You are on page 1of 14

The Effects of Urban Spatial Structure on Travel Demand in the United States Author(s): Antonio M. Bento, Maureen L.

Cropper, Ahmed Mushfiq Mobarak, Katja Vinha Source: The Review of Economics and Statistics, Vol. 87, No. 3 (Aug., 2005), pp. 466-478 Published by: The MIT Press Stable URL: http://www.jstor.org/stable/40042942 . Accessed: 15/06/2011 16:49
Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=mitpress. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.

The MIT Press is collaborating with JSTOR to digitize, preserve and extend access to The Review of Economics and Statistics.

http://www.jstor.org

THE EFFECTS OF URBAN SPATIAL STRUCTURE ON TRAVEL DEMAND IN THE UNITED STATES
and Ahmed Antonio Bento,Maureen Cropper, M. L. Mobarak, Katja Vinha* Mushfiq
to Previousattempts answerthesequestionshave relied or observations on studiesof household either city-level on data in which measuresof urban formare endogenous. use of correlate measures automobile studiesthat City-level or withpopulationdensity density (Levinson & gradients 1989; Malpezzi, Kumar, 1997; Newman & Kenworthy, 1999) oftenfail to controlfor othervariablesthataffect and automobile ownership mode choice. Analysesof vehiinclude data often cle ownership VMT usinghousehold and I. Introduction but measures urbanform, ones thatare clearlysubjectto of of theSecondWorld Warthepredominant of householdchoice. For example,the population density pattern lives is or urban growthin the United States has been one of thecensustract ZIP code in whichthehousehold SINCE and decentralization,oftenused as a measure of urban sprawl (Train, 1986; low-density development employment accompanied a rapidincreasein automobile by ownership Levinson& Kumar,1997; Boarnet& Crane,2001), and the or and vehicle miles traveled(VMTs) (Mills, 1992; Miesz- distanceof a household'sresidencefrompublic transit kowski& Mills, 1993; Glaeser& Kahn,2001). The last 15 fromthe centralbusinessdistrict (CBD) as a measureof & of have witnessed reaction urban a to (Train,1980; Boarnet however, years, sprawlin availability publictransportation & to the formof "smartgrowth"initiatives. 1998; Crane& Crepeau,1998; Boarnet Crane, Attempts limit Sarmiento, estimates obtainedin thesestudiesare urbangrowth to changeits form motivated three 2001).2 Coefficient or are by - to preserve locate in if people who dislikedriving urbandevel- likelyto be biased concerns open space and foster is transit morelikelyto be provided. thatis moreaesthetically appealing,to reducethe areas wherepublic opment of measures We addresstheseissues by addingcity-wide and costof providing publicservices, to reducedependence Perto the 1990 Nationwide associated with sprawland transit on the automobileand the externalities availability sonal Transportation use automobile thathave accompaniedurbansprawl.1 Survey(NPTS). The surveycontains automobileownershipand annual miles on This naturallyraises the question: how does urban information of measured thespatialdistribution pop- drivenforover 20,000 U.S. householdsand on the comform whether by withinthese households.For behaviorof workers network affect muting or or ulation employment thepublictransit NPTS householdsliving in the urbanizedportionof 114 vehicle ownershipand the numberof miles driven by meaareas (MSAs)3 we construct statistical in households theUnitedStates?This paperaddressesthis metropolitan suresof urbanform measuresof cityshape (how close to and transit measuresof urbanform questionby combining the supplyin 114 urbanareas in the U.S. withdata fromthe circularthe city is), the densityof the road network, of distribution population(how close to the CBD 1990 NationwidePersonalTransportation Survey.We ask spatial describe the populationis located), and thejobs-housingbalance. of measures urbansprawl measuresthat whether networkwe computecityof city shape, spatial distribution population,and jobs- To characterizethe transport - and thesupplyof publictransit bus route wide measures of transit the affect balance supply specifically, housing miles supplied and rail routemiles supplied,normalized mode choices of U.S. households. VMT and commute by city area. two We use thesedatato estimate setsofmodels.The first Receivedforpublication May 6, 2003. Revisionacceptedforpublication December29, 2004. mode choice (McFadden, 1974), in is a model of commute * University Maryland; and of of University Maryland The WorldBank; which we fouralternatives driving, walking/ distinguish de of University Colorado,Boulder;and Universidad Los Andes,Colomand commuting rail. We by commuting bus, by bicycling, bia, respectively. and The findings, from NPTS who live the expressedin thispaperare estimate thismodelusingworkers interpretations, conclusions the thoseof theauthors. represent views in one ofthe26 citiesin theU.S. that They do notnecessarily entirely of have someform rail or oftheWorldBank,itsExecutiveDirectors, thecountries they represent. form. of measures urban as well as data on ourother Matthew Richard We wouldliketo thank Kahn,Jordan Arnott, Rappaport, transit, ZmarakShalizi, Ken Small, Margaret JulioRotemberg, Walls, two anon- We also estimate logitmodel to explainwhether nota or a
at and participants seminarsat Columbia University, ymous referees, of StanfordUniversity, University Californiaat Berkeley,the NBER SummerInstitute 2002, WorldConJuly2001, ASSA MeetingsJanuary and gress of Environmental Resource Economists2002, and Colorado 2002. and Environmental ResourceEconomicsWorkshop 1For a discussionof the effects these externalities urbanspatial on of structure Brueckner(2001) and Bento and Franco (forthcoming). see Kahn (2000) discusses theenvironmental impactof suburbanization. and TheReview Economics Statistics, 2005, 87(3): 466-478 August of 2 For a reviewof the literature, Badoe and Miller (2000). see 3 We use the 1990 boundaries urbanized areas associatedwitnme 114 or are These boundaries defined theU.S. areas in our study. by metropolitan Census Bureau. Urbanizedareas are thosethathave a population density of than1,000 persquaremile and a totalpopulation at least thatis greater 50,000. Abstract We examinetheeffects urbanform of and publictransit supply on thecommute mode choices and annualvehiclemiles traveled (VMTs) ofhouseholds of livingin 114 urbanareas in 1990.The probability driving to work is lower the higherare populationcentrality and rail miles centrality, suppliedand theloweris road density. Population jobs-housing have a significant effect annual on balance,cityshape, and road density household VMTs. Althoughindividual elasticitiesare small absolute values (<0.10), movingsample householdsfroma citywiththe characteristics Atlantato a city withthe characteristics Boston reduces of of annualVMTs by 25%.

2005 by the Presidentand Fellows of Harvard College and the Massachusetts Instituteof Technology

THE EFFECTS OF URBAN SPATIAL STRUCTURE worker drivesto work,using data fromthe 114 cities for data.The secondset whichwe have bothsprawland transit of of modelsexplainsthenumber vehiclesownedand miles driven vehicleforhouseholdslivingin thesesame 114 per urbanareas. The restof thepaperis organizedas follows.SectionII between urban formand travel reviews the relationship describesour demand in the urban economics literature, empiricalmeasures of urban form,and compares these It measures withtraditional sprawlmeasures. also describes our city-widetransitvariables. Section III presentsthe modechoicemodels,and sectionIV results ourcommute of our modelsof automobile and ownership VMTs. SectionV examinestheimplication movingour samplehouseholds of to citieswithdifferent vectors transit sprawlcharacof and and teristics, sectionVI concludes. II. The Relationship BetweenUrban Form and Travel Demand

467

The monocentric modelcan also be modified allow for to Assumethatthehousehold can travel publictransportation. either publictransit private or auto (Bento et al., 2003). by Thereis a marginal timecostand a marginal pricefor public transit each pointin the city, at just as thereis a marginal dollarcostof driving a marginal and timecostthat depends, on of and Wheaton, thedensity theroad network following of population each pointin thecity. is an easy matter at It to construct model in whichthe householddetermines a the of number trips makeby private to auto and thenumber of The frequency transit of tripsto make by public transit. service and the numbers transit of stops and routemiles the timecost of public supplied,by influencing marginal will commute modechoice and thenumber of transit, affect miles drivenannually. B. Measuresof UrbanForm The previous subsection that suggests VMT and commute modechoicedepend three on of form theroad aspects urban the of land network, pattern residential use,andthedistribution ofemployment a proxy thedistributionservices) for of (also the area.The question How should is: these throughout urban dimensions urban formbe measuredempirically? of Our choiceamong alternative measures eachdimension urban of of form guidedby twoprinciples: setof measures is the should different of form should not is, (that they capture aspects urban be too highly correlated witheach other), and, to facilitate our it to interpreting results, shouldbe possible, conceptually, whileholding others the constant. varyone measure Road Network: A complete description the road of in network a circularcity would include describing road in density successiveannuliaroundtheCBD, as well as the of with or pattern roads (for example, a radial network without is in ringroads). The situation morecomplicated a We to citythatis notradially symmetric. use two measures describethe road network. The first a measureof city is for shape. The secondis a measureof average road density theurbanarea. thattrip distances shouldbe Cityshape. Theorysuggests longerin long, narrowcities than in circularcities with radialroad networks. measurehow muchan urbanized To area deviatesfroma circularcity,we have circumscribed each citywithan ellipseequal in area to theurbanized area of thecity, and have measured majorand minor the axes of the ellipse. The ratioof the minorto the major axis, our measure of city shape, ranges between0 and 1, with 1 a indicating circular city.5 Road density. For each urban area, miles of road are of multiplied averageroadwidth by (fordifferent categories
(1977) shows under plausible conditionsthat female workersin twoearnerhouseholdsshould have shorter commutesthan eithersingle or married male workers. 5 See the Data Appendix in Bento et al. (2003) for a more complete of description the measures.

