.Alternatively,morecul-de-sacsimplygreatercircuity(i.e.,longerdistancestotravelbetweenanytwolocations),andhence,agreaterPMTcanbeexpected.Theempiricalresultsareconsistentwiththeseexpectations.ThemarginalimpactofnumberofintersectionspermileofroadwayalsovariesacrossPalmBeach,Broward,andMiami-Dade.Theeffectofthenumberofcul-de-sacswasfoundtobestatisticallythesameacrossthethreecounties.Althoughanextensivesetofland-usevariableswereexplored(SeeTable 3-1 ),onlytheabove-mentionedfactorsturnedouttobestatisticallysignicant.Forinstance,wendthatresidentialdensitywasstatisticallyinsignicant.Onepossibleexplanationisthattheotherland-use/urbanformvariablesincludedinthemodelarealsoreectingtheresidentialdensityofthehouseholdlocation(forinstance,lowerdensityareasaremorelikelytohavemorecul-de-sacs). 52

PAGE 53

Themarginaleffectsoftheland-usevariablesweredeterminedaftercontrollingfortheeffectsofseveralsocio-economicfactors.Adultslivinginnon-car-sharinghouseholds(householdshavingatleastasmanycarsasadults)havehigherPMT.AccessibilitytocarsfacilitatespeoplestravelwhichresultsinhigherPMT.Ontheotherhand,theneedtoshareacarcanrestrictmobility.FulltimeandparttimeemployeeshavehigherPMTcomparedtounemployedadults.ButretiredpeoplehavelesserPMTcomparedtounemployedadults.Theseresultsreectthesignicantcontributionofcommutingtoaperson'soverallPMT.Peoplelivinginhigher-incomehouseholdstravelmore(greaterPMT)comparedtoothers.Ingeneral,theeffectsofallsocio-economicvariableswerefoundtobethesameacrossthethreecountiesintheexogenouslyspatially-segmentedmodel(exceptinthecaseofhouseholdincome,whichwasnotsignicantforPalmBeach). 3.4.3SpatialVariationintheMarginalEffectsTheanalysisthusfarfocusedonthestationarycoefcientsandthemeaneffectsfromtheGWRmodels.NextweexaminethespatialvariationofthemodelcoefcientsfromtheGWRmodels.Table 3-5 presentsasummaryofthevariationoftheGWRmodelcoefcientsoverthesample(ThemeanvaluesinthistableisthesameasthemeanvaluesfortheGWRmodelpresentedinTable 3-4 ).Itisimportanttonotethatnotalllocalizedcoefcientsarestatisticallysignicant.ThecolumntitledPercentageofsignicanteffectsidentieshowmanyoftheestimatedcoefcientsaresignicant.Forexample,theeffectofthefactionofdevelopedareathatiscommercialissignicantatonlyabout86%ofthelocationswhereastheeffectofdistancetonearestactivitycenterissignicantatall(100%)locations.Allsummarymeasuresonthemarginaleffects(suchasmean,range,variance,andquartiles)inTable 3-5 arebasedonstatistically-signicantvalues(95%signicance)only.Whereversignicant,themarginaleffectsretainthesamedirectionalityofimpact.However,thereisconsiderablenumericalvarianceintheestimatedcoefcients.F3 53

PAGE 54

Table3-5. SummaryofGWRmodelresults VariablesPercentsignicantMinLQMedianMeanUQMaxVar.F(3)TEST Fractionofdevelopedareathatiscommercial86.34-1.150-1.061-1.020-1.006-0.964-0.8010.00650.21Distancetonearestregionalactivitycenter100.000.0080.0250.0310.0290.0360.0390.000110.31*Numberofintersectionspermileofroadway75.05-0.070-0.058-0.03-0.043-0.033-0.0270.00023.35*Numberofcul-de-sacspermileofroadway59.130.0530.0830.0960.0920.1050.1120.00022.87*Nocarsharing(carsadults)65.280.0940.1040.1110.1100.1170.1250.00010.36Fulltimeemployed100.000.2280.2540.2770.2870.3170.3610.00171.03Parttimeemployed37.150.1750.2080.2320.2250.2420.2480.00040.73Retired39.70-0.210-0.213-0.200-0.201-0.193-0.1710.00010.39Highincome(income80K)80.820.0880.1430.1910.1760.2110.2250.16804.20*Constant2.2822.3432.4152.5292.7083.063Numberofobservations3499SumofSquareError(SSE)2877.70AkaikeInformationCriterion(AIC)9268.60GlobalMoranCoefcient0.001 *-IndicatessignicantspatialvariationsinGWRcoefcientsofthesevariablesat5%levels 54

PAGE 55

statisticaltestwasappliedtodeterminewhethertheestimatedGWRparametersexhibitsignicantvariationoverthestudyregion.Specically,anFtestisusedtotestthenullhypothesisofequalityofcoefcientsoverspace.TheF3(notethatthesubscript3issimplyusedtodistinguishthisFtestfromthetwopreviousFtests)valueforeachcoefcientispresentedinTable 3-5 .Onlyfouroftheestimatedmarginaleffectshavestatistically-signicantvariabilityoverspace.Threeoftheseareland-use/urbanformvariables(distancetonearestregionalactivitycenter,numberofintersectionspermileofroadway,andnumberofcul-de-sacspermileofroadway)andoneisasocio-economicvariable(income).Thispresenceofstatistically-signicantspatialvariationintheimpactsofkeyland-usedescriptorsonPMTisimportantfromapolicystandpoint.Theonlyland-usevariablethatdidnothaveastatistically-signicantspatialvariationintheGWRmodelisthefractionofdevelopedareathatiscommercial.Itisusefultonotethatthisvariablewasestimatedtohavesamecoefcientsacrossthethreecountiesintheexogenously-segmentedmodelaswell(SeeTable 3-4 ).Theimpactofregionalaccessibilityshowsstrongvariabilityoverthestudyarea.Interestingly100%ofthelocalcoefcientsonthisvariableturnedouttobestatisticallysignicantaswell.Inter-countyvariationsinthemarginaleffectswerealsoobservedintheexogenously-segmentedmodel.Theeffectofthenumberofintersectionspermileofroadwayhasasignicantspatialvariationthatisalsoevidentfromtheexogenously-segmentedmodel.Whileimpactofnumberofcul-de-sacspermileofroadwaywasshowntovaryspatiallybytheGWRmodel,theexogenously-segmentedmodeldoesnotcapturethisvariability.Theonlysocio-economicvariableintheGWRmodelthatisestimatedtohaveaspatialvariabilityinitseffectishouseholdincome.Thiswasalsothecaseintheexogenously-segmentedmodel.Figure 3-2 presentsthespatialvariationofthemarginalimpactsofthevariablesfromtheGWRmodel.Thegureincludesonlythefoureffectswhichwereestimated 55

PAGE 56

tohaveastatistically-signicantspatialvariation.Thecolorsbreakonquartilevalueofmarginalimpactsanddarkershadesrepresentgreaterimpacts.Onexamining Figure3-2. Spatialproleofmarginalimpactsforstatistically-signicanteffects thespatialvariationontheimpactofregionalaccessibility,wendthatthehighestcoefcientsareobservedinBrowardCountyandthelowestinPalmBeachCounty.Quiteinterestingly,theexogenous-segmentationmodelalsosuggeststhatsametrend(themarginalimpactinBrowardandMiami-Dadearerespectively.041and.026andtheeffectisinsignicantinPalmBeach). 56

PAGE 57

Inthecaseoftheimpactofintersections,theGWRmodelsindicatethatthegreatesteffectsareinPalmBeachCountywithBrowardandMiami-Dadehavinggenerallycomparableeffects.Again,theexogenoussegmentationmodelreectsthesameprole(Thecoefcientsare-0.1,-0.024,and-0.021inPalmBeach,Broward,andMiami-DadeCountiesrespectively).Unlikeinthecaseoftheprevioustwoeffects,theGWRandtheexogenously-segmentedmodelsdonotreectthesametrendsintheimpactsofthenumberofcul-de-sacs.Accordingtotheformermodel,thelocationsofhighestimpactsareinpartsofBrowardandMiami-DadewiththeeffectbeinggenerallyinsignicantinPalmBeachCounty.BrowardCountyhaslocationswithalllevelsofimpact.Theexogenously-segmentedmodelshowednostatistically-differentvariationsacrossthecounties.Thisindicatesthatinsomecases,thecounty-boundariesmaynotbethenaturalbreak-pointtoallowforstructuralchangesinthemodel.Theonlysocio-economicvariablewithanestimatedspatialvariationinitsimpactfrombothexogenous-segmentationandGWRmodelsishouseholdincome.Intheformermodel,theeffectofincomeisinsignicantinPalmBeach,andsignicantandcomparableinBrowardandMiami-DadeCounties.TheGWRalsosuggestsasimilartrend;however,agradualprogressionofincreasingeffectsfromNorthtoSouth(PalmBeachtoMiami-Dade)isalsoobserved. 3.4.4VarianceDecompositionandImplicationsThesimple,global-regressionmodelcomputesstationarymodelparametersbyassumingthattheelasticity/marginalresponsestotheexplanatoryvariablesarexedoverthespaceofthestudyarea.Whilethisisagoodbenchmarkforaggregate/generalresponseevaluation,itmightnotreectvariabilityinlocalizedimpactsoverspace.SincetheGWRmodelscapturethislattereffect,itwouldbeusefultoquantifytheextenttowhichcapturingspatialvariationincoefcientsisimportanttoquantifyinglocalizedpredictionsofPMT. 57

PAGE 58

Toaddressthisissue,thevariance-decompositionanalysisusedbyAlietal.(2007)[ 3 ]isadopted.ThevariabilityinthepredictedPMTfromaGWRmodelisaffectedbybothvariationsintheexplanatoryfactorsandthevariationsinthemodelcoefcients.Inaglobalregression,thePMTvariabilityisonlyaffectedbyvariationinexplanatoryfactors(thecoefcientsarexed).Therefore,decomposingthevarianceoftheimpactsofeachexplanatoryvariableintoaparametereffectandvariableeffectisuseful.IftheproportionofpredictedvariabilityinPMTduetothespatialheterogeneityofcoefcients(theparametereffect)isgreater,thisisevidencefortheneedforGWRoveraglobal-regressionmodel.Table 3-6 showsthedecompositionofthepredictedimpactonPMTintotheparametereffectandthevariableeffect.Therowsofthetablecorrespondingtovariablesthathaveastatisticallysignicantspatialvariancearearemarkedwith(*).Theratiooftheparametereffecttothevariableeffectisthehighestforthesefourvariables.Inparticular,theratiosarehighforthethreeland-usevariables(regionalaccessibilityandnumberofintersectionsandcul-de-sacs).TheseresultshighlightthatallowingparameterstovaryacrossspacedocontributesubstantiallytoexplainingthevariationinPMTacrosstheregionbeyondwhatisexplainedbyvariationsintheexplanatoryfactors.Therefore,allowedfortheestimationoflocalizedcoefcientsdocontributetobetterestimatesofPMTthansimplerglobalregressionmodels. 3.5SummaryUnderstandingtherelationshipbetweenland-useandtravelbehaviorisessentialtowardsquantifyingtheimpactsofland-usepolicies.Inparticular,examiningtheeffectsoflandusepatternsonPMTiscriticalfromthestandpointofassessinggoalsaimedatreducingenergyconsumptionsandemissions.Whilethereisasignicantbodyofliteratureonthissubject,almostallofthestudiesemploystationarymethodsthatdonotallowforvariationsofmarginaleffectsoverspace.Ourstudycontributestotheliterature 58

PAGE 59

Table3-6. Decompositionofvariance Variable(X)Parameters()Ratioofparametertovariableeffect VariablesMeanSDMeanSDFractionofdevelopedareathatiscommercial0.060.05-1.010.080.02Distancetonearestregionalactivitycenter10.935.80.030.010.31*Numberofintersectionspermileofroadway8.531.74-0.040.012.34*Numberofcul-de-sacspermileofroadway1.520.750.090.010.54*Nocarsharing(carsadults)0.770.420.110.010.01Fulltimeemployed0.690.460.290.040.07Parttimeemployed0.100.30.220.020.01Retired0.130.33-0.200.010.01Highincome(income80K)0.210.410.180.040.07* 59

