UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Urban labour markets and transportation Tyndall, Justin 2019

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2019_september_tyndall_justin.pdf [ 12.8MB ]
Metadata
JSON: 24-1.0379885.json
JSON-LD: 24-1.0379885-ld.json
RDF/XML (Pretty): 24-1.0379885-rdf.xml
RDF/JSON: 24-1.0379885-rdf.json
Turtle: 24-1.0379885-turtle.txt
N-Triples: 24-1.0379885-rdf-ntriples.txt
Original Record: 24-1.0379885-source.json
Full Text
24-1.0379885-fulltext.txt
Citation
24-1.0379885.ris

Full Text

URBAN LABOUR MARKETS AND TRANSPORTATIONbyJustin TyndallA DISSERTATION SUBMITTED IN PARTIAL FULFILLMENT OF THEREQUIREMENTS FOR THE DEGREE OFDOCTOR OF PHILOSOPHYinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Business Administration)The University of British Columbia(Vancouver)July 2019c© Justin Tyndall, 2019The following individuals certify that they have read, and recommend to the Faculty ofGraduate and Postdoctoral Studies for acceptance, the dissertation entitled: Urban LabourMarkets and Transportation, submitted by Justin Tyndall in partial fulfillment of therequirements for the degree of Doctor of Philosophy in Business Administration.Examining Committee:Dr. Thomas Davidoff, Sauder School of BusinessCo-supervisorDr. Sanghoon Lee, Sauder School of BusinessCo-supervisorDr. Tsur Somerville, Sauder School of BusinessSupervisory Committee MemberDr. Robin Lindsey, Sauder School of BusinessUniversity ExaminerDr. Marit Rehavi, Vancouver School of EconomicsUniversity ExaminerAdditional Supervisory Committee Members:Dr. Joshua Gottlieb, Vancouver School of EconomicsSupervisory Committee MemberiiAbstractThis thesis will study how transportation systems facilitate commuting, affect labourmarket outcomes and alter the urban spatial equilibriums of workers and firms. Workersliving in cities benefit from spatial proximity to local job opportunities. The ability ofworkers to access the labour market is enabled by existing public and privatetransportation networks. Transportation policy that expands the set of opportunities facedby workers could help overcome spatial matching problems and generate welfareimprovements. The first study in this thesis estimates the causal effect of New York City’ssubway system on neighbourhood unemployment rates. I show that a reduction in publictransportation access leads to a rise in the local unemployment rate. The second studyanalyses light rail transit systems in four US cities. While light rail can generate benefits interms of improving the transportation network, induced household sorting has importantconsequences regarding the distribution of benefits. The third study examines the generalconsequences of increasing commuter mobility in US cities. Results show exogenousincreases in mobility increase urban sprawl and do not result in improvements in aggregatemetropolitan labour market outcomes.iiiLay SummaryWorkers try to live close to where they work in order to limit the costs of commuting. Jobsthat are too far from the home are unlikely to be viable options for employment. Thiscollection of papers studies the spatial accessibility of jobs in US cities and the role oftransportation networks in facilitating employment. I find that workers benefit when urbantransportation systems connect them to more opportunities. However, places with bettertransportation amenities have higher real estate prices, meaning that richer residents areoften able to capture the benefits of transportation infrastructure. I use statistical modelsand data analysis to understand these relationships.ivPrefaceThis dissertation is original and independent work by the author, Justin Tyndall. Thework comprising Chapter 2 was published by Sage Journals as Tyndall (2017), Waiting forthe R train: Public transportation and employment, Urban Studies, Volume 54, Issue 2.The remainder of the work in this thesis is unpublished.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xDedication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 Waiting for the R Train: Public Transportation and Employment . . . . 32.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32.2 Related Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.4 Hurricane Sandy and the R Train . . . . . . . . . . . . . . . . . . . . . . . . 92.5 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132.7 Results: Effect of Private Vehicle Access . . . . . . . . . . . . . . . . . . . . 172.8 Results: Differences Across Race Groups . . . . . . . . . . . . . . . . . . . . 182.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203 The Local Labour Market Effects of Light Rail Transit . . . . . . . . . . . 223.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223.2 Light Rail Investment in Four US Cities . . . . . . . . . . . . . . . . . . . . 243.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263.4 Neighbourhood Effects of LRT . . . . . . . . . . . . . . . . . . . . . . . . . . 273.4.1 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273.4.2 Neighbourhood Change Results . . . . . . . . . . . . . . . . . . . . . 313.4.3 Effect of Light Rail on Commute Flows . . . . . . . . . . . . . . . . . 353.5 Urban Structural Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 37vi3.5.1 Modelling Neighbourhood Choice . . . . . . . . . . . . . . . . . . . . 373.5.2 Workers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 393.5.3 Firms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.5.4 Estimation Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423.5.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453.5.6 Cost Benefit Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 523.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534 Going Nowhere Fast: Urban Mobility, Job Access and Employment Out-comes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 554.2 Measuring Mobility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574.3 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 594.4 A Recent History of Commuter Velocity . . . . . . . . . . . . . . . . . . . . 614.5 Commuter Velocity has an Ambiguous Effect on Job Access . . . . . . . . . 644.6 Identification Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 704.7.1 Commuter Velocity as a Cause of Sprawl . . . . . . . . . . . . . . . . 704.7.2 Labour Market Impacts of Commuter Mobility . . . . . . . . . . . . . 724.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76Appendices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86A Google API Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86B Structural Results - Change in Potential Income . . . . . . . . . . . . . . . . 88C Policy Extension: Local Transit Pass Requirement . . . . . . . . . . . . . . . 89D Policy Extension: “Bus Rapid Transit” . . . . . . . . . . . . . . . . . . . . . 91viiList of Tables1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92 Probability of Being Unemployed, Full Workforce . . . . . . . . . . . . . . . 153 Shifts in Demographic Characteristics . . . . . . . . . . . . . . . . . . . . . 174 Probability of Being Unemployed, Subpopulations . . . . . . . . . . . . . . . 195 Summary Statistics, Metropolitan Areas . . . . . . . . . . . . . . . . . . . . 256 First Stage Results, Predicting Station Locations . . . . . . . . . . . . . . . 317 Neighbourhood Change Results . . . . . . . . . . . . . . . . . . . . . . . . . 328 Controlling for Post Treatment Demographics . . . . . . . . . . . . . . . . . 359 Shifts in Commuter Flows Towards LRT Treated Routes . . . . . . . . . . . 3810 Structural Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4311 CBSA Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6012 Large Metropolitan Areas (>1,000,000), by 2014 Commuter Velocity (Ω) . . 6213 Effect of Committee Representation on Home State Allocations, 2005-2015 . 6814 First Stage Regressions, Predicting Commuter Velocity from Lagged Commit-tee Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6915 Impact of Commuter Velocity (Ω) on Mean Localised Job Density (ζ) . . . . 7116 Impact of Commuter Velocity (Ω) on Average Distance to a Job . . . . . . 7117 Impact of Commuter Velocity (Ω) on Labour Market Outcomes . . . . . . . 73viiiList of Figures1 Areas Affected by R Train Service Interruption . . . . . . . . . . . . . . . . 102 Unemployment Rate Within Treated and Control Neighbourhoods . . . . . . 133 The Proliferation of LRT Stations . . . . . . . . . . . . . . . . . . . . . . . . 254 LRT Treated Tracts and Instrumental Variable . . . . . . . . . . . . . . . . 285 Share of Large US Metros with a Rail Link from Downtown to the LargestAirport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Mapping Commuter Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377 Structural Results, Change in Rent . . . . . . . . . . . . . . . . . . . . . . . 478 Structural Results, Change in Employment Rate . . . . . . . . . . . . . . . 489 Structural Results, Change in Public Transit Mode Share . . . . . . . . . . 4910 Structural Results, Distribution Across Potential Income Percentiles . . . . 5011 Correlation Between a Common Mobility Measure and Commuter Velocity (Ω) 5812 Summary of Commuter Velocity . . . . . . . . . . . . . . . . . . . . . . . . . 6313 Partial Effect of Increase in Commuter Velocity on Job Accessibility . . . . . 6514 Structural Results, Change in Potential Income (Skill) . . . . . . . . . . . . 8815 Transit Pass Structural Results, Distribution Across Potential Income Per-centiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9016 BRT Structural Results, Distribution Across Potential Income Percentiles . . 92ixAcknowledgementsI would like to thank my supervisors, Thomas Davidoff, Sanghoon Lee and Tsur Somervillefor their assistance and support throughout my PhD. Thank you to the University ofBritish Columbia, where I have enjoyed nine years of my career. Thank you to my wifeKelli for her love and support.xDedicated to Wendell T. Stamps of the Department of Labor.xi1 IntroductionCities contain dense agglomerations of labour market opportunities. The structure ofa city’s transportation network dictates the spatial accessibility of these opportunities toworkers. This collection of papers studies the role of transportation networks in the abilityof workers’ to access and secure jobs.Transportation infrastructure is a substantial component of public expenditure. In2017, the US spent $177 billion on highway infrastructure and an additional $70 billion onmass transit projects.1 Much of this investment was intended to shorten commuting times,an activity that comprises 51 minutes of the average American’s workday.2 There ispotentially significant room for welfare improvements if transportation systems canincrease the speed or efficiency of urban commuting.Transportation amenities are both a cause and consequence of economic development.Estimating the causal effect of infrastructure must consider that locations receivinginfrastructure may have unique economic characteristics. For workers, while qualitytransportation might improve labour outcomes, workers with more wealth or income tobegin with may preferentially move to well connected locations. The below papers willconsider how transportation providers allocate services across space, how workers optimisetheir decisions when the transportation network changes and how the causal effects oftransportation infrastructure on labour markets can be estimated.The first two studies of this dissertation are concerned with the role of publictransportation systems in the US. While only a small minority of US workers use publictransit for commuting (4.5% of commuters in 20173), populations dependent on transit aremore likely to be unemployed or on the margins of the labour market. Improving publictransit networks may therefore have a large impact on aggregate urban labour marketoutcomes. Chapter 2 of this thesis estimates the role of New York City’s subway system indetermining neighbourhood level unemployment levels. The chapter makes use of anexogenous shock to transportation infrastructure caused by a hurricane. The event allowsfor the isolation of causal neighbourhood effects, demonstrating a loss of transit accessgenerates a rise in local unemployment. The third chapter of this dissertation investigatesthe effect of light rail transit systems on US cities, estimating both neighbourhood effectsand aggregate metropolitan effects. The chapter proposes a new instrumental variable forthe otherwise endogenous spatial allocation of rail infrastructure, and also provides a1Congressional Budget Office (2018), Public Spending on Transportation and Water Infrastructure, 1956to 2017.2American Community Survey, 2017, 1-year estimates.3ibid.1methodological contribution by extending structural estimation approaches in urbaneconomics. The use of structural estimation can directly account for the endogenoussorting decisions of urban workers and firms.Chapter 4 of this dissertation studies the relationship between urban mobility ingeneral, job accessibility and labour market outcomes in US metropolitan areas. Whileexpanded transportation infrastructure can allow for the average urban commuter to travelat a higher speed, I find that high mobility also causes urban sprawl, lowering the spatialdensity of jobs. I provide evidence that increasing urban mobility fails to improve theaccessibility of firms to workers and does not generate significant labour marketimprovements. I use random variations in the US congressional committee appointmentprocess to instrument for endogenous infrastructure allocation across US states.The ambition of this thesis is to extend understanding of urban commuting networksand to inform urban policy pertaining to transportation infrastructure and regional labourmarkets. I find transportation networks play a vital role in urban labour market outcomes.The general equilibrium effects of transportation infrastructure are complex, necessitatingcareful empirical study to determine impacts.22 Waiting for the R Train: Public Transportationand Employment42.1 IntroductionThere has been substantial interest in public policy circles in recent years regardingstrategies of “job creation” and fostering “job access.” These terms are rarely providedwith concrete definitions and are instead meant to capture an alleged capacity on the partof government to decrease unemployment, or increase the quality of jobs that are availableto individuals. Making job opportunities more spatially accessible represents a plausiblepolicy lever to improve employment outcomes amongst urban residents. If governmentimproves transportation networks, the number of jobs available to a typical individual willbe increased, potentially improving the speed with which workers match to firms andimproving the quality of matches.Job access is closely related to issues of urban sprawl and the theory of spatialmismatch, both of which consider spatial gaps between workers and jobs. For populationsreliant on public transportation, the ability to access employment is closely tied to theusability and extent of the region’s public transportation network. There is empiricalevidence that populations with better access to jobs through public transportationnetworks also enjoy lower rates of unemployment (Holzer et al., 2003; Kain, 1992; Sanchezet al., 2004). Contrastingly, studies have argued that private vehicle ownership is thedominant transportation variable driving differences in employment outcomes (Baum,2009; Ong and Miller, 2005; Raphael et al., 2001; Taylor and Ong, 1995).Locations that occupy geographically central locations, or have exogenously developedas centers of economic activity or affluence, will be more likely to see local public transitinvestment due to the higher economic returns to transit infrastructure in such areas. Theeffect of Hurricane Sandy on New York City’s public transportation infrastructure presentsan unprecedented natural experiment to investigate a causal relationship. Using the trulyexogenous reduction in public transportation that occurred in particular neighbourhoods,this study will provide evidence for a causal relationship between public transportationaccess and local unemployment.4The research in this section has been published as Tyndall, J. (2017). Waiting for the R train: Publictransportation and employment. Urban Studies, 54(2), 520-537.32.2 Related ResearchThis section forms a basis for the current study in the foundational contribution ofKain (1968). Subsequently, this section will describe more recent and closely related works,particularly those that look specifically at transportation and the accessibility of jobswithin US metropolitan areas.Post-war North American cities have undergone suburbanization of employment,accompanied by a decrease in the relative importance of the manufacturing economy(Wilson, 2011). This shift in the labour market arose concurrently with rising inner-cityunemployment. Kain (1968) represents the first effort in the literature to draw an empiricalconnection between location of housing and the propensity to be unemployed. Kain (1968)specifically attempted to explain the unusually high rates of unemployment that persistedin black neighbourhoods in the inner cities of Chicago and Detroit. While previousinvestigations had exclusively blamed job market discrimination for the gap in employmentbetween white and black neighbourhoods, Kain (1968) suggested that the locationalcharacteristics of neighbourhoods with respect to job centers might also play a strong rolein determining job market outcomes.Since Kain (1968), several papers have reviewed or extended the spatial mismatchhypothesis (Brueckner and Zenou, 2003; Coulson et al., 2001; Gobillon et al., 2007;Raphael, 1998; Rogers, 1997; Smith and Zenou, 2003; Taylor and Ong, 1995). To this pointthere have been conflicting findings regarding whether spatial mismatch is a primary driverof high unemployment amongst inner-city populations.Harrison (1972) looked at large US metros and found no conclusive evidence thatspatial isolation was causing unemployment amongst black populations. The study cededthe difficulty of identifying a causal relationship, calling for a longitudinal study to trackthe movement of households through time. Farley (1982) presented empirical evidence thatspatial dimensions of employment markets cause higher rates of unemployment amongstblack populations, particularly in northern US metros. The study used controlledregressions and estimated that 15% of the gap in white-black unemployment could bederived as a consequence of housing segregation and suburban employment location.Immergluck (1998) applied an investigation of spatial mismatch to Chicago, finding astrong correlation between localized job market opportunities and the likelihood of beingemployed.The importance of spatial dimensions of labor markets can be generalized beyondinner-city minority populations to investigations into the importance of job accessibility toworkers generally. It is important to recognize that isolation from opportunities is aconsequence of inaccessibility, rather than distance. If workers have efficient transportation4that connects them to jobs then distance can plausibly be overcome. This reasoning hasled to the growth of so-called “transportation mismatch” literature, which purports toshow that if isolated populations are extended transportation opportunities, unemploymentgaps may be abated.Numerous prior papers have demonstrated that increasing rates of private vehicleownership amongst low-income or minority populations may be an effective means ofreducing unemployment amongst these populations (Baum, 2009; Gordon et al., 1989;Kawabata, 2003; Ong and Miller, 2005; Raphael et al., 2001; Taylor and Ong, 1995).Taylor and Ong (1995) found that commute times amongst minority workers were actuallyshorter than for white workers, in apparent contradiction to the spatial mismatchhypothesis. Driving to work alone -as apposed to alternative modes- was shown to be asignificant predictor of a short commute time across race groups and neighbourhoods.Raphael et al. (2001) examined the gap in employment outcomes between minority andwhite workers, finding that lower rates of car ownership amongst minority populationsexplained 45% of the black-white employment gap, and 17% of the Hispanic-whiteemployment gap. Ong and Miller (2005) argued that there is surprisingly scant evidencesupporting spatial mismatch being a major driver of differential unemployment rates forblack workers, and presented compelling empirical evidence that the lack of a privatevehicle significantly limits job prospects of black workers in the context of Los Angeles.Despite evidence of a beneficial marginal effect of car ownership, it is not clear thataggregate regional accessibility is well served by increasing private vehicle use. In denseurban environments in the US, road networks are typically filled past their designedcapacity during peak hours, resulting in congestion. Although providing a car to amarginal