A. Theory Urban economics predictsthatthe numberof miles a household travels themodeitchooses fordifferent and trips will depend on the structure the city in which the of householdresides on the distribution populationand of within city, thesize ofthecity(in square the on employment networks. miles),and on its road and transit In thesimplemonocentric model (Muth,1969) in which all employment located in the CBD and the number is of worker fixed, number miles a household is the of trips per travelsis proportional how farfrom CBD it locates to the it (t). This dependson the rentgradient faces and on the marginalcost of travel,which, in general, varies with distancefrom CBD. To allow forcongestion, the Wheaton timecost of travelat a (1998) suggeststhatthe marginal withpopulation and inversely pointvariesdirectly density withtheproportion land devotedto roads at thatpoint. of The household's travel whichequals thenumber of demand, one-waytripsto the CBD timest, thusdepends(through choice of t) on theroad network on thedistribution and of the population throughout city. Modifications themonocentric to model(MichelleWhite, to 1977, 1986, 1988) have allowed firms move out of the thatemployment locatedthroughout is the CBD, implying area. The spatial distribution firms(and of metropolitan associatedwage gradient) affects commutelengths afby wherehouseholdschoose to live and where they fecting chooseto work.By determining locationof servicesand the retailestablishments, spatial distribution firms the of also affects length nonwork the of trips.4
4 This literature also has in pointedout how differences incomeand in tastes(due to differences family in commute composition) may influence of lengths. Assumingthatthe incomeelasticity the demandforhousing exceeds theincomeelasticity commuting of incomepersons costs,higher should locate farther fromthe CBD and have longercommutes.White

468

THE REVIEW OF ECONOMICS AND STATISTICS


Table 1.- Examplesof the Population Centrality Measure Distance from CBD 2 6 10 1 3 5 % of Total Distance 20 60 100 20 60 100 Actual Population 5 3 2 2 3 5 DistanceWeighted Population 10 18 20 2 9 25 Cumulative Actual Population 50 80 100 20 50 100 DistanceCumulative Weighted Population 21 58 100 6 31 100 | J } J Measureof Population Centrality 0.1694

City A

0.1130

the CBD as livingat a distanceof 2 miles,all personsliving2-5 miles fromthe CBD as livingat a distanceof 5 miles,and all personsliving 6 The correlation of betweenland area and percentage the 5-10 miles fromthe CBD as livingat a distanceof 10 miles. coefficients 9 It is measuretorCityA is 0.166 when that the variousdistancesfrom CBD are as follows: easy to verify thecentrality livingwithin population as Adding 5 km (-0.61), 10 km (-0.66), 15 km (-0.73), 20 km (-0.73). We also a sprawledcityis defined one withuniform density. population has the for similar measures citieswithmultiple CBDs, wheredistances a millionpeople 20 miles from CBD of CityA, however, a much computed measure(it increasesit by 81%) thanon our from CBDs. These measures, largerimpacton the latter the a from pointequidistant weremeasured measure(which it increasesby 21%). withcityarea. correlated too, were highly

for distance.The corresponding figures CityB are 20% of and 50% of the within 20% of thedistance, thepopulation within 60% of thedistance. population of distribution actual Pattern Residential Land Use: In a circular Withwhat should the cumulative of citythe A natural measureof thepattern residential of land use is the population comparedto measurecentrality? possible be of (McDonald, 1989). The density definition a sprawledcityis one witha uniform populapopulation density gradient lives describesthe centralization populationaround tiondensity, of that gradient implying x% of thetotalpopulation theCBD. The density withthecityradius betweentheCBD and x% of thedistanceto thecitycenter. gradient, together to is of this (or city area) and the intercept the densityfunction, Unfortunately, definition quite sensitive thedefidescribes distribution population the of within nitionof the city boundary:adding a small numberof a completely monocentric An alternative thedensity to as the city. gradient a people farfrom CBD will lowertheaveragepopulation of is of and make thecityappearmuchless sprawled. measure centrality thepercentage population living density the of We instead a definition a sprawled in which use at variousdistances 5 10 (within km,within km,and so on) city from CBD is weighted the at from CBD (Glaeser & Kahn, 2001). Both measuresof actualpopulation each distance the a of Our definition sprawlweights person courserequire thatone identify singleCBD. The popula- by distance. a living of tion densitygradientis the more restrictive the two on thecityedge 10 timesas muchas a person living1 mile that the as measures, theconventional by gradient from CBD. Thismaybejustified thefact theperson negative-exponential the with assumesthat declinesmonotonically increasing living10 milesfrom CBD musttravel10 timesas farto density the as of distance from CBD. Because of thepoorfit exponen- reachthecitycenter a person the living1 milefrom CBD. is measure computed averaging in manycities (Malpezzi, 1999), we Our population tial density by centrality gradients n in the between cumulative population annulus reject the populationdensitygradientas a measure of thedifference and of as We distribution. also reject as a measure of (expressed a percentage total population) thecumupopulation as n in of the decentralization percentage population population annulus (expressed distance-weighted livingwithin lative of Formally, 5, 10, 15, and 20 km of the CBD. The correlation distance-weighted population). among a percentage total is measure givenby these measuresand betweeneach measureand city area ourcentrality measuresof urbanform that violatedour criterion different correlated.6 notbe too highly To Populationcentrality. createa measureof population withcityarea, we compare that centrality is less correlated of distribution populathe in percentage terms cumulative of where/,/ = 1, . . . , N, indexesannuliaroundtheCBD, d{ from CBD (as a percentage the distances tionat different maximumdistance fromthe edge of the city) with the is the distanceof annulusi fromthe CBD, and P, is the in at of distribution population thesedistances a populationof annulus iP In our example, the centrality cumulative sprawled city. Table 1 compares two cities with equal measureequals 0.1694 forCityA and 0.1130 forCity B, cities.89 more centralized CityA, witha radiusof 10 miles,and CityB, with largernumbersindicating populations: with a radius of 5 miles. The populationin City A is, 7 If we were to definea however,more centralized:Half of the populationlives population sprawledcityas one withuniform would be -dnldN. in the from CBD to theedge of the density, termon the right the parentheses the within 20% of thedistance 8 In the 2 all we examplecalculation, treat personslivingwithin milesof city,and 80% of the populationlives within60% of that road) and divided by the size of the urbanizedarea (in squarekilometers).

THE EFFECTS OF URBAN SPATIAL STRUCTURE

469

traveledby all (bus) represent railroadcars (buses) duringa year. 10 also 13 We the distance jobs from of Table 2 of Bento et al. (2003) presents between computed averageweighted housing pairwisecorrelations in each urbanarea Galsteret al.'s (2000) proximity measure,originally measuresof sprawland transit supply. 14 it correlated The correlation population of withrailsupplyis 0.48; withbus by proposed MichaelWhite(1986); however, was veryhighly density withcityarea (r = 0.80). it supply, is 0.73. 11 15 statistics for We use the 1990 NPTS because it is the closest NPTS to the years AppendixB ot Bento et al. (2003) presentssummary variablesforall cities in our sample. used to computemeasuresof sprawland transit sprawland transit supply.