PAGE 60

bydevelopingageographically-weightedregressionmodelforPMTanddemonstratingitsbenetsoversimplermethods.Travelsurveyandland-usedatafromSouth-EastFloridawereusedinthisanalysis.TheempiricalresultsreconrmthestrongimpactofregionalaccessibilityonPMT.Further,themodelsdosupportthevalueofmixed-usedevelopmentsasresidenceslocatedinareasofhighercommercialestablishmentsdohavealowerPMT.Connectivityprovidedbytheroadwaysystemalsoimpactsthemagnitudeoftravelgeneratedbythehouseholds.Theseland-useeffectswereestimatedtobesignicantaftercontrollingforsocio-economicvariablessuchashouseholdincome,carownership,andemploymentstatus.ThesubstantivecontributionofthisstudyisindemonstratingthatthemarginalsensitivitiesofPMTtovariousland-useattributesdovaryoverspace.Thisspatialvariationwasparticularlystronginthecaseoftheeffectofregionalaccessibility.WhilesomeofthebroadtrendsinspatialvariationscapturedbyGWRcouldbereplicatedwithasimplerexogenous-segmentedmodel,theempiricalresultsalsoshowthatallowingforexibletrendsintheparametereffectsdoesimprovethemodelsandexplainagreaterproportionofthevarianceinPMTacrosstheregion.ThestudyalsohighlightsthestatisticalsuperiorityoftheGWRmodelovertheglobalregressionmodels.Thestudyalsoshowsthatnotalleffectshaveastatistically-signicantvariationoverspace.Theglobal-andexogenouslysegmentedmodelsareeasytodevelop,interpret,andapply.However,thedevelopmentofGWRmodelsinvolvesconsiderableeffortandcaretoensurethattheresultsareintuitivelyreasonableandstatisticallyrobust.Therefore,futurestudiesshouldseektodevelopmodelsthataccommodateanyregionalvariations(orlackthereof)viaacombinationofglobal,exogenouslysegmented,andGWRparameters. 60

PAGE 61

CHAPTER4VEHICLETIMEANDDISTANCETRAVELEDMODELING 4.1OverviewIntherecentyears,thereisanincreasinginterestintheapplicationofspatialmodelsfortransportationproblems.Onesuchpopularmodelisthegeographicallyweightedregression(GWR)whichextendstheconventionalregressionmodelbyallowingtheparameterstovaryoverspace.TheneedtouseGWRmodelsisgenerallymotivatedbytheinherentspatialclusteringintravel-patternsandthesuperiorityoftheGWRmodelsoverconventionalregressionisestablishedusingstatisticalmeasures(mostlyF1andF2).However,GWRmodelsrequiretheestimationofaverylargenumberofparametersandassuchnotefcient.Further,toourknowledge,practicallyallstudiesthatreportGWRmodelssimplyreportallcoefcientstobespatially-varyingwithoutexaminingwhetherthespatial-differencesareindeedstatisticallysignicant.Atthesametime,ourpastresearch[ 86 ]hasshownthatsensitivitiestocertainvariables(especiallysocio-economicfactors)donothavestatistically-signicantspatialvariationswhileothers(especiallyland-useandtransportationfactors)do.Inthiscontext,thisstudypresentsmixed-GWRmodelswhichallowforacombinationofspatially-xedandspatially-varyingparameters.Themethodologicalinnovationisappliedinthecontextofmodelingvehicle-distance-traveled(VDT)andvehicle-time-traveled(VTT),twoimportantmeasuresofdailyvehiculartravelvolumes.Whiletheeffectsofland-useonthesetransportationsystemperformancemeasureshavebeenstudiedveryextensivelyusingglobal/aspatialmodels,empiricalinsightsintospatialvariabilityinthesensitivityofparametersarelimited.Therefore,theapplicationofmixed-GWRmodelsinthiscontextistimelyandwouldsupporttheevaluationoflocalor“context-specic”strategiesaimedatreducingvehiculartrafcvolumes,fuelconsumption,andemissions. 61

PAGE 62

4.2DataDescriptionThestudyareacoversthethree-countyregion(Miami-Dade,Broward,andPalmBeach)inSouthEastFlorida.The2009NationalHouseholdTravelSurvey(NHTS)istheprimarysourceoftravelbehaviordata.Table 4-1 presentsasummaryofthesesocioeconomicmeasures.Thissurveyincludesdetailedinformationonalltripsmadebytherespondentsincludingthemode,thepresenceofotherhouseholdmembersastripcompanions,trip-endlocations(latitudeandlongitude)andthestart-andend-times.Thedailyvolumeofvehiculartravelgeneratedbyahouseholdcanbemeasured Table4-1. Socioeconomicexplanatoryanalysis HouseholdCharacteristicsFreq% 1vehiclehousehold129537.92vehiclehousehold148043.43ormorevehiclehousehold63618.6Presenceofchild0-52768.1Presenceofchild6-1541612.2Presenceofchild16-211775.2Householdrace:White280182.1AfricanAmerican2667.8Other34410.1Highincomehousehold(income80K)101929.9Mediumincomehousehold(40Kincome<80K)94627.7Detachedsinglehouse218163.9Duplex3279.6Rowhouseortownhouse84924.9Apartment,condominium441.3Otherhousingunit100.3Householdowningthehouse307390.1Householdinurbanizedarea(notrural)329296.5Travelsurvey:Saturday47113.8Sunday48014.1Monday47513.9Tuesday50614.8Wednesday48614.2Thursday49814.6Friday49514.5 intermsofbothtraveltimeandtraveldistancebysummingupthecorrespondingvaluesforeachhouseholdvehicle.Sincethetraveltimesaredirectlyreportedinthesurveys 62

PAGE 63

foreachtrip,andthevehicleusedforeachtripisknown,estimatesofhouseholds'vehicle-time-traveled(VTT)canbeeasilyobtainedbyaggregation.Itisensuredthatthelengthsoftripsmadejointlybyhouseholdmembersarenotdouble-counted.Inordertocalculatethetriplengthsbasedontherecordedtrip-endlocations,supplementaldataontheroadway-networkcharacteristicsareneeded.Forthisstudy,weusedthe2010statewideNAVTEQroadnetworkle.Thetrip-endlocationswereoverlaidontheNAVTEQroadnetworktodeterminetheshortest-distancetravelpathsbetweeneachorigin-destinationpair.The“NetworkAnalyst”toolsetwithinArcGISwasusedandextendedforthispurpose.Thedistancesofalltripsmadebyhouseholdmembersasdriver(usinghouseholdcar)wereaggregatedtodeterminethemotorizedhouseholdvehicle-distance-traveled(VDT).ViolinplotsoftheVDTandVTTareprovidedinFigure 4-1 .Aviolinplotisadatavisualizationtechniquethatcombinestheboxplotandkerneldensityplot(see[ 58 ]formoredetails).TheaverageVDTinthesampleis33.91miles(SD=38.84)andtheaverageVTTinthesampleis94.93minutes(SD=78.47minutes).TheverticaldashedlinerepresentsthemeanofVDTandVTT.ThebivariatedistributionofVDTandVTTisplottedinFigure 4-2 throughbagplot.ABagplotisabivariategeneralizationofboxplottoexplorelocation,spread(theareaofthebag),correlation(theorientationofthebag),skewness(theshapeofthebag),andtail(outliers)ofthedata[ 93 ].Halfofthedataaroundmedianliesinthedarkbluebag.Therestofthepointsexcludingoutliersarelocatedinlightbluebagcalledfence(similartowhiskersinboxplot).Tocontrolforoutliersthespeedofeachobservationwascalculatedsotheobservationswithspeedsover85milesperhourandlessthan3milesperhourfromtheoutlierswereremoved.Checkingthespeedofmostoutliersdidnotrevealanyunusualspeedrangesoforouranalysiswekeptmostofthedata.Thespeedrangeofthenaldatasampleisintherangeof3.4to83.65milesperhour.TheVDT,VTTbagplotisnicelybalanced(notskewed),anditsshapesuggestanellipticdistribution.Theorientation(slopesupward) 63

PAGE 64

andshape(area)ofhebagimpliesthatVDTandVTTarehighlylinearlyassociated.The0.82Pearsoncorrelationbetweenthemveriesit. Figure4-1. ViolinplotsofVDT(inmile)andVTT(inminute) Anotherthingtobeconsideredisthatmoststudiesonlyconsiderimpactofurbandesignondistancetraveled.Whiledecreasingdistancetraveledmightnotbeassociatedwithdecreaseinvehicleusage.Asinmoredenseareasduetotrafccongestionand 64

PAGE 65

Figure4-2. BagplotofVDTandVTT lowerspeedlimittravelersspendtimeintrafcwithoutmobility.Figure 4-10 veriesthatasdensityincreases,vehicletimetraveldecreases.Therefore,itiscanbeinferredthatinhigherdensity,thevehicleuse(bothtimeanddistance)decreases.Florida2009NHTSadds-onwassampledinalldaysofweekincludingweekdaysandweekends.Figure 4-3 showshowhouseholdVDTandVTTvariesacrossthedays.Itisevidentthat 65

PAGE 66

onaveragehouseholdsusecaronweekendslessthanweekdays.Sundayisthedaythathouseholdsusecartheleast. Figure4-3. HouseholddailyVTTandVDTatdifferentdays Landusepatternscanbeclassiedbasedonscaleofanalysistomesoandmicrolevel.Mesolevel(intra-metropolitanscale)measuresdealwithrelativelocation(suchasdistancetotheCBDorcitycenter)whilemicrolevelmeasuresdifferentiateamong 66

PAGE 67

neighbours.Inthisstudybesidesmicrolevelurbandesign(neighborhoodlevelurbandesign),mesoscaleurbandesignindicator,thenearestnetworkdistancetomajorregionalactivitycenterswerederivedaswelltoanalyzeimpactofbothneighborhoodlevelandregionalaccessibilityofhouseholdlocationontheirtravelpattern. 4.2.1MicroLevelUrbanDesignPatternVariousmicrolevelbuiltenvironmentmeasureswerecreatedtoexaminetheirimpactsonhouseholdVDT,andVTT.Thesemeasurescanbeclassiedintodensity,design,diversity,andtransitaccessibilityfactors.Sincethereisnotastandardformeasuringeachfactor,differentmeasureswerecreatedineachgroupforourstudy.Differentspatialscaleshavealsobeenused(censusblock,censusblockgroup,tract,1 4,1 2,and1milebufferaroundthehouseholdlocations).Someofpreviousstudiespointedoutthatlandusepatternsatdifferentspatialscaleandresolutioncanresultindifferentestimatesandsomeurbandesignfactorsturntobesignicantatspecicspatialscale[ 115 , 116 ].Therefore,inadditiontoderivingdifferentmeasuresineachclass,eachmeasurewascreatedatdifferentspatialresolutions. 4.2.1.1DensityLandusevariableswerecompiledfromdifferentsources.The2010censuspopulationandhousingunitblockandtractshapelewhichprovidesnumberofpopulationandhousingunitsatcensusblockandtractlevelistheprimarysourceofpopulationandresidentialdensitydata.Densityreferstopopulation,employment,dwellingunits,buildingoorareapereithernetorgrossunitofarea.Itcanbemeasuredatdifferentspatialscales:national,regional,county,TAZ,censustract,blockgrouporblock.Density,themostcommonvariablebeingused,istheproxyofbigspectrumofotherlandusevariables.Forinstance,ithasbeenusedsolelyinthestudiesatlargegeographicalscale(e.g.cities,counties,...)basedontheassumptionthatitisacombinatorialindicatorofalllandusemeasures.Higherdensitytendstoincreaseproximityandnumberofdestinationsinanareawithhigherlandcostandreduced 67

PAGE 68

parkingsupplyandhigherparkingpricing.Higherdensityincreasesthecostefciencyofnon-motorizedinfrastructureandfacilitiessuchassidewalk,bikelane,andtransitservices.Thereforeresidentsofhigherdensitycommunitieshavebettertransitaccessibility,bikelane,andsidewalk.Tosumup,densityiscorrelatedwithotherlandusepatterns.EwingandCervero(2010)[ 42 ]calleddensityasintermediatevariabledenedbyotherurbandesignfactors.Whileresidentialorpopulationdensityhasbeenusedinliteratureasproxyforurbandesignforminlackofotherurbanformmeasures,thisapproachhasbeencriticizedbysomescholarsthatitisnecessarytoincludedifferentfactorstohavearealisticimpactofeachlandusefactor.TheybelieveelasticityofdensityinimpactingtravelbehaviorsuchasVDTismostlybecauseofotherfactors[ 25 , 94 ].Indeedwithoutconsideringotherfactors,derivedelasticityfordensityisoverestimated.Inthisstudybyusingdensitymeasuresalongwithotherurbandesignmeasuresthetrueorunbiasedimpactofdensityisexplored.Employmentdensityatcensusblock,blockgroup,andtractlevelisobtainedfrom2009LongitudinalEmployerHouseholdDynamics(LEHD).Alongwithemploymentandresidentialdensity,MixedDensityIndex(MDI)wascalculatedby:MDIi=EDiRDi EDi+RDi (4)WhereEDiandRDirepresentstheemploymentdensityandresidentialdensityrespectively.MDIiwasderivedatcensusblockandtractlevelforeachhousehold.Ingeneralthehigherresidentialandemployment,arerelatedtothehigherlandprice,parkingcarcost,andhighertransituse.Itrepresentstheimpactoflandprice,andaccesstotheactivitylocations,andjob-housingbalanceonhouseholdautomobileusagepatterns. 68