household may increase the job prospects of that household, increased carownership also inflicts a cost on existing commuters through higher congestion. The effectof increased car ownership on aggregate mobility and accessibility is therefore ambiguous,with the negative effects being more pronounced in high congestion cities.The provision of public transit provides a plausible means to increase employmentaccess while not contributing to road congestion. Thomas Sanchez furnishes the literaturewith US case studies looking directly at public transit characteristics-such as the nearnessto a bus or subway stop, or transit service frequency-relating high transportation access tolower levels of unemployment. Sanchez (1999) analyzed access to public transportation forpoor black communities in Portland, Oregon and Atlanta, Georgia, finding thatunemployment is higher for those residents who live more than 400 meters from a publictransit node. Sanchez et al. (2004) looked at a wider sample of cities and found transitaccess to be negatively related to the likelihood of a household being on government5assistance. Despite these seemingly strong findings, Sanchez et al. (2004) admits topossible identification issues, pointing out that the locational choices of households “resultfrom complex and intricate factors” that may be codetermined with economic success.Bollinger and Ihlanfeldt (1997, 2003) estimated local employment growth attributableto the construction of rail infrastructure in the Atlanta region. The authors found thatneighbourhoods adjacent to a new rail station had no significant increase in localemployment. Bollinger and Ihlanfeldt (1997) suggest this is attributable to the lowridership experienced by the rail system in the highly auto-oriented environment of Atlanta.Kain (1992) provided an exhaustive review of prior spatial mismatch research at thetime. Kain (1992) contains a direct discussion of the merits of promoting mobility as ameans to overcome the problems of spatial mismatch, particularly directing the discussionat Hughes and Madden (1991). Hughes and Madden (1991) advocated integrative housingpolicies that would end spatial isolation amongst black populations, suggesting thattransportation based policies are untenable because they accept persistent segregation.Kain (1992) responded that increasing suburban access to inner-city residents is actuallypro-integrative as it reduces the daily experience of isolation. Gobillon et al. (2007)investigated transportation-based solutions to spatial mismatch and found evidence for theefficacy of such policies to be mixed.Identifying the causal effect of transportation investment on employment mustovercome the potentially endogenous processes by which transit provision and localeconomic growth are determined. Specifically, it is unclear if transportation infrastructurecauses changes in local labour market outcomes or rather the siting of transportation isdetermined by preexisting local economic conditions. A related argument put forward byKnight and Trygg (1977) is that the impact of rail infrastructure on accessibility is limitedbecause urban rail is almost exclusively sited in areas that can be easily accessed by car tobegin with. Ihlanfeldt and Sjoquist (1998) provide a discussion of the potential endogeneitythat occurs when using household location as a predictor of employment outcomes: “Theproblem with this approach is that while job access may affect employment, employmentmay also affect the magnitude of the measure of job access.” Ihlanfeldt and Sjoquist (1998)considered past evidence to be inconclusive regarding the presence of a causal role oftransportation access on employment. A primary focus of this paper will be to establishthat the accessibility provided by public transportation has a causal relationship withneighborhood unemployment, particularly amongst those without access to a privatevehicle.Exclusively studying youth populations has provided a partial solution to theidentification problem because the location of a youth’s home is more plausibly exogenous.6If it is assumed that youth have no influence over household locational choice then ayouth’s location may be orthogonal to their employability; however, job market ability hasbeen shown to be highly stable across familial generations (Clark, 2014), suggesting thatthe neighbourhood choice of parents may be spatially stratifying youth populations byability. Ellwood (1986) examined youth employment outcomes in Chicago, finding thatalthough isolated black populations commuted significantly farther on average than whites,spatial isolation had only a small effect on their ability to actually secure employment.Ihlanfeldt and Sjoquist (1990) delivered compelling evidence that “nearness to jobs”through the transportation network is strongly correlated with unemployment amongstyouth populations in Philadelphia. O’Regan and Quigley provided a series of papers on theconnection between neighbourhood accessibility and youth employment rates (O’Reganand Quigley, 1996, 1998), generally finding that more centrally located neighbourhoodsprovide superior job market outcomes for youth.Holzer et al. (2003) supplies a study that is closely related to the current paper in itsattempt to overcome the identification problem through use of a natural experiment.Holzer et al. (2003) recognized, “No study has identified a clear exogenous source ofvariation in spatial access to employment opportunities.” Holzer et al. (2003) used the1997 expansion of the San Francisco region’s Bay Area Rapid Transit (BART) system,which extended service to a particular suburb, as an exogenous shock to the labor poolavailable to firms located along the new BART route. Holzer et al. (2003) found that firmsalong the extension hired more Hispanic residents from the inner city after the extensionwas completed, although found no significant impact on black employment. The exogeneityassumption made by Holzer et al. (2003) is suspect because it ignores that the siting oftransit infrastructure may be codetermined with economic activity (Knight and Trygg,1977). The siting of the new BART line was not random, but was specifically located alonga route where planners foresaw a future demand for commuting. Furthermore, exogeneityof the event assumes that the extension played no factor in firm locational choice in theyears leading up to the actual opening of the new rail route. It is likely that firms thatchose to locate along the new line were predisposed to taking advantage of the inner-citylabour pool, as accessible labour is a natural consideration in firm locational choice.Therefore, the finding that firms along the rail extension hired a higher percentage ofHispanic workers cannot be simply attributed to increased mobility of these residents. Acentral contribution of the present paper will be to utilize an unplanned, and unforeseenvariation in infrastructure to avoid these barriers to causal inference.Several of the aforementioned papers share the same barrier to identification:disentangling the endogenous relationship between localized economic development and the7siting of public transit. In order to infer a causal impact of public transportation onneighbourhood employment, the effect of employment on transportation infrastructureconstruction must be removed. The remainder of this paper will explicitly address thisconfounding relationship.2.3 DataThis paper relies on American Community Survey (ACS) data, collected by the USCensus Bureau. In order to identify trends through time, one-year estimates are used foryears 2010 through 2013. The ACS Public Use Microdata Sample (PUMS) provides annualindividual level observations, including employment status, for a randomly selected 1% ofthe US population. Variables are taken directly from the Integrated Public Use MicrodataSeries (IPUMS) data products (Ruggles et al., 2014).All individuals living outside of New York City are dropped from the sample. Onlyindividuals at least 16 years of age and in the labor force are retained for analysis. Ofindividuals at least 16 years of age and living in New York City, 60.6% are in the laborforce. This creates a sample of 136,726 individual level observations, split roughly evenlyacross the four years observed.The smallest identifiable geographic unit is the Public Use Microdata Area (PUMA).Each PUMA contains at least 100,000 persons. PUMA boundaries are coterminous withNew York City boundaries, covering the entire city.Income is represented by the individual respondent’s total pre-tax personal income,normalized to 2013 dollars. Income is represented in regressions in logged form. Dummyvariables for race are generated for each observation. White, black and Asian areconsidered mutually exclusive race groups; however, Hispanic ethnicity is derived from aseparate survey question and therefore an individual can be considered as both Hispanicand either white, black or Asian.Regressions also use Federal Emergency Management Agency (FEMA) insurancepayout data, which cover payouts to housing owners and renters under the Individuals andHouseholds Program (IHP). Observations can be identified as resulting from damagecaused by Hurricane Sandy. All payments resulting from other events are dropped. IHPdata is used as a proxy for the intensity of local storm damage. FEMA reports these dataat the Zip Code Tabulation Area. Data has been cross-walked to PUMAs using theMissouri Census Data Center’s Geographic Correspondence Engine, providing an estimatefor the average individual’s FEMA payout by PUMA.Summary statistics for the entire sample of New York City workers are provided in8Table 1.Table 1: Summary StatisticsVariable Mean Std. Dev.Unemployed 0.108 0.310Adjacent to Brooklyn R Train in 2013 0.015 0.122Adjacent to Brooklyn R Train 0.060 0.237Year: 2010 0.243 0.429Year: 2011 0.248 0.432Year: 2012 0.253 0.435Year: 2013 0.256 0.436Age 41.433 13.821Log of Annual Income 9.964 2.297Access to a Vehicle 0.577 0.494High School Graduate or GED 0.864 0.343Some College Completed 0.648 0.478College Graduate 0.405 0.491Master’s or Professional Degree 0.163 0.37PhD 0.016 0.125White 0.496 0.500Black 0.250 0.433Hispanic 0.230 0.421Asian 0.156 0.363Storm Damage per Person (100s) 0.694 1.903N 136,726Average characteristics of New York City workers, American Community Survey.2.4 Hurricane Sandy and the R TrainHurricane Sandy made landfall in New York City on October 29th, 2012. Amongstinstances of infrastructure damage across the region, the storm resulted in the flooding ofthe Montague Street Tunnel, which lies beneath the East River and connects Brooklyn toManhattan (shown in Figure 1). The Montague Street Tunnel is primarily used by the RTrain subway, which represents the major public transit connection linking neighborhoodson the western edge of Brooklyn to Manhattan. After the hurricane, R Train service wasimmediately interrupted due to the flooding of the Montague Street Tunnel. Over thesubsequent two months the tunnel was drained and repaired, reopening for R Train serviceon December 21st, 2012. In the months after the tunnel reopened it was determined by theMetropolitan Transportation Authority (MTA), who operates the tunnel, that additionalrepairs would be required. It was announced to the public on June 5th, 2013 that a9long-term closure would be necessary. The tunnel was subsequently closed for repairsbetween August 2nd, 2013 and September 14th, 2014. During closure the R Train was splitinto two routes, divided by the East River.Figure 1: Areas Affected by R Train Service InterruptionThe closure of the R Train created significant commuting delays for workers who livedalong the R Train route in Brooklyn and worked in Manhattan, potentially putting strainon their ability to maintain employment. Additionally, unemployed individuals living alongthe Brooklyn R Train route may have altered their job search strategy in response to the10altered transit service. Manhattan job centers that were previously within commutingdistance ceased to be viable options for employment because they could not be reached in areasonable commuting time. This point is central to the identification strategy: the removalof R train service for Brooklyn represents an exogenous shock to job access for particularresidents. If it is true that public transportation access has a causal impact on job marketoutcomes, then a sudden and unexpected reduction in public transportation service shouldresult in reduced job market success for the residents of impacted neighborhoods.The context of New York City is unique within the US in regards to high rates ofpublic transportation use and low rates of private vehicle ownership. In 2012, 57% of theNew York City workforce reported using public transportation as their primary mode ofcommuting to work and only 44% of households reported owning a private vehicle.Amongst the 387 other principal cities of defined metropolitan areas in the US in 2012, theaverage rate of public transit use for commuting was only 4.3%, while the average rate ofhousehold vehicle ownership was 89%. Findings for New York City therefore reflect thespecific realities of a dense urban environment with significant public transit use.2.5 MethodologyThis study uses a difference-in-difference approach to identify the impact the R Trainclosure had on employment outcomes for affected neighborhoods. A difference-in-differenceapproach allows for the estimation of how employment outcomes changed in the affectedneighborhoods, while controlling for regional employment trends through time, as well asfixed characteristics that may differentiate the affected and unaffected neighborhoods. Tofurther reduce bias in estimation, borough fixed effects are used in all regressions, tocontrol for localized employment conditions.Of the four years of observations used in analysis, only 2013 is considered as havingbeen “treated” by the closure of the R Train. The R Train was closed for a five-monthperiod in 2013, whereas it was closed for less than two months in 2012. The effect of theextended 2013 closure on employment should have the greatest magnitude and so is thetopic of empirical investigation. Estimation will rely on representing the 2013 closure as anegative shock to employment prospects. The brief 2012 closure does not pose a barrier toestablishing a causal effect, as any negative employment effects on treatmentneighborhoods in 2012 will result in an underestimation of the true 2013 effect.Three PUMAs are identified that are primarily reliant on the R Train in order toreach job centers in Manhattan and beyond: Community Districts 6, 7, and 10 (see Figure1), all of which are in western Brooklyn. Although PUMAs represent a coarse geographic11unit, this study is fortunate that these three PUMAs align very closely to theneighborhoods one would hypothesize to be heavily reliant on the R Train. The boundariesof treatment PUMAs are consistent through time.Community Districts 6, 7, and 10 represent a highly diverse population. The medianincome of New York City in 2013 was $35,000. The median income of Community District6 was substantially higher at $65,000, Community District 7 was substantially lower at$25,000, and Community District 10 was somewhat higher than the city median at$42,000. These three community districts had a greater share of white residents (67%) anda lower share of black residents (4%) than the citywide shares, which were 50% and 24%respectively.In testing the impacts the tunnel closure had on these neighborhoods it is useful toconfirm that these areas were not disproportionately affected by the direct impacts ofHurricane Sandy; otherwise, the observed negative employment effects may be simplyattributable to general damage from the storm. FEMA provides data on insurance payoutsunder the IHP, providing a reasonable proxy for comparing hurricane damage betweenPUMAs. The average individual in New York City received $69 in hurricane recoveryassistance under IHP, while the average individual in the treatment neighborhoods receivedonly $15. It appears that if anything, the treatment neighborhoods received less damagethan was typical. An exception to this average is the coastal neighborhood of Red Hook inCommunity District 6, which suffered significant damage to building stock as a result ofthe hurricane. Red Hook residents only comprise 2.7% of the total population of treatedPUMAs; furthermore, results are robust if Community District 6 is considered as a controlrather than a treatment neighborhood. To guard against the potential direct impacts ofstorm damage on local economic conditions, the average FEMA payout within theindividual’s PUMA is controlled for in regressions.As discussed in the previous section, 2013 brought a substantial reduction in transitservice for the treated neighborhoods. Observations that are both located within atreatment neighborhood and were recorded in 2013 are referred to as the treatment group.Defined in this way, there are 2,082 treatment observations, within a sample of 136,726.The methodology presented here overcomes potential codetermination betweeneconomic conditions and transportation infrastructure by exploiting random variation ininfrastructure. There is no possibility that Hurricane Sandy’s impact on the MontagueStreet Tunnel was planned or predicted, meaning initial local economic conditions wereorthogonal to this variation.122.6 ResultsThe impact of the R Train disruption on the treatment neighborhoods can beglimpsed in Figure 2. Between 2010 and 2012 the rate of unemployment decreasedsignificantly within the treatment neighborhoods, falling from 9.8% to 8.0%.Unemployment in the remainder of New York City was relatively stable over this period,falling slightly from 11.4% to 11.1%. This trend was starkly reversed in 2013 wherein theneighborhoods along the Brooklyn R Train saw a rise in unemployment (8.0% to 8.4%)while in other neighborhoods unemployment fell substantially (11.1% to 9.7%). Thisdeparture in outcomes in 2013 is suggestive that job market conditions changed in thetreatment neighborhoods in 2013, making employment more difficult to secure or maintain.Figure 2: Unemployment Rate Within Treated and Control NeighbourhoodsThe impact of the R Train disruption on commuters would only be felt within thetreated Brooklyn neighborhoods if there were a substantial proportion of the localworkforce using the R Train to reach employment, the most probable destination being jobcenters in Manhattan. IPUMS provides workplace location data at the borough levelwithin New York City. Of those living within the treatment neighborhoods in 2012, 43% ofthe workforce commuted to Manhattan for employment. Within the treatmentsubpopulation that commuted to Manhattan, 84% reported using the subway to commute.This suggests the R Train represented an important link to employment for the treatmentneighborhoods. The Hispanic population departs somewhat from these statistics, with only1332% of the Hispanic workforce commuting to Manhattan. Those earning high incomes weremore likely to commute to Manhattan. Workers earning above the median income workedin Manhattan at a rate of 53%, compared to 33% for workers earning below the median.For workers earning in the bottom 10th percentile of incomes, the probability of working inManhattan was similarly 33%.In accordance with Angrist (2001) and Angrist and Pischke (2008) OLS is usedthroughout this study rather than an estimation method specific to limited-dependentvariables. 