Because ourpopulation measure does notcapture balance, and the 92nd least sprawled in termsof road centrality this San Diego, whichis the 13thmostdenselypopucitysize, we supplement withthesize of theurbanarea density. in squaremiles. latedcityin our sample,is themostsprawledcityin terms of job-housing balance. Miami, the2nd mostdenselypopDistribution Employment: The set of possible em- ulated city in the sample,is the least circularcity.These of in the locations an urban area clearly affects commute comparisonsillustrate fact thatour measurescapture ployment in the of of that lengths. Similarly, distribution employment com- dimensions urbanstructure are missingin thepopmercial and retail measure. relative thedistribution ulationdensity to of occupations, residences,is likely to affectthe distance traveledfor Thereare severalways in whichthedistri- C Measuresof Transit nonwork Supply trips. bution employment of could be measured. One is a measure Reliance on public transportation, whether commute for of employment similarto our measuresof popucentrality or fornoncommute on theextent thetransit of trips, depends is lationcentrality; another the employment density gradi- network. measurethe extentof the We netpublic transit that ent.We believe,however, forstudying determinants the workby the number bus routemiles suppliedin 1993 of of driving to behavior,it is more important measurethe divided thesize of theurbanarea (in squarekilometers), by in locationof employment relative population a way that to and by the numberof rail route miles supplied in 1993 is independent the numberor location of CBDs. To of dividedby the size of theurbanarea.12 measure spatialbalance ofjobs versushousingwe have the Not surprisingly, measuresof transit our supplyare corborrowed measurefromthe residential a litersegregation relatedwitheach other, well as withmeasuresof urban as ature(Massey & Denton,1988), whichwe computeusing form.13 Cities that are largerin area and more densely data from1990 Zip Code Business Patterns employment tendto have a greater populated supplyof bothrailand bus (U.S. Census Bureau).10 transit. supplyof nonrail The transit twiceas greatin the is Balance ofjobs versushousing.To measurehow evenly 26 railcitiesin oursampleas in theother cities.Thismay 88 relativeto population, orderZIP we jobs are distributed indicatean attempt link rail and bus networks that to or codes in each cityfrom one havingthesmallest the number similarfactors(such as, populationdensity)favorboth.14 ofjobs to theone havingthelargest plotthecumulative and (r Higher road densityis also correlated = 0.39) with of percentage jobs ( j-axis) againstthecumulative percentof however, greater supply bus transit; population centrality, (jt-axis) to obtaina Lorenz curve.The age of population correbalance, and cityshape are not highly jobs-housing linerepresents even distribution jobs versus an of 45-degree lated withpublictransit or withroad density. supply population.Our balance measure (Massey and Denton's Gini coefficient) thearea betweentheLorenz curveand is III. CommuteMode Choice Models the 45-degreeline, expressedas a proportion the area of underthe 45-degreeline. Larger values of this measure In thissectionwe link the measuresof urbanform and of transit in implya less even distribution jobs versushousing. supplydescribed theprevioussectionto the 1990 How different ourmeasures are from traditional measures NationwidePersonalTransportation Survey(U.S. Departof urban sprawl? Urban sprawl is most oftenmeasured mentof Transportation, 1990) to explainthe"usual mode" area. of commute workof workers to using average populationdensityin a metropolitan livingin the 114 citiesfor Thisis clearly blunt a measure sprawl, is onlyweakly which we have data on urban formand transit of and supply.15 correlated with populationcentrality = 0.16), jobs- Specifically, the 26 cities withrail transit, estimate for we (r multinomial housingbalance (r = 0.06), or cityshape (r = -0.10). logitmodelsof mode choice in whichworkers Thatpopulation and balancecapture choose among(a) driving work, taking transit, to rail (b) centrality jobs-housing (c) different nonrail and For transit, (d) walkingor bicycling. all aspectsof sprawlthanaveragepopulation density taking is illustrated comparing ranking urbanareas by 114 citieswe estimate logitmodel to explainwhether the of a or by thesethreemeasures.11 drivesto work. Using a rankof 1 to indicatethe nota worker urbanized area in our sample,theNew York least-sprawled Mode Choice urbanarea (whichincludesnorthern New Jersey Long A. Modelsof Commute and the urbanized Island) is, notsurprisingly, 3rdleastsprawled Our empirical model of commutemode choice is a area in terms population of It density. is also the 5th least random modelin whichtheobservable of utility component of it however, sprawled cityin terms population centrality; is the 95th least sprawledcity in termsof jobs-housing 12 Rail routemiles the number miles of

470

THE REVIEW OF ECONOMICS AND STATISTICS

relative this to hence all coefficients shouldbe interpreted of The tabledisplaysthecoefficient each explancategory. to atoryvariableforeach mode,theratioof thecoefficient its standard and (for significant the marerror, variables) of each mode with on ginal effect theprobability selecting in Because workers theNew York to respect thevariable.17 of we urbanarea constitute sucha largefraction oursample, results withand without New York. present The effectsof household characteristics commute on mode choice accord withthe literature. Income,race and education all have statistically effectson the significant thata commuter takestransit walksto work. or probability In both samples higherincome workers less likelyto are walk to workor take public transit thantheyare to drive. The incomeelasticity bus, rail,and nonmotorized of modes are well below 1 in absolutevalue in thefullsample(-0.50, a -0.25, -0.46, respectively), resultsimilarto McFadden (1974). The elasticitiesare somewhathigherwhen New York is removed fromthe sample: -0.74, -0.83, and Whitesare -0.49 forbus, rail,and walking,respectively. thanareother less likely ridethebus or train to significantly groups.A 10% increase in years of schoolingraises the of pointsin both probability ridingrail by 1. 1 percentage in elasticities thisimpliesquitedifferent samples;however, in each sampledue to thebaselinedifferences thepercentage of commuters takingrail to work.Resultsforgender B. TheNPTS Worker Sample whichaccords are and householdcomposition not robust, The 1990 NPTS consistsof 22,317 householdslivingin withmuchof theliterature mode choice.18 on urban and rural areasof theU.S. Of thesehouseholds, 9,719 and In examining effects urbanform transit the of supply, livedin the 114 urbanareas forwhichwe have data on both two results effect is the standout.The first that mostrobust These households constitute of urbanform, measured measures. sprawland transport and as by populationcentrality the ourcoresample.To estimate multinomial logitmodelof jobs-housing of is to increasetheprobability walkbalance, commute modechoice,we focuson the26 citieswithsome the increases to centrality ingor bicycling work.Population in railtransit. 6,470 workers our samplehouseholdsin The withelasticities 1.7 of walksto work, a chancesthat worker the mode choice thesecitiesare used to estimate commute with and 2.3 withoutNew York. In cities with greater modes private four model.We distinguish usual commute are balance, workers more likelyto walk to and nonrail railtransit, nonmotorized jobs-housing transit, transportation, for is of the work;however, magnitude thiseffect lowerthan of transit. percentage workers The transportausingprivate < 0.5 in bothsamples). centrality (|elasticity| tionin our sample(79.7%) is lowerthantheaverageforall population The second resultis that increasingrail (bus) supply in in workers the NPTS (85.4%). This is because workers the 30% of increases modalshareforrail(bus) in bothsamples.The theNew Yorkurbanarea constitute approximately of 6% in our sampleof workers rail cities.Approximately of elasticity the rail mode with respectto rail supplyis, large(over 6) whenNew Yorkis includedin the New however, our rail city sample commuteby bus (5% without of New York); approxi- sample,and is no doubtan artifact thehighmodal share York) and 8% by rail (2% without New for rail in the New York area. When New Yorkersare bikeor walkto work(withor without 6% either mately of decision, we use excluded fromthe sample, the elasticity the share of York). In modelingthe drive-no-drive 85.4% of workersin all 114 metropolitan areas, 11,426 17 whomdriveto work. the are Marginaleffects computedby increasing value of an explanindirect from commute modew forhousehold ( Viw) / utility dependson income,on travelcosts,on measuresof urban form and transit and and household availability, on worker characteristics influence that We utility. includein Viwthe and of the age, race,education, gender theworker, numbers of adultsand children thehousehold, in and thehousehold income.The cost per mile of driving calculatedas the is city-specific gasoline price divided by the average fuel of in efficiency cars ownedby households thesame income class as thecommuter.16 on thepriceofrailor bus trips Data were available fortoo few cities to make these variables usable. Also includedin Viware measuresof urbanformand transit We thesemeasuresas exogenousto the supply. treat individual an thatis more difficult to worker, assumption justifyfor road densityand rail and bus supplythanfor measures urbanform of such as cityshape.The problem is thatcity attributes we do not measure(for example, that crimenearpublictransit) bothpeople's propenmayaffect and the supplyof routemiles. Ideally, sityto ride transit measuresof transport supplyshould be modeled together withindividual mode choice. The same measuresof urban formand transit supply are included in both the mode choiceand driving variableaddedto models,witha dummy thelatter capturewhether cityhas rail serviceat all. to a Mode ChoiceResults C. Commute Resultsforour multinomial logitmodelsappearin table 2. In both models the omittedmode is drivingto work;
16 Details are given in AppendixC of Bento et al. (2003). the in variableforeach worker thesampleand predicting probability atory that the worker selects each mode. The average of these predicted is probabilities probabilities comparedwiththe average of the predicted and dummyvarivariable.For integer beforechangingthe explanatory ables a 1-unitchange is evaluated; for continuousvariables,a 10% change. 18 Sarmiento (2UUU)in a reviewor me ettectsor genderana nousenoia of compositionon travel,notes thatthe effect genderon mode choice from one studyto another. variesconsiderably