PAGE 69

4.2.1.2DesignThe2009TIGER/Lineshapeleswasusedtocalculateroadwaydesignandconnectivityindicatedbyintersectiondensity,cul-de-sacsdensityandConnectedNodeRatio(ratioofnumberofintersectionstothenumberofcul-de-sacsplusnumberofintersections)within1 4,1 2,1milebuffer,ofeachhouseholdlocation.Lesscul-de-sacs,higherintersections,andhigherCNRareproxyofhigherconnectivity,grid-like,shortblocklengthdecreasetriplengthbyloweringdistancebetweendestinationsandencouragenon-motorizedtripmodes(walking,biking)comparedtostreetlayoutswithmanycul-de-sacsandlongblocklengthneighborhoods. 4.2.1.3DiversityLandusemixreferstoallocatingdifferenttypeoflanduseinacloseproximity.NewUrbanismandneo-traditionalplanningconceptsadvocatesmixoflandusedevelopmentresultsinasustainablecommunitythatreducemotorizetriplength.Inthisstudyseveralmeasuresoflandusearoundtherespondentresidencesweredetermined.Parcel-levellandusedatafromthe2009FloridaDepartmentofRevenue(FDOR)wereusedtoderivediversityindicators.TheFDORlandusecategoryisclassiedinto100classes.Thelandparcelswereaggregatedintosixmajorcategories:residential(singlefamily,multi-family,mobilehomes),commercial(largeretail,regularretail,conveniencestore,drive-through),ofce(professionalandnon-professionalservicesbuilding),industrial(light,heavy,warehousing),institutional/publicadministration(privateschoolandcolleges,hospitals,military),andother(forests,parks,recreationalarea,touristattractions,nightclubs).Landusetypeareawascalculatedforeachhouseholdina1 4milebufferzonearoundeachhouseholdlocation.EntropyIndex(EI)wasusedtomeasurethedistributionofthelandusesaroundthehousehold.EIisdenedasEI=Xkpkln(pk) ln(k)k=1,...,6 (4) 69

PAGE 70

Wherepkistheproportionofthedevelopedlandinthekthlandusetype.ln(k)normalizestheEIindexsoitvariesintherangeofzerotoone. 4.2.1.4TransitaccessibilityNumberofbusstopsandtotaltransitlinelengthwithina1 2and1milebufferzonearoundeachhouseholdlocationandEuclideandistancetothenearestbusstationwerecreatedtorepresenttransitaccessibilityofeachhouseholdinthestudyarea. 4.2.2MesoLevelUrbanDesignPatternMesolevelurbandesigncapturesimpactofthelocationoftheneighborhoodwithintheregion.Thenetworkdistanceofeachhouseholdlocationtoeachoffourregionalactivitycenters(oneineachofthefourmajorcitiesincludingMiami,FortLauderdale,BocaRaton,andWestPalmBeach)wascalculatedasameasureofregionalaccessibility.Theregionalactivitycentersweredenedasneighborhoodswiththehighestcommercialsquarefootage(includes,retail,ofce,andentertainment).Tables 4-2 ,and 4-3 presentastatisticalsummaryofurbandesignmeasures. 4.3ModelStructureTheVDTandVTTareeachlinkedtothehousehold-andtheland-use-characteristicsusingtheMixedGeographicallyWeightedRegression(MGWR)model.TheMGWRextendstheGWRbyallowingsomecoefcientstobexed(global)whileothersvaryacrossthestudyarea[ 78 , 79 , 110 ].TheMGWRhasincreaseddegreesoffreedomandimprovedefciencyofestimatorsofcoefcients[ 110 ]relativetotheGWR. 4.4EmpiricalResultsForeachofVDTandVTT,threemodelswereestimated:theglobalregression,theGWR,andthemixed-GWR.TheGWRwereusedasthestartingpointfortheestimationofthemixed-GWR.Tables 4-4 ,and 4-5 showtheresultofGWRandF3teststatisticsforVDTandVTT.IntheestimationoftheGWR,andMGWR,theGaussianadaptiveweightingschemewasusedgiventhatthedataweredenselyclusteredinsomeareasandsparselyinothers.ToestimateaMixedGeographicallyWeighted 70

PAGE 71

Table4-2. Meanurbandesigncharacteristics UrbanDesignIndicators1 4milebuffer1 2milebuffer LandUseMixResidentiallanduse(acre)37.50106.30Commerciallanduse(acre)3.4025.40Ofcelanduse(acre)1.5010Industriallanduse(acre)1.609.40Institutionallanduse(acre)8.7062.10Otherlanduse(acre)7.9051.10Fractionoflandareaintheneighborhoodthatisdeveloped0.480.53Fractionofdevelopedareathatisresidential0.620.42Fractionofdevelopedareathatiscommercial0.060.10Fractionofdevelopedareathatisofce0.030.04Fractionofdevelopedareathatisindustrial0.020.03Fractionofdevelopedareathatisinstitutional0.150.23Fractionofdevelopedareathatisother0.120.18EntropyIndex(EI)0.370.60DesignNumberofintersections28.90112.50Numberofcul-de-sacs6.1023.81ConnectedNodeRatio(CNR)0.820.82AccessibilitytotransitNumberofbusstop8.70Totaltransitlinelength(ft)10623Euclideandistancetonearestbusstop(meter)1464.70RegionalaccessibilityDistancetonearestactivitycenter(mile)11.40Distancetonearestresidentialcenter(mile)9.30 Table4-3. Meanurbandensityexplanatoryvariables UrbanDensityIndicatorsCensusblockCensustract Residentialdensity(unitspersquaremile)3020.62907.8Populationdensity(persquaremile)64185895Employmentdensity(jobspersquaremile)2754.52224.3MixedDensityIndex(MDI)1188.11179.1 71

PAGE 72

Regression,rstitshouldbespeciedwhichcoefcientsarexed.InthisstudyweusedadatadrivenapproachbyapplyingLeungetal.(2000)[ 68 ]teststatisticstoexaminethespatialvariationofthemodelcoefcients.Tables 4-4 ,and 4-5 presentasummaryofthevariationoftheGWRmodelcoefcientsoverthesample.Leungetal.(2000)[ 68 ]developedstatisticaltestsfortestingwhetherestimatedGWRparametersexhibitsignicantvariationoverthestudyregion.Theirteststatisticsisbasedonthevarianceoftheestimatedcoefcientsandtheestimatederror.F3teststatisticsexaminesspatialvariationofcoefcientsofeachexplanatoryvariableandtestwhetherestimatedcoefcientsforanexplanatoryvariableisxedacrossthestudyareaornot.F3valuesgreaterthanoneshowsignicantspatialvariabilityofestimatedmarginalimpacts.Theresultshighlightpresenceofstatisticallysignicantspatialvariationintheimpactsofland-usedescriptorsonVDTandVTT.Theonlyland-usevariablethatdidnothaveastatistically-signicantspatialvariationintheGWRmodelisconnectednoderatioforVDTanddistancetonearestregionalactivitycenterforVTT.TheotherlandusevariablesinbothVDTandVTTdepictspatialvariationacrossregion.EntropyIndexandfractionofdevelopedareathatisinstitutionalshowsspatialvariabilityforbothVDTandVTT.ByxingtheimpactofthevariablesthatdidnotshowasignicantspatialvariationtheMixedGeographicallyWeightedRegressionisspeciedandestimated.TheglobalandMGWRresultsareshowninTables 4-6 ,and 4-7 .Forparametersthatwereestimatedtovaryspatially,thetablepresentstherangeofvalues.Forparametersthatwereestimatedtohavenostatistically-signicantspatialvariability,thesingleglobalparameterisreported.TheresultsshowthatbothVDTandVTTareaffectedbyseveralland-useandsocio-economicvariablesand,ingeneral,thedirectionalityoftheeffectsisconsistentacrossthemodelsandisintuitivelyreasonable.ThespatialproleofthemarginalimpactofEntropyIndexonVDTandVTTispresentedinFigures 4-4 ,and 4-5 .Figure 4-6 showsvariationofdistancetonearestregional 72

PAGE 73

Table4-4. GWRresultsforVDT VDTGWRMinLQMedianUQMaxF(3)ExplanatoryVariables 1workerhousehold7.757.837.897.598.110.382workerhousehold20.7420.7820.8120.9320.980.173workerhousehold30.2832.8332.9933.7133.961.344workerhousehold44.3144.3844.4846.3248.072.50Numberofhouseholdvehicles:27.957.998.078.148.180.25Numberofhouseholdvehicles:3ormore14.0414.4314.8214.8614.92.31Presenceofchild6-15Presenceofchild16-214.164.304.454.594.670.38Surveyday:Sunday-7.9-7.87-7.83-7.60-7.500.83Surveyday:SaturdayHighincome(income80K)3.123.393.443.513.560.74EntropyIndex(0.5mile)-7.06-6.86-6.52-5.94-5.841.80Fractionofdevelopedareathatisinstitutional(0.5mile)-6.39-6.17-5.65-5.20-4.982.23Distancetonearestregionalactivitycenter(mile)0.430.440.460.470.471.34Distancetonearestbusstop(kilometer)0.320.330.340.390.402.84ConnectedNodeRatio(1mile)-14.49-14.20-13.81-13.57-12.580.41Constant26.5027.3328.0028.0528.190.67F10.99F21.48 73

PAGE 74

Table4-5. GWRresultsforVTT VTTGWRMinLQMedianUQMaxF(3)ExplanatoryVariables 1workerhousehold10.7111.0712.1815.5117.111.732workerhousehold42.7645.0647.2948.1350.611.153workerhousehold81.1183.384.7193.2697.551.464workerhousehold82.8288.3891.5497.13115.901.23Numberofhouseholdvehicles:218.5819.2621.522.1122.380.61Numberofhouseholdvehicles:3ormore33.8937.7839.3441.6743.171.50Presenceofchild6-159.149.9612.0117.3819.102.59Presenceofchild16-2111.8313.8917.7921.2223.231.33Surveyday:Sunday-37.88-36.73-36.05-31.66-28.832.21Surveyday:Saturday-20.23-18.53-17.11-16.31-14.640.56Highincome(income80K)EntropyIndex(0.5mile)-28.24-25.54-16.64-9.96-8.803.23Fractionofdevelopedareathatisinstitutional(0.5mile)-22.26-19.16-10.50-6.18-4.412.64Distancetonearestregionalactivitycenter(mile)0.430.440.470.470.47Distancetonearestbusstop(kilometer)ConnectedNodeRatio(1mile)Constant63.3565.2971.2179.6682.152.84F10.99F22.18 74

PAGE 75

activitycenteronVDT.ThespatialproleofproportionofdevelopedareathatisinstitutionalonVTTandVDTaredepictedinFigures 4-7 ,and 4-8 . Figure4-4. SpatialproleoftheimpactofEntropyonVDT 4.4.1AverageMarginalEffectsTables 4-6 and 4-7 presentasummaryofempiricalresultsfortheregression(aspatial)andMGWR(spatial)modelsforbothVDTandVTT.Theresultsshowthatboth 75

PAGE 76

Table4-6. VTTMGWRresults GlobalVTTCoefcientVIFMinLQMedianUQMaxVIF ExplanatoryVariables1workerhousehold14.01.411.0711.512.4715.3116.811.02workerhousehold46.71.743.5245.9447.5948.3950.221.83workerhousehold90.81.281.8984.0185.2593.1397.132.14workerhousehold103.61.183.8288.7892.0197.53116.42.1Numberofhouseholdvehicles:220.71.520.702.1Numberofhouseholdvehicles:3ormore35.81.732.9237.9639.6941.3542.492.2Presenceofchild6-1514.11.19.410.4212.5617.5619.184.7Presenceofchild16-2119.61.112.5714.4418.0721.5523.455.7Surveyday:Sunday-33.41.0-37.6-36.2-35.6-31.6-292.7Surveyday:Saturday-16.11.0-15.643.3Highincome(income80K)EntropyIndex(0.5mile)-15.11.1-29.7-27.5-19.30-11.20-9.804.3Fractionofdevelopedareathatisinstitutional(0.5mile)-10.81.0-23.30-20.10-11.20-6.40-4.682.5Distancetonearestregionalactivitycenter(mile)0.51.20.452.4Distancetonearestbusstop(kilometer)Constant69.266.5268.0276.1983.1784.78OptimalBandwidth(numberofnearestneighbors)13423245AIC C38286.4638257GlobalMoranCoefcient.006SumofSquareError1483498214612974 76

PAGE 77

Table4-7. VDTMGWRresults VDTGlobalMGWRCoefcientVIFMinLQMedianUQMaxVIF ExplanatoryVariables1workerhousehold8.081.48.052.02workerhousehold20.91.720.862.33workerhousehold33.41.232.8632.9033.0733.6733.862.44workerhousehold46.51.144.3344.4244.5546.3048.062.6Numberofhouseholdvehicles:281.58.012.8Numberofhouseholdvehicles:3ormore14.31.713.9914.3714.8414.8714.882.8Presenceofchild6-15Presenceofchild16-214.71.14.635.4Surveyday:Sunday-7.71.0-7.746.5Surveyday:SaturdayHighincome(income80K)3.21.23.202.7EntropyIndex(0.5mile)-6.61.2-7.104-6.96-6.67-5.96-5.814.1Fractionofdevelopedareathatisinstitutional(0.5mile)-6.11.1-6.36-6.16-5.70-5.18-5.007.2Distancetonearestregionalactivitycenter(mile)0.451.30.420.440.460.470.474.8Distancetonearestbusstop(kilometer)0.371.30.330.330.340.390.400.3ConnectedNodeRatio(1mile)-12.51.2-13.783.0Constant26.927.3827.4727.8128.0628.21OptimalBandwidth(numberofnearestneighbors)34113245AIC C32440.1832431GlobalMoranCoefcient.003SumofSquareError26710002666233 77