5 Table 2 provides a regression for the full sample of New York City workers. Aseconomic conditions improved following the financial crisis of 2008, the probability ofunemployment amongst New York City workers fell. Year fixed effects show the probabilityof being unemployed decreased across the four years studied. Column 2 adds individuallevel controls for age, age squared, logged annual income, and whether the individual hadaccess to a private vehicle. Subsequent regressions include a full set of controls foreducation level. Kasarda (1989) provides evidence from US inner cities demonstrating thata great deal of variation in employment outcomes amongst spatially isolated householdscan be explained by education level. Column 4 in Table 2 adds controls for race group.Finally, column 5 controls for the PUMA level variation in local storm damage. Theinterpretation of control coefficients is not of particular interest because many of thecontrols are codetermined. The central result, indicating the impact of living in thetreatment neighborhoods in 2013, is robust to the inclusion of these controls.After controls are in place, the estimated effect on an individual attributable to livingadjacent to the R Train in 2013 is an increase in the probability of being unemployed of 1.4percentage points (Table 2, column 5). The result is highly statistically significant, as wellas representing a large effect in practical terms. This finding is completely consistent witha transportation mismatch hypothesis, in which a lack of mobility results in inferioremployment outcomes. This observed effect cannot be explained by the codetermination oftransit location and economic activity, demonstrating a causal connection between transitprovision and neighborhood employment outcomes.Estimated neighborhood level effects can be attributed to a reduction in neighborhoodconnectivity. However, neighborhood self-selection may still exist, which presents a barrierto inferring a causal effect on individual workers. If households are able to participate in5Angrist (2001) and Angrist and Pischke (2008) argue for dispensing with limited dependent variablemodels in favour of conventional OLS approaches in cases where the research interest is the estimation of aparticular causal effect. The ostensible advantage of limited dependent variable methods is tied to reconcilingstructural parameters rather than isolating causal effects. The statistical significance of the estimated partialeffect of interest in the current study holds in the case of estimation using a logit or probit model; however,the partial effect provided by OLS is more readily interpretable.14Table 2: Probability of Being Unemployed, Full Workforce(1) (2) (3) (4) (5)Adjacent to Brooklyn R Train in 2013 .013∗∗ .015∗∗ .015∗∗ .014∗∗ .014∗∗(.002) (.002) (.002) (.002) (.002)Adjacent to Brooklyn R Train -.028∗∗ -.0006 -.004∗∗ .004∗ .005∗∗(.0005) (.001) (.001) (.002) (.002)Year: 2011 .0003 -.008∗∗ -.008∗∗ -.008∗∗ -.008∗∗(.004) (.002) (.002) (.002) (.002)Year: 2012 -.006∗∗ -.011∗∗ -.012∗∗ -.012∗∗ -.012∗∗(.002) (.002) (.002) (.002) (.002)Year: 2013 -.019∗∗ -.019∗∗ -.019∗∗ -.019∗∗ -.019∗∗(.003) (.004) (.003) (.003) (.003)Age .003∗∗ .003∗∗ .003∗∗ .003∗∗(.0005) (.0005) (.0005) (.0005)Age (Squared) -.00002∗∗ -.00002∗∗ -.00002∗∗ -.00002∗∗(4.82e-06) (4.99e-06) (4.61e-06) (4.61e-06)Log of Annual Income -.085∗∗ -.086∗∗ -.087∗∗ -.087∗∗(.002) (.002) (.002) (.002)Access to a Vehicle .0009 -.002 -.002 -.003(.004) (.004) (.003) (.003)High School Graduate or GED .008 .002 .002(.005) (.005) (.005)Some College Completed .014∗∗ .013∗∗ .013∗∗(.003) (.003) (.003)College Graduate .013∗∗ .015∗∗ .015∗∗(.002) (.003) (.003)Master’s or Professional Degree .007∗ .006 .006(.003) (.004) (.004)PhD -.003 -.002 -.002(.004) (.005) (.005)White .008∗∗ .008∗∗(.0002) (.0003)Black .019∗∗ .019∗∗(.004) (.004)Hispanic -.012∗∗ -.012∗∗(.003) (.003)Asian -.023∗∗ -.022∗∗(.004) (.003)Storm Damage per Person (100s) .0009∗∗(.0003)Const. .116∗∗ .886∗∗ .879∗∗ .879∗∗ .878∗∗(.002) (.032) (.030) (.030) (.030)Obs. 136,726 136,726 136,726 136,726 136,726R2 .0009 .375 .376 .378 .378Significance levels: ∗ : 5% ∗∗ : 1%. Robust standard errors in parenthesis.15neighborhood sorting in response to the R Train closure between the time when thehurricane damaged the Montague Street Tunnel (October 29th, 2012), and the end of 2013(marking the last day in which 2013 ACS data may have been collected) then particularhouseholds could be self-selecting into neighborhoods based on the R Train closure.The ACS contains a variable displaying how long ago an individual moved into theircurrent residence. One possibility that would suggest neighborhood sorting is thatunemployed individuals moved to the treatment neighborhoods preferentially. This can beempirically investigated. The controlled difference-in-difference regression was repeated,partitioning the sample into those who moved within the past year (recent movers), andthose who did not. As would be predicted by a neighborhood sorting argument, the“impact” of the R Train closure is far greater for those who recently arrived in thetreatment neighborhoods, suggesting that the unemployed are preferentially moving toaffected neighborhoods. The effect of living in a treatment neighborhood amongst recentmovers is an increase in the unemployment rate of 3.2 percentage points. The effect onlonger-term residents is only 1.1 percentage points. Recent movers claim a 13.7%population share within the treatment neighborhoods in 2013. The effect of the R Trainclosure on non-recent removers is still highly statistically significant.A second potential source of individual level bias is the possibility that employedresidents disproportionately left the treatment neighborhoods or dropped out of theworkforce in response to the R Train closure. Any large exodus from the neighborhoodcould be observed as a drop in average home value or rent in 2013. Table 3, columns 1 and2 provide difference-in-difference estimates for average home value and rent paid. There isno significant effect on either of these variables from being in a treated neighborhood,although the point estimates demonstrate a negative effect.Workers who left the treatment neighborhoods in 2013 cannot be observed andtherefore their employment characteristics cannot be directly examined. Changes inneighborhood demographics can be observed and used to check for shifts in neighborhoodcomposition. Table 3, columns 3-7 present the partial effect of being in the treatmentgroup on the likelihood or magnitude of relevant demographic characteristics. There is anobservable shift in workforce demographic characteristics in treatment neighborhoodsincluding an increase in the proportion of black and Hispanic residents, and a decrease inthe probability of holding a graduate degree. Although these observations provide someevidence of a shift in neighborhood demographics tied to the R Train closure, they do notpose a direct problem for estimation as they are all explicitly entered into thedifference-in-difference model as controls. Therefore, changes in socioeconomic make-up ofthe neighborhood cannot be cited as an explanation of the increase in unemployment found16Table 3: Shifts in Demographic CharacteristicsHome Monthly Log College Masters Black HispanicValue Rent Income Degree Degree(1) (2) (3) (4) (5) (6) (7)Adjacent to Brooklyn R -300.160 -12.649 .024 .001 -.009∗∗ .014∗∗ .013∗∗Train in 2013 (2574.968) (10.812) (.026) (.007) (.003) (.002) (.002)Adjacent to Brooklyn 183092.400∗∗ 219.502∗∗ .327∗∗ .155∗∗ .095∗∗ -.339∗∗ .041∗∗R Train (725.377) (2.818) (.007) (.002) (.0008) (.0006) (.0004)Year: 2011 -3768.625 -9.609 -.096∗∗ .005 .005 .011 -.006(4649.019) (30.192) (.029) (.008) (.008) (.011) (.007)Year: 2012 -14589.150 45.512 -.063∗∗ .012 .005 .004 -.005∗(7647.874) (25.110) (.015) (.009) (.006) (.009) (.002)Year: 2013 -7767.520 103.668∗∗ .006 .029∗∗ .013∗ -.007 -.008∗∗(4514.000) (19.969) (.032) (.010) (.006) (.008) (.003)Const. 520045.600∗∗ 1225.348∗∗ 9.982∗∗ .384∗∗ .152∗∗ .268∗∗ .233∗∗(3731.423) (17.901) (.017) (.006) (.004) (.007) (.003)Obs. 54,948 76,158 136,726 136,726 136,726 136,726 136,726R2 .018 .012 .001 .006 .003 .031 .0006Significance levels: ∗ : 5% ∗∗ : 1%. Robust standard errors in parenthesis.in this study. However, the potential for shifts in latent ability characteristics that are notcontrolled for prevents the translation of clear neighborhood level effects to individual levelimpacts.The following section will exploit variation in a worker’s dependence on publictransportation to look for evidence that the observed rise in unemployment can be linkedto reductions in mobility.2.7 Results: Effect of Private Vehicle AccessSeveral prior studies have found a relationship between vehicle ownership and anincreased propensity to secure employment (Baum, 2009; Gordon et al., 1989; Kawabata,2003; Ong and Miller, 2005; Raphael et al., 2001; Taylor and Ong, 1995). If theemployment effect found in the current study is in fact a result of the loss of R Trainservice, this impact should be larger amongst those who are most dependent on publictransit. The ACS asks respondents how many vehicles are kept at the household, and areavailable to the household member. In this section the sample is split in two: those with novehicles available at all, and those with at least one. 57.7% of the workforce has access toat least one vehicle.Table 4 (columns 1 and 2) conforms to expectations regarding the role of vehicleownership. Education and income are controlled in all regressions so the impact of carownership can be interpreted as independent of an income effect. Individuals with access toa vehicle were found to experience a significant increase in unemployment of 0.7 percentage17points as a result of the transit disruption, while individuals without access to a vehiclewere found to suffer a much larger increase of 2.2 percentage points. The effect of reducedtransit is clearly more pronounced amongst those who lack an outside option fortransportation.This section’s findings strengthen that of the previous section by drawing a clear line inthe data between job access through the public transit system and employment outcomes.2.8 Results: Differences Across Race GroupsExploring differences in US employment outcomes between race groups has receivedsignificant attention in the literature; furthermore, investigations into the spatial mismatchhypothesis are often predicated on the observed spatial isolation of urban blackpopulations. It is therefore of interest whether the impact of spatial isolation is particularlyacute amongst minority populations.This section divides observations into groups of race and ethnicity. Table 4 (columns3-6) shows how the impact of job access is highly variable across race groups. For theentire population (Table 2, column 5), the impact of the R Train closure was estimated tobe an increase in unemployment in affected neighbourhoods of 1.4 percentage points.Amongst white residents, the estimated effect is significantly less (0.7 percentage points),falling short of statistical significance. For Asian residents, the effect is not significantlydifferent from the aggregated estimate, and is estimated as a highly significant increase inunemployment of 1.3 percentage points. For black residents the estimated effect is 1.7percentage points, also statistically indistinguishable from the aggregate estimate. Inagreement with Holzer et al. (2003) and Andersson et al. (2018), this study finds theunemployment rate within the Hispanic population to be most affected by job accessibility.The closure of the R Train is associated with a 3.4 percentage point increase in theunemployment rate of Hispanic residents. In practical terms, the increase in unemploymentattributable to the R Train closure represents a precipitous drop in the employmentprospects of Hispanic residents in the treated neighbourhoods.Spatial mismatch has often dealt exclusively with the lagging employment outcomes ofisolated black populations. The current study finds that, although black residents of thetreatment neighbourhoods experienced higher rates of unemployment in each of the fouryears studied relative to Hispanic residents, the impact of transit access on employment ismuch larger amongst Hispanics. Holzer et al. (2003) put forward a number of possibleexplanations for a pronounced effect amongst Hispanics. Firstly, if a particular subgroup ismore dependent on employee referrals to secure employment, the loss of jobs within the18Table 4: Probability of Being Unemployed, SubpopulationsVehicle No-Vehicle White Black Hispanic AsianAccess Access(1) (2) (3) (4) (5) (6)Adjacent to Brooklyn R Train in 2013 .007∗∗ .022∗∗ .007 .017∗∗ .034∗∗ .013∗∗(.003) (.002) (.004) (.003) (.004) (.004)Adjacent to Brooklyn R Train .011∗∗ -.002 .012∗∗ .005∗∗ -.024∗∗ .004∗∗(.001) (.003) (.002) (.002) (.001) (.0008)Year: 2011 -.008∗∗ -.008∗∗(.002) (.003)Year: 2012 -.012∗∗ -.011∗∗ -.014∗∗ -.005 -.014∗∗ -.010∗∗(.002) (.002) (.002) (.006) (.004) (.003)Year: 2013 -.017∗∗ -.022∗∗ -.016∗∗ -.021∗∗ -.020∗∗ -.019∗∗(.004) (.004) (.005) (.005) (.006) (.004)Age .003∗∗ .003∗∗ .004∗∗ .003∗∗ .002∗∗ .004∗∗(.0006) (.0007) (.0006) (.0008) (.0007) (.001)Age (Squared) -.00003∗∗ -.00002∗∗ -.00003∗∗ -.00003∗∗ -.00002∗ -.00003∗(6.44e-06) (5.67e-06) (5.65e-06) (8.45e-06) (7.56e-06) (1.00e-05)Log of Annual Income -.086∗∗ -.088∗∗ -.081∗∗ -.091∗∗ -.090∗∗ -.084∗∗(.001) (.003)(.003) (.0002) (.0005) (.002)Access to a Vehicle -.003 -.008∗ .003 .007∗(.002) (.004) (.007) (.003)High School Graduate or GED .0008 .003 .010 -.016∗∗ .005 .012∗∗(.007) (.002) (.012) (.002) (.004) (.004)Some College Completed .010∗ .019∗∗ .015∗∗ .007∗∗ .014∗ .020∗∗(.005) (.005) (.004) (.002) (.006) (.004)College Graduate .018∗∗ .011 .013∗∗ .004 .006 .031∗∗(.004) (.007) (.003) (.005) (.006) (.003)Master’s or Professional Degree .006 .005 -.001 .014∗∗ .013∗∗ .009∗∗(.003) (.005) (.005) (.002) (.002) (.002)PhD -.002 -.002 -.006 .003 -.004 -.006(.004) (.010) (.006) (.012) (.007) (.006)White .009∗∗ .007∗∗(.002) (.002)Black .019∗∗ .020∗(.004) (.009)Hispanic -.008∗∗ -.018∗(.002) (.009)Asian -.019∗∗ -.027∗∗(.003) (.007)Storm Damage per Person (100s) .00003 .004∗ .001∗∗ -.0002 .004∗∗ .001∗∗(.0001) (.002) (.0002) (.0008) (.0004) (.0004)Const. .857∗∗ .901∗∗ .787∗∗ .988∗∗ .939∗∗ .797∗∗(.023) (.039) (.039) (.015) (.023) (.035)Obs. 78828 57898 67882 34173 31494 21287R2 .381 .373 .313 .422 .402 .41Significance levels: ∗ : 5% ∗∗ : 1%. Robust standard errors in parenthesis.19community may have a multiplier effect in which a lost job lowers the prospects of securingemployment for those within the unemployed individual’s social network. Prior researchhas found that referrals play a disproportionately powerful role in securing employmentamongst Hispanics and new immigrants (Elliott, 2001; O’Regan, 1993). Holzer et al. (2003)also suggests Hispanics may be more willing to travel a greater distance to secureemployment; however, this seems a less plausible explanation for the results found here, asthe treated Hispanics are observed to be less likely to travel outside of Brooklyn foremployment than other groups.An alternate explanation for the pronounced effect among Hispanics is the high rate ofsubway use in New York City amongst Hispanic commuters: 43.9%, compared to 40.0%amongst blacks and 39.2% for whites. These findings are consistent with nationalinvestigations that show high public transit use amongst foreign-born populations(McKenzie and Rapino, 2011).In the current study, spatial isolation from jobs appears to exert a larger influence onblack communities than white communities. However, the impact within the Hispanicpopulation is significantly greater than within either white or black populations, suggestingthat future investigations into spatial mismatch should pay greater attention to theapparently large impacts of job accessibility amongst Hispanics.2.9 ConclusionThe advent of Hurricane Sandy flooding the Montague Street Tunnel represents aunique natural experiment for investigating the impact of job accessibility through publictransportation on employment outcomes. There is compelling evidence that a suddendecrease in public transportation triggered a significant hardship for the job marketprospects of affected workers. This finding provides an argument against eliminatingexisting urban transit services, as reductions in service may have significant and costlyeffects realized through increased local joblessness. An inability for agencies to fundcurrent transit levels and to contemplate service reductions is not an uncommon scenario(Gomez-Ibanez, 1996; Nelson et al., 2007).Household locational choice is not exclusively determined by current employment oremployment prospects. Locational choice is instead the result of a complex decisionfunction of which one element is employment. It is therefore not sensible to invoke anoverriding theory of spatial equilibrium with respect to jobs. Workers may be compelled byfinances, family ties or community networks to remain in a neighborhood even if it doesnot perfectly suit their needs for employment or mobility. This reality opens policy space20for efficiency gains through maintaining transit to neighborhoods with otherwise poor jobaccessibility. Establishing a deeper understanding of the extent and speed with whichhouseholds relocate in response to transit alterations would be a fertile area for futureresearch.This study finds strong evidence that public transportation access plays a meaningfulrole in setting the level of local unemployment. During contemplation of publictransportation policy the localized employment effects are rarely explicitly considered;however, the impact appears to be large. In New York City the effect is particularlypronounced within the Hispanic population, and amongst those without access to a privatevehicle. Maintaining public transit service to job centers should be prioritized intransportation and labor policy, specifically within congested metropolitan areas whereincreasing private vehicle ownership may be inefficient due to negative externalities.213 The Local Labour Market Effects of Light RailTransit3.1 IntroductionUS cities have made significant investments in Light Rail Transit (LRT) in recentyears, with current annual expenditures exceeding six billion dollars.6 A commonjustification for LRT is that transit infrastructure will improve urban commutingnetworks.7 I test the contention that LRT improves labour market outcomes. First, Iestimate the neighbourhood level effects of LRT stations. I introduce a new instrumentalvariable that establishes orthogonality between station location and pretreatment localeconomic conditions. I find that gaining a LRT station increases the local employmentrate. Second, I estimate a structural neighbourhood choice model to uncover themechanisms that generate employment changes and estimate welfare effects. My analysisspans four US cities over the 2000-2013 period.LRT has become a popular form of transit due to low construction costs relative tosubway systems and large perceived economic benefits. LRT systems are typically builtalong existing roads, removing the need for expensive tunnelling or elevated infrastructure.While LRT shares road space with vehicles and pedestrians, portions of routes are giventraffic priority, enabling faster speeds and fewer delays than experienced by buses. Incontrast to bus transit, the need for rails, an overhead power source and station platformsensures that LRT represents a long term local investment.Transit is not allocated randomly within a city, it is directed toward neighbourhoodswith specific characteristics. Comparing the economic outcomes of areas with transit tothose without will not provide causal estimates of project impacts due to the effect ofdiffering pretreatment conditions and economic trends. I propose a new instrumentalvariable to estimate the causal effect of LRT stations on neighbourhoods. An inclinationamong transportation planners to extend light rail to the airport provides a naturalexperiment that introduces an element of randomness to station location. Neighbourhoodsbetween downtown and the airport were much more likely to receive a LRT station thansimilar neighbourhoods located elsewhere in the metro. I exploit a preference for airportconnections to estimate local effects. The endogeneity of transit location is a well known6American Public Transportation Association, 2017 Public Transportation Fact Book.7For example, the environmental impact assessment for Seattle’s LRT system claimed the project wouldresult in, “improved access to employment opportunities” (Sound Transit, 1999). A chief political advocatefor the Minneapolis LRT system stated: “We’re trying to reconnect people, particularly people with highlevels of unemployment, to the job market” (Peter McLaughlin, Hennepin County Commissioner, from TheTrain Line That Brought the Twin Cities Back Together, by E. Trickey, Politico Magazine, March 16, 2017).22issue from prior literature (Baum-Snow and Kahn, 2000; Holzer et al., 2003; Ihlanfeldt andSjoquist, 1998). For example, affluent neighbourhoods have been found to resist railinfrastructure due to concerns that transit may lead to a rise in local crime (Kahn, 2007).After correcting for endogenous transit allocation, I find LRT generates large improvementsin neighbourhood level employment outcomes.Using reduced form estimates as model inputs, I estimate a structural neighbourhoodchoice model and conclude that LRT systems fail to raise aggregate metropolitanemployment. LRT stations increase demand for local housing, raising rents. LRT istypically built in accessible, central locations. As a result, low skilled workers are displacedfrom central locations by rising prices. As labour force participation is more elastic amongthe low skilled, the mechanism leads to an aggregate decrease in metropolitan employment.LRT may, counterintuitively, exacerbate the spatial isolation of low skilled workers througha process of household displacement. The ability of local amenities to drive up land valuesand alter a neighbourhood’s composition is a familiar mechanism from literature on placebased urban policies (Hanson, 2009; Kline, 2010; Kline and Moretti, 2014). Thismechanism has been known to undermine spatially targeted policies. I show that the samemechanism is relevant to LRT projects. Taking account of household sorting, I find thatthe welfare benefits of LRT are positive and exceed typical project costs. Welfare benefitsare generated through reductions in the commuting costs