THE EFFECTS OF URBAN SPATIAL STRUCTURE


Table 2.- Mode Choice Models WholeSample Bus Variable Age of worker Coeff.f e* -3.65 -0.061 (3.97)*** 0.001 Age squared (4.33)*** Indicator female for 0.577 worker (2.33)** of 0.004 Number adultsin thehousehold (0.06) No. of children 0.035 aged 5-21 (0.41) Indicator female -0.936 for workers with children (1.94)* -0.640 Log of income (4.13)*** Years of education -0.018 (0.62) -0.812 White household (6.49)*** Black household 0.609 (3.18)*** Annual rainfall 0.002 (0.10) Annual snowfall -0.215 (1.68)* Gasolinecostof 0.474 (1.13) driving mile per Road density 0.406 (1.80)* of Supply railtransit 6.990 (0.29) of Supply bus transit 40.946 (2.00)** Population centrality 18.837 (2.02)** balance -3.404 Jobs-housing (1.80)* 0.735 Cityshape (1.23) Population density -2.134 (2.21)** Land area 0.112 (0.74) Constant -0.840 (0.13) Observations Rail Coeff.t -0.056 (4.89)*** 0.000 (3.72)*** 0.303 (2.47)** 0.041 (0.85) -0.046 (2.48)** -1.562 (7.30)*** -0.460 (4.08)*** 0.116 (2.47)** -0.935 (4.13)*** 0.092 (0.70) -0.047 (2.16)** 0.169 (1.01) 0.071 (0.19) -0.773 (1.98)** 163.218 (4.75)*** 36.893 (0.93) -46.633 (2.59)*** -3.396 (0.85) 0.465 (0.43) 0.300 (0.21) -0.142 (0.63) 15.913 (2.85)*** 6476 et Nonmotor Coeff.t et -2.78 -0.080 (4.84)*** 0.001 (4.70)*** +17.2 0.148 (1.18) -0.018 (0.55) 0.026 -3.80 (0.45) -67.5 -0.387 (0.89) -0.25 -0.521 (4.23)*** +1.27 0.012 (0.65) -45.8 -0.046 (0.17) -0.09 0.082 (0.38) -0.003 (0.52) -0.055 (1.22) 0.288 (1.40) -5.19 0.183 (1.90)* +6.58 16.048 (1.96)** +0.63 11.543 (1.40) -5.35 8.496 (1.79)* -0.76 -1.823 (3.07)*** 0.129 (0.55) +0.89 -0.659 (1.58) -0.031 (0.54) 1.095 (0.35) Bus Coeff.f et -4.44 New YorkCity Excluding Rail Coeff.t -0.039 (0.89) 0.000 (0.66) 0.118 (0.41) 0.249 (3.03)*** -0.080 (0.85) -0.490 (0.89) -0.917 (4.47)*** 0.304 (6.08)*** 0.185 (0.48) 0.909 (2.89)*** 0.032 (1.53) -0.070 (0.34) -0.800 (1.75)* -0.042 (0.14) 193.274 (5.60)*** -35.599 (1.34) 27.036 (1.37) 1.059 (0.41) -0.883 (1.14) -0.646 (0.52) 0.234 (0.88) -0.045 (0.01) 4468 et

471

Nonmotor Coeff.t et -5.57 -2.50 -0.091 (3.49)*** 0.001 (3.21)*** 0.066 (0.45) +25.5 -0.014 (0.23) -8.33 -0.073 (1.14) 0.328 (1.63) -0.83 -0.665 (5.78)*** +4.58 0.024 (0.91) +25.0 0.301 (1.02) +109 0.324 (1.29) 0.005 (0.92) -0.061 (1.72)* 0.021 (0.15) -1.25 0.305 (3.46)*** +2.92 27.354 (3.67)*** -0.83 -7.092 (1.09) +3.75 17.161 (3.08)*** +0.42 -1.315 (3.00)*** 0.026 (0.12) -0.83 -0.779 (2.08)** -0.006 (0.12) 1.201 (0.56)

+58.5

+3.17 -46.1 -0.50 -0.48 -46.1 +70.4

-4.92 -0.071 (2.84)*** 0.001 (3.07)*** +7.1 0.270 (1.20) -0.017 (0.13) +3.08 -0.121 (2.04)** -14.9 -0.129 (0.39) -0.46 -0.848 (6.75)*** +0.02 -0.021 (0.46) +11.7 -0.925 (5.47)*** +1.54 0.459 (1.58) 0.011 (0.71) -0.200 (1.61) 0.126 (0.32) +2.0 0.529 (1.84)* -0.46 23.742 (0.95) +0.07 19.484 (0.64) +1.69 28.466 (2.07)** -0.46 -2.715 (1.45) 0.539 (0.89) -0.77 -2.238 (2.28)** 0.125 (0.89) 0.885 (0.17)

-2.23 -9.50

-1.79 -4.92

-0.74 -0.56 -58.2 +40.8

-0.49 +0.16 +42.6 +26.7

+5.08 -1.03 +0.63 +4.13 -0.95

+5.0 -0.56 +0.37 +4.26 -0.93

+2.6 +0.16 -0.16 +2.30 -0.33

-2.70

-2.41

-0.66

to Driving workis theomitted category. * Significant 10%; ** significant 5%; *** significant 1%. at at at thecoefficient thevariable for the t The column Standard errors corrected heteroskedasticity clustered each city. are for and at and,in parentheses, z-statistic. reports for and of the to variables. variables, thepercentage $ The column elasticity continuous changein probability choosing modein response a unitchangein discrete reports

rail to commuters fallsto 2.9. taking with respect railsupply The elasticity bus ridership of with respectto bus route milesis 0.63 in thefullsampleand 0.37 without New York. transit have Although supplyand populationcentrality effects rail,bus, and nonmoon nonnegligible percentage torized modalshares, their effect milesdriven workis on to of take small,because onlya small percentage commuters in or transit walk to work.The coefficients table 2 may be effects our measuresof of used to calculatethe marginal of urbanform and transit supplyon theprobability driving on to work.Rail supplyhas thelargest effect driving all of variables.A 10% increase in rail our sprawl and transit

of supplyreducestheprobability driving 4.2 percentage by whichtranslates a 5.3% decreaseat themean,or into points, an elasticity 0.53. The effects 10% increasesin bus of of transit and jobs-housingbalance are more modest, supply in of of resulting a decreasein theprobability driving 1.3% and an increaseof 1.9%, respectively. To summarizethe quantitative effectsof sprawl and transit variableson theprobability driving workusing of to data fromall cities in our sample, table 3 presents logit models of the drive-no-drive decision thatare estimated in areas. These models usingworkers all 114 metropolitan are used to calculatethemarginal effect a 10% changein of

472

THE REVIEW OF ECONOMICS AND STATISTICS The effects rail and bus routemileson thedecisionto of driveto work,though are statistically significant, generally smallerin magnitude thaneitherpopulationcentrality or road density.The elasticityof drivingwith respectrail supplyis -0.11 in the full sample but only -0.03 in the New York. The corresponding elasticities sample without forbus routemiles are -0.07 withNew York and -0.03 without New York.The magnitudes theseresults quite of are in mode choice plausiblein lightof findings thecommute literature. milessupplied should Changesin bus or railroute affect mode choice through theirimpacton waitingtimes forbus and rail.McFadden(1974) reports elasticities the of of to to wait probability driving workwithrespect transfer timesof 0.07 forbus and 0. 11 forrail,whicharein linewith our findings. The effects driving computeusingthe sampleof on we workers all 114 citiesare usuallysmaller - butin the in than samedirection - thecorresponding as effects from mode the choice models, which used workersin the 26 cities with some railtransit. example,a 10% increasein thesupply For ofrailtransit reducesdriving 4.2 percentage in by points the 26 citysample,and by onlyabout 1 percentage in the point 114-city sample.A 10% decrease in jobs-housingbalance increases probability driving 1.5 percentage the of by points in thesmallersample,butby only0.4 percentage pointsin thelargersample. IV. Models ofAutomobile Ownership and Annual VMTs

Table 3.- BinaryLogit Models of Workers'DrivingDecisions WholeSample Drive 0.056 (5.95)*** -0.001 Age squared (5.81)*** Indicator female for worker -0.315 (3.12)*** Number adultsin the of -0.01 1 household (0.48) No. of children -0.050 aged 5-21 (1.89)* Indicator female for 0.781 workers with children (1.76)* 0.554 Log of income (6.02)*** Years of eduation -0.023 (1.39) White household 0.558 (2.93)*** Black household -0.359 (2.75)*** Annual rainfall 0.009 (2.07)** Annual snowfall -0.060 (3.13)*** Gasolinecostof driving per -0.1 17 mile (0.74) Road density 0.061 (2.20)** Presence railtransit of -0.291 (1.94)* of -38.137 Supply railtransit (5.30)*** - 26.762 of Supply bus transit (2.54)** -4.407 Population centrality (1.50) balance 0.894 Jobs-housing (1.28) -0.113 Cityshape (0.37) 0.178 Population density (0.63) Land area 0.039 (0.70) Constant -4.051 (2.14)** Observations 11,541 Age of worker ef New York Excluding Drive 0.057 (4.16)*** -0.001 (4.52)*** -0.193 (2.15)** -0.025 (0.65) 0.001 (0.04) 0.016 (0.09) 0.696 (11.10)*** -0.021 (0.81) 0.209 (1.54) -0.547 (3.66)*** 0.009 (2.20)** -0.069 (3.58)*** 0.112 (0.61) 0.055 (2.08)** -0.232 (1.72)* -51.250 (4.92)*** - 2 1.022 (1.83)* -6.137 (2.17)** 0.593 (0.88) -0.133 (0.43) 0.067 (0.25) 0.053 (1.07) -5.827 (3.34)*** 9533 ef