PAGE 78

Figure4-5. SpatialproleoftheimpactofEntropyonVTT VDTandVTTareaffectedbyseveralland-useandsocio-economicvariablesand,ingeneral,thedirectionalityoftheeffectsisconsistentacrossthemodelsandisintuitivelyreasonable.Theimpactofthefractionofdevelopedareaaroundtherespondentsresidencethatisinstitutionalisnegativeandsignicantinallmodels.Themoretheneighborhood 78

PAGE 79

Figure4-6. SpatialproleofdistancetonearestregionalactivitycenteronVDT isinstitutional,thegreateraretheopportunitiesforstudentstoregisterinaschoolclosetotheirresidenceleadingtoareductionofoverallVDTandVTTduetoshorteddistancetravelfortheirparentstodropoff/pickupthem.EntropyIndexinahalfamilebufferaroundhouseholdresidentsplaysasignicantroleindecreasingVDTandVTT.HigherEIprovidesgreateropportunitiestoparticipateinactivitieslocallyleadingtoa 79

PAGE 80

Figure4-7. SpatialproleoffractionofdevelopedareathatisinstitutionalonVTT reductionofoverallVDTandVTT.Thedistancetothenearestmajoractivitycenter(i.e.,ameasureofregionalaccessibility)wasfoundtobestrongdeterminantofVDTandVTT.Specically,householdsthatarelocatedfurtherfrommajoractivitycenterstravelmore.ConnectedNodeRatioinaonemilebufferaroundhouseholdresidenthasastatistically-signicantimpactonVDT.Thegreatertheproportionofstreetintersectionto 80

PAGE 81

Figure4-8. SpatialproleoffractionofdevelopedareathatisinstitutionalonVDT cul-de-sacsimpliesgreaterconnectivity(i.e.,shorterdistancestotravelbetweenanytwolocations)and,hence,alowerVDTmaybeexpected.ThedistancetonearestbusstophasapositiveimpactonVDT.Itisinterestingthatevenaftercontrollingforagenericpreferencetolocateneartransit,public-transitaccessibilityhadasignicantimpactontheVDT.ThisresultfavorsTransitOrientedDevelopment(TOD)policiestoencourage 81

PAGE 82

residentsonmorepublictransportusagethandrivingprivatevehicle.WhilebothCNRandtransitaccessibilitydecreasesVDTbutthesetwomeasuresdidnotturnouttobesignicantinreducingVTT.Maybeitisassociatedtolowerspeedlimitandmobilityingrid-linestreetlayoutsandTODareas.Althoughanextensivesetofland-usevariableswereexplored(SeeTable 4-2 ),onlytheabove-mentionedfactorsturnedouttobestatisticallysignicant.Forinstance,wendthatresidential/populationdensityatdifferentspatialresolutions(censusblock,blockgroup,andcensustract)wasstatisticallyinsignicant.Onepossibleexplanationisthattheotherland-use/urbanformvariablesincludedinthemodelarealsoreectingtheresidentialdensityofthehouseholdlocation(forinstance,lowerdensityareasaremorelikelytohavemorecul-de-sacs).ThesendingsverifytheneedtoincludedifferenturbandesignfactorsanditisnotjustdensitythatreduceVDT/VTTbutwhatcomeswithdensitysuchasdiversity,accessibility,etc.Themarginaleffectsoftheland-usevariablesweredeterminedaftercontrollingfortheeffectsofseveralsocio-economicfactors.NumberofworkersinahouseholdisasignicantindicatorinexplainingvariabilityofVDT,andVTT.Theseresultsreectthesignicantcontributionofcommutingtoahouseholdsoverallcaruse.IncreasednumberofhouseholdcarownershipfacilitatesaccessibilitytocarsandhouseholdmemberstravelwhichresultsinhigherVDTandVTT.Theresultshighlighttheimportanceoftheweekendversusweekdaysinthehouseholdstravelpattern.OnSundayshouseholdsdriveless(timeanddistance)andaremorelikelytomakeshortertripthanotherdaysmaybesincethedayafteristhebeginningoftheweekdaystheyarelesslikelytomakelongvehicletrips.Householdstravelpatternalsoshowsthathouseholdtraveltimedecreasecomparedtotheweekdays.Presencesofchildrenalsoincreasehouseholdvehicleusage.Also,higher-incomehouseholdstravelmore(greaterVDT)comparedtoothers.Themodelresultsverifythattofullycapturetheimpactofurbandesignontravelbehavior(hereVDT,VTT),differenturbandesignmeasures(forinstancedifferentdiversitymeasures) 82

PAGE 83

atdifferentspatialresolutionandaggregation(censusblock,censustract,1 2mile,or1mile)shouldbeanalyzedtoderiveinuentialurbandesignindicatorsinlackofastandardwaytomeasureit. 4.4.2StatisticalComparisonsThecorrectedAkaikeInformationCriterion(AICC)measuresfortheglobalregression,andMGWRmodelsarerespectively32440,and32431forVDTand38286,and38257forVTT.ThesemeasuresgiveindicationofthesuperiorityoftheMGWRoverglobalregression.Anotherissuerelatedtoincludinglandusevariablesisthatdensity,diversityandtransitaccessibilitymeasurestendtobehighlycorrelated.Inthisstudydifferentmeasuresatdifferentspatialresolutionwereincluded.Wedidnotfoundanyevidenceonthecorrelationsacrossthenalvariablesinourmodels.VIFisacommoncollinearitydiagnosticforregressionmodels.Specicallyinthecontextofgeographicallyweightedregressionsomescholars[ 87 , 112 , 113 ]raisedimportantquestionsaboutthesuitabilityofgeographicallyweightedregressionmodelsforinferentialpurposesduetolocalcollinearity.Thespatialsubsamplingmethodingeographicallyweightedregressioncanpronouncecollinearityforeachlocalregressionmodel.ForMGWRlocal-basedVIFforassessingcollinearityateachpointisderived.ThemeanvaluesofVIFarepresentedinTables 4-6 ,and 4-7 .ThedistributionoflocalVIFdoesnotshowanyseriouscollinearity.TheGlobalMoranIndex(I)valuesforglobalregressionforVMT,andVTTarerespectively0.003and0.006andstatisticallysignicantindicatingspatialauto-correlationandspatialclusteringintheerrors.TheGlobalMoranfortheMGWRmodelresidualsturnouttobestatisticallyinsignicantwhichindicatesmodelresidualsarerandomlydistributedacrossspace.ToexamineimpactofdensityonVDTandVTT,boxplotofhouseholddailyVDTandVTTatdifferentresidentialdensity(tractlevel)classesareshowninFigures 4-9 ,and 4-10 .Theclassesarebasedonthecorrespondingdensitypercentileofournalsample.ThehorizontallinerepresentsthetotalmeanofVDTandVTT.Consistentwith 83

PAGE 84

literatureitisclearthatdailyVDTandVTTdecreaseasdensityincrease.Itisinterestingtoseethatthesedensityindicatorswhicharemostcommonusedinliterature,arenotsignicantinthenalmodelsaftercontrollingwithsocioeconomicvariablesandotherurbandesignfactors.Thisresultsverifythatitisnecessarytoincludedifferentfactorstohavearealisticimpactofeachlandusefactor.Whileresidentialorpopulationdensityhasbeenusedinliteratureasproxyforurbandesignforminlackofotherurbanformmeasures,thisapproachhasbeencriticizedbysomescholarsthatitisnecessarytoincludedifferentfactorstohavearealisticimpactofeachlandusefactor.TheybelieveelasticityofdensityinimpactingtravelbehaviorsuchasVDTismostlybecauseofotherfactors[ 25 , 94 ].Indeedwithoutconsideringotherfactors,derivedelasticityfordensityisoverestimated.EwingandCervero(2010)[ 42 ]calleddensityasintermediatevariabledenedbyotherurbandesignfactors. 4.4.3SummaryInsummary,theapplicationofmixed-GWRmodelswillsupporttheevaluationoflocalor“context-specic”strategiesaimedatreducingvehiculartrafcvolumes,fuelconsumption,andemissions.Atthesametime,thesemodelsalsoensurethatthevariabilityintheparametersareindeedstatisticallysignicanttherebyimprovingtheefciencyoftheestimates.TheresultsverifythatlandusevariablesnotonlydecreaseVehicle-Distance-TraveledbutalsoVehicle-Time-Traveledathouseholdlevel.Inthenalmodelbothmicroandmacrolevellandusevariablesturnouttobesignicant.Ithighlightsbothneighborhoodlevelandregionalpoliciestoreducedailyvehicleusage.Themodelresultsverifythattofullycapturetheimpactofurbandesignontravelbehavior(hereVDT,VTT),differenturbandesignmeasures(forinstancedifferentdiversitymeasures)atdifferentspatialresolutionandaggregation(censusblock,censustract,1 2mile,or1mile)shouldbeanalyzedtoderiveinuentialurbandesignindicatorsinlackofastandardwaytomeasureit. 84

PAGE 85

Figure4-9. HouseholddailyVDTatdifferentresidentialdensity(tractlevel)percentiles 85

PAGE 86

Figure4-10. HouseholddailyVTTatdifferentresidentialdensity(tractlevel)percentiles 86

PAGE 87

CHAPTER5CAROWNERSHIPMODELING 5.1OverviewCarownershipmodelsareappliedtoderivetraveldemand,oilandenergyconsumption,andemissions.Althoughtheliteratureoncar-ownershipmodelingisextensive,practicallyallexistingmodelsemployaspatialmodelingtechniques.Consequently,issuesofspatialdependencyandspatialnon-stationarityhavenotbeenadequatelyaddressedinthecontextofcarownershipmodeling.Theaimofthisstudyistocontributetoimprovingourunderstandingofcarownershippatternsatadisaggregate(household)levelbyusingamodelingstructurethatrecognizesthecountnatureofthedataandthataccommodatesspatialheterogeneityintheeffectofdeterminantvariables.Specically,aQuasiGeographically-WeightedPoissonModelisestimatedthatcapturesspatialdependencyandnonstationaryandtheresultsarecomparedtoaQuasi(Global)PoissonRegressionModel(quasi-PM).DatafromtheNationalHouseholdTravelSurveysfromthethreecountyregioninSouthEastFloridaareusedforthemodelestimations.Theresultsdemonstratetheneedforspatialmodelsandestablishthebettertofthespatialmodelscomparedtoaspatialmodels.Further,theGaussiandistance-decayfunctionisalsofoundtotthecurrentmodelingcontextbetter.Overall,theresultsverifysignicantimpactofurbandesignfactorssuchasdensityandtransitaccessibilityoncarownershipaftercontrollingsocio-economicandattitudinal/preferencefactors.Theeffectsofthesefactorsoncarownershiparealsofoundtovaryspatially. 5.2DataAssemblyThestudyareacoversSouthEastFlorida(SE)metropolitanareawhichincludesthreecounties:Miami-Dade,Broward,andPalmBeachCounty.The2009NationalHouseholdTravelSurvey(NHTS)isprimarysourceofdataforhouseholdsocio-economiccharacteristicsandcarownershiplevels. 87

PAGE 88

Thereare3980householdssurveyedfromthisregion.Theaveragecarownershipinthesampleis1.71perhousehold(withStandardDeviation(SD)equalto0.98).Among3980households272(6.83%)householdsdonothaveaprivatecarand1487(37.36%),1560(39.19%),492(12.36%),127(3.19%),42(1.07%)householdsown1,2,3,4,5ormorevehiclesrespectively.Tables 5-1 and 5-2 presentastatisticalsummaryofsocioeconomicandattitudinalcharacteristicsobtainedfromthesurvey.Whilethesocio-economicfactorsaregenerallyself-explanatory,weusetheresponsetothefollowingsurveyquestionasindicatoroftheresidentsattitudetowardstheresidentiallocation“Whatisthemostimportantreasonyouchoseyourcurrenthomelocation?”.Thereaderwillnotethatresponsesfortheabovequestionincludestatementsaboutconvenienceandaccessibilitytodifferentmodesandlocations.Webelievethatfactorsthathavecontributedtotheselectionoftheresidentiallocationsuchasclosetopublictransportationtosomeextentcontrolsforself-selectionandresultsinrelativelyunbiasedimpactsderivedforhouseholdspreferenceforcarownershiplevels.Variousbuiltenvironmentmeasureswerecreatedtoexaminetheirimpactsonhouseholdcarownership.Thesemeasurescanbeclassiedintodensity,design,diversity,andtransitaccessibilityfactors.Sincethereisnotastandardformeasuringeachfactor,differentmeasureswerecreatedineachgroupforourstudy.Differentspatialscaleshavealsobeenused(censusblock,censusblockgroup,tract,1 4,1 2,and1milebufferaroundthehouseholdlocations).Someofpreviousstudiespointedoutthatlandusepatternsatdifferentspatialscaleandresolutioncanresultindifferentestimatesandsomeurbandesignfactorsturntobesignicantatspecicspatialscale[ 115 , 116 ].Therefore,inadditiontoderivingdifferentmeasuresineachclass,eachmeasurewascreatedatdifferentspatialresolutions.Table 5-3 presentsastatisticalsummaryofurbandesignmeasures. 88