of some workers but also throughLRT acting as a local amenity that enhances consumption. I also find LRT is effective atraising aggregate transit use, as it appeals to higher income workers who would be unlikelyto take other forms of public transit.Poor spatial access to job opportunities can hinder employment outcomes due to highcommuting costs (Kain, 1968). Numerous studies have expanded upon the spatial mismatchhypothesis to explain heterogeneity in urban labour market outcomes and particularly toexplain the lagging outcomes of racial minorities and youth (Gobillon et al., 2007; Holzer,1991; Holzer et al., 2003; Immergluck, 1998; Sanchez et al., 2004; Stoll, 1999; Tyndall,2017). Past research has found that unemployed and poor workers tend to live in placesthat are isolated from relevant job opportunities. However, the literature has not shownconclusively whether the relationship between accessibility and employment is the result ofunemployed workers self-selecting into isolated neighbourhoods, or if there is a causal effectof neighbourhood connectivity on individual employment outcomes. If the effect is causal,expanded access to transit may raise equilibrium employment by reducing spatial isolation.Some prominent papers have directly analysed local effects of rail stations(Baum-Snow and Kahn, 2000; Kahn, 2007). Results pointed towards localized increases inhome values and increased transit use. Few studies have attempted to estimate the23neighbourhood effects of LRT stations specifically. Cao and Schoner (2014) studiedridership effects of LRT in Minneapolis. Residents moving towards new transit were foundto be less likely to use LRT than the original residents, suggesting a gentrification effect.Recent work by Severen (2018) investigates the effect of LRT construction in Los Angeles,finding that LRT has a positive effect on labour supply. The paper addressed empiricalidentification challenges related to neighbourhood choice models. Otherwise, the literatureprovides sparse guidance on the overall effects of LRT systems, which is striking given therapid propagation of such systems in the US.There is strong evidence that proximity to transit is an important consideration inhousehold location choice (Glaeser et al., 2008). LeRoy and Sonstelie (1983) provided adynamic model of transportation induced urban change, where heterogeneity in workerearning ability gives rise to heterogeneity in transportation mode choice andneighbourhood composition. Wasmer and Zenou (2002, 2006) propose a general urbancommuting model that leads to unemployed workers voluntarily occupying inaccessibleareas due to infrequent travel. I extend the intuition of these models by incorporating apolycentric city, which generates more complex patterns of neighbourhood sorting.I contribute to the literature in a number of ways. First, I provide policy relevantestimates of the labour market effects of LRT. Second, I supply a new instrumental variablefor endogenous station location. Third, I extend the neighbourhood choice literature bydeveloping a structural sorting model that includes preference parameters for transit.The study will proceed as follows. Section 3.2 will summarize the LRT projects underanalysis. Section 3.3 introduces data sources. Section 3.4 estimates the neighbourhoodeffects of new LRT stations. Section 3.5 proposes and estimates a structural neighbourhoodchoice model providing estimated welfare effects and section 3.6 concludes.3.2 Light Rail Investment in Four US CitiesLRT has become a popular transportation and economic development strategy acrossthe US. Between 2000 and 2016 the number of LRT stations in the US grew by 60%(Figure 3). The empirics of this study will focus on four metropolitan areas: Minneapolis,Minnesota; Portland, Oregon; Salt Lake City, Utah; and Seattle, Washington. These fourmetropolitan areas are similar in that they all completed substantial LRT construction overthe period of study. Minneapolis and Seattle had no LRT stations prior to 2000, whilePortland and Salt Lake City had already completed a portion of their systems. Following apopular trend in transportation planning, these four cities all extended rail access to themetro’s largest airport. The metros range in population from 1.1 million (Salt Lake City)24to 3.6 million (Seattle). Table 5 displays metropolitan level characteristics as contrastedwith the full sample of US metropolitan residents. The four selected metros containresidents with higher median household income than the US urban population as a whole.Figure 3: The Proliferation of LRT Stations8Table 5: Summary Statistics, Metropolitan AreasMinneapolis Portland Salt Lake City Seattle All US MetrosMetro Population (2013) 3,458,513 2,312,503 1,141,510 3,609,617 .Median Household Income (2013) 73,585 63,891 67,808 75,391 54,516Public Transit Mode Share (2000) 4.4% 5.5% 3.3% 6.6% 4.6%Public Transit Mode Share (2013) 4.8% 6.1% 3.6% 8.4% 5.1%LRT Stations (2000) 0 53 16 0 601LRT Stations (2013) 19 87 56 19 887Public transit comprises only a small share of total commutes in these metros. Seattlehad the highest rate in 2013, with 8.4% of commuters using public transit. Salt Lake Cityhad the lowest public transit mode share in the sample at 3.6%. Across the entire urbanpopulation of the US, 5.1% of workers commute by public transit. The sample of cities istherefore fairly representative of transit uptake in a typical US metro. Public transit modeshare increased in all four metros during the 2000-2013 period. Seattle experienced thelargest increase, expanding public transit mode share among commuters by 27%.Populations who depend on public transit are more likely to be on the margin of the labourmarket (Sanchez, 1999; Sanchez et al., 2004), suggesting transit expansions may havesignificant labour market effects.8Station counts obtained from the annual American Public Transportation Association Fact Book. Be-tween 2000 and 2016, the number of LRT stations in the US grew by 60% while the number of heavy railsubway stations grew by 2%.253.3 DataI use census tract level data from the 2000 US Decennial Census as well as the 2015American Community Survey (ACS), five-year estimates. Census data from 2000 arecrosswalked to 2010 boundaries using the Missouri Census Data Center’s GeographicCorrespondence Engine. Metropolitan areas will be bounded according to 2013 Bureau ofLabor Statistics core-based statistical areas. Census microdata on worker characteristicswill be used in structural estimation to provide joint distributions of worker income anddemographic characteristics. Worker microdata is taken from the the 2000 US CensusIntegrated Public Use Microdata Sample (IPUMS). All income and price variables areinflation adjusted to 2013 dollars.In addition to census data on home prices I make use of the US Federal HousingFinance Agency (FHFA) Annual House Price Index (HPI). FHFA HPI estimates arederived from a repeat sales index constructed from multiple public and proprietary datasources on home sales and are reported at the census tract level. A description of the HPImethodology can be found in Bogin et al. (2016).Job flow data are obtained from the Longitudinal Employer-Household Dynamics,Origin-Destination Employment Statistics (LODES) data products. LODES provideslinked workplace and residence data that provide a matrix of commute flows at the censustract level. The data is compiled by the US Census Bureau. LEHD data coverage extendsto 95% of wage and salaried employment nationally (Graham et al., 2014). Omittedworkers include self-employed individuals and US military personnel (Graham et al., 2014).2013 data are used for post-treatment commute flows, and 2002 data are used forpre-treatment commute flows. Data from 2000 are not available from LODES.Empirically identifying the effect of commuting costs requires detailed data oncommute times. Internet based route planning services, such as Google Maps, publiclydisseminate travel instructions and estimates of travel duration. To approximate thecommute options faced by urban travellers I automate a process to collect Google travelestimates within my sample of metros. Specifically, I use the Google Maps ApplicationProgramming Interface (API) to scrape data on relevant trips. I constructed a full matrixof potential tract to tract commute routes for each metro, resulting in 1,412,602origin-destination pairs. I queried routes through the API for travel instructions for bothdriving and public transit for an 8 am departure on a Wednesday. The API provided thetrip time and distance as estimated by Google’s algorithm. The resulting data set providesthe precise travel time and distance for any possible commute executed through thenetwork, with granularity at the census tract level.I further process the Google data to identify all trips that make use of new LRT26infrastructure. I first compile a list containing the name of every LRT station builtbetween 2000 and 2013, as it is identified within the API. Using step-by-step navigationinstructions for public transit trips, I run a text search program to identify all of theorigin-destination pairs that make use of the new LRT infrastructure. Identifying routesthat use new LRT allows for the approximation of a commute matrix from before LRT wasconstructed. Furthermore, this novel data set allows for the direct estimation of LRT’spower to redirect commute flows (section 3.4.3). Further details on the API generated dataare relegated to Appendix A. I find that straight line distances are a poor proxy for publictransit trip durations, as public transit trips often follow circuitous routes. The use ofGoogle routing data provides a much more realistic matrix of travel times than would bepossible using traditional commute flow data sets.3.4 Neighbourhood Effects of LRT3.4.1 MethodologyI consider a census tract to be “treated” by LRT if a new station was built within onekm of the tract’s population weighted centroid, between the pre and post treatmentperiods (2000-2013). There are 70 treatment tracts identified across the four metros. 22 arein Seattle, 18 are in Salt Lake City, 15 are in Minneapolis and 15 are in Portland. Toisolate a valid control group, a number of tracts are dropped from analysis. Tracts withinone kilometre of the central business district (CBD) or the airport are dropped, whereCBD location is proxied by city hall. Any tract that contained or was within one km of aLRT station prior to 2000 is also omitted from analysis. Additionally, all untreated tractsthat are within three km of a new station are omitted to avoid tracts that were partiallytreated by local spillovers.9 The resulting data set contains 1,974 tracts. The location ofnew LRT stations and treated tracts are shown in Figure 4.It is important to note that the introduction of the LRT systems normallycorresponded with an adjustment of the existing bus service. In particular, LRT routestypically replace bus routes. Therefore, results should be interpreted as the effect ofupgrading from standard bus service to LRT service.Equation 1 presents the general regression approach.∆Yi = β0 + β1LRTi + ΓXi + Θm(i) + εi (1)∆Yi is the change in a neighbourhood characteristic between 2000 and 2013. LRTi is a9Minneapolis opened a second LRT line in 2014. To avoid potentially confounding partial treatment oranticipation effects, tracts within three km of a 2014 station are also omitted.27Figure 4: LRT Treated Tracts and Instrumental VariableA. Minneapolis B. PortlandC. Salt Lake City D. Seattle - Treated tracts  - Airport corridor - New station - Pre 2000 stationdummy variable that takes a value of one if the tract is treated with LRT. Xi is a vector ofcontrol variables. The list of control variables comprising X can be found below Table 7.Control variables are drawn from 2000 and 1990 values to control for differences in bothlevels and trends. Θm(i) captures core based statistical area (CBSA) fixed effects. i indexestract and m indexes metropolitan area.28OLS results are expected to be biased due to the endogenous placement of rail stationsrelative to local economic trends (Ihlanfeldt and Sjoquist, 1998; Holzer et al., 2003;Tyndall, 2017). To identify a causal relationship, researchers have equipped past empiricalinvestigations with exogenous network shocks. Notably, Holzer et al. (2003) analysed anextension of the San Francisco subway system, focussing on firms located along thesuburban portion of the extension. A key identifying assumption was that transportationplanners never anticipated a local demand for reverse commuting, which would mean theeffect of the infrastructure was plausibly exogenous to future job growth among reversecommuters. The exogeneity is unclear as relevant policy documents called for a “specialemphasis on off peak and reverse commute trips” (Bay Area Rapid Transit District, 1987).Tyndall (2017) made use of station closures triggered by a hurricane event as an exogenousshock to the New York City subway system, finding neighbourhoods that lost subwayaccess experienced increased rates of unemployment. I propose a new instrument forderiving random variation in LRT placement: straight lines connecting the CBD to themetro’s primary airport.Consider the ideal randomized experiment to identify the effect of a LRT station onneighbourhood outcomes. Among a set of neighbourhoods, a lottery determines whichsubset of tracts gain LRT stations. After treatment is applied, any differences in outcomesobserved between the treated and control neighbourhoods would be attributed to thecausal effect of LRT. This is true because the mechanism that assigned treatment statuswas orthogonal to pretreatment characteristics. The proposed instrument aims to capture acase of analogous random allocation.In 1975, of the 25 largest US metros, only Boston and Cleveland had a direct rail linkfrom the CBD to the largest metropolitan airport, today 16 of the largest 25 metros havesuch a link. Figure 5 plots this progression. The economic development motivation forconstructing rail links from city centres to airports is based on very little economicliterature. Case studies have generally been unable to provide compelling arguments infavour of such projects (Stubbs and Jegede, 1998; Widmer and Hidber, 2000).Contrastingly, the political motivations for constructing such “mega-projects” appear to bestrong (Altshuler and Luberoff, 2004). The origins of these large rail projects are oftenattributed to state or regional governments who are promoting broad economicdevelopment goals and are unlikely to be apprised of, or motivated by, differences inneighbourhood level transit demand. As such, tracts treated by LRT by virtue of theirlocation relative to the airport can be assumed to have local economic trends that areorthogonal to the mechanism assigning treatment status. I assume an exclusion restrictionwherein changes to local economic conditions are unaffected by being en route to the29airport except through differential LRT allocation. The assumption is imposed aftercontrolling for pretreatment observables, including distance to the airport.Figure 5: Share of Large US Metros with a Rail Link from Downtown to the LargestAirportThe composition of the largest 25 metros is allowed to change through time according to US Censuspopulation estimates. Additional US cities are currently planning LRT extensions to their airports,including Buffalo, Pittsburgh and Sacramento.I instrument the LRT dummy variable in equation 1 with a dummy variable for beingwithin the “airport corridor.” A tract is considered to be in the airport corridor if itscentroid is within two km of a straight line drawn between the airport and the pre 2000station that is closest to the airport, creating a corridor that is four km wide. If there is nopre 2000 station (Seattle and Minneapolis) then city hall is used. Airport corridors aremapped in Figure 4. First stage regression results are shown in Table 6. A tract in theairport corridor is 46.7 percentage points more likely to receive a LRT station relative to asimilar tract located outside of the corridor, according to first stage OLS results. Thecorridor variable alone explains 22.2% of the variation in station assignment probability.First stage results demonstrate that the instrument is a strong predictor of LRT stationlocation over the period of study. Standard weak instrument tests strongly reject theproposition that the instrument is weak.The four km wide corridor is selected because it maximizes the explanatory power ofthe first stage. However, results were estimated using corridors ranging from two to eightkm and results changed very little. Corridors wider than eight km experience weakinstrument issues.I use robust standard errors throughout the analysis. I have repeated analysis withstandard errors clustered at the CBSA level and find results that are very simlar. I follow30Table 6: First Stage Results, Predicting Station LocationsVariableAirport Corridor (dummy) .467∗∗(.079)Tract level controls? YCBSA fixed effects? YR2 0.334Cragg-Donald Wald F statistic 300.12Kleibergen and Paap Wald F statistic 35.24Significance levels: ∗ : 5% ∗∗ : 1%. N = 1,974. Robust standard errors in parenthesis. See notesunder Table 7 for full list of control variables. The outcome variable is a dummy variable for LRTtreatment status.the advice of Imbens and Kolesar (2016) as well as Angrist and Pischke (2008) and do notcluster standard errors due to the small number of clusters.3.4.2 Neighbourhood Change ResultsNeighbourhood change results are summarized in Table 7. Initially, a naive OLSapproach is used to estimate the effect of LRT stations on neighbourhood characteristics(equation 1). Table 7 also provides IV results, where treatment status (LRTi) isinstrumented with a binary variable for being within the airport corridor. The partialeffects of control variables are excluded from the table. All estimates are executed in firstdifferences to focus analysis on changes in neighbourhood characteristics between 2000 and2013. In general, a local LRT station is found to significantly improve local employmentoutcomes. The below summary of results will focus on the IV specification. The IV resultswill be important to the subsequent structural estimation approach. A comparison of IVand OLS results suggests that rail infrastructure was directed towards neighbourhoods withlatent economic trends that were below average. The effect is consistent with the findings ofKahn (2007), wherein transit infrastructure is routed through less affluent neighbourhoods.Labour market outcomes have a strong positive response to the introduction of LRT(Table 7). The local employment rate among adults rose by a highly significant 8.3percentage points, relative to a 2000 treatment tract average of 62.1%. Correspondingly,the share of the local adult population participating in the workforce rose by 6.4percentage points and the local unemployment rate fell by 4.0 percentage points. A portionof the positive labour market effects may be attributable to public transit providing accessto new labour market opportunities for the local population. However, the shift in labourmarket outcomes may be the result of a large change in the characteristics of the localworkforce. A rise in characteristics that are correlated with strong employment outcomes31Table 7: Neighbourhood Change Results∆ Employment ∆ Labour Force ∆ UnemploymentRate Participation Rate RateOLS IV OLS IV OLS IVGained LRT Station .009 .083∗∗ .006 .064∗∗ -.010 -.040∗∗(.008) (.025) (.007) (.024) (.005) (.013)Mean 2000 value .621 .675 .081(treated obs)∆ Public Transit ∆ Private Vehicle ∆ CommuteMode Share Mode Share Time (min)OLS IV OLS IV OLS IVGained LRT Station .018∗∗ .031 -.024∗ -.058∗ .201 .237(.006) (.016) (.009) (.023) (.364) (1.042)Mean 2000 value .132 .738 22.422(treated obs)∆ Log Home ∆ Repeat Sales ∆ Log HousingValue Index UnitsOLS IV OLS IV OLS IVGained LRT Station -.018 -.016 5.200 16.378 .187∗∗ .320∗(.031) (.080) (2.870) (9.945) (.048) (.148)Mean 2000 value 12.181 100 7.269(treated obs)∆ White ∆ College ∆ Log MedianPop. Share Degree IncomeOLS IV OLS IV OLS IVGained LRT Station .005 .039 .015 .026 .020 .106(.012) (.036) (.010) (.031) (.027) (.070)Mean 2000 value .632 .246 10.731(treated obs)Significance levels: ∗ : 5% ∗∗ : 1%. N = 1,974. Robust standard errors in parenthesis. Controlvariables include CBSA fixed effects, the distance to city hall, the distance to the airport, as well as the1990 and 2000 values of the following variables: private vehicle mode share, public transit mode share, whitepopulation share, black population share, Asian population share, share of population with a high schooldiploma, share of population with a college degree, share of population with a graduate degree, employmentrate, unemployment rate, median household income, share of population between age 18 to 30, share ofpopulation over age 65 and share of local housing units that are detached homes.could be considered as evidence of LRT induced gentrification. Additionally, workers