-3.99

-1.86

+0.07 -0.04 +8.22 -4.89 +0.06 -0.01 -0.07 +0.08 -3.64 -0.11 - 0.07 -0.09

+0.06 -0.03 +2.15 -5.50 +0.03 -0.01 +0.06 +0.04 -2.23 -0.03 - 0.03 -0.10

Urbanformand transit household supplymay influence VMTs by affecting either number cars ownedor the the of number mileseach car is driven. therefore of We estimate a +0.00 +0.00 modelto explainthenumber carsownedand thedemand of +0.04 +0.01 forVMTs per vehicle(Train,1986; Goldberg,1998; Walls, & 2000; West,2004). 19The modelis Harrington Krupnick, +0.02 +0.01 estimated two parts.The first in partis a multinomial logit modelthatexplainswhether householdowns zero,one, the or the two,or three morevehicles.We thenstudy determinants annualVMTs pervehicleseparately households of for in Standard errors corrected heteroskedasticity clustered each are for and z statistics parentheses. at that own one, two, or threeor more vehicles. Because city. * Significant 10%; ** significant 5%; *** at at at significant 1%. unobservable factorsthatexplain the number vehicles of t The column for and of variables, thepercentage reports elasticity continuous changein probability the to variables. choosing modein response a unit changein discrete owned maybe correlated withtheerror in terms theVMTcorrection per-vehicleequations, we use the selectivity Dubin and McFadden (1984) to approach developed by the each variable on the probability thata randomly chosen estimate demandforVMT equations. worker drivesto work,whichis expressedas an elasticity. A. Specification theEconometric Model of Of all measuresof urbanform, and population centrality road density have the largestimpacton whether worker a We estimate multinomial a of logitmodelof thenumber drivesto work.Theireffects, in magni- vehiclesowned and an equationfortheaveragenumber though comparable of tude to the effects income and education,are small in of absolute terms.A 10% increase in populationcentrality 19Train (1986) and Goldberg(1998) have also examinedthe determilowersthe probability driving approximately per- nantsof vehicle ownership, theirfocus differs 1 of but from ours.Trainuses by VMTs. = -0.09). If the average worker models of vehicle ownershipand utilizationto forecastfuture centagepoint(elasticity Goldberg models vehicle ownershipand use to study the effectsof drives6000 miles to workeach year,thistranslates intoa corporateaverage fuel economy standards.Neitherstudyexaminesthe reduction 54 miles annually. of of on effects urbanform traveldemand.
+0.05 +0.02

THE EFFECTS OF URBAN SPATIAL STRUCTURE miles driven,conditionalon owning a vehicles. Let the indirect household receivesfrom / utility owninga vehicles be written thesumof an unobservable as uia component and an observable Via component thatincludeshouseholdcharacteristics (whichmayaffect the Z/ utility); pricepermileof <p; driving, incomenetof thefixedcosts of car ownership, of yt Fa\ and characteristics theurbanarea in whichthe household that owns lives,S;. The probability thehousehold a vehiclesis givenby B. Characteristics Vehicle of Ownership

473

In our sample including New York,approximately 14% of householdsown no passengervehicles,33% own one and 14% own three more or vehicle,39% own twovehicles, vehicles.The percentage households of owningno vehicles fallsto 10% whenNew Yorkis excluded,and thepercentage owningone, two, or threeor more increasesslightly. for Averagemilesdriven vehicleare highest twovehicle per households(12,264 miles per year; 12,428 without New vehicle (11,719 miles; 11,836 = P(Via + uia > Vih uib)9 all b*a, York), and higherfor one+ Pa New York) thanforthree- more-vehicle without or house~ Fa). holds (11,218 miles; 11,260 without New York). The dif= B,Z, + rflSf- |Jfl<p/ + + 7(j/ where Via ferencein average miles drivenper vehicle betweenone and on onlyabout600 milesper Conditional a, thenumber miles thata household category thenextis, however, of The accordswiththefactthat substantial the increases drives,per vehicle,will depend on the same variablesas year. in vehicle miles traveled U.S. householdsover the last enter indirect the function by Vi(n utility two decades have occurred largelybecause of increasesin = D.Z, + HflSf. afl<p/ K(yi ~ Fa) + e/fl. (3) the number vehiclesowned rather of thanin miles driven + + {Mia), vehicle.Finally, difference driving the in habits between per Because thesame unobservable variables that affect vehicle the full sample and the sample includingthe New York are miles driven, is reasonable urbanarea is small.AverageVMTs pervehiclebetween it the ownership likelyto affect to assume that the errortermin the average-miles-per-two samplesare significantly different fortwovehicle only vehicleequation, will be correlated withuia. We handle households. e,, thisby addingthe selectivity correction factorderivedby Dubin and McFadden to equation(3). C. Modelsof Vehicle Ownership To estimate equations(2) and (3) we mustmeasurethe costpermiledriven thefixed and costsof vehicleownership Table 4 presentsthe vehicle ownershipmodels. The foreach household.The fixedcosts of vehicle ownership omitted categoryin each model is "owns no cars." In includethecostsof interest depreciation thevehicle, and on addition reporting multinomial to the and logitcoefficients as well as the cost of automobileinsurance.The make, their standard as errors, effects, marginal expressed elasticof model,and vintage each vehiclethehouseholdowns are ities,are calculatedforvariableshavinga statistically sigrecordedin the NPTS. However, to avoid endogeneity nificant effect vehiclechoice.20 on problems(forexample,the chosen make and model may The effects householdcharacteristics vehicleownof on reflect household's the for we preferences driving), estimate as ershipare largely expectedand agreewiththeliterature. thecost permileand fixedcostsof vehicleownership a for Householdsize and composition have a significant effect on in household household s incomeclass. [Appendix f typical thenumber vehiclespurchased, found Train(1980, of as by C of Bento et al. (2003) describesour calculationof the and Mannering Winston and are (1985). These results fixedcosts of vehicle ownershipand the price per mile 1986) robust theinclusionor exclusionof New Yorkfrom to the The priceper mile is thepriceof gasolinein the traveled.] have a smaller chanceof owning household'sMSA divided by the average fuel efficiency sample.Whitehouseholds no vehiclesand larger chancesof owningone, two,or three (miles per gallon) of vehiclesowned by householdsin the or more vehicles than nonwhite households.When New household'sincome group. Household characteristics (Z) Yorkis excludedfrom sample,black households the have a includethenumber persons thehousehold in of classified by chanceof owningzero or one vehiclethannonblack larger the head,and the age and workstatus, race of thehousehold an member a has family of number yearsof schoolingcompleted the mostedu- households.In general, additional by effect theprobability owningtwo vehicles(or on of larger cated personin thehousehold.S includesour measuresof or urban form transit and as and three morevehicles)thandoes race. supply, well as annualrainfall Incomeand education have small butstatistically signifannualsnowfall. icanteffects car ownership. on Increasesin income(netof These models are estimated using all householdsin the costsof car ownership) reducetheprobability a of 1990 NPTS livingin the 114 urbanareas forwhichcity- thefixed household owning one or no vehicles, but increase the wide sprawland transit measureshave been computed and forwhomcomplete data on VMTs are available.The subset 20 are the Marginaleffects calculatedas in table2, by computing effects of thesehouseholdsforwhichall householdvariablesare of a unitchange in race and in the number family of members and of a availablenumbers our 8,367.As above,we estimate models 10% changein othervariableson theprobability a householdselects that withand without householdsin theNew Yorkurbanarea. each alternative. These changesare averagedacross households.