PAGE 89

Table5-1. Socioeconomicexplanatoryanalysis HouseholdCharacteristicsOverallFreq% Householdrace:White322481.0AfricanAmerican3308.3Hispanic/Mexican872.2Other3398.5Highincome(income80K)107026.9Mediumincome(40K
PAGE 90

Landusevariableswerecompiledfromdifferentsources.The2010censuspopulationandhousingunitblockandtractshapeleswhichprovidesnumberofpopulationandhousingunitsatcensusblockandtractlevelistheprimarysourceofpopulationandresidentialdensitydata.Residentialorpopulationdensityhasbeenusedinliteratureasproxyforurbanforminlackofotherurbanformmeasures. 5.3ModelStructureThisstudyadoptsaquasiPoissonregressionmodelanditsextension,thequasigeographicallyweightedPoissonregressionmodelsforanalysis.ThePoissonregressionmodelisappropriateformodelingcountdataandfallswithintheframeworkofgeneralizedlinearmodels[ 85 ].However,thisapproachassumesequidispersion(meanisequaltothevariance),whichhasbeenfoundtobeuntrueoften.Toaddressthisissue,negative-binomialregressionmodelsarepopularlyused.Thesemodelsassumethatthevarianceisaquadraticfunctionofthemeanandarethereforeappropriateforoverdisperseddata(variance>mean).Analternativeapproachisthequasi-Poissonmodelthatisappropriateforbothover-andunder-dispersedcases.TheQuasi-Poissonmodelassumesalinearrelationshipbetweenmeanandvariance:Var(YjjXj)=j (5)If0<<1underdispersionoccurs,andwhen>1overdispersionhappens.Incase=1thespecicationwouldbethesameasstandardPoissonregression.Foracompletediscussionofquasi-PoissonmodelsrefertoMcCullaghandNelder(1989)[ 82 ].IntransportationcontextOrazioetal.(2011)[ 46 ]andYannisetal.(2007)[ 114 ]haveusedquasi-Poissonmodelinthecontextofsafetyanalysis.Thequasi-GeographicallyWeightedPoissonRegressionmodelrelaxesthespatialstationaryassumptionbyallowinggreaterexibilityandensuringsmoothnessinthevariationofthemodelcoefcientsoverspace.Specically, 90

PAGE 91

Table5-3. Landusevariablesummary OverallMeanStd.Deviation NeighborhoodLandUseCharacteristicsLandUseMix(1 4milebuffer)Fractionoflandareaintheneighborhoodthatisdeveloped0.480.22Fractionofdevelopedareathatisresidential0.610.28Fractionofdevelopedareathatiscommercial0.060.12Fractionofdevelopedareathatisofce0.030.07Fractionofdevelopedareathatisindustrial0.150.19Fractionofdevelopedareathatisinstitutional0.020.09Fractionofdevelopedareathatisother0.120.21EntropyIndex(EI)0.380.24Density(censusblocklevel)Residentialdensity(unitspersquaremile)3094.524590.22Populationdensity(persquaremile)6526.387868.08Employmentdensity(jobspersquaremile)2789.253366.84MixedDensityIndex(MDI)1210.2914.67.41Density(censusblockgrouplevel)Employmentdensity(jobspersquaremile)2432.702223.49Density(censusTractlevel)Residentialdensity(unitspersquaremile)3026.963468.67Populationdensity(persquaremile)6092.334835.18Employmentdensity(jobspersquaremile)2280.931726.10MixedDensityIndex(MDI)1217.291010.37Design(1 4milebuffer)Numberofintersections29.4915.38Numberofcul-de-sacs6.086.10ConnectedNodeRatio(CNR)0.820.14Design(1 2milebuffer)Numberofintersections114.4448.16Numberofcul-de-sacs23.7817.70ConnectedNodeRatio(CNR)0.820.10Design(1milebuffer)Numberofintersections428.40160.72Numberofcul-de-sacs86.6048.83ConnectedNodeRatio(CNR)0.820.08AccessibilitytotransitNumberofbusstop(1 2milebuffer)9.5012.98Euclideandistancetonearestbusstop(meter)1394.102295.79Totaltransitlinelength(ft)(1 2milebuffer)11406.0016185.51Numberofbusstop(1milebuffer)36.0143.74 91

PAGE 92

Inquasi-Poissonmodel,thecoefcientsareconstantacrossthestudyarea(spatialstationaryassumption).Specically,ln(E(YjjXj))=0+pXk=1(kXjk)+j (5)Thequasi-GeographicallyWeightedPoissonRegressionmodelrelaxesthespatialstationaryassumptionbyallowinggreaterexibilityandensuringsmoothnessinthevariationofthemodelcoefcientsoverspace.Specically,ln(E(Y(uj,vj)))=0(uj,vj)+pXk=1(k(uj,vj)Xjk)+j (5)Intheaboveequation,Y(uj,vj)isthecarownershiplevelforhouseholdj,(uj,vj)denotesthecoordinatesresidentiallocationofthejthhouseholdinthespace,Xjkareurbandesign,transportationandhouseholdcharacteristics,and0(uj,vj)andk(uj,vj)representsthevalueoftheparameters0andkatestimationpointj,jiserrortermcapturingeffectofunobservedfactorsoncarownership.Suchmodelshavebeenappliedinthecontextofcrashanalysis[ 54 , 71 ].Theestimationofcoefcientsspecictoeverylocationisaccomplishedbyperforminglocalquasi-GWPM,eachusingasub-sampleofdatainthevicinityofthelocationunderconsideration.ForacompletediscussionseeFotheringhametal.(2002)[ 45 ]andNakayaetal.(2005)[ 84 ].Therearetwoaspectsofinterestwithrespecttodeningtheweight:thebandwidth(i.e.,thenumberofdatapointsconsideredtobenearthelocationunderconsideration)andtherateofdecayintheinuenceofadjacentdatapointswithincreasingdistancewithinthebandwidth(Thedatapointsoutsidethebandwidthhavenoinuence).Therearetwoapproachestospecifyingthebandwidth:xedandadaptive.Thexed-bandwidthapproachassumesaconstantvalueofdistanceindeterminingthevicinityofanyestimationlocation.Theadaptive-bandwidthapproachvariesthisdistanceofinuencebasedonthesampledensityensuringthatthetotalnumberofdatapoints 92

PAGE 93

(neighbors)inthebandwidthisthesameirrespectiveofthedistance.Inthisstudyanadaptivebandwidthisemployed.Specically,regionsofhighdensityofobservationsusesmallerdistanceforbandwidthsandthoseoflowdensityofobservationsuselargerdistanceforbandwidths.Itisalsoensuredthatthetotalnumberofdatapoints(neighbors)inthebandwidthisthesameirrespectiveofthedistance.Theoptimalsizeofthebandwidth(numberofdatapoints)isobtainedusingacross-validationprocedure.ThereaderisreferredtoNowrouzianandSrinivasan(2013)[ 86 ]foradiscussionoftheseitemsinthecontextofgeographicallyweightedlinearregression.Gaussianorbisquarearethemostcommondistancedecaykernelfunctions.Gaussiandistancedecayweightingfunctionisdenedas:W(uj,vj)=8>><>>:1)]TJ /F8 11.955 Tf 11.95 0 Td[(exp((dij hj)2),ifdijhj0,otherwise,wheredijisEuclideandistancebetweenpointsiandj,andhjisthebandwidthforlocationj.Notethattheinuencedecaysexponentiallyinthiscase.Theweightforpointsfartherthanhjisdenedzeroforpointjastheseexertnoinuenceinthedeterminationofthelocalparameter.Bisquaredistancedecayfunctionisdenedas:W(uj,vj)=8>><>>:[1)]TJ /F8 11.955 Tf 11.96 0 Td[((dij hj)2]2,ifdijhj0,otherwise,Theweightforpointsfartherthanhjisdenedzeroaswell.ToourknowledgeallapplicationsofgeographicallyweightedregressionsintransportationhaveselectedGaussianweightingfunction.Chascoetal.(2007)[ 23 ]chosebisquareadaptivespecicationoverGaussiansincebettertnessandadaptivetodifferentspatialcongurationsinmodelinghouseholddisposableincome.InthisstudybothGaussianandbisquarespecicationswillbeexamined. 93

PAGE 94

5.4EmpiricalResultsThreemodelsforanalyzingimpactofurban-designandsocio-economicfactorsoncarownershipwereestimated.TherstistheaspatialQuasi-Poissonmodel.ThenexttwoarethespatialGeographicallyWeightedQuasi-Poissonmodels.Thesetwomodelsdifferinthedistance-decayfunctionsused(Gaussianversusbisquare).Theadaptivebandwidthschemewasusedinbothspatialmodelsandthecross-validationapproachwasemployedtoidentifytheoptimalbandwidths.Thisoptimalbandwidthis78%ofalldatawhentheGaussiandistance-decayfunctionisusedand100%ofalldataforthebi-squaredecayfunction.Thedispersionparameterestimatedforthequasi-Poissonmodelis0.28(lessthan1)whichmeansunderdispersionoccurs.TheseanalyseswereimplementedinRwhichisanopensourcestatisticalandgraphicssoftwareandlanguage. 5.4.1OverallFitMeasuresandSpatialAutoCorrelationsThesumofsquareerrorwasthelargestfortheaspatialquasiPoissonmodel(1295.61)indicatingtheimprovedtofthespatialmodelsovertheaspatialmodel.ThiserrormeasurewaslowerforthespatialmodelwithGaussianfunctioncomparedtothemodelwiththebisquarefunction(1271.29versus1287.23).AsecondmeasureofmodeltnessistheLeave-One-OutCross-Validation(LOOCV).ThisisaspecialcaseofK-FoldCross-Validationwithnumberoffoldsisequaltonumberofobservations.Inthiscase,3980experimentswereperformedandineachexperimentoneobservationisremovedandthedifferencebetweenttedandobservedvalueoftheremovedobservation,determinestheerroroftheout-of-sampleestimations.Thelargenumberofexperiments(fold)inLOOCVreducesthebiasofthetrueerrorratealthoughthecomputationaltimeincreases.Onceagaintheaspatialmodelhadthehighesterror(1928.7)andtheGaussianspatialmodelperformedbetterthanthebisquaremodel(1925.79versus1926.32). 94

PAGE 95

Next,weundertakeadiagnosisofthepresenceofspatialclusteringofresiduals(spatialcorrelation)inthemodelspecication.Existenceofspatialautocorrelationamongmodelresidualscancausebiasedparameterestimateandfalsiedstatisticalinferenceresults[ 69 ].Moran(1950)[ 81 ]developedateststatisticsforanalyzingspatialautocorrelationamongneighbors.GlobalMoransIisthemostcommonglobalmeasureofspatialautocorrelationwhichisapplicabletothepointsorpolygonscontainingcontinuousdata.Itvariesbetween-1to+1.GlobalMoranformodelresidualsiscalculatedbydividingcrossproductoftheerroranditsspatiallagtotheerrorcrossproductadjustedforspatialweights.Toevaluatespatialautocorrelationofmodelresiduals,valuesofGlobalMoransIatdifferentdistancebands(spatialcorrelograms)werecomputed(Figure 5-1 ).Thisplotshowsthatthereissignicantspatialautocorrelationintheaspatialmodelresidualswhichdecreasewithincreasingdistanceband.Thesecorrelationsvanishonlybeyond40KMs.Incontrast,theresidualsfromthespatialmodelsshownoautocorrelationevenatlowerdistancebands.Theresultsdiscussedthusfardemonstratetheneedforspatialmodelsandestablishthebettertofthespatialmodelscomparedtoaspatialmodels.Further,theGaussiandistance-decayfunctionisalsofoundtotthecurrentmodelingcontextbetter. 5.4.2AverageMarginalEffectsTable 5-4 showsasummaryofempiricalresultsforthethreemodels.Forthespatialmodels,themeaneffectsandthevariabilityinthecoefcients(Standarddeviation)overtheestimationspacearepresented.Overallthedirectionalityoftheeffectoftheexplanatoryvariablesisconsistentacrossthethreemodelsandintuitivelyreasonable.Thereaderwillalsonotethatthemeaneffectsfromthebisquaremodelareclosertotheaspatial(global)modelresults.Householdswithmorelicenseddriversandworkershavemorecarsaswouldbeexpected.Atthesametimethepresenceofoneormoreretiredpersonsinahousehold 95

PAGE 96

Table5-4. Comparisonsofmodelestimationresults Quasi-PoissonGaussianAdaptiveBi-SquareAdaptive ExplanatoryVariablesCoefcientMedianSDMedianSDNumberoflicenseddrivers0.3530.3640.0080.3580.006Numberofworkers0.0420.0340.0100.0420.005Presenceofretired-0.063-0.0530.017-0.6200.008Mediumincome(40K
PAGE 97