maybe sorting on employment status itself, as employed workers will value the commutingbenefits of LRT (Wasmer and Zenou, 2002, 2006).Viewed as a place-based policy, LRT appears to be a powerful tool to advance averagelocal labour market outcomes. The local employment effects I find are larger than thosetypically reported in evaluations of government run place based economic developmentpolicies such as Empowerment Zones and Enterprise Zones (Ham et al., 2011; Neumarkand Kolko, 2010).32Policy makers may be interested in LRT as a means to increase the use of publictransit and decrease reliance on privately owned vehicles. The partial effect of gaining alocal station on the transportation mode share of local commuters is shown in Table 7.There is only weak evidence that LRT stations increase the share of the local workforcecommuting by public transit, relative to control tracts. However, structural estimation willconclude that LRT causes an aggregate increase in public transit use, but the effects aredispersed across the metro (Section 3.5). A LRT station is found to increase the share ofthe workforce using public transit by 3.1 percentage points from a baseline of 13.2%. Theeffect is statistically insignificant. However, there is strong evidence that the share of thelocal workforce using private vehicles to commute is reduced. The share using a privatevehicle falls by a significant 5.8 percentage points, from a treatment tract average of 73.8%.While a portion of the effect may be attributable to commuters switching from cars to thenew LRT option, a portion is potentially due to induced changes in neighbourhooddemographics and preferences, brought about by endogenous household sorting. I test foran effect on the share of commuters either walking or biking to work and find a statisticallysignificant increase. Why local LRT would cause individuals to walk or bike is not obvious,suggesting mode changes are influenced by changing characteristics of local workers.Redevelopment around LRT stations that include pedestrian and cyclist friendlyinfrastructure may also contribute to this result. I also find that a local LRT station causesthe share of individuals who report having access to a private vehicle to fall from 93.6% to90.2%, the effect is significant at the 10% level.It is informative to consider the comparative magnitudes of the employment effectsand public transit use effects. If there is no endogenous sorting, new public transitcommuters are comprised of employed workers switching modes and workers who werepreviously not working, securing employment as a result of improved transit. Even if everynew public transit commuter had previously been unemployed, the public transit effects areinsufficient to explain the employment effects. This suggests that neighbourhoodemployment effects are largely driven by workers sorting within the metro.Lower income residents may disproportionately benefit from public transit as thisgroup is less likely to own a private vehicle and more likely to be reliant on transit.However, lower income residents may also be more sensitive to local increases in housingcosts. Prior research has generally found transit amenities have a positive effect on localhome values. Kahn (2007) found new “walk-and-ride” transit stations increased averagelocal home values by 5.4%, 10 years after station construction. I find no evidence that LRTraises the average local home value. However, this estimate is not adjusted for potentiallychanging housing size or quality. Of particular concern is that LRT station constructions33may be accompanied by the displacement of local single family homes with condominiums,as cities attempt to coordinate transportation and land use policies. The tract level FHFAHPI is used to control for changes in housing characteristics. Using this repeat sales index,LRT stations are found to cause a 16.4% increase in local housing values. The repeat salesestimate is executed on a reduced sample (1,852 tracts rather than 1,974) due to missingobservations in the FHFA data. Data constraints render the home value estimatestentative, but the evidence points towards LRT induced home value increases.I test for an effect of LRT on shifts in racial composition. No statistically significanteffects are found but point estimates suggest that LRT led to an increase in the local whitepopulation share of 3.9 percentage points (Table 7). I also test for an effect on localeducation rates. Similarly, I find no statistically significant shifts, but point estimatessuggest a rise in local education levels, including an 11% increase in the share of the localpopulation with a college degree. Both of these results provide some evidence of LRTinduced neighbourhood gentrification.While LRT constructions are often proposed as a means of reducing commute times, Ifind no evidence that average commute durations are reduced when a tract gains a LRTstation (Table 7). Private vehicle commuting is the fastest mode for virtually every route,so reductions in private vehicle commuting lead to longer average commute times.However, monetary commuting costs likely fell due to reduced use of private vehicles.LRT allocation is often accompanied by local real estate development and reductions inzoning restrictions (Atkinson-Palombo, 2010). This reality will be important for structuralestimation as the introduction of LRT may increase demand for a neighbourhood, but maysimultaneously increase the supply of housing in that neighbourhood. Relative to controltracts, tracts that received LRT saw an average increase in local housing stock of 37.7%,over the 13 year period. The result suggests that cities followed an approach of TransitOriented Development (TOD), directing new housing towards transit stations.Despite the relatively small sample size available to conduct neighbourhood estimates,LRT has a clear effect on labour market outcomes and commuter mode share at theneighbourhood level. There are two mechanisms that potentially contribute to LRT’s effecton neighbourhood labour market outcomes, (1) the beneficial expansion of localindividual’s access to transit and labour market opportunities and (2) endogenousneighbourhood sorting in response to local transit infrastructure.Table 8 directly controls for local post-treatment demographic characteristics (raceand education) as a way to reduce the statistical effect of neighbourhood sorting. Themagnitude of employment effects are reduced by roughly 10% when post-treatmentdemographics are controlled for. While observable demographics may be controlled for, a34more intractable form of sorting may persist. As noted in prior works (LeRoy andSonstelie, 1983; Wasmer and Zenou, 2002, 2006), workers are likely to sort on employmentstatus itself. Even with identical demographics, an employed individual has a motivation tolocate close to their workplace while a worker without a job lacks this incentive. Withoutthe ability to observe the locational decisions of individuals through time, it is not possibleto accurately decompose neighbourhood changes into sorting effects and individualincentive effects. Section 5 will propose and estimate a microfounded structural model thatincorporates both mechanisms.Table 8: Controlling for Post Treatment DemographicsIV IV with Post Treatment Controls∆ Employment Rate .083∗∗ .074∗∗(.025) (.022)∆ Labour Force Participation Rate .064∗∗ .057∗∗(.024) (.021)∆ Unemployment Rate -.040∗∗ -.037∗∗(.013) (.012)Significance levels: ∗ : 5% ∗∗ : 1%. N = 1,974. Robust standard errors in parenthesis. Controlvariables from Table 7 are included. Additionally, the following post treatment controls, calculated as thedifference between 2013 and 2000 values, are added: white, black and Asian population shares; and highschool, college and graduate education rates.3.4.3 Effect of Light Rail on Commute FlowsAn observed increase in commuting along LRT routes would be an indication thatworkers are responding to the infrastructure. I find that, once LRT is in place, workers aremore likely to execute commutes along routes that directly benefit from LRT. I use Googlerouting data to identify all home-work pairs for which the fastest transit route involves asegment of newly added LRT infrastructure. Of 1,412,602 possible home-work pairs, 84,063(6.0%) were connected through new LRT infrastructure. These newly connectedhome-work pairs cover 6.2% of 2013 jobs. Figure 6 maps 2013 observed commute flowsacross the four metros. Commute routes that are populated by at least one commuter(according to LEHD LODES data) are captured on the map, with heavier lines indicatingmore commuters. The left column of Figure 6 displays all commutes in 2013. The rightcolumn displays only the commutes that make use of a LRT station added between 2000and 2013. In Salt Lake City, the large expansion of LRT meant that 26.5% of 2013commuters had new LRT infrastructure as a component of their quickest transit route. InPortland, 9.8% of commuters became connected through new LRT. Relative to metro size,35LRT construction in Minneapolis and Seattle was less expansive. In Minneapolis, 2.4% ofcommuters were covered by new LRT infrastructure and in Seattle the figure is 2.0%.[LRT Route Share]jt = β0 + β1[Post LRT]jt + Φj + Ψt + εjt (2)LRT Route Share is the share of the tract’s employed workers who commute along aroute that received new LRT service. j indexes home tract. t indicates year; either2000 or 2013. Φ is a home census tract fixed effect. Ψ is a year fixed effect. β1 is theparameter of interest.I compare the share of commutes executed along LRT treated routes before and afterthe LRT infrastructure was built. A home-work route is considered to be treated if thefastest transit link between home and work, according to Google data, includes a LRTstation built between 2000 and 2013. I regress the share of a tract’s workers who commutealong a LRT treated route against a dummy variable for post treatment status, as shown inequation 2. Fixed effects are included for the tract as well as the year. Table 9, column 1displays the effect of a LRT connection on the LRT route share. Results show a small butsignificant increase in the overall share of the local workforce commuting along LRTserviced routes. In the average tract in 2000, 7.13% of a tract’s workers were commutingalong a route that would be treated with LRT, in 2013 the fraction had risen to 7.24%.The relatively modest public transit use across the three cities (4.94% of commuters in2000) is consistent with LRT having only a minor effect on overall commute flows.Limiting the analysis to populations who are more likely to be reliant on public transityields results that are significantly larger, suggesting that the role of LRT in choosingwhere to work and live is more significant to particular groups. Column 2 limits analysis tojobs that pay less than $15,000 annually –capturing part time and low wage employment–and shows an effect 2.6 times higher than the aggregate, with the share of low income jobstraversing LRT treated routes increasing from an average of 6.4% to 6.7%. Column 3 testsfor an effect among jobs that pay over $40,000 annually, but finds no effect. Column 4 testsfor an effect among workers under 30 years of age. Workers under 30 had a response 2.5times larger than the overall effect. The strong effect on the commute flows of young peopleis consistent with prior literature that has found youth to be more sensitive to the spatialproximity of jobs, for example O’Regan and Quigley (1998). The stronger effect among lowincome workers is consistent with this group’s higher rate of reliance on public transit.36Figure 6: Mapping Commuter FlowsAll Commutes LRT Connected CommutesA. MinneapolisB. PortlandC. Salt Lake City| - Commute Flow - Pre 2000 Station - New StationThe figure continues on the following page.3.5 Urban Structural Estimation3.5.1 Modelling Neighbourhood ChoiceI rely on the above estimated causal neighbourhood effects to estimate a structuralneighbourhood choice model. By assigning workers a utility function, the observed changesin the LRT system can be reconciled with observed neighbourhood impacts through thedecisions of individual workers. The model will yield parameters that govern worker37Figure 6: Mapping Commuter FlowsAll Commutes LRT Connected CommutesD. Seattle| - Commute Flow - Pre 2000 Station - New StationHeavier lines indicate more popular commute routes. LODES data does not include interstate commutes,therefore the Washington state portion of the Portland metro and the Wisconsin portion of theMinneapolis metro are omitted from this portion of analysis.Table 9: Shifts in Commuter Flows Towards LRT Treated RoutesAll Jobs Low Income Mid-High Income Under Age 30(1) (2) (3) (4)Treatment .0011∗∗ .0029∗∗ -.0005 .0028∗∗Effect (.0004) (.0006) (.0006) (.0006)Home Tract FE? Y Y Y YYear FE? Y Y Y YMean LRT Route Share, .0713 .0642 .0770 .0694Treated Routes (2000)Significance levels: ∗ : 5% ∗∗ : 1%. N = 3,592. Robust standard errors in parenthesis. Low incomejobs pay below $15,000 annually. Mid-High jobs pay above $15,000 annually.preferences for LRT. Model results can answer whether LRT is successful at catalysingwelfare improving labour market linkages. Preference parameters will also enablecounterfactual welfare analysis, facilitating estimation of how the benefits of LRT are38spread across skill groups. The microfounded model overcomes the issue of endogenoussorting by modelling worker choice explicitly.The practice of estimating structural neighbourhood choice models is becomingincreasingly popular due to advances in methodology and the ubiquity of computationalpower. Structural estimation provides an important advantage in its ability to recover theparameter estimates that account for complex and endogenous choice. In regards to thecurrent research question, the ability of new rail infrastructure to advance neighbourhooddevelopment is of general interest. However, a more fundamental question is how theseinvestments affect individual behaviour and translate into a distribution of societal welfarechanges.The contributions of Alonso et al. (1964), Muth (1969), Mills (1967) and Fujita andOgawa (1982) provide a basis for modelling urban spatial structure. From this early workit was clear that the rational decisions of utility maximizing agents who face differentcommuting costs give rise to spatial heterogeneity in the characteristics of residents.Epple and Sieg (1999) pioneered an estimation methodology for structuralneighbourhood choice models. The focus of Epple and Sieg (1999) as well as Bayer andMcMillan (2012) was primarily on reconciling observed data with the predictions ofTiebout (1956). Bayer et al. (2004) further developed a framework of discreteneighbourhood choice. Sieg et al. (2004) applied a neighbourhood choice model to estimatethe welfare effects of a change in air quality in southern California, demonstrating thatreduced form methods were not able to capture the distribution of welfare effects due toresidential sorting. The exploration of parental schooling decisions and neighbourhoodchoice has become an interesting application for neighbourhood sorting structural models(Bayer et al., 2007; Ferreyra, 2007). Ahlfeldt et al. (2015) implemented a structuralapproach to investigate agglomeration and amenity forces in an urban environment. Recentapplications to transportation amenities are found in Severen (2018) (LRT in Los Angeles)and Tsivanidis (2018) (bus rapid transit in Bogota, Columbia). The common challengeshared by these papers and the current task is to estimate the benefits of a spatiallydelineated amenity in the presence of sorting.3.5.2 WorkersThe model of worker choice will take the following general form. The utility of aworker is represented by a Cobb-Douglas style function (equation 3).Uijkv = (ρjC)γH(1−γ) + ξijkv (3)39Workers derive utility from numeraire consumption (C) and the consumption ofgeneric units of housing (H). The share of income a worker spends on housing is set by1− γ. i indexes the worker, j indexes home tract, k indexes work tract and v indexes atransportation mode. In addition to the effect on commute times, the presence of a localLRT station may improve utility through its ability to enhance numeraire consumption.LRT may allow workers to consume a wider variety of local goods and services due to theirimproved mobility or through enjoyment of local economic development adjacent tostations. Indeed, LRT routes are often oriented to provide access to consumptionamenities, for example airports, professional sports venues and retail. ρj takes a value ofone if no LRT station was built in the tract. If a LRT station was built in the tract ρj is auniform consumption multiplier that will be endogenously determined. A Type 1 extremevalue distributed error term (ξijkv) captures a worker’s idiosyncratic preferences over homelocation, work location and mode choice. All workers are renters and pay rent to a landlordoutside of the local economy.Workers experience iceberg commuting costs. Commuting costs (equation 4) varyaccording to the particular home-work pair, wage, and mode choice. The cost containsboth a pecuniary and non-pecuniary component. Use of a private vehicle carries a flatrental fee (r) and a per km use fee (g), such that the pecuniary cost of commuting (θ) byprivate vehicle is r + gdjkv. Where djkv is the trip distance (km) between locations j and k.djkv is taken from Google routing data and corresponds to the actual road distancecovered. The worker can forgo renting a car and instead incur a flat pecuniary commutingcost equal to the cost of a monthly transit pass (t), such that θjkv = t.cijkv = θjkv︸︷︷︸pecuniary+ ζvwiτjkv︸ ︷︷ ︸non-pecuniary(4)The non pecuniary cost of commuting is calculated according to the time duration ofthe trip (τjkv) multiplied by the worker’s wage and a value of time constant (ζv). ζv isindexed by travel mode to allow for the possibility that distaste for travel in privatevehicles, bus and LRT may be different. I use the term bus to denote all public transittrips that do not include new LRT.Utility maximization is subject to a budget constraint (equation 5).wi = HpHj + C + cijkv (5)wi is worker income, pHj is the location specific price of a unit of housing and cijkv isthe transportation cost incurred from commuting. There is no saving, and workers exhaust40their budget constraint.Workers make the following choices simultaneously:(1) the location of the home tract(2) the location of the work tract(3) whether to rent a private vehicleEquations 3 and 5 result in an indirect utility function which governs worker choice(equation 6).Vijkv = (wi − cijkv)γγργj (1− γpHj)1−γ + ξijkv (6)Workers sell their labour in a competitive market. Microdata is used to construct adistribution of potential labour income. I do not observe the market price of labour forindividuals who are not employed, so I estimate the full distribution based on observablecharacteristics, imputing potential labour income for those not employed. The valuation ofan individual’s labour is determined by estimating a Mincer equation on a vector ofobserved demographics among employed workers (equation 7).ln(wi) = β0 + β1Xi + i (7)Xi is a vector of individual characteristics including age, age squared and dummyvariables for educational attainment (high school, college, graduate school), race andethnicity (black, white, Hispanic, Asian), gender, and home metro. Every worker isassigned a potential income (wpotentiali ) that is calculated based on their characteristics andthe partial effects estimated in equation 7. Potential income is a measure of worker skill, asvalued by employers. The wage earned through employment (wi) is equal to potential wage(wpotentiali ) multiplied by a tract specific multiplier. Wages may differ across space.Agglomeration economies provides one theoretical source of wage heterogeneity throughspace. Average wages in a CBSA are matched to data, while spatial wage differentials areendogenous to the model.In equilibrium, every worker selects the home, work, and mode that maximizes Vijkv. Iassume a closed city, where workers must stay within their current metro. A worker canforgo employment, in which case they receive a set government transfer (η), earn zeroemployment income and pay zero commute costs. Pecuniary relocation costs within ametro are assumed to be zero.413.5.3 FirmsI introduce two simplifying assumptions to model the role of firms. First, I assumeeach tract possess one representative firm, located at the tract centroid. Second, I assumethat individual firms have a perfectly elastic demand for labour. Firms are willing to hireas many workers as are willing to accept employment at a constant wage level. The secondassumption will be true if firms exist in a competitive market, possess constant returns toscale production technology and have access to a perfectly elastic external capital market.Firms offer wages that differ depending on the skill level of the worker. The wagevariation across skill groups