474

THE REVIEW OF ECONOMICS AND STATISTICS


Table 4.- MultinomialLogit Models of Vehicle Choice WholeSample 1-Car Coeff.t e* -0.030 -23.0 (0.42) 0.683 -46.4 (7.95)*** 0.493 -36.5 (3.83)*** -0.079 -38.8 (0.80) -0.366 -38.9 (3.02)*** 0.015 -6.9 (0.22) 0.201 -0.3 (1.94)* 0.167 (8-26)*** 1.137 +12.9 (6.02)*** -0.140 (0.56) -0.000 (0.09) -0.042 (1.52) -0.063 +1.1 (0.27) 1.415 (0.17) -0.217 +0.1 2-Car Coeff.t e$ 0.526 +16.4 (4.84)*** 1.968 +28.0 (12.94)*** 1.416 +14.5 (6.20)*** 0.798 +13.5 (8.30)*** 0.181 -16.8 (1.13) 0.239 +12.2 (2.86)*** 0.931 +0.2 (6.23)*** 0.207 (9.97)*** 1.344 +14.2 (4.30)*** -0.417 (1.43) -0.004 (0.79) -0.007 (0.22) -0.564 -0.9 (1.73)* -0.021 (0.00) -0.240 +0.4 3f-Car Coeff.t e$ 0.618 +20.2 (4.87)*** 2.528 +86.8 (15.73)*** 2.084 +89.1 (9.15)*** 1.356 +72.3 (11.20)*** 1.413 +137 (6.60)*** -0.015 -12.2 (0.16) 1.196 +0.4 (4.31)*** 0.193 (7.84)*** 1.581 +34.0 (4.10)*** -0.494 (1.29) -0.010 (1.47) -0.024 (0.52) -0.606 -0.8 (1.76)* -2.142 (0.23) -0.220 +0.1 1-Car Coeff.t -0.058 (0.61) 0.757 (5.13)*** 0.654 (4.00)*** -0.160 (1.45) -0.227 (1.79)* -0.069 (1.39) 0.332 (4.95)*** 0.186 (8.53)*** 0.820 (4.82)*** -0.536 (2.90)*** 0.001 (0.35) -0.049 (1.92)* 0.139 (0.90) 0.539 (0.06) -0.290 +9.5 +4.7 e$ -27.6 -64.5 -41.5 -42.7 -39.5 -7.4 -0.3 +0.09 New YorkCity Excluding 2-Car Coeff.t e* 0.597 +17.6 (4.44)*** 2.131 +26.6 (11.86)*** 1.715 +16.2 (9.45)*** 0.801 +13.2 (5.66)*** 0.360 -18.4 (2.06)** 0.133 +10.7 (2.34)** 1.117 +0.2 (11.17)*** 0.221 (8.21)*** 0.880 (4.01)*** -0.812 (3.09)*** -0.003 (0.55) -0.017 (0.55) -0.265 (1.31) -0.911 (0.12) -0.299 +4.6 -9.2 +0.3 3t-Car Coeff.t e* 0.710 +22.3 (4.83)*** 2.691 +82.3 (13.84)*** 2.369 +85.8 (11.44)*** 1.366 +71.0 (8.10)*** 1.666 +144 (9.04)*** -0.122 -13.2 (1.74)* 1.534 +0.5 (11.35)*** 0.199 -0.05 (5.90)*** 1.019 +15.3 (4.27)*** -0.983 -18.7 (3.08)*** -0.008 (1.34) -0.035 (0.78) -0.284 (1.20) -3.378 (0.36) -0.290 (1.58) -29.965 (2.01)** -19.336 (1.14) -12.580 (3.11)*** -0.722 (1.79)* 0.913 (0.94) -0.141 (1.90)* 0.492 (1.32) -17.522 (7.63)*** -0.02 +0.05 -0.2 -0.09

Elderly adult males Working adult Working females adults Nonworking Children aged 17-21 Children 0-16 aged Log ofadjusted income Yearsofschooling memter6'11103^ White household Blackhousehold Annual rainfall Annual snowfall Costofautomobile travel mile per Roaddensity Presence rail of transit

(1.46) (1.42) of -11.751 Supply railtransit -10.862 (1.34) (1.21) of +0.01 -30.347 Supply bustransit -17.551 (1.76)* (2.21)** -0.2 -10.569 Population centrality -9.402 (3.45)*** (2.88)*** -0.479 -0.657 City shape (1.31) (1.55) balance 0.149 0.889 Jobs-housing (0.23) (1.08) Landarea -0.098 +0.04 -0.199 (1.66)* (2.70)*** 0.578 +0.1 0.701 Population density (2.02)** (2.23)** Constant -1.988 -8.683 (0.79) (2.74)*** Observations 8367

-0.1 -0.2

-0.12 +0.15

(1.03) (2.38)** (2.16)** -14.422 -16.825 +0.004 -25.394 (1.27) (1.91)* (2.19)** -20.636 +0.04 -17.853 +0.02 -29.338 (1.16) (1.94)* (2.38)** -11.359 -0.2 -9.705 -0.09 -11.462 (2.83)*** (3.41)*** (3.45)*** -0.763 -0.419 +0.02 -0.610 (1.61) (1.34) (1.72)* 0.809 0.958 0.238 (0.76) (1.39) (0.42) -0.177 -0.05 -0.069 +0.05 -0.163 (2.75)*** (2.03)** (1.49) 0.496 -0.1 0.610 +0.06 0.709 (2.29)** (2.46)** (1.26) -12.765 -4.371 -11.974 (6.80)*** (2.86)*** (2.96)*** 6878

-0.01 -0.09 -0.14 -0.04

-0.06 +0.11

-0.01 -0.1

* Significant 10%; ** at at at significant 5%; *** significant 1%. and at for the are for the Standard errors corrected heteroskedasticity clustered each city. t The column and,in parentheses, z statistic. reports coefficient thevariable variables. for and of the to variables, thepercentage $ The column changein discrete reports elasticity continuous changein probability choosing modein response a unit

that or probability it owns two,or three more,vehicles.An Households in less sprawledcities (cities withmorecenare increasein the years of schoolingof the most educated tralized populations) less likelyto own one vehicle,two increasesthe chances thata household vehicles, or threeor more vehicles. A 10% increase in householdmember of reducestheprobability owningtwo owns one or two vehicles.The factthatthevehicleowner- population centrality of 1.5% and the probability is inelasticwithrespectto income agrees withother vehicles by approximately ship 2.1% in & U.S. studiesbased on householddata (Mannering Win- owningthreeor more vehicles by approximately citiesreducetheodds of ownbothsamples.More circular ston,1985; Train,1980). two or more vehicles, althoughthe effectis only of Of ourmeasures urban form, onlypopulation centrality ing The of has a significant balance, significant. effect jobs-housing impact on the odds of car ownership. marginally

THE EFFECTS OF URBAN SPATIAL STRUCTURE

475

is different from0 at con- increasesannualmiles driven thesehouseholds, does as by contrast, neversignificantly by ventional a decreaseinjobs-housing balance.The morecircular city, a levels,noris thatof road density. a of the vehiclehouseholds. rail In Amongmeasures transit, 10% increasein bus supply thefewer milesdriven oneby reducestheodds of owningtwo vehiclesby approximately citiesan increasein rail routemilesreducesannualVMTs; or 1%, whether not New York is includedin the sample. however,the magnitude this effectis sensitiveto the of When New Yorkis excludedfrom sample,greater the rail inclusionof New Yorkin the sample. supplyreducesthe likelihoodof vehicle purchase,condiWhatare thecombinedeffects our measuresof urban of tionalon a cityhavinga rail system beginwith.21 to form and transit on totalmilesdriven The supply annually? effectof a 10% change in city shape, road density, rail D. Modelsof VMTper Vehicle Effects TotalVMTs and on supply (for rail cities), and jobs-housingbalance is to annual miles drivenby at most 0.7% for Table 5 presents demandfunctions VMTs pervehicle, change average for each variable.Populationcentrality, whichaffects average estimatedseparatelyfor one-, two-, and three or more VMTs only through effecton vehicle choice, has a its vehiclehouseholds. The selectivity correction term addedto but still modest,effect. 1% increasein A larger, each equationis based on table 4. Inasmuchas thedepen- slightly reducesaverage annual miles driven population centrality dent variable is the logarithm VMTs per vehicle, the of the by 1.5% whenNew Yorkis removedfrom sample.As coefficients table5 represent proportionate in the changein we report elsewhere(Bento et al., 2003), the 10% increase annualhousehold VMTs corresponding a 1-unit to changein in New York each variable(withtheexceptionof the incomevariable), in populationcentrality the sample without reducesannualaverageVMTs by approximately miles 300 holdingthehousehold'svehiclestockconstant. - approximately thesize (in absolutevalue) of half The number personsin a householdhas a significant peryear of in income.Individually, effect the effect annualVMTs per vehicle;however, on thiseffect is a 10% increase household notas great theeffect an additional as of supplyis on of changingmeasuresof urbanformand transit generally person VMTs thatoccurs through vehicle choice.22For example, small indeed, smaller than the predictedimpact of a and on without New York,addingan adult changein thegasolinetax on vehicleownership miles focusing theresults & male to a household raisesaverageVMTs by approximately driven(Mannering Winston, 1985; West,2004). This is, not the of with most of this effectoccurring however, necessarily case ifmeasures urbanform 6,000 miles annually, vehiclechoice (5,000 miles) rather thanmiles per and transit supplyare considered jointly. through vehicle(1,000 miles).Addinga working adultfemaleto the household a youngadultaged 17-21 increases or driving by V. The Effect ChangingAll Sprawl and Transit of In 5,000 miles annually. each case approxiapproximately Measures Simultaneously occursthrough increase an 4,000 milesof thiseffect mately in the number vehicles owned rather of than through an The resultspresented above suggestthat measuresof increasein milesdrivenper vehicle. urban sprawl and transitavailabilitymay have modest Previous studies (Mannering& Winston,1985; Train, effects thecommute on mode choices and annualVMTs of on 1986) suggestthatincome has a small effect vehicle U.S. households. This is, however, necessarily case not the of usage, holdingnumber vehiclesconstant. Regardlessof ifseveralmeasures urbanform transit of and supply change thenumber vehiclesowned,theelasticity VMTs with of of To all simultaneously. examinethe impactof changing of to the of respect incomeis small,although incomeelasticity our measuresof sprawland transit we availability, predict annualVMTs is approximately twiceas highin one-vehicle households 0.30 without New York as in two-or three- thevehiclechoicesand VMTs pervehicleof all households in our sample, assuming that they live in a city with vehiclehouseholds (table 5). measuresof urbanform and transit identical to availability Our sprawland transit measureshave statistically signifthosein each of six U.S. cities:Atlanta, Boston,Chicago, icanteffects miles drivenper vehicleprimarily oneon in vehicle households(table 5). An increasein road density Houston,New Yorkand San Diego. We also use thelogit model from table 3 to predictthe probability that the 21Note that workers our sample householdswill driveto work,for in equations (2) and (3) include a dummy variable (rail is and of dummy) equal to 1 ifa railsystem present 0 if it is not.Rail supply each of thesix cities.As a checkon theconsistency our be as of maytherefore interpreted theproduct rail miles suppliedand the results, commutemiles drivenare calculatedby average rail dummy. 22The effect a variableon annualVMTs can be the of to of as computed follows. multiplying average probability driving workby Let P\MX + P2M2 + P3M1 be the average household miles traveled the number workers our sample households of in average beforea variable is altered,where Pi is the proportion households of in (1.04) timestheaverageannualcommute length the1990 owning/ vehicles and Mj is the annual average VMTs for households 6,000 miles per year). The ratioof owning/ vehicles,/ = 1,2,3. Let primesdenotethevalue of each term NPTS (approximately after variableis altered. a The changein averageannualVMTs, 2 PJM*milesdriven averageVMTs shouldbe to averagecommute 1 P,Mh can be decomposedas I (Pj - Pi)M, + 2 P* (Mj - M,), one-third, given thatworktripsaccountfor wherethe first termrepresents effect vehicle ownership an on and the approximately second an effect miles traveled vehicle. on 34% of annualVMTs in the 1990 NPTS. per