Figure5-1. Spatialcorrelogramsofmodelresiduals decreasescarownership.Incomeisanotherfactorthatdeterminesthenumberofvehiclesthatahouseholdowns.High(income80K)andmedium(40K
PAGE 98

Dwellingtypehasasignicantimpacttoo.Thoselivinginadetachedsinglehouseownmorecarscomparedtothoselivinginapartments.Thiscouldreectdifferencesinthelife-cyclestagesoftheresidentsofsuchunitsandotherfactorssuchasavailabilityofparking.Hispanichouseholdshavefewercarscomparedtohouseholdsofotherraces.GoetzkeandWeinberger(2012)[ 49 ]alsofoundthatHispanichouseholdsarelesslikelytoownacarrelativetoWhites,Black,andAsianhouseholds.TheNHTSalsocollecteddataonimportantreasonsforthehouseholdschoiceofcurrentresidentiallocations.Thosewhohadindicatedthatclosenesstotransitorschoolwereimportantreasonswerealsofoundtohavefewercars.Thenumberofbusstopswithinhalfamilebufferaroundhouseholdlocationhadanegativeimpactoncarownership.Itisinterestingthatevenaftercontrollingforagenericpreferencetolocateneartransit,public-transitaccessibilityhadasignicantimpactonthecarownership.ThisresultfavorsTransitOrientedDevelopment(TOD)policiestoencourageresidentsonmorepublictransportusagethanowningprivatevehicle.Theimpactofhousingdensityoncarownershipiscapturedviaaquadraticspecication.Thisimpliesthattheelasticityofcarownershiptodensitydecreaseswithincreasinghousingdensity.Theresultsindicatethepotentialfordensicationtodecreaseautoownershipuptoacertainlevelofdensity.Thenumberofintersectionsinaonemilebufferaroundhouseholdsresidentdidnotturnouttobesignicantintheaspatialmodel.Overall,theresultsverifysignicantimpactofurbandesignfactorsoncarownershipaftercontrollingsocio-economicandattitudinal/preferencefactors. 5.4.3SpatialVariationintheMarginalEffectsNextwemovetoadiscussionofthespatialvariationinthemarginaleffects.Ingeneral,allthevariablesthatweresignicantatthegloballevelwerealsosignicantinthespatialmodels.InquasiPoissonmodel,thenumberofintersectionsinonemile 98

PAGE 99

bufferaroundhouseholdslocationwasnotsignicant.However,inthespatialmodelsthisimpactwassignicantforalmost30percentofobservations.Table 5-4 alsopresentstheStandardDeviation(acrossspace)oftheestimatedparametersfromthespatialmodels.ItisinterestingtonotethatthespreadoftheparametersestimatedusingtheGaussianfunctionisgreaterthanthespreadestimatedusingthebi-squarefunction.Thesearealsoevidentfromthedensityplotsofthecoefcients.TheplotsforlandusevariablesareinFigure 5-2 andthoseforsocio-economicvariablesareinFigure 5-3 .Formanyofthevariables(andespeciallythesocio-economicvariables),thedensityplotshaveabi-modaldistribution.Thissuggeststhatthereisapreponderanceoftwopopulationsintheregionwithdifferentialsensitivitiestothesameexplanatoryvariable.Insuchcases,theuseofasingleglobalaverageeffectofthesevariablesinpredictionsandpolicyanalysiscanbemisleading.Figure 5-4 andFigure 5-5 presentsaspatialdistributionoftheestimatedcoefcientsfromthespatialmodelusingtheGaussiandistancedecayfunction. 5.5SummaryCarownershipmodelsareappliedtoderivetraveldemand,oilandenergyconsumption,andemissions.Althoughtheliteratureoncar-ownershipmodelingisextensive,practicallyallexistingmodelsemployaspatialmodelingtechniques.Consequently,issuesofspatialdependencyandspatialnon-stationarityhavenotbeenadequatelyaddressedinthecontextofcarownershipmodeling.Theaimofthisstudyistocontributetoimprovingourunderstandingofcarownershippatternsatadisaggregate(household)levelbyusingamodelingstructurethatrecognizesthecountnatureofthedataandthataccommodatesspatialheterogeneityintheeffectofdeterminantvariables.Specically,aQuasiGeographically-WeightedPoissonModelisestimatedthatcapturesspatialdependencyandnonstationaryandtheresultsarecomparedtoaQuasi(Global)PoissonRegressionModel(quasi-PM). 99

PAGE 100

Figure5-2. Densityplotofcoefcientsonlandusevariables DatafromtheNationalHouseholdTravelSurveysfromthethreecountyregioninSouthEastFloridaareusedforthemodelestimations.Theresultsdemonstratetheneedforspatialmodelsandestablishthebettertofthespatialmodelscomparedtoaspatialmodels.Further,theGaussiandistance-decayfunctionisalsofoundtotthecurrentmodelingcontextbetter.Overall,theresultsverifysignicantimpactofurban 100

PAGE 101

Figure5-3. Densityplotsofcoefcientsonsocioeconomicvariables designfactorssuchasdensityandtransitaccessibilityoncarownershipaftercontrollingsocio-economicandattitudinal/preferencefactors.Theeffectsofthesefactorsoncarownershiparealsofoundtovaryspatially. 101

PAGE 102

Figure5-4. Spatialproleofmarginalimpactsoflandusevariables 102

PAGE 103

Figure5-5. Spatialproleofmarginalimpactsofsocioeconomicvariables 103

PAGE 104

CHAPTER6SUMMARYANDCONCLUSIONS 6.1OverviewUnderstandingtherelationshipbetweenland-useandtravelbehaviorisessentialtowardsquantifyingtheimpactsofland-usepolicies.Inparticular,examiningtheeffectsoflandusepatternsonPMT,VDT,VTTandcarownershipiscriticalfromthestandpointofassessinggoalsaimedatreducingenergyconsumptionsandemissions.Whilethereisasignicantbodyofliteratureonthissubject,almostallofthestudiesemploystationarymethodsthatdonotallowforvariationsofmarginaleffectsoverspace.Thisdissertationcontributestoimprovingourunderstandingoftheeffectsoflandusepatternsonautomobileownershipandusage(distanceandtime)byusingspatial-modelingtechniquesthataccommodatesspatialheterogeneityintheeffectofdeterminantvariablesandrecognizesthenatureoftheoutcomemodeled(continuousversuscount).ThestudyareacoversSouthEastFlorida(SE)metropolitanareawhichincludesthreecounties:Miami-Dade,Broward,andPalmBeachCounty.Inthisdissertationbothmicrolevel,andmesoscaleurbandesignindicatorswerecreated.TheNationalHouseholdTravelSurveysof2008/2009andthe1999SouthEastFloridaHouseholdTravelSurveysconstitutedtheprimarysourceoftravelandsocio-economicdata.VariousmicrolevelbuiltenvironmentmeasureswerecreatedtoexaminetheirimpactsonPMT,VDT,VTTandcarownership.Thesemeasurescanbeclassiedintodensity,design,diversity,andtransitaccessibilityfactors.Sincethereisnotastandardformeasuringeachfactor,differentmeasureswerecreatedineachgroupforourstudy.Differentspatialscaleshavealsobeenused(censusblock,censusblockgroup,tract,1 4,1 2,and1milebufferaroundthehouseholdlocations).The2010statewideNAVTEQroadnetworklewasusedtodeterminetheshortest-distancetravelpathsbetweeneachorigin-destinationpair.Chapter 3 presentsaGeographicallyWeightedRegression(GWR)modelforcapturingtheimpactsoflanduseonperson-milestraveled 104

PAGE 105

anddemonstratingitsbenetsoversimplermethodssuchastheglobalregressionandtheexogenously-segmentedmodel.Chapter 4 extendstheGWRtoamixed-GWRwhichallowsforacombinationofspatially-xedandspatially-varyingparameters.Thismethodologicalinnovationisappliedinthecontextofmodelingvehicle-distance-traveled(VDT)andvehicle-time-traveled(VTT),twoimportantmeasuresofdailyvehiculartravelvolumes.Chapter 5 contributestoimprovingourunderstandingofcarownershippatternsatadisaggregate(household)levelbyusingamodelingstructurethatrecognizesthecountnatureofthedataandthataccommodatesspatialheterogeneityintheeffectofdeterminantvariables.Specically,aQuasiGeographically-WeightedPoissonModelisestimatedthatcapturesspatialdependencyandnon-stationarityandtheresultsarecomparedtoaQuasi(Global)PoissonRegressionModel(quasi-PM). 6.2ContributionsThesubstantivecontributionofthisstudyisindemonstratingthatmarginalsensitivitiesofpersonmiletravel,vehicle-mileand-timetraveled,andcarownershiptovariousland-useattributesdovaryoverspace.Theglobalmodelsassumespatialstationarymodelparametersbyassumingtheelasticity/marginalresponsestotheexplanatoryvariablesarexedoverthespaceofstudyarea.Whileitisagoodbenchmarkforaggregate/generalresponseevaluation,itmightnotreectvariabilityinlocalizedimpactsoverspace.Forinstance,theplotsforimpactoflanduseandsocio-economicvariablesoncarownershipareshowninFigures 5-2 ,and 5-3 respectively.Formanyofthevariables(andespeciallythesocio-economicvariables),thedensityplotsofdistributionofexplanatoryvariablesacrossstudyareaforspatialmodel(QuasiGeographicallyWeightedPoisson)haveabi-modaldistribution.Thissuggeststhatthereisapreponderanceoftwopopulationsintheregionwithdifferentialsensitivitiestothesameexplanatoryvariable.Asshownbyplots,theglobalmodelmarginalimpacts,inmostcases,isnotrepresentativeofthemostfrequentestimated 105

PAGE 106

parameters.Insuchcases,theuseofasingleglobalaverageeffectofthesevariablesinpredictionsandpolicyanalysiscanbemisleading.Spatialdiscreteclassicationofthespatialunitsinexogenoussegmentedmodeloreachhierarchylevelinspatialmultilevelmodelisnotagoodpracticeinmodelingspatialprocess.AsshowninFigures 3-2 ,and 5-4 impactofexplanatoryvariablesonestimatingPersonMileTravel,andcarownership,thespatialboundaryfordifferentexplanatoryvariablesaredifferent.Deningsamespatialsegmentationorspatialhierarchylevelforallexplanatoryvariablesdoesnotsoundrealistic.Besides,spatialprocessiscontinuousandvariesgraduallyratherthanabruptly.Howeverinexogenoussegmentedandspatialmultilevelmodelscouldresultintwohouseholdssituatedclosetoeachotherbutoneithersideofthepre-denedspatialboundarytohaveverydifferentsensitivitiestothesamelanduseattribute.GeographicallyWeightedRegressionallowsgreaterexibilityandensuringsmoothnessinthevariationofthemodelcoefcientsoverspace.TheFigures 4-4 ,and 4-5 showssmoothnessinthevariationoftheimpactofEntropyIndexonVehicleTime/DistanceTraveloverthespace.TheestimatedmodelsverifysignicantimpactofurbandesignfactorsonPMT,VDT,VTT,andcarownershipaftercontrollingsocio-economicandattitudinal/preferencefactors.ImpactoffractionofdevelopedareathatiscommercialorinstitutionalonPMT,VDT,andVTT;numberofintersections,cul-de-sacs,andCNRoncarownership,PMTandVDT;distancetonearestbusstopandnumberofbusstopsonVDTandcarownership;EntropyIndexonVDTandVTT;andresidentialdensityonPMT,andcarownershipveriesthatsmartgrowthwithhigherdensity,higherconnectivity,landusemix,andtransitorienteddevelopmentdecreasecarownershipandusage.Besidesneighborhoodlevelindicators,distanceofhouseholdlocationtonearestregionalactivitycenterwhichcapturesimpactofneighborhoodwithintheregionfoundtobesignicantonPMT,VDT,VTT.Thisndinghighlightsbothneighborhoodandregionallevelurbandesignfactorsimpactvehicleusage(timeanddistance). 106

PAGE 107

Wefoundsomeurbandesignfactorsturnouttobesignicantatspecicspatialscale.EntropyIndexat0.5milebufferaroundhouseholdlocationonVDTandVTT;connectednoderatioin1milebufferaroundhouseholdlocationonVDT;fractionofdevelopedareathatisinstitutionalat0.5milebufferaroundhouseholdlocationonVDTandVTT;housedensityatcensustractoncarownership.Therefore,inadditiontoderivingdifferentmeasuresineachclass,itisnecessarytocreatedifferentspatialresolutions.ExaminingthespatialvariationofthelandusecoefcientsfromtheGWRmodels,revealedthatmostoftheurbandesignlocalizedcoefcientsarestatisticallysignicant.OnlyonelandusevariableineachofPMT,VTT,andVDTdidnotshowspatialvariationinthemarginaleffectoverthestudyarea.Distancetonearestregionalactivitycenter,numberofintersectionspermileofroadway,numberofcul-de-sacspermileofroadwayonestimatingPMT;fractionofdevelopedareathatisinstitutional,EntropyIndex,distancetonearestregionalactivitycenter,distancetonearestbusstoponestimatingVDT;andfractionofdevelopedareathatisinstitutional,EntropyIndexonestimatingVTThadastatisticallysignicantspatialvariance.Theseresultsemphasizeapplicationofspatialnon-stationarymodelsinderivingimpactoflanduseontravelbehaviorsuchasPMT,VDT,VTT,andcarownership.TheGWRmodelswerecomparedtotheglobalmodelsusingFtests.TheteststatisticsF1,F2,andAICwereusedtocomparespatialmodelsspecicationsversusglobalmodels.Inallmodelstheteststatisticsshowsuperiorityofspatialmodelscomparedtoglobalmodelsinttingthedata.GlobalMoranalsorevealedspatialclusteringofresidualsofglobalmodels,whilethegeographicallyweightedmodelresidualsweredistributedrandomlyacrossthestudyarea. 6.3FutureWorkFutureresearchinthisareamayfocusonestimatingcombinationofGeographicallyWeightedRegressionanddimensionreductiontechnique,PrincipalComponent 107