is recovered from equation 7. Wages also differ across firms(tracts). I introduce a tract level wage multiplier. In the pretreatment period, the numberof workers employed by each firm is set according to observed LEHD data from the year2002. When solving the model, I raise and lower the tract specific wage multipliers in orderto match each firm’s capacity to the share of workers selecting that firm. After theintroduction of the LRT system, firms may endogenously hire more or fewer workers, butthe offered wages remain constant.If LRT induces a reduction in local commuting costs the number of workers willing toaccept employment at the persistent wage rate may increase because the costs of acceptingemployment at that location have been reduced. Credit (2018) estimated a similar effectfor the Phoenix, Arizona LRT system, finding a large increase in firm formation adjacentto new LRT stations. Locations that do not benefit from LRT become less competitive andexperience pressure to endogenously shrink. The aggregate quantity of jobs in a metro mayrise or fall endogenously in response to LRT.3.5.4 Estimation MethodThe neighbourhood causal effects estimated in Section 3.4 will be reconciled withinthe structural model. This methodology compels structural estimation to be grounded inobservable effects, diminishing the reliance of results on imposed functional formassumptions, which is a common concern for structural estimation models. Within treatedLRT tracts, LRT caused the share of local workers commuting by private vehicle to fall by5.80 percentage points, and the local employment rate to rise by 8.28 percentage points.The share of workers commuting by private vehicle in the pretreatment period was 87.2%.The structural model will be solved to precisely match these three terms.To enable estimation, one parameter will be taken directly from prior literature; thevalue of time for car commuting (ζcar). Significant prior research has attempted to estimatethe value of time for private vehicle commuting. Estimation will proceed by using the42estimated value from Small et al. (2005). Using data from drivers in the Los Angeles areaSmall et al. (2005) estimated the parameter to be 0.93. This suggests workers would bewilling to undertake an additional hour of car commuting if they were compensated by acash transfer equal to 93% of one hour’s wages.Equation 3 assumes that workers spend a constant fraction of income on housing(1− γ). Davis and Ortalo-Magne´ (2011) provided evidence that this share is relativelyconstant across US cities. The central estimate of Davis and Ortalo-Magne´ (2011) showshouseholds spend 24% of income on housing (γ = .76). Estimation will rely on microdata ofreported rent expenditure and income. According to 2000 microdata from the four metros,the mean share of income spent on rent is 25.6%, consistent with Davis and Ortalo-Magne´(2011). The average masks heterogeneity across income groups. At the 10th percentile ofthe income distribution workers report spending 50.9% of income on rent, while at the 90thpercentile the figure is 16.1%. Estimation will proceed by setting each worker’s γ to accordwith their position on the potential income distribution as estimated in equation 7.Pecuniary transport costs are generally observable. r is assumed to be $471 which isan industry estimate of monthly fixed vehicle costs (American Automobile Association,2007). g is assumed to be $3.96/km, which is derived from the industry estimate ofvariable vehicle costs (gas and maintenance), scaled up with the assumption that workerscomplete 22 round trip commutes per month. (American Automobile Association, 2007).The pecuniary cost of public transit (t) is parametrized as the price of a monthly transitpass in the relevant CBSA. t ranges from a low of $83.75 in Salt Lake City to a high of$110 in Minneapolis. Structural parameters are summarized in Table 10.Table 10: Structural ParametersSymbol Value Source Descriptionζcar 0.93 Small et al. (2005) Time value as share of wage rate, private vehicleζbus 1.34 (estimated within model) Time value as share of wage rate, bus transitζLRT 1.03 (estimated within model) Time value as share of wage rate, new LRT transitg 3.96 American Automobile Association (2007) Variable vehicle cost per commute km per month ($)r 471 American Automobile Association (2007) Monthly rental fee for a vehicle ($)tMinneapolis 110 Local transit information Metro specific monthly transit pass ($)tPortland 100 “ ” “ ”tSeattle 99 “ ” “ ”tSalt Lake City 83.75 “ ” “ ”η 1000 . Out of labour force monthly income ($)ρj 1.096 (estimated within model) Amenity value of transit (ratio of consumption utility inLRT tract to non LRT tract)The model’s pre and post treatment periods are differentiated in that a subset oftracts gain LRT stations and an updated commute time matrix is used to capture theeffect of LRT. Additionally, I rely on IV estimates of housing expansion and increase the43relative share of available housing units in LRT treated tracts by 37.8%.10 Changes to thecommute time matrix can be estimated with the Google travel data. The pretreatmentcommute time matrix is adjusted to approximate the state of affairs before the LRTexpansions were in place. Pre LRT transit commute times are not observed in the data.The reduction in commute costs that can be attributed to the introduction of LRT is thecombined effect of a reduction in travel time and a change in the valuation of time whilecommuting. LRT may be more desirable than bus service in terms of reliability, comfort, ordiminished social stigma. LRT treated commute routes are assumed to experience areduction in trip duration of 30%. The time value of bus commuting (ζbus) and LRTcommuting (ζLRT ) will be estimated within the model.Calculating the probability that a particular worker will select a particular home,work, mode combination is enabled by the assumed extreme value distributed idiosyncraticerrors, which results in the following multinomial logit probability function, where Pijkv isthe probability of worker i selecting a specific home location, work location and vehiclerental decision (equation 8). Upper bar notation indicates the maximum value in the set.Pijkv =eVijkv∑j1∑k1∑v0 eVijkv(8)Computationally, local rents, local wages, and structural parameters (ζbus, ζLRT , ρj) aresolved through iteration using contraction mapping, until the system reaches anequilibrium wherein every tract contains the share of employees and residents dictated bythe data. The model is identified by restricting the possible equilibriums to generate theobserved pretreatment private vehicle mode share of 87.2%, a fall in private vehiclecommuting in treatment tracts of 5.80 percentage points and a rise in the localemployment rate of 8.28 percentage points. Neighbourhood changes are relative to thecontrol tracts denoted in Section 3.4. An equilibrium is further defined by a Nashequilibrium where the decision of every worker is optimal, taking into account the decisionof all other workers. The uniqueness of the equilibrium follows naturally from Brouwer’sfixed-point theorem. A proof of equilibrium uniqueness for this class of model can be foundin Bayer and Timmins (2005). The uniqueness of the current model solution is clear as thedecision of each worker only affects other workers through equilibrium prices and wagesand not through endogenous neighbourhood amenity characteristics. Intuitively,neighbourhood rents and wages must be at a level that exactly attracts the correct numberof residents and employees. Ceteris paribus, the share of workers using public transit10As a robustness check, I rerun the model while assuming zero endogenous housing growth. Results donot change significantly, suggesting results are driven by LRT changes rather than spatial changes to housingallocation.44decreases monotonically with ζbus, the change in local transit commuting in tracts gainingLRT decreases monotonically in ζLRT and the increase in the local employment rate intracts gaining LRT increases monotonically in ρj. Therefore, there is a unique set ofparameters that map to the unique moments in the data.3.5.5 ResultsSolving the model yields the necessary preference parameters. The time valueparameter for bus commuting (ζbus) is found to be 1.34, substantially higher than theimposed time value of private vehicle commuting (0.93). This parameter aligns with theperception that bus travel is unpleasant relative to private vehicle travel, particularly dueto uncertainty in trip duration (Kou et al., 2017; Tyndall, 2018). The value of ζLRT isestimated as 1.03, suggesting workers consider a unit of time spent on LRT as less costlythan riding a bus, but more costly than driving in a private vehicle. According toestimates, a worker earning $20 per hour would be willing to forgo $26.80 to avoid one hourof commuting by bus and $20.60 to avoid one hour of commuting by LRT. These figuresare pure time costs, independent of the accompanying monetary costs of transportation.The model also recovers the average amenity preference for living in a tract with LRT.The value of ρj is estimated as 1.096. ρj suggests that numeraire consumption provides9.6% more utility per dollar when the household is located in a LRT tract. The resultindicates that there is a substantial positive amenity value of a local LRT station. Theparameter captures not only consumption mobility benefits, such as access to shopping andleisure amenities, but also changing neighbourhood characteristics induced by LRT.It is worth considering why the positive amenity value of local LRT must exist, giventhe estimated neighbourhood changes. A local LRT station causes more individuals tobecome employed than it causes to begin commuting by public transit. This is onlyplausible if LRT increases local employment through a channel unrelated to commutingcosts. Modelling LRT as a local amenity that improves consumption provides such amechanism. The local amenity raises local rents, repelling low skilled workers and workerswithout employment, pushing up the local employment rate.An advantage of the discrete neighbourhood choice model is that it can predictneighbourhood changes across the metro. Figures 7, 8 and 9 show the spatial predictions ofthe model. A clear prediction of the model is an increase in rent per unit of housing inneighbourhoods gaining access to LRT (Figure 7). The valuation of LRT as a local amenityincreases the utility of any worker able to locate in a LRT tract. The commuting benefit ofLRT further increases the utility of local employed workers commuting by transit, but hasno additional effect on car commuters or those out of the labour force. Given these45mechanisms, LRT unambiguously increases demand for a neighbourhood and thereforelocal rents. Areas not gaining LRT generally experience rent reductions as they becomeless desirable relative to the areas gaining LRT (Figure 7).Residing in a LRT tract requires paying higher rents, in part, for the opportunity toreduce commuting costs. Therefore, the benefits of LRT to an individual who is not in thelabour force are comparatively low. Figure 8 maps the changes in local employment levelsinduced by LRT. Employment rates rise in neighbourhoods that gained LRT stations, andfall marginally in other areas. I also observe a significant increase in average labour marketabilities (potential income) for neighbourhoods that gain a LRT station (Appendix B).Figure 9 shows neighbourhood impacts on the share of local workers commuting bypublic transit. While the average public transit mode share in tracts gaining LRTincreased more than other areas, the effect is quite heterogeneous across treatmentneighbourhoods. The large increase in local amenities encourages high skilled workers tomove towards the treated neighbourhoods, and these workers are less likely to use publictransit. Positive public transit effects are spread throughout the metropolitan area,including neighbourhoods very far from the new infrastructure. The broad spatialdistribution of mode shift is a result of low skilled, transit dependent populationsrelocating out of the urban core to avoid local rent increases. The new suburban locationof these workers does not enable them to afford a private vehicle, and they remain transitcommuters in less transit accessible areas.Attempts to expand urban public transit use often consider “captive” and “choice”riders (Krizek and El-Geneidy, 2007). The former have public transit as their only option,while the latter only choose public transit if it provides better service than their alternative(private vehicle). Local amenity effects of LRT stations repel captive users and attractchoice users. This may be an effective method to raise aggregate metropolitan transit usebecause the mode choice elasticity of choice riders is higher and they sort towards the highquality transit. However, the process degrades the spatial access of captive transit riders,undermining the progressivity of transit investment.There is some evidence that neighbourhoods with pre 2000 stations benefit from thenew stations constructed elsewhere. Network effects generate this outcome, asneighbourhoods along the LRT extensions become more accessible from other parts of theLRT system. However, the local amenity effects of LRT appear to dominate the networkeffects, with positive changes much larger in neighbourhoods that actually gained a newstation.The distribution of behavioural and welfare effects across income groups can berecovered from the model and are graphed in Figure 10. I scale results to correspond to the46Figure 7: Structural Results, Change in RentA. Minneapolis B. PortlandC. Salt Lake City D. Seattle< −4% (−4%,−2%) (−2%, 0%) + (0%, 2%) + (2%, 4%) + > 4%- New rail station - Pre 2000 rail station(“+” symbols are redundant with colours, but enable interpretation of figures if viewed in grayscale.)predicted effect of constructing ten new LRT stations within a metro of average size (1.25million workers). Panel A indicates that the probability of a worker switching to publictransit is distributed uniformly across potential income groups. While low skilled workersare generally more willing to take public transit, high skilled workers are more likely to47Figure 8: Structural Results, Change in Employment RateA. Minneapolis B. PortlandC. Salt Lake City D. Seattle< -4pp (-4pp,-2pp) (-2pp,0pp)+(0pp,2pp)+(2pp,4pp)+> 4pp- New rail station - Pre 2000 rail stationpp = percentage points(“+” symbols are redundant with colours, but enable interpretation of figures if viewed in grayscale.)relocate adjacent to the new LRT infrastructure. The two opposing effects result in thenew uptake of transit being spread uniformly across groups. Every ten LRT stations areestimated to increase metro wide public transit mode share by 0.43 percentage points(3.4% rise from the 2000 baseline).48Figure 9: Structural Results, Change in Public Transit Mode ShareA. Minneapolis B. PortlandC. Salt Lake City D. Seattle< -4pp (-4pp,-2pp) (-2pp,0pp)+(0pp,2pp)+(2pp,4pp)+> 4pp- New rail station - Pre 2000 rail stationpp = percentage points(“+” symbols are redundant with colours, but enable interpretation of figures if viewed in grayscale.)Panel B graphs the effect of ten LRT stations on the employment rate. Overall theintroduction of LRT is estimated to marginally reduce overall metro employment. Theshare of workers who are employed falls by one tenth of a percentage point, from a baselineof 69.9%. While reduced transit times may encourage employment, increased rents in49Figure 10: Structural Results, Distribution Across Potential Income PercentilesA. Public Transit Mode Share B. Employment RateOverall effect: +0.43 pp Overall effect: -0.10 ppC. Rent per Square Foot D. Commute DurationOverall effect: +0.18% Overall effect: +0.03 minutesE. WelfareOverall effect: +0.16%Results are scaled to represent the effect of ten LRT stations in a metro of average size (1.25 millionworkers).50accessible neighbourhoods cause the displacement of low skilled workers to low accessareas. Low skilled workers are likely to be on the margin of the labour market. The higherskilled workers who move into the central locations -encouraged by the potentialconsumption benefits of LRT- are very likely to be employed with or without the new LRTinfrastructure. I find LRT exacerbates spatial mismatch through the gentrification ofaccessible areas. The result demonstrates that if the consumption amenity effects of publictransit are sufficiently large, the consequent sorting may completely eliminate intendedemployment increases. The result is counterintuitive, given new transit is often built withthe explicit goal of improving labour market accessibility.The mechanism of displacement is also highlighted by rent effects shown in panel C.Residents with low potential income pay lower rents per unit of housing as a result of LRTinfrastructure, while residents with high potential income pay higher rents. This effect isgenerated by higher income residents moving to more central locations to reapconsumption benefits of LRT and low income workers moving out of the central city. Themechanism mirrors modern accounts of higher skilled groups returning to denser urbanlocations due to consumption preferences (Couture and Handbury, 2017).Given transit’s relatively slow speed relative to private vehicles, the average workerexperiences a small increase in commute duration (panel D). The lowest skilled workers seea small reduction in average commute time, as low skill workers who retain their residentiallocation close to LRT stations reap time savings as they switch from bus to LRT. Whenaveraged across the labour force, the effect of the transit infrastructure on averagecommute duration is negligibly small.Panel E illustrates the distribution of changes in the deterministic portion of workerutility. Every ten LRT stations increases the average resident’s utility by 0.16%. A largeshare of the welfare gains are derived from the direct consumption amenity benefits of LRTbut also through monetary savings from forgone car ownership. Critically, welfare benefitsare not generated by increased employment. While LRT may reduce monetary travel costsand raise the enjoyment of travel for those using transit, local rent increases cause sortingthat eliminates the potential for increased aggregate employment. All groups are found toexperience positive welfare effects, with low and high skill groups capturing the largestbenefits. I revisit the implications of these welfare effects in Section 3.6.The ability of LRT to expand transit commuting and increase employment is undercutby sorting. In Appendix C, I propose a complementary policy regarding transit passes thatreduces sorting, limiting these negative impacts. Appendix D considers the impact ofimplementing a bus system, assuming buses can accomplish similar mobility improvementsto LRT but do not bring consumption benefits to local neighbourhoods.513.5.6 Cost Benefit AnalysisI compare the estimated welfare benefits of LRT with typical project costs. Theprevious section estimated that a hypothetical ten station LRT line would increase thewelfare of the metro’s average worker by 0.16%. Achieving an equivalent variation througha uniform cash transfer program would require every worker to receive $92.50 annually.The average metro in the sample contains 1.25 million workers, suggesting this cashtransfer program would cost $116 million annually in the average sized metro. Assumingan annual discount rate of 5%, the present value of this cash transfer program is equal to$2.32 billion. Therefore, a LRT project in the average sized metro cannot be justified onthe basis of worker welfare benefits if project costs exceed $2.32 billion. All figures in thissection are in inflation adjusted 2013 dollars.The cost of building and maintaining LRT varies across metros, but is in many casesbelow $2.32 billion per 10 stations. The Minneapolis project had capital costs of $880million,11 for a line with 19 stations. Annual operational expenses in 2010 were reported tobe $27.5 million.12 Using the same 5% discount rate suggests that the cost of theMinneapolis light rail system in present value terms was $1.43 billion. Scaling the expensedown to a hypothetical 10 station system yields costs of $0.75 billion, substantially lessthan the estimated benefits.Costs of the Seattle system were higher than Minneapolis. Reported capital costs were$2.40 billion for a 13 station line.13 The annual 2010 operating budget for the LRT line was$51.6 million.14 The present value of these costs under a 5% discount rate is $3.43 billion.Scaling to a hypothetical 10 station line yields costs of $2.64 billion. The Seattle systemhad costs slightly above the estimated benefits accruing to workers. Seattle may be anoutlier in terms of high capital costs, as costs were inflated by the decision to put adowntown portion of the line underground.I calculate benefits that directly accrue to workers. I do not account for numerousother potential benefits of