476

THE REVIEW OF ECONOMICS AND STATISTICS


Table 5.- OLS Models of ln( Vehicle Miles Traveled) per Vehicle WholeSample 1-Car Households 2-Car Households -0.194 (5.44)*** 0.071 (1.78)* -0.011 (0.22) -0.056 (1.40) 0.046 (0.81) -0.012 (0.67) 0. 126 (3.73)*** 0.028 (5.05)*** 0.078 (1.34) -0.080 (0.81) 0.001 (0.63) -0.009 (1.47) 0.012 (0.19) 1.014 (0.70) 0.043 (1.18) -4.702 (2.32)** -3.150 (0.82) -0.036 (0.04) 0.093 (0.80) 0.324 (1.43) -0.020 (0.28) 0.017 (0.97) 0.007 (0.33) 7.163 (12.47)*** 3247 0.07 3-Car Households -0.024 (0.52) 0.085 (2.11)** 0.044 (0.96) 0.025 (0.39) 0.165 (2.39)** -0.041 (1.81)* 0. 134 (2.97)*** 0.031 (2.45)** -0.024 (0.29) 0.010 (0.08) 0.002 (0.74) 0.008 (0.50) -0.081 (0.80) 0.734 (0.26) -0.026 (0.38) -1.046 (0.32) -6.210 (0.85) -0.389 (0.28) -0.240 (1.33) 0.204 (0.69) 0.104 (0.75) -0.002 (0.08) -0.023 (1.12) 7.408 (8.32)*** 1208 0.04 New YorkCity Excluding 1-Car Households -0.197 (2.13)** 0.539 (4.05)*** 0.262 (2.15)** 0.086 (0.71) 0.369 (2.88)*** 0.040 (0.99) 0.299 (4.70)*** 0.031 (2.90)*** -0.023 (0.21) -0.198 (1.49) 0.003 (1.25) -0.011 (0.47) 0.056 (0.61) 6.632 (2.16)** -0.026 (0.35) -9.281 (1.69)* -3.091 (0.59) 1.503 (0.98) -0.286 (2.23)** 0.734 (1.64) -0.092 (0.73) 0.002 (0.08) 0.096 (2.22)** 4.730 (4.30)*** 2296 0.12 2-Car Households -0.217 (6.47)*** 0.056 (1.39) -0.037 (0.79) -0.079 (2.00)** 0.077 (1.29) -0.023 (1.36) 0.123 (3.13)*** 0.028 (4.45)*** 0.073 (1.09) -0.108 (0.96) 0.001 (0.78) -0.009 (1.42) 0.023 (0.33) 1.099 (0.75) 0.034 (0.98) -3.366 (0.78) -3.112 (0.78) 0.076 (0.09) 0.096 (0.83) 0.336 (1.50) -0.020 (0.27) 0.020 (1.20) 0.019 (0.97) 7.195 (11.90)*** 2824 0.08 3-Car Households 0.000 (0.00) 0.053 (1.16) 0.010 (0.20) -0.028 (0.47) 0.122 (1.65) -0.037 (1.42) 0. 12 1 (1.93)* 0.034 (2.39)** -0.055 (0.59) -0.001 (0.01) 0.002 (1.03) 0.010 (0.56) -0.070 (0.63) 1.141 (0.41) -0.054 (0.75) 2.929 (0.45) -7.248 (0.92) -0.029 (0.02) -0.233 (1.30) 0.250 (0.80) 0.116 (0.82) 0.003 (0.11) -0.006 (0.27) 7.529 (7.16)*** 1048 0.04

Elderly adultmales Working adultfemales Working adults Nonworking Children aged 17-21 Children aged 0-16 income Log of adjusted Years of schooling most of educated member White household Black household Annual rainfall Annual snowfall Cost of automobile travel permile Road density Presence railtransit of of Supply railtransit of Supply bus transit Population centrality Cityshape balance Jobs-housing Population density Land area correction factor Selectivity Constant Observations R2

-0.188 (2.74)*** 0.480 (5.87)*** 0.248 (3.32)*** 0.021 (0.28) 0.323 (3.62)*** 0.043 (1.31) 0.232 (3.54)*** 0.042 (3.44)*** 0.063 (0.59) -0.096 (0.69) 0.003 (1.34) -0.007 (0.31) 0.012 (0.14) 6.954 (2.30)** -0.031 (0.41) -5.160 (1.35) -4.148 (0.85) 1.899 (1.28) -0.275 (2.15)** 0.762 (1.66)* -0.077 (0.60) 0.000 (0.01) 0.091 (3.37)*** 5.334 (4.87)*** 2762 0.11

/-Statistics parentheses. in Standard errors corrected heteroskedasticity clustered each city. are for and at * Significant 10%; ** significant 5%; *** significant 1%. at at at

The results ourexperiment of appearin table6.23Several resultsstand out. The probability drivingto work is of lowest in the oldest threecities in the table New York (0.40), Boston (0.73), and Chicago (0.74), each of which of has an extensiverail and bus system.The probability
23The calculations table6 are based on themodelsin tables3, 4, and in 5 thatincludeNew York. Using the models excludingNew York alters of predictions averageannualVMTs by at most2%. Usingthelogitmodel of to estimated without New Yorklowersthepredicted probability driving workby 1-2 percentage points.

drivingis highestin Houston (0.90), which has no rail of road density the six cities.The systemand the highest behaviorare in line with our of predictions commuting of predictions total householdVMTs- average commute 26% of annualVMTs in New York milesdriven rangefrom to 36% of annualVMTs in Houston. of The effects all measuresof urbanformand transit If supply on average household VMTs are striking. the measures in households oursamplewereto livein a citywith to of urbanformidentical thosein Atlanta, averageannual

THE EFFECTS OF URBAN SPATIAL STRUCTURE


Table 6.- Impactson Drivingof MovingOur Sample Households to Various MetropolitanAreas UrbanArea Lane density (area of roads per 100 squaremilesof land) Land area (km2) Population Density (peoplepersquare kilometer) Presence railtransit of Rail transit (10,000 supply mi/km2) Nonrail transit supply (10,000mi/km2) balance Jobs-housing centrality Population Cityshape Predicted averageannual vehiclemilestraveled perhousehold Predicted average of to probability driving workby workers Predicted commute average milesdriven Minimum Maximum Atlanta, GA MA Boston, Chicago,IL TX Houston, New York,NY