PAGE 108

Analysis,toderiveGeographicallyWeightedPrincipalComponentAnalysis.Thesub-selectionofcombinationofexplanatoryvariablesinprincipalcomponentanalysiswouldreducethelocalcollinearityissuethatmightariseinGeographicallyWeightedRegression.Thelocalcollinearitycanalsobereducedusingshrinkage,penalization,orregularizationmethodssuchasridgeregressionandthelasso.GeographicallyWeightedRidgeRegression(GWRR)orGeographicallyWeightedLasso(GWL)canhelptoreducetheissueofcollinearityinGWRmodels.EstimatingspatialmodelsinStructuralEquationModeling(SEM)frameworktounraveltheinuenceoflandusepatternonPMT,VDT,VTT,andcarownership.SEMestimatesthecorrespondingtravelbehaviorsjointlysoaccountsforself-selectioneffectsandderivesmorereliableoutcomewithrespecttocausality.Italsoalleviatesandresolvesendogeneityissue. 108

PAGE 109

REFERENCES [1] Ahn,Jiwoon,GicheolJeong,YeonbaeKim.2008.Aforecastofhouseholdownershipanduseofalternativefuelvehicles:Amultiplediscrete-continuouschoiceapproach.EnergyEconomics30(5)2091. [2] Akar,Gulsah,Jean-MichelGuldmann.2012.Anotherlookatvehiclemilestraveled.TransportationResearchRecord:JournaloftheTransportationResearchBoard2322(1)110. [3] Ali,Kamar,MarkDPartridge,MRoseOlfert.2007.Cangeographicallyweightedregressionsimproveregionalanalysisandpolicymaking?InternationalRegionalScienceReview30(3)300. [4] Anowar,Sabreena,ShamsunnaharYasmin,NaveenEluru,LMiranda-Moreno.2012.AnalyzingcarownershipintwoQuebecmetropolitanregions:acomparisonoflatentorderedandunorderedresponsemodels.TechnicalPaper,DepartmentofCivilEngineeringandAppliedMechanics,McGillUniversity. [5] Anselin,Luc.1988.Spatialeconometrics:methodsandmodels,vol.4.Springer. [6] Anselin,Luc.1995.Localindicatorsofspatialassociation-LISA.Geographicalanalysis27(2)93. [7] Anselin,Luc.2002.Underthehoodissuesinthespecicationandinterpretationofspatialregressionmodels.Agriculturaleconomics27(3)247. [8] Anselin,Luc.2003.Spatialexternalities,spatialmultipliers,andspatialeconometrics.InternationalRegionalScienceReview26(2)153. [9] Anselin,Luc.2010.Thirtyyearsofspatialeconometrics.PapersinRegionalScience89(1)3. [10] Anselin,Luc,RaymondFlorax,SergioJRey.2004.Advancesinspatialecono-metrics:methodology,toolsandapplications.Springer. [11] Anselin,Luc,DanielA.Grifth.1988.Dospatialeffectsreallymatterinregressionanalysis?PapersinRegionalScience651. [12] Besag,Julian,CharlesKooperberg.1995.Onconditionalandintrinsicautoregressions.Biometrika82(4)733. [13] Blainey,Simon.2010.TripendmodelsoflocalraildemandinEnglandandWales.JournalofTransportGeography18(1)153. [14] Boarnet,MarlonG.2010.Planning,climatechange,andtransportation:Thoughtsonpolicyanalysis.TransportationResearchPartA:PolicyandPractice44(8)587. 109

PAGE 110

[15] Bowman,AdrianW.1984.Analternativemethodofcross-validationforthesmoothingofdensityestimates.Biometrica71(2)353. [16] Brownstone,David,HaoAudreyFang.2010.Avehicleownershipandutilizationchoicemodelwithendogenousresidentialdensity.UniversityofCaliforniaTransportationCenter,UCBerkeley. [17] Brownstone,David,ThomasFGolob.2009.Theimpactofresidentialdensityonvehicleusageandenergyconsumption.JournalofUrbanEconomics65(1)91. [18] Brunsdon,C.1998.Geographicallyweightedregression:anaturalevolutionoftheexpansionmethodforspatialdataanalysis.EnvironmentandplanningA301905.URL http://www.envplan.com/epa/fulltext/a30/a301905.pdf . [19] Burt,RonaldS.1997.Thecontingentvalueofsocialcapital.Administrativesciencequarterly42339. [20] Cardozo,OsvaldoDaniel,JuanCarlosGarca-Palomares,JavierGutierrez.2012.Applicationofgeographicallyweightedregressiontothedirectforecastingoftransitridershipatstation-level.AppliedGeography34548.URL http://www.sciencedirect.com/science/article/pii/S0143622812000070 . [21] Cauleld,Brian.2012.Anexaminationofthefactorsthatimpactuponmultiplevehicleownership:ThecaseofDublin,Ireland.TransportPolicy19132. [22] Chandra,BhatR.,VamsiPulugurta.1998.Acomparisonoftwoalternativebehavioralchoicemechanismsforhouseholdautoownershipdecisions.Trans-portationResearch32(1)61. [23] Chasco,Coro,IsabelGarca,JosVicns.2007.Modelingspatialvariationsinhouseholddisposableincomewithgeographicallyweightedregression.MunichPersonalRePEcArkhive(MPRA)16821. [24] Chatman,DanielG.2003.Howdensityandmixedusesattheworkplaceaffectpersonalcommercialtravelandcommutemodechoice.TransportationResearchRecord:JournaloftheTransportationResearchBoard1831(1)193. [25] Chatman,DanielG.2008.Deconstructingdevelopmentdensity:quality,quantityandpriceeffectsonhouseholdnon-worktravel.TransportationReseach421008. [26] Chattopadhyay,Sudip,EmilyTaylor.2012.Dosmartgrowthstrategieshavearoleincurbingvehiclemilestraveled?afurtherassessmentusinghouseholdlevelsurveydata.TheBEJournalofEconomicAnalysis&Policy12(1). [27] Chu,You-Lian.2002.Automobileownershipanalysisusingorderedprobitmodels.TransportationResearchRecord180560. 110

PAGE 111

[28] Clark,Stephen.2004.Estimatingcarownershipusinggeographicallyweightedregression.TrafcEngineeringandControl45(11)416. [29] Clark,Stephen.2007.Estimatinglocalcarownershipmodels.JournalofTransportGeography15184. [30] Clark,Stephen,AndrewO.Finley.2010.Spatialmodelingofcarownershipdata:acasestudyfromtheUnitedKingdom.AppliedSpatialAnalysisandPolicy3(1)45. [31] Cleveland,WilliamS.1979.Robustlocally-weightedregressionandsmoothingscatterplots.JournaloftheAmericanStatisticalAssociation74829. [32] Cleveland,WilliamS,SusanJDevlin.1988.Locallyweightedregression:anapproachtoregressionanalysisbylocaltting.JournaloftheAmericanStatisticalAssociation83(403)596. [33] Cliff,Andrew,KeithOrd.1972.Testingforspatialautocorrelationamongregressionresiduals.GeographicalAnalysis4(3)267. [34] Cliff,AndrewDavid,JKeithOrd.1981.Spatialprocesses:models&applications,vol.44.PionLondon. [35] Cohen,Jeffrey,KristenMonaco.2008.Portsandhighwaysinfrastructureananalysisofintra-andinterstatespillovers.InternationalRegionalScienceReview31(3)257. [36] Cohen,JeffreyP.2010.Thebroadereffectsoftransportationinfrastructure:Spatialeconometricsandproductivityapproaches.TransportationresearchpartE:logisticsandtransportationreview46(3)317. [37] Cohen,JeffreyP,CatherineJMorrisonPaul.2003.Airportinfrastructurespilloversinanetworksystem.JournalofUrbanEconomics54(3)459. [38] Cressie,NoelAC,NoelACassie.1993.Statisticsforspatialdata,vol.900.WileyNewYork. [39] deAbreueSilva,Joao,CatherineMorency,KonstadinosGGoulias.2012.Usingstructuralequationsmodelingtounraveltheinuenceoflandusepatternsontravelbehaviorofworkersinmontreal.TransportationResearchPartA:PolicyandPractice46(8)1252. [40] Durbin,James,GeoffreySWatson.1950.Testingforserialcorrelationinleastsquaresregression:I.Biometrika37(3/4)409. [41] Ewing,Reid,RobertCervero.2001.Travelandthebuiltenvironment:asynthesis.TransportationResearchRecord:JournaloftheTransportationResearchBoard1780(1)87. 111

PAGE 112

[42] Ewing,Reid,RobertCervero.2010.Travelandthebuiltenvironment.JournaloftheAmericanPlanningAssociation76(3)265. [43] Fang,HaoAudrey.2008.Adiscrete–continuousmodelofhouseholdsvehiclechoiceandusage,withanapplicationtotheeffectsofresidentialdensity.Trans-portationResearchPartB:Methodological42(9)736. [44] Fotheringham,AStewart,ChrisBrunsdon.1999.Localformsofspatialanalysis.GeographicalAnalysis31(4)340. [45] Fotheringham,A.Stewart,ChrisBrunsdon,MartinCharlton.2002.GeographicallyWeightedRegression:TheAnalysisofSpatiallyVaryingRelationships.Wiley. [46] Giuffra,Orazio,AnnaGrana,MarinoRobertab,FerdinandoCorriereb.2011.Handlingunderdispersionincalibratingsafetyperformancefunctionaturban,four-leg,signalizedintersections.JournalofTransportationSafety&Security3(3)174. [47] Giuliano,Genevieve,JoyceDargay.2006.Carownership,travelandlanduse:acomparisonoftheusandgreatbritain.TransportationResearchPartA:PolicyandPractice40(2)106. [48] Giuliano,Genevieve,JoyceDargay.2006.Carownership,travelandlanduse:acomparisonoftheUSandGreatBritain.TransportationResearch40106. [49] Goetzke,Frank,RachelWeinberger.2012.Separatingcontextualfromendogenouseffectsinautomobileownershipmodels.EnvironmentandPlan-ningA441032. [50] Gomez-Ibanez,DrJose,MarlonGBoarnet,DianneRBrake,RobertBCervero,AndrewCotugno,AnthonyDowns,SusanHanson,KaraMKockelman,PatriciaLMokhtarian,RolfJPendall,etal.2009.Drivingandthebuiltenvironment:theeffectsofcompactdevelopmentonmotorizedtravel,energyuse,andco2emissions.Tech.rep.,OakRidgeNationalLaboratory(ORNL). [51] Guo,Zhan.2013.Doesresidentialparkingsupplyaffecthouseholdcarownership?thecaseofNewYorkcity.JournalofTransportGeography2618. [52] Gutierrez,Javier,OsvaldoDanielCardozo,JuanCarlosGarca-Palomares.2011.Transitridershipforecastingatstationlevel:anapproachbasedondistance-decayweightedregression.JournalofTransportGeography19(6)1081. [53] Gutierrez,Javier,AnaCondeco-Melhorado,ElenaLopez,AndresMonzon.2011.Evaluatingtheeuropeanaddedvalueoften-tprojects:amethodologicalproposalbasedonspatialspillovers,accessibilityandgis.JournalofTransportGeography19(4)840. 112