LRT. In particular, LRT may generate substantial positiveenvironmental benefits. I find strong evidence that LRT reduces the use of private vehiclesfor commuting, which are a large source of emissions. Benefits accruing to those out of thelabour force, such as children and retired persons, are also unaccounted for in the aboveanalysis. The simplified calculation in this section suggests that, in many cases, benefitsaccruing directly to the workforce may be sufficient to justify LRT project expenditures.11Metro Transit, Facts About Light-Rail Trains and Construction, 2010.12MetroTransit, Blue Line Operations, Financial Results by Calendar Year, 2013.13The Seattle Times, Light-rail contract dispute is resolved, June 23, 2011.14Sound Transit, Adopted 2010 Budget, 2009.523.6 ConclusionBetween 2000 and 2017, an average of 20 new LRT stations opened per year in theUS. The potentially significant economic consequences of this large infrastructureinvestment has received relatively little economic study. I test whether LRT hassignificantly affected urban labour markets across four US metros. I find strong evidencethat LRT improves neighbourhood level employment outcomes.I provide a structural model that can capture the complexities of neighbourhoodsorting that result from new transit amenities. Model results provide a nuancedunderstanding of the mechanisms that relate LRT to local labour markets. LRT improvescommuting networks and raises public transit use but also raises demand for transitaccessible areas. Lower skilled residents are more likely to directly consume the mobilitybenefits of public transit, but are also more likely to be displaced by local rent increases.Overall, I find that LRT causes a modest reduction in overall metropolitan employment, asthe local gentrification caused by LRT stations forces workers on the margin of the labourforce to locate in areas that are not transit accessible, exacerbating spatial mismatch. Theeffect is driven by the relatively high employment elasticity among low skilled workers:pushing lower skilled workers out of transit accessible locations can degrade their ability toobtain employment, undoing the accessibility benefits of LRT. The result iscounterintuitive, given that public transit projects are often constructed with the explicitintention of improving labour market access for socially vulnerable populations. Despitethe effect of sorting, I find that LRT systems provide positive welfare benefits across themetropolitan population. A simplified cost benefit analysis suggests that the welfarebenefits of LRT exceed typical project costs.The mechanisms described in this paper provide some explanation for efforts to resistLRT projects. For example, a second LRT line that was recently constructed inMinneapolis faced significant resistance from local populations along the planned routewho were concerned that the gentrification induced by LRT may be sufficiently harmful tocompletely offset mobility benefits.15 Resistance included a lawsuit filed by the NationalAssociation for the Advancement of Colored People that aimed to halt the project. Thispaper aimed to provide some description of the complicated economic impacts of LRT onurban residents. While I find LRT creates positive welfare effects, I do find that local homeprice increases reduce the mobility benefits accruing to workers with lower earning ability.High quality bus transit could potentially provide similar mobility improvements to LRTwithout inducing the same level of gentrification, potentially yielding a more progressive15The Train Line That Brought the Twin Cities Back Together, by E. Trickey, Politico Magazine, March16, 2017.53distribution of benefits. Given that high earning workers are able to capture significantwelfare benefits from LRT transit, even though transit commuting among high earners isextremely low, provides a partial explanation as to why LRT projects are proliferatingrapidly while bus transit systems have not undergone similar expansions over this timeperiod. Rich households may wield outsized control over public policy. These householdswould support using public money for LRT transit over bus because LRT directs significantconsumption benefits towards the rich.Current analysis is limited by a lack of neighbourhood level microdata. Tracking anindividual’s response to new transit infrastructure through time would allow for therelevant behavioural effects to be estimated directly. The absence of such data necessitatesinnovative approaches to modelling worker choice and the introduction of novelinstruments.544 Going Nowhere Fast: Urban Mobility, Job Accessand Employment Outcomes4.1 IntroductionUrban vehicle congestion is the topic of significant public concern. An annual reportfrom the Texas Transportation Institute argues that the costs of congestion to the USeconomy are approximately $160 billion annually (Schrank et al., 2015). A typical policyresponse has been to construct infrastructure that allows commuters to move at fasterspeeds through urban environments, such as new highways or transit systems. If workerand firm locations are taken as exogenous, such projects may successfully reduce commutetimes. However, the higher order consequences of mobility, such as inducing urban sprawl,make the equilibrium consequences of transportation infrastructure on commute timesunclear. Sprawl may act to increase the average distance between workers and firms,counteracting the benefits of mobility. This paper will analyse the relationship betweenurban mobility, sprawl and labour market connections. I first implement an explicitmeasure of mobility, attempting to capture the rate with which a commuter can traverseurban space. Subsequently, I test for a causal effect of mobility on sprawl and jobaccessibility across US metros. Finally, I test for a link between mobility and metropolitanlabour market outcomes.Though many authors have investigated the mechanics of vehicle congestion (Coutureet al., 2018; Duranton and Turner, 2011; Meyer, 1959; Walters, 1961; Lindley, 1987),limited work has been undertaken to determine the broad impacts of urban mobility oneconomic or labour market outcomes. Prior investigations have attempted to connectcongestion and economic growth, but unlike the current paper, do not explicitly considermobility’s propensity to induce urban sprawl. Boarnet (1997) was able to show arelationship between reductions in road congestion and economic growth across counties inCalifornia. However, Boarnet (1997) found no direct relationship between the expansion ofroad infrastructure and economic growth. Prud’homme and Lee (1999) investigated theimpact of commuter speed on urban productivity. In the context of France, the studyestimated that a 10% increase in urban transport speed was related to a 2.9% increase incity-wide productivity. The study stressed the necessity of limiting urban sprawl in orderfor these returns to be realized.While increasing commuter speed has intuitive benefits in terms of time savings, theliterature has described potential negative consequences of high mobility urbanenvironments, such as auto dependence and sprawl (Cervero, 1997; Crane, 2000; Grengs55et al., 2010; Jacobs, 1961; Levine et al., 2012; Yang, 2008). Some past work has arguedthat travelling at a high speed should not be considered a public benefit (Cervero, 1997,2001). Rather, public benefits arise from improving the accessibility of destinations, whichrelates to both speed and distance.Urban sprawl is often considered as a sign of poor urban management by planners.However, dispersed development does hold some benefits. If peripheral land is madeaccessible, homes and firms are provided with new location options. Furthermore, sprawldoes not necessarily increase average commute time in a polycentric environment becauseworkers can strategically locate close to their place of work. Anas (2011) found that USareas that cover twice as much land have average commutes that are only 10% longer.Anas (2015) provides further analysis regarding the complex relationship between sprawl,travel times and vehicle congestion.Expanding the capacity of a transportation system also encourages more travellers touse that system, a phenomenon known as induced demand. Particularly for highwaycapacity, past studies have found that the allotment of road space leads to more drivers,potentially to an extent that new capacity does nothing in abate congestion (Duranton andTurner, 2011).Some research has attempted to estimate the causal effect of mobility on jobaccessibility. Grengs et al. (2010) compared San Francisco and Washington, DC, findingthat the higher vehicle speeds in San Francisco were largely counteracted by greateraverage trip distances. Levine et al. (2012) expanded analysis to 38 large metropolitanregions in the US, finding that high travel speeds related to reductions in job accessibility,consistent with high mobility causing urban sprawl. Baum-Snow (2007) provided a seminalpaper relating highway construction to sprawl, establishing that the construction of the USinterstate highway system explains a significant portion of population decline in centralcities through the mid 1900s.Clear benefits exist in enabling residents to access a wider set of destinations andexpend less time on travel. Kain (1968) provided the framework of spatial mismatch toexplain why the inaccessibility of work locations may result in diminished employmentoutcomes, a theory expanded upon by numerous studies (Coulson et al., 2001; Gobillonet al., 2007; Holzer, 1991; Immergluck, 1998; Kawabata, 2003; Ong and Miller, 2005;Rogers, 1997). Cities with employment centres isolated from population centres putsworkers with high transportation costs at a disadvantage through limiting theiremployment opportunities and increasing the costs of job search. Prior work has identifiedblack (Zax and Kain, 1996) and youth (O’Regan and Quigley, 1996) populations to beparticularly sensitive to spatial isolation from jobs.56This paper will proceed as follows. Section 4.2 describes a metric for metropolitanmobility. Section 4.3 introduces data sources. Section 4.4 describes how mobility haschanged through time in US metros. Section 4.5 shows theoretically how increased mobilitycould reduce the accessibility of jobs. Section 4.6 proposes an instrumental variableidentification strategy. Section 4.7 presents results and Section 4.8 concludes.4.2 Measuring MobilityIn order to study urban mobility empirically an objective measure of mobility must beapplied. The proposed metric is adapted from Prud’homme and Lee (1999). The measuretakes the existing built environment as an enabler of mobility and attempts to measure thespeed with which an individual overcomes urban space. This study will be specificallyconcerned with home-work commuting.Provided individual level observations such that an individual’s residence andworkplace can be geographically identified, it is possible to calculate the “as the crow flies”distance separating these two locations. Such a route follows a geodesic line. Commutersare limited to travelling through the built transportation network, forcing them to deviatefrom the geodesic. The level of commuter deviation from the geodesic is a function of thecircuity of the transportation network (Levinson and El-Geneidy, 2009; Giacomin andLevinson, 2015). By using geodesic distance, the current paper will allow for variation innetwork circuity to impact the measure of mobility. The measurement of the geodesic lineis therefore an intentional abstraction and is not meant to capture the actual routeexecuted. The departure from network analysis to the conception of urban space as acontinuous velocity field has some commonality with Angel and Hyman (1970).The speed with which an urban resident navigates a given geodesic distance is afunction of (1) the average ground speed with which they travel and (2) the extent towhich the existing transportation network prompts the commuter to deviate from theirideal geodesic. Metropolitan conditions that enable high ground speeds and direct routeswill enable higher mobility. The velocity with which a traveller can overcome the geodesicdistance spanning their origin-destination pair will be referred to as commuter velocity.The commuter velocity of an individual is given simply by equation 9.Ωi = Di / ti (9)Where Ωi is the commuter velocity of individual i, Di is the geodesic distance between individual i’s originand destination pair, and ti is the time taken to complete individual i’s trip.Unfortunately, there is no publicly available data set that reports both Di and ti at57Figure 11: Correlation Between a Common Mobility Measure and Commuter Velocity (Ω)the level of individual commuters. I therefore base subsequent analysis on the averagemetropolitan level commuter velocity. This metric can be calculated because metropolitanlevel average distance travelled and average commute duration can be recovered from theLEHD LODES and ACS data sets respectively. These rich data sets should provideaccurate estimates of the respective parameters at the metropolitan level. Taken over theentire workforce (N) of a metro, commuter velocity (Ω) provides an indication of theregional level mobility provided by the transportation network (equation 10).Ω =N∑i=1DiN∑i=1ti(10)Commuter velocity is distinct from more commonly applied mobility metrics derivedfrom vehicle kilometres travelled (VKT) data, notably used in Couture et al. (2018) andDuranton and Turner (2011). Such metrics do not account for the role of route circuity.Figure 11 displays the correlation between the average “ground speed” of a commuter andthis study’s metric of commuter velocity. The ground distance between home and work isrecovered from the 2009 National Household Travel Survey–the same survey used inCouture et al. (2018) as well as Duranton and Turner (2011). Only 47 large metros havesufficient data to calculate the statistic. The correlation between the two metrics is .44,suggesting the metrics are measuring unique variations in mobility.58The remainder of this study will apply commuter velocity calculations to USmetropolitan areas. Commuter velocity displays a high level of variability across USmetropolitan areas and across US regions.Transportation investment in the US has been historically dominated by theaccommodation of private vehicles (Giuliano and Dargay, 2006; Grengs et al., 2010).Investment is normally directed towards either creating new roadways that provide moredirect access to destinations, or building out existing roadways to accommodate moretraffic at a higher speed. Therefore, much of transportation investment in the US can becharacterized as an attempt to increase Ω. It is not clear that this is a desirable policygoal, given the secondary effects on urban sprawl discussed below.4.3 DataIn order to calculate Ω, the current study aggregates data from multiple sources.Consistent data from all sources are obtained for the period of 2005 through 2014, whichcomprises the period of study. US metropolitan Core Based Statistical Areas (CBSA) areused as the unit of analysis. CBSA boundaries are held consistent through time andconform to the 2010 Office of Management and Budget delineations. This paper makes useof the Longitudinal Employer-Household Dynamics (LEHD) data products, compiled bythe US Census Bureau. The LEHD Origin-Destination Employment Statistics (LODES)data set provides located home-work pairs for employees across the US. Home and worklocations are identified at the census block level, providing a high degree of locationalprecision. LEHD data records approximately 95% of wage and salary employmentnationally, notable omissions are self-employed individuals and US military personnel(Graham et al., 2014). Any CBSA resident who works outside of their home CBSA isdropped from analysis, this applies to less than 1% of the sample. An additional limitationof the data is the exclusion of workers who live and work in different states. LODEScontains incomplete data for the District of Columbia, 2005-2009; Massachusetts,2005-2010; and Wyoming, 2014. Metro-year observations affected by these instances ofmissing data are omitted from analysis.The LEHD assigns work locations according to the physical mailing address of theemployer. This method may systematically misrepresent actual commuting patterns,particularly within industries that have inconsistent work locations, such as theconstruction industry (Graham et al., 2014). Misreported work locations could introducebias in the estimation of commuter velocity. However, assuming these inaccuracies areconsistent through time, this bias will generally cancel out in the identification strategy,59which will rely on year-to-year variation in commuter velocity within a particular metro.This paper will incorporate American Community Survey (ACS) data, which is alsocollected by the US Census Bureau. To identify trends through time, one-year estimatesare used. The Public Use Microdata Sample (PUMS) provides annual individual levelobservations for a randomly selected 1% of the US population. Variables are taken directlyfrom the Integrated Public Use Microdata Series data products (Ruggles et al., 2014). TheACS asks workers to report the number of minutes taken to commute one way to work,“door-to-door,” which will be used as the measure of commute travel time. PUMS alsocontains a wide array of individual level demographic characteristics that will be used inanalysis. 17 CBSAs lack ACS labour force data by race group for some or all years and aredropped from analysis. These 17 CBSAs are small, resulting in the ACS failing to survey asufficient sample. The final sample includes 361 metropolitan CBSAs, of which 355 areavailable across all ten years.Processed LODES and ACS data are collapsed to the CBSA level. PUMS observationsare identified at the Public Use Micro Data Area (PUMA) and are crosswalked to CBSAsusing the Missouri Census Data Center’s Geographic Correspondence Engine. CBSAsummary statistics are provided in Table 11. The combining of LODES and ACS dataallows metropolitan mobility to be measured annually. This method captures richervariation through time than provided by alternative data sources for mobility, for instancethe US National Household Travel Survey (NHTS), which is only conducted once every fiveto seven years and is only available for a subset of metropolitan areas.Table 11: CBSA Summary StatisticsVariable Mean Std. Dev.Commuter velocity (km/hour) 31.485 7.790Employment rate 0.560 0.061Labour force participation 0.605 0.057Weekly hours worked 37.994 1.357White 0.829 0.112Black 0.092 0.096Hispanic 0.090 0.129High school completion 0.862 0.053College completion 0.259 0.080Graduate school completion 0.098 0.040Mean age 47.598 2.628Population 703,464 1,598,416Metropolitan GDP (millions) 35,959 100,666N 3,58560Data for the instrumental variables on congressional representation are collected frompublicly available directories of the US House of Representatives standing committees. Imanually digitized these directories. Highway Trust Fund allocations to states are publiclyavailable from the Federal Highway Administration Office of Highway Policy Information.Data for metropolitan level GDP controls are gathered from the US Department ofCommerce, Bureau of Economic Analysis.4.4 A Recent History of Commuter VelocityIn 2014, the commuter velocity (Ω) of the average US commuter was 36.4 km/hr.There exists significant variation across US Census regions. In 2014, commuter velocity inthe South averaged 39.4 km/hr, but in the Northeast was only 30.5 km/hr. Commuters inthe Midwest and West regions travelled at 37.0 km/hr and 35.8 km/hr respectively (Figure12A).Between 2005 and 2014, growth in commuter velocity was highest in the South whereresidents experienced an annualized growth rate of 0.36% (Figure 12B). The Midwest andWest experienced annualized growth rates of 0.11% and 0.10% respectively. The Northeastexperienced essentially no change in commuter velocity between 2005 and 2014. The periodof study encompasses a sustained period of national growth in average commuter velocityfrom 35.8 km/hr in 2005 to 37.4 km/hr in 2011, followed by a decline to the 2014 value of36.4 km/hr (Figure 12C). Commuter velocity trends roughly mirror contemporaneous jobgrowth patterns in the US, consistent with literature demonstrating declining mobilityduring times of employment growth (Morrison and Lawell, 2016).Larger metros have higher commuter velocity values on average. For every 100,000person increase in population, Ω was 0.13 km/hr higher on average in 2014 (Figure 12D).Population is able to explain 7% of the variation in commuter velocity. There is no clearrelationship between metro size and commuter velocity growth over the period of study(Figure 12E).Table 12 provides a list of 2014 commuter velocity values for all CBSAs with apopulation over one million. The highest commuter velocity values occur in metros thatare notoriously highway and automobile