477

San Diego,CA

1.6 135 158,553 446 0 0 0.1 0.115 0.114 0.038

10.6 7,683 16,044,012 2,240 1 5.7 4.3 0.58 0.218 0.994

3.9 2,944 2,157,806 733 1 0.7 1 0.443 0.114 0.264

4.3 2,308 2,775,370 1,202 1 1.8 1.3 0.284 0.171 0.816

4.7 4,104 6,792,087 1,655 1.0 1.9 2.75 0.35 0.15 0.48

5.2 3,049 2,901,851 952 0 0.0 1.42 0.44 0.13 0.80

5.3 7,683 16,044,012 2,088 1 5.7 3 0.412 0.197 0.727

4.2 1,788 2,348,417 1,314 1.0 0.2 1.64 0.58 0.20 0.36

16,899 0.87 5,450

12,704 0.73 4,565

14,408 0.74 4,620

15,685 0.90 5,641

9,453 0.40 2,496

16,493 0.84 5,247

a 10% change in rail and bus miles suppliedare approximatelyhalfas large.Urban formand transit supplyaffect annualmilesdriven influencing thenumber cars both of by ownedand themilestraveled vehicle.In citieswhere the per of is spatialdistribution population morecompact,householdsare less likelyto own a car.The quantitative effect of thesevariables annualaverageVMTs is, however, on small: a 10% increasein population its on centrality, through effect vehiclechoice,reducesannualVMTs by only 1.5%. Other measuresof urbanformand transit supply jobs-housing road density, and the supplyof rail balance, city shape, transit all affect averagemiles drivenper vehiclebut the notthenumber vehiclesowned.The elasticity annual of of VMTs withrespect each of thesevariables less than 1 to is 0. in absolutevalue. The resultspresentedabove suggestthat measuresof urbansprawland transit availability mayhave onlymodest effects thecommute on mode choices and annualVMTs of U.S. households. This is, however, necessarily case not the ifseveralmeasures urbanform transit of and supply change To for simultaneously. examinethepotential suchmeasures we use themodelsof vehicleownership milesdriven and to VI. Conclusions theannualmilesthat each of oursamplehouseholds predict Our resultssuggestthat individualmeasuresof urban would drive if theywere to live in a city withthe same form publictransit and of and transit supplyhave a smallbutstatistically measures urbanform supplyas six U.S. cities effect traveldemand.For example, a 10% (Atlanta,Boston, Chicago, Houston,New York,and San on significant increasein population lowers the chance thata Diego). We perform similarexercise for commuters a centrality to worker drivesto workby 1 percentage of predictthe probability drivingto work.This exercise point.The effects of and transit suggeststhatmeasuresof urbanform supply 24 we Formally, calculate( 1/N) I, I, P(iJ)M(iJ), whereP(i,j) is the takentogether have a significant effect travel on demand. thathousehold i purchasesvehicle bundlej, and predicted probability of is the numberof miles the household is predictedto travel The effect movingour sample householdsfroma city M(/, j) conditional owningbundlej. on withmeasures urbanform of and transit supplythesame as

VMTs per householdwould equal 16,899.24 This number drops to 12,704 miles annuallyif the householdsin our samplemove to a citywithurbanformand transit supply variablesidentical Boston a reduction annualVMTs to in of 25%. This result driven differences publictransit is in by in cityshape,and,especially, population in centralsupply, betweenthe two cities. Atlantais almost 2 standard ity deviations below the mean of all 114 cities in population whereas Bostonis 0.66 standard deviations above centrality, the mean. Jobs-housing balance is also greater Boston in than in Atlanta.When we move the households in our sample to New York, the effectis even more striking annualVMTs per householdfall to 9,453. This is average the resultof large differences populationcentrality in betweenAtlantaand New York (New York is almost two standard deviations above themeanforall U.S. cities),and of differences between twocitiesin thesupplyof public the rail transit. Annual VMTs in Chicago transit, especially - thehighest ourthree of oldercities are still8% (14,408) below annual VMTs in a city with the characteristics of Houstonand 15% below annualVMTs in Atlanta.

478

THE REVIEW OF ECONOMICS AND STATISTICS


"The Environmental Journal Kahn,Mathew, Impactof Suburbanization," of PolicyAnalysisand Management19:4 (2000), 569-586. to Levinson, D. M., and A. Kumar,"Density and the Journey Work," Growth and Change 28:2 (1997), 147-172. of and Malpezzi, Stephen,"Estimatesof the Measurement Determinants of UrbanSprawl in U.S. Metropolitan Areas,"University Wisconsin mimeograph (1999). "A Fred,and Clifford Winston, DynamicEmpirical Analysis Mannering, of HouseholdVehicleOwnership Utilization," and RandJournal of Economics 12:2 (1985), 215-236. Massey, Douglas, and Nancy Denton, "The Dimensionsof Residential Social Forces 67:2 (1988), 281-315. Segregation," "Econometric Studies of UrbanPopulationDensity:A McDonald, John, Journalof Urban Economics26:3 (1989), 361-85. Survey," of McFadden,Daniel, "The Measurement UrbanTravel Demand,"Journal of Public Economics3 (1974), 303-328. Mieszkowski, Peter, and Edwin Mills, "The Causes of Metropolitan Journalof Economic Perspectives7:3 (1993), Suburbanization," 135-147. and of Mills, Edwin,"The Measurement Determinants Suburbanization," Journalof Urban Economics32:3 (1992), 377-387. Muth, Richard E., Cities and Housing. The Spatial Patternof Urban ResidentialLand Use (Chicago: The University Chicago Press, of 1969). Cities and Automobile Newman,Peter,and Jeffrey DepenKenworthy, dence: An International Sourcebook(Aldershot: Gower, 1989). Sarmiento, Sharon,"Household,Genderand Travel,"in Proceedings from on 's theSecond National Conference Women TravelIssues Federal (2000). HighwayAdministration "A and Train,Kenneth, Structured Logit Model of AutoOwnership Mode Choice," Reviewof EconomicStudies47:2 (1980), 357-370. MA: MIT Press, 1986). QualitativeChoice Analysis(Cambridge, of Federal HighwayAdministration, U. S. Department Transportation, Nationwide Personal Transportation Survey (WashingtonDC, 1990). WinstonHarrington, Alan Krupnick, and Walls, Margaret, "Population and Vehicle Travel: Results froma Density,TransitAvailability NestedLogitModel of VehicleChoice and Use," Resourcesforthe Futuremimeograph (2000). Vehicle Pollution Effectsof Alternative West, Sarah, "Distributional ControlPolicies," Journalof Public Economics 88 (2004), 735757. WilliamC, "Land Use and Densityin Cities withCongestion," Wheaton, Journalof Urban Economics43:2 (1998), 258-272. White,Michael J., "Segregationand Diversity:Measures in Population Distribution," PopulationIndex 52 (1986), 198-221. LocationChoice and CommutWhite,Michelle,"A Model of Residential Journalof Regional Science ing by Men and WomenWorkers," 17:1 (1977), 41-52. American "Sex Differencesin Urban CommutingPatterns," EconomicReview76:2 (1986), 368-372. Behavior in Cities withDe"Location Choice and Commuting Journal of Urban Economics 24:2 centralized Employment," (1988), 129-152.

to thoseofAtlanta a citywithmeasures same as those to the ofBostonis to reduceannualVMTs by25% and theaverage of to 0.87 to 0.73. probability driving workfrom Our resultsshould howeverbe interpreted withcaution forat leasttworeasons.First, sprawlmeasures our could be capturing aspects of cities that are correlatedwith our measuresof urbanformbut are different fromthem.(For example,a citywithextensivebus routesmay have good sidewalks.)Second, thereis a large gap betweenthe measuresofurban form ourmodelsand policiesto alter in urban form: may take severalyearsuntilmeasuresthatchange it urban Nonetheless, spatialstructure produceanyrealeffect. ourstudy even in a country theU.S. that like that, suggests is heavilydependent the automobile, on urbanformdoes affect traveldemand.
REFERENCES Use Badoe, Daniel, and Eric J. Miller,"Transportation-Land Interaction: for EmpiricalFindingsin NorthAmerica,and Their Implications ResearchPart D 5 (2000), 235-263. Modeling,"Transportation AhmedMushfiqMobarak,and Bento,AntonioM., MaureenL. Cropper, on Katja Vinha, "The Impact of Urban Spatial Structure Travel Demand in the U.S.," WorldBank policy researchworking paper no. 3007 (2003). PoliBento,AntonioM., and Sofia Franco, "The Costs of Anti-Sprawl cies," Journalof Urban Economics(forthcoming). MarlonG., and Randall Crane, "The Influence Land Use on of Boarnet, TravelBehavior:Specification and Estimation TransStrategies," ResearchPart A 35:9 (2001), 823^45. portation MarlonG., and SharonSarmiento, "Can Land Use PolicyReally Boarnet, Affect Travel Behavior?A Studyof the Link betweenNon-eork Traveland Land Use Characteristics," UrbanStudies35:7 (1998), 1155-1169. Jan K., "Urban Sprawl: Lessons fromUrban Economics," Brueckner, (2001). Brookings-Wharton paperson urbanaffairs Crane, Randall, and RichardCrepeau, "Does Neighborhood Design InfluenceTravel? A BehavioralAnalysis of Travel Diary and GIS ResearchPart D 3 (1998), 225-238. Data," Transportation and Dubin, Jeffrey, Daniel McFadden, "An EconometricAnalysis of EconoResidential Electric ApplianceHoldingand Consumption," metrica52:2 (1984), 53-76. Harold Wolman, Galster,George, Royce Hanson, Michael Ratcliffe, StephenColeman, and JasonFreihage,"Wrestling Sprawl to the Ground:Defining and Measuringand Elusive Concept,"Housing Policy Debate 12:4 (2000), 681-717. and Glaeser,Edward,and MathewE. Kahn,"Decentralized Employment theTransformation theAmericanCity,"NBER working of paper no. 8117(2001). Goldberg,Pinelopi K., "The Effectsof the CorporateAverage Fuel in Economics Standards theU.S.," JournalofIndustrial Efficiency 46:1 (1998), 1-33.

You might also like