PAGE 113

[54] Hadayeghi,Alireza,AmerSShalaby,BhagwantNPersaud.2010.Developmentofplanningleveltransportationsafetytoolsusinggeographicallyweightedpoissonregression.AccidentAnalysis&Prevention42(2)676.URL http://www.sciencedirect.com/science/article/pii/S0001457509002954 . [55] Handy,SusanL.2005.Smartgrowthandthetransportation-landuseconnection:whatdoestheresearchtellus.InternationalRegionalScienceReview28(2)146. [56] Hepple,LeslieW.1998.Exacttestingforspatialautocorrelationamongregressionresiduals.EnvironmentandPlanningA30(1)85. [57] Heres-Del-Valle,David,DebNiemeier.2011.Co2emissions:Areland-usechangesenoughforCaliforniatoreducevmt?specicationofatwo-partmodelwithinstrumentalvariables.TransportationResearchPartB:Methodological45(1)150. [58] Hintze,JerryL.,RayD.Nelson.1998.Violinplots:aBoxPlot-densitytracesynergism.TheAmericanStatistician522. [59] Imhof,JP.1961.Computingthedistributionofquadraticformsinnormalvariables.Biometrika48(3/4)419. [60] Karlaftis,Matthew,JohnGolias.2002.Automobileownership,householdswithoutautomobiles,andurbantrafcparameters:aretheyrelated?”.TransportationResearchRecord179229. [61] Kim,HongSok,EungcheolKim.2004.EffectsofpublictransitonautomobileownershipanduseinhouseholdsoftheUSA.ReviewofUrban&RegionalDevelopmentStudies16(3)245. [62] Kim,Sanghong,RyotaOkajima,ManabuKano,ShinjiHasebe.2013.Developmentofsoft-sensorusinglocallyweightedplswithadaptivesimilaritymeasure.ChemometricsandIntelligentLaboratorySystems12443. [63] Kirby,DustinK,JamesPLeSage.2009.Changesincommutingtoworktimesoverthe1990to2000period.RegionalScienceandUrbanEconomics39(4)460. [64] Koerts,J,AdriaanPieterJohannesAbrahamse.1968.Onthepoweroftheblusprocedure.JournaloftheAmericanStatisticalAssociation63(324)1227. [65] Legendre,Pierre.1993.Spatialautocorrelation:troubleornewparadigm.Ecology64(6)1659. [66] LeSage,JamesP,RKelleyPace.2001.Spatialdependenceindatamining.DataMiningforScienticandEngineeringApplications439. 113

PAGE 114

[67] LeSage,JamesP.,R.KelleyPace.2009.IntroductiontoSpatialEconometrics.Chapman&Hall/CRC. [68] Leung,Yee,MChang-Lin,ZWen-Xiu.2000.Statisticaltestsforspatialnonstationaritybasedonthegeographicallyweightedregressionmodel.En-vironmentandPlanningA32(1)9. [69] Leung,Yee,Chang-LinMei,Wen-XiuZhang.2000.Testingforspatialautocorrelationamongtheresidualsofthegeographicallyweightedregression.EnvironmentandPlanningA32(5)871. [70] Li,Jieping,JoanLWalker,SumeetaSrinivasan,WilliamPAnderson.2010.ModelingprivatecarownershipinChinainvestigationofurbanformimpactacrossmegacities.TransportationResearchRecord219376. [71] Li,Zhibin,WeiWanga,PanLiua,JohnM.Bighamb,DavidR.Ragland.2013.Usinggeographicallyweightedpoissonregressionforcounty-levelcrashmodelinginCalifornia.SafetyScience5889. [72] Lloyd,Chris,IanShuttleworth.2005.Analysingcommutingusinglocalregressiontechniques:scale,sensitivity,andgeographicalpatterning.EnvironmentandPlanningA37(1)81.URL http://envplan.com/epa/fulltext/a37/a36116.pdf . [73] MacCullagh,Peter,JohnAshworthNelder.1989.Generalizedlinearmodels,vol.37.CRCpress. [74] Malczewski,Jacek,AnneliesePoetz.2005.ResidentialburglariesandneighborhoodsocioeconomiccontextinLondon,Ontario:globalandlocalregressionanalysis.TheProfessionalGeographer57(4)516.URL http://www.tandfonline.com/doi/abs/10.1111/j.1467-9272.2005.00496.x . [75] Matas,Anna,Jose-LuisRaymond,Jose-LuisRoig.2009.CarownershipandaccesstojobsinSpain.TransportationResearchPartA43607. [76] McMillen,DanielP.2010.Issuesinspatialdataanalysis.JournalofRegionalScience50(1)119. [77] McMillen,DanielP,ChristianLRedfearn.2010.Estimationandhypothesistestingfornonparametrichedonichousepricefunctions.JournalofRegionalScience50(3)712. [78] Mei,C,NingWang,WZhang.2006.Testingtheimportanceoftheexplanatoryvariablesinamixedgeographicallyweightedregressionmodel.EnvironmentandPlanningA38(3)587.URL http://envplan.com/epa/fulltext/a38/a3768.pdf . [79] Mei,Chang-Lin,Shu-YuanHe,Kai-TaiFang.2004.Anoteonthemixedgeographicallyweightedregressionmodel.JournalofRegionalScience44(1)143. 114

PAGE 115

[80] Moore,AdrianT,SamuelRStaley,RobertWPooleJr.2010.Theroleofvmtreductioninmeetingclimatechangepolicygoals.TransportationResearchPartA:PolicyandPractice44(8)565. [81] Moran,P.A.P.1950.Notesoncontiniousstochasticphenomena.Biometrika3717. [82] Myers,Raymond,DouglasMontgomery,GeoffreyVining.2002.GeneralizedLinearModels:withApplicationsinEngineeringandtheScience.2nded.WileySeriesinProbabilityandStatistics. [83] Nss,Petter.2010.Residentiallocation,travel,andenergyuseintheHangzhoumetropolitanarea.Journaloftransportandlanduse3(3). [84] Nakaya,T.,A.S.Fotheringham,C.Brunsdon,M.Charlton.2005.Geographicallyweightedpoissonregressionfordiseaseassociationmapping.StatisticsinMedicine24(17)2695. [85] Nelder,J.A.,R.W.M.Wedderburn.1972.Generalizedlinearmodels.JournaloftheRoyalStatisticalSociety135(3)370. [86] Nowrouzian,Roosbeh,SivaramakrishnanSrinivasan.2013.Modelingtheeffectofland-useonperson-mile-traveledusinggeographically-weightedregression.TransportationResearchRecord2728. [87] Paez,Antonio,StevenFarber,DavidWheeler.2011.Asimulation-basedstudyofgeographicallyweightedregressionasamethodforinvestigatingspatiallyvaryingrelationships.EnvironmentandPlanning-PartA43(12)2992. [88] Parent,Olivier,JamesPLeSage.2010.Aspatialdynamicpanelmodelwithrandomeffectsappliedtocommutingtimes.TransportationResearchPartB:Methodological44(5)633. [89] Parry,IanWH,MargaretWalls,WinstonHarrington.2007.Automobileexternalitiesandpolicies.Journalofeconomicliterature373. [90] Pereira,AlfredoMarvao,OriolRoca-Sagales.2003.Spillovereffectsofpubliccapitalformation:evidencefromtheSpanishregions.JournalofUrbaneconomics53(2)238. [91] Potoglou,Dimitris,PavlosS.Kanaroglou.2008.Modelingcarownershipinurbanareas:acasestudyofHamilton,Canada.JounalOFTransportGEOGRAPHY1642. [92] Ritter,Nolan,ColinVance.2013.Dofewerpeoplemeanfewercars?populationdeclineandcarownershipinGermany.TransportationResearch5074. [93] Rousseeuw,PeterJ.,IdaRuts,JohnTukey.1999.Thebagplot:abivariateboxplot.TheAmericanStatistician534. 115

PAGE 116

[94] Salon,Deborah.2009.Neighborhoods,cars,andcommutinginNewYorkcity:adiscretechoiceapproach.TransportationResearch43180. [95] Salon,Deborah,MarlonG.Boarnet,SusanHandy,StevenSpears,GilTala.2012.HowdolocalactionsaffectVMT?acriticalreviewoftheempiricalevidence.TransportationReseach17495. [96] Shay,Elizabeth,AsadJ.Khattak.2005.Automobileownershipanduseinneotraditionalandconventionalneighborhoods.TransportationResearchRecord190218. [97] Shay,Elizabeth,AsadJ.Khattak.2007.Automobiles,trips,andneighborhoodtype:comparingenvironmentalmeasures.TransportationResearchRecord201073. [98] Shay,Elizabeth,AsadJ.Khattak.2012.Householdtraveldecisionchains:residentialenvironment,automobileownership,tripsandmodechoice.Interna-tionalJournalofSustainableTransportation62. [99] Spissu,Erika,AbdulRawoofPinjari,RamMPendyala,ChandraRBhat.2009.Acopula-basedjointmultinomialdiscrete–continuousmodelofvehicletypechoiceandmilesoftravel.Transportation36(4)403. [100] Srinivasan,Sivaramakrishnan,RussellProvost,RuthSteiner.2013.Modelingtheland-usecorrelatesofvehicle-triplengthsforassessingthetransportationimpactsoflanddevelopments.JournalofTransportandLandUse6(2)59. [101] Susilo,DimitrisPotoglouYusakO.2008.Comparisonofvehicleownershipmodels.TransportationResearchRecord207697. [102] Tiefelsdorf,Michael.2002.Thesaddlepointapproximationofmoran'sI'sandlocalmoran'sI'sreferencedistributionsandtheirnumericalevaluation.GeographicalAnalysis34(3)187. [103] Tiefelsdorf,Michael,BarryBoots.1995.Theexactdistributionofmoran'sI.EnvironmentandPlanningA27985. [104] Tobler,WaldoR.1970.AcomputermoviesimulatingurbangrowthintheDetroitregion.Economicgeography46234. [105] Tong,Tingting,Tun-HsiangEdwardYu,Seong-HoonCho,KimberlyJensen,DanielDeLaTorreUgarte.2013.Evaluatingthespatialspillovereffectsoftransportationinfrastructureonagriculturaloutputacrosstheunitedstates.JournalofTransportGeography3047. [106] VanAcker,Veronique,FrankWitlox.2010.Carownershipasamediatingvariableincartravelbehaviourresearchusingastructuralequationmodellingapproachtoidentifyitsdualrelationship.JournalofTransportGeography18(1)65. 116

PAGE 117

[107] Vance,Colin,RalfHedel.2007.Theimpactofurbanformonautomobiletravel:disentanglingcausationfromcorrelation.Transportation34(5)575. [108] Vyas,Gaurav,RajeshPaleti,ChandraRBhat,KonstadinosGGoulias,RamMPendyala,Hsi-hwaHu,ThomasJAdler,AnissBahreinian.2012.Jointvehicleholdings,bytypeandvintage,andprimarydriverassignmentmodelwithapplicationforCalifornia.TransportationResearchRecord:JournaloftheTransportationResearchBoard2302(1)74. [109] Wang,Ning,Chang-LinMei,Xiao-DongYan.2008.Locallinearestimationofspatiallyvaryingcoefcientmodels:animprovementonthegeographicallyweightedregression.EnvironmentandPlanningA40986. [110] Wei,Chuan-Hua,FeiQi.2012.Ontheestimationandtestingofmixedgeographicallyweightedregressionmodels.EconomicModeling292615. [111] Weinberger,Rachel,FrankGoetzke.2010.Unpackingpreference:howpreviousexperienceaffectsautoownershipintheUnitedStates.UrbanStudies47211. [112] Wheeler,David,MichaelTiefelsdorf.2005.Multicollinearityandcorrelationamonglocalregressioncoefcientsingeographicallyweightedregression.JournalofGeographicalSystems7(2)161. [113] Wheeler,DavidC.2007.Diagnostictoolsandaremedialmethodforcollinearityingeographicallyweightedregression.EnvironmentandPlanningA39(10)2464. [114] Yannisa,George,ConstantinosAntonioua,EleonoraPapadimitrioua.2007.Roadcasualtiesandenforcement:distributionalassumptionsofseriallycorrelatedcountdata.TrafcInjuryPrevention8(3)300. [115] Zegras,Christopher.2010.Thebuiltenvironmentandmotorvehicleownershipanduse:evidencefromSantiagodeChile.UrbanStudies47(8)1793.URL http://usj.sagepub.com/content/47/8/1793.short . [116] Zhang,Lei,JinHyunHong,ArefehNasri,QingShen.2012.Howbuiltenvironmentaffectstravelbehavior:acomparativeanalysisoftheconnectionsbetweenlanduseandvehiclemilestraveledinUScities.THEJOURNALOFTRANSPORTANDLANDUSE5(3)40. [117] Zhang,Lianjun,HaijinShi.2004.Localmodelingoftreegrowthbygeographicallyweightedregression.ForestScience50(2)225. [118] Zhao,Fang,NokilPark.2004.Usinggeographicallyweightedregressionmodelstoestimateannualaveragedailytrafc.TransportationResearchRecord:JournaloftheTransportationResearchBoard1879(1)99.URL http://trb.metapress.com/index/k214576381126g46.pdf . 117

PAGE 118

[119] Zhou,Bin(Brenda),KaraMKockelman.2008.Self-selectioninhomechoice:Useoftreatmenteffectsinevaluatingrelationshipbetweenbuiltenvironmentandtravelbehavior.TransportationResearchRecord:JournaloftheTransportationResearchBoard2077(1)54. 118

PAGE 119

BIOGRAPHICALSKETCH RoosbehNowrouzianwasbornin1982inUS.Hereceivedhisbachelor'sdegreein2006andhismaster'sdegreein2008fromtheDepartmentofCivilEngineeringatSharifUniversityofTechnology,Tehran,Iran.InAugust2009,hejoinedthePh.D.programintheDepartmentofCivilEngineeringattheUniversityofFlorida.Hisresearchinterestslieindifferentareasofdesigningandimplementingstatistical/predictivemodelsandcutting-edgealgorithmsutilizingdiversesourcesofdata.HereceivedhisDoctorofPhilosophydegreeincivilengineeringwithaminorinappliedstatisticsfromtheUniversityofFloridainthesummerof2014. 119