dependent. Dallas had the highest commutervelocity amongst large cities in 2014 at 50.6 km/hr, followed by Phoenix at 50.0 km/hr.Nashville, Houston and Atlanta were also among the fastest five large cities. San Jose hadthe slowest commuter velocity amongst large cities at 26.7 km/hr. Giacomin and Levinson(2015) found San Jose’s road network to be amongst the most circuitous in the US. SanJose also experienced a significant decline in commuter velocity over the study period61Table 12: Large Metropolitan Areas (>1,000,000), by 2014 Commuter Velocity (Ω)Rank* CBSA name Ω Annualized Avg Avggrowth in km timeΩ, 05-144 Dallas-Fort Worth-Arlington, TX 50.6 0.4 23.3 27.65 Phoenix-Mesa-Scottsdale, AZ 50.0 1.3 21.6 25.96 Nashville-Davidson–Murfreesboro–Franklin, TN 49.0 0.3 21.9 26.810 Houston-The Woodlands-Sugar Land, TX 47.4 0.0 23.4 29.712 Atlanta-Sandy Springs-Roswell, GA 47.2 0.4 24.7 31.413 Birmingham-Hoover, AL 47.0 0.6 20.3 25.915 Oklahoma City, OK 45.4 0.9 16.8 22.217 Detroit-Warren-Dearborn, MI 45.0 0.2 20.3 27.121 Columbus, OH 44.1 -1.4 17.0 23.124 St. Louis, MO-IL 43.1 0.5 18.2 25.427 Rochester, NY 42.7 0.4 15.0 21.228 Indianapolis-Carmel-Anderson, IN 42.5 0.5 17.7 24.929 Riverside-San Bernardino-Ontario, CA 42.4 0.9 22.5 31.830 Minneapolis-St. Paul-Bloomington, MN-WI 42.3 -0.6 17.9 25.433 Charlotte-Concord-Gastonia, NC-SC 41.9 0.3 18.3 26.234 Jacksonville, FL 41.8 0.8 17.9 25.735 Richmond, VA 41.4 0.7 17.2 25.037 Miami-Fort Lauderdale-West Palm Beach, FL 41.1 0.2 19.6 28.538 San Antonio-New Braunfels, TX 40.9 0.5 17.4 25.646 Kansas City, MO-KS 39.9 0.1 15.5 23.451 Tampa-St. Petersburg-Clearwater, FL 39.6 -0.1 17.8 27.053 Cleveland-Elyria, OH 39.4 0.4 15.9 24.254 San Diego-Carlsbad, CA 39.3 0.3 16.7 25.558 Virginia Beach-Norfolk-Newport News, VA-NC 39.1 0.7 15.5 23.860 Austin-Round Rock, TX 39.0 0.2 17.2 26.462 Los Angeles-Long Beach-Anaheim, CA 38.8 -0.4 19.3 29.864 Memphis, TN-MS-AR 38.7 0.4 15.4 23.867 Raleigh, NC 38.4 0.0 16.3 25.668 Pittsburgh, PA 38.2 -0.2 17.2 26.970 Sacramento–Roseville–Arden-Arcade, CA 38.1 0.4 16.8 26.573 Chicago-Naperville-Elgin, IL-IN-WI 37.7 0.4 20.0 31.876 Seattle-Tacoma-Bellevue, WA 37.5 -0.6 18.6 29.879 Buffalo-Cheektowaga-Niagara Falls, NY 37.3 0.6 13.0 21.083 New Orleans-Metairie, LA 37.1 1.4 15.6 25.284 Orlando-Kissimmee-Sanford, FL 37.0 0.8 17.1 27.785 Cincinnati, OH-KY-IN 37.0 -0.1 15.2 24.696 Milwaukee-Waukesha-West Allis, WI 36.0 -0.3 14.1 23.4103 Hartford-West Hartford-East Hartford, CT 35.7 0.0 14.0 23.5115 Louisville/Jefferson County, KY-IN 34.9 0.1 14.0 24.0127 Denver-Aurora-Lakewood, CO 34.5 -0.5 15.8 27.6151 Baltimore-Columbia-Towson, MD 33.2 -0.4 16.7 30.2153 Las Vegas-Henderson-Paradise, NV 33.1 1.1 13.5 24.5162 Salt Lake City, UT 32.2 0.1 12.1 22.5193 San Francisco-Oakland-Hayward, CA 30.7 -0.9 16.4 32.2199 Portland-Vancouver-Hillsboro, OR-WA 30.2 -0.3 13.0 25.9206 New York-Newark-Jersey City, NY-NJ-PA 29.7 -0.3 17.8 36.0217 Philadelphia-Camden-Wilmington, PA-NJ-DE-MD 29.1 0.1 14.2 29.3252 San Jose-Sunnyvale-Santa Clara, CA 26.7 -1.1 12.3 27.5*Rank indicates the CBSA’s commuter velocity (Ω) ranking amongst the full set of 359 CBSAs for whichthere is 2014 data. Avg km is the average geodesic distance covered by a one-way commute. Avg time isthe average number of minutes elapsed during a one-way commute.62Figure 12: Summary of Commuter VelocityAnnual growth rates are taken over 2005-2014. Figures A-C are weighted by CBSA population. FiguresD-E display individual metros. Ω is commuter velocity measured in km/hour.(-1.1% per year), consistent with very high local job growth contributing to congestion.New York City, which is often cited for its high vehicle congestion, had the third lowestcommuter velocity amongst large metros at 29.7 km/hr.634.5 Commuter Velocity has an Ambiguous Effect on Job AccessTo establish that commuter velocity (Ω) could theoretically raise or lower jobaccessibility I present a simplified model. For this section, I assume a worker is located ona flat, featureless plain such that commute velocity is equal in all directions and jobs aredistributed uniformly. In this environment, the area accessible to a worker will be equal tothe area of a circle with a radius matching the distance the worker can travel in time τ . τis set exogenously and corresponds to the maximum time a worker is willing to spend incommute. The density of jobs on the flat, featureless plain is some positive number ζ.Equation 11 captures the quantity of jobs accessible to a worker.Ji = pi(Ωτ)2ζ (11)Where Ji is the number of jobs accessible to worker i within time τ and ζ is the density ofjobs.For a static city, in which workers and firms do not move in response to changes in Ω,any increase in Ω increases the number of accessible jobs to the worker. Equation 12captures the partial effect of a change in Ω on job accessibility in a static city. The partialderivative is strictly positive.∂Ji∂Ω= 2piΩτ 2ζ (12)Equation 13 considers the case in which ζ is a function of Ω; wherein local firm densityaround a worker responds to changes in mobility. For example, an increase in Ω mightreduce the density of jobs by encouraging urban sprawl.∂Ji∂Ω= 2piΩτ 2ζ︸ ︷︷ ︸Static Effect > 0+ pi(Ωτ)2∂ζ∂Ω︸ ︷︷ ︸Dynamic Effect < 0(13)In equation 13, the first term of the derivative is identical to the partial derivative inthe static case and takes a positive value. If firms alter locational decisions in response toincreases in Ω such that local job density decreases (i.e. ∂ζ∂Ω< 0) the second term ofEquation 13 will be negative and depending on the magnitude of ∂ζ∂Ω, ∂Ji∂Ωmay be negative.Equation 14 identifies a critical inequality of the partial effect of mobility on sprawl( ∂ζ∂Ω). When Equation 14 holds, a marginal increase in Ω will lead to a decrease in thenumber of accessible jobs.∂ζ∂Ω<−2ζΩ(14)64This toy model illustrates why the effect of mobility (Ω) on accessibility istheoretically ambiguous: if mobility’s impact on locational dispersion is sufficiently large inmagnitude, the net impact of increased mobility will be a decline in accessibility. Equation14 is a codification of the central question of Levine et al. (2012): whether access requiresdensity or speed.Figure 13 decomposes equation 13 into the static and dynamic effects. Although theparameters are in general unobserved, reasonable values can be estimated from data andare used in Figure 13. A sufficiently strong job density effect breaks the equation 14threshold and leads to a negative partial effect of commuter velocity on accessibility.Figure 13: Partial Effect of Increase in Commuter Velocity on Job AccessibilityParameterization: τ = 1 hour, ζ = 56.3 jobskm2, Ω = 31.5 kmhour . τ is set according to the 95th percentile ofcommute duration observed in micro ACS data. ζ is set according to the average density of jobs within a30 km radius of a worker according to LEHD data. Ω is set according to the average CBSA-yearobservation. In this parametrization, if job density falls by more than 3.6 jobs per km2 for every 1km/hincrease in Ω, a rise in Ω reduces job access.The proposed model has notable limitations. The assumption of uniformly distributedjobs is incompatible with the monocentric city model. Modelling ζ as a decreasing functionof distance to the city centre would increase realism, though complicate derivations.Additionally, an exogenously assumed limit to commute time (τ) is restrictive. A morecomplex model might consider τ to be determined endogenously, or as a transport costthat causes the probability of employment to be decreasing in τ .654.6 Identification StrategyInferring a causal relationship between mobility, labour market access and labourmarket outcomes must overcome two major econometric hurdles. Firstly, omitted variablebias is likely to confound analysis. Idiosyncratic characteristics of metros maysimultaneously influence both labour market performance and the travel behaviour ofworkers. For example, very dense cities might be both slow and productive. Several papersare able to explain travel behaviour through observed metro characteristics (see, forexample Cervero and Kockelman (1997); Schwanen and Mokhtarian (2005)). Althoughsome metropolitan characteristics can be directly controlled for, some remain unobserved.The current study overcomes this barrier by leveraging the panel format of the data.Omitted variable bias can be handled by including metropolitan level fixed effects in allregressions, and basing estimation on variation that occurs within metros, across years.Additionally, year fixed effects are included to account for national economic changesacross years. The basic regression model is represented by equation 15.Kmt = β0 + β1Ωmt +Xmt + Fm + Yt + mt (15)Where Kmt is a labour market outcome of interest for metro m in year t, Ωmt is the commutervelocity of metro m in year t, Xmt is a vector of metro-year specific demographic and economiccontrols (mean age, white population share, black population share, Hispanic populationshare, high school completion rate, college completion rate, graduate degree completion rateand log of metropolitan GDP), Fm is a vector of metro dummy variables and Yt is a vectorof year dummy variables.A second barrier to identification arises due to the possibility that labour marketchanges through time and changes in observed mobility are endogenous (Cervero, 2001;Morrison and Lawell, 2016). Labour market growth may generate congestion, which lowersmobility. Conversely, a labour market contraction may free up transportationinfrastructure, which could lessen congestion and increase mobility. The current study seeksto estimate the effect of mobility on labour outcomes, and must therefore remove the effectsof reverse causation. Identification is achieved through the use of instrumental variables.To be valid, the chosen instrument must influence the endogenous variable (mobility),but cannot wield influence on the dependent variable (labour market outcome) other thanthrough mobility. The proposed instrument is lagged political representation on the USHouse of Representatives Transportation and Infrastructure standing committee, anapproach successfully implemented in Knight (2002).A congressperson being appointed to the committee is a result of a political process66which is plausibly orthogonal to future economic conditions of the state, as argued inKnight (2002). Committee members possess outsized power to influence the allocation ofearmarked transportation spending, and have political incentives to direct funding towardsthe state of their constituents. Investment in transportation infrastructure has an intuitiverelationship with commuter velocity. Firstly, building new roadways may lower the circuityof particular routes. Secondly, improving road conditions or capacity may increase groundspeed. Thirdly, new infrastructure may prompt mode switching as commuters shift toimproved modes. The dominance of private vehicle transportation in the US suggests thatmode choice is a secondary issue for all but a few large metros in which public transit is acommon alternative.The US Department of Transportation (DOT) allocates state transportation funds inaccordance with acts of the US Congress. The formula for allocation is complex, andincludes many so-called bonuses to particular states. The political process is such that thefederally controlled DOT has power to allocate money to states, while recipient statessubsequently have significant discretion regarding where these funds are spent within stateboundaries. The complexities of allocation have led to criticisms of the apparent arbitraryand politically capricious nature of allocations “that have little or nothing to do with astate’s transportation needs” (Cooper and Griffith, 2012). Much of the variation in fundallocation between states is a function of state size, in terms of population, geography andeconomic activity. However, year-to-year variation in funding is considerable, and isprimarily the result of political negotiations, which largely take place within theTransportation and Infrastructure Committee.Knight (2002) investigated whether federal highway grants crowd out stateinfrastructure spending, arguing that OLS regressions of state transportation spending onfederal grant amounts will be biased, as federal grants are endogenous to local demand forinfrastructure. Knight (2002) implemented cross-state variations in congressionalcommittee representation as instruments for federal grants. The key identifying assumptionis that while grants may be endogenous to economic conditions, the political happenstancethat leads to particular committee appointments is orthogonal to state economicconditions. Under instrumentation, Knight (2002) found crowd out estimates weresubstantially reduced.For the current analysis, the lagged number of congress members a metro’s “homestate” has appointed to the Congressional Transportation and Infrastructure Committeewill serve as the instrument for commuter velocity. In cases where a CBSA spans multiplestates, the state in which the highest portion of the CBSA population resides is consideredas the home state. Within a lagged sample period (2002-2011), the average metro was67Table 13: Effect of Committee Representation on Home State Allocations, 2005-2015Funding (million $) Log of Funding (million $)(1) (2)Committee Reps 23.03∗∗ .007∗(4.41) (.003)CBSA fixed effects? Y YObs. 445 445Significance levels: ∗ : 5% ∗∗ : 1%. Standard errors shown in parenthesis are clustered at the CBSAlevel. Each regression includes standard controls and fixed effects for CBSA and year.represented by 2.32 “home state” congresspeople on the committee, with a standarddeviation of 1.84 and a range of 0-7.Due to the inclusion of metropolitan fixed effects, the variation preserved in theinstrument represents the year-to-year vagaries in committee representation, independentof persistent state effects. If exogenous variation is in fact coming from arbitraryfluctuations in a state’s political clout over funding legislation there should be ameasurable effect of committee representation on allocations. The amount of federalhighway trust fund money allocated to a metro’s home state is regressed againstcontemporaneous committee representation, with standard fixed effects and controls.Having one additional congressperson on the committee relates to a highly statisticallysignificant increase in federal highway trust fund allocations of $23 million to the homestate (Table 13). Converting the regression to a log-level form yields the estimate that eachadditional congressperson increases allocations to the state by 0.7%.Though it is clear that the instruments must be lagged so that mobility and land useconditions may respond to the shock, it is unclear what magnitude of lag is appropriate.The marginal changes in federal funding resulting from the instrument are likely insufficientto spur entirely new infrastructure projects. Rather, the mechanism proposed is that theadditional funding allows for planned projects to be completed more quickly, or for smallscale projects to be implemented, for example repavings or the addition of new lanes.Table 14 provides first stage results for the IV analysis. The estimated partial effect ofcommittee representation on commuter velocity is estimated for a range of lags. A threeyear lag is statistically significant, while two, four and five year lags are positive but fallshort of significance. The largest effect for congressional representation is a three year lag,for which an increase in one committee member produces a 0.144 km/hr increase in Ω. Totake advantage of the predictive power of each instrument, and to allow flexibility in thedynamics of how political representation affects mobility, multiple lags are usedsimultaneously in the first stage. Table 14, column 5 displays the first stage regression that68will be used in subsequent analysis. The instruments are highly jointly significant. Thejoint F-statistic reported for the first stage regression rejects the null hypothesis that weakinstrument bias is more than 10% of the bias existing in OLS estimates (Stock and Yogo,2005). Using multiple instrument lags simultaneously allows for the implementation of anoveridentification test. The Sargan-Hansen statistic fails to reject instrument validityacross main model specifications. As a robustness check, I repeat all IV analysis using onlythe three year lagged instrument (shown in column 2), results are essentially unchangedunder this alternative specification.Table 14: First Stage Regressions, Predicting Commuter Velocity from Lagged CommitteeRepresentationOutcome: Commuter Velocity(1) (2) (3) (4) (5)Committee Reps: 2 year lag .093 .018(.059) (.057)Committee Reps: 3 year lag .144∗∗ .121∗∗(.055) (.044)Committee Reps: 4 year lag .104 .010(.054) (.057)Committee Reps: 5 year lag .069 .026(.048) (.052)Obs. 3585 3585 3585 3585 3585F-test for Joint Sig. of Instruments 9.41Significance levels: ∗ : 5% ∗∗ : 1%. Standard errors shown in parenthesis are clustered at the CBSAlevel. Each regression includes standard controls and fixed effects for CBSA and year.A significant first stage may at first seem at odds with Duranton and Turner (2011),which found the construction of new road infrastructure in US cities to have no impact onroad congestion. However, the effect of road space on congestion is not being tested in thisregression. Vehicle crowding on road space may remain, while Ω may increase throughchanges in commuter mode choice or lowered route circuity. This highlights an advantageof using commuter velocity (Ω), as the measure can capture all sources of increased urbantravel speed.A potential obstacle to the IV strategy is that additional highway funding toparticular states represents a direct economic contribution through constructionemployment and related local economic activity (Melo et al., 2013), this would underminethe instrument’s exclusion restriction and bias estimation. However, the partial effect ofthe instrument on annual funding ($23 million per congressperson) is small relative to astate economy. Furthermore, annual metropolitan GDP is controlled throughout, whichshould absorb direct stimulus effects. Controlling for GDP could bias estimates of69employment outcomes as it absorbs variation that is highly correlated with labour marketchanges. I rerun employment IV regressions without controlling for GDP and find resultsare robust, with the main point estimate changing by less than 1%.Instrumentation offers the additional important benefit of overcoming attenuationbias. Year-to-year variation in Ω is generally small, and a significant portion of thisvariation is likely measurement error. By relying on predicted Ω through the instruments,rather than observed Ω, the potentially biasing effects of measurement error are removed.Main results will be consistent with instrumentation undoing attenuation bias occurring inOLS regressions.4.7 Results4.7.1 Commuter Velocity as a Cause of SprawlThe ambiguity of mobility’s impact on employment access is a consequence of thepotential for mobility to increase the spatial dispersion of employment relative to workers.One mechanism generating this phenomenon would be the induced demand for driving ifmore highway infrastructure is supplied, as use of a private vehicle can allow largerdistances between work and home. This section will empirically test for a relationshipbetween commuter velocity (Ω) and local job density (ζ). The precise spatial identificationof workers and firms in the LEHD data allows for the local density of jobs to be calculatedfor individual workers, and subsequently averaged across a CBSA’s workforce.Table 15 regresses the density of local jobs (ζ) that surround a metro’s average workerwithin various geodesic radii on commuter velocity (Ω), including CBSA and year fixedeffects as well as CBSA-year demographic controls as detailed in section 4.6. OLSestimation suggests that increased Ω is related to a significant decrease in local job densityacross a range of search radii. Setting the commute radius to 30 km yields the result that a1 km/hr increase in Ω relates to a drop in the density of jobs around the average worker of0.26 jobs/km2. Turning to IV estimation, the magnitude of the estimated effects increasesubstantially. Under instrumentation, the estimated partial effect evaluated for a 30 kmcommute radius is a drop of 3.74 jobs/km2. The average metro in the sample provides itsaverage worker access to 56.3 jobs/km2 within a 30 km radius. The direction of results arerobust to the choice of search radius. The magnitude of results are decreasing in the searchradius, as larger radii lead to the inclusion of comparatively more undeveloped peripheralland. Regarding the relevant search radius, a worker willing to commute no more than 1hour (τ = 1) and whose metro offers the sample average commuter velocity (31 km/hr),will have an implied search radius of 31 km.70The -3.74 jobs/km2 estimate in column 4 is very close to the -3.6 jobs/km2 identifiedas the approximate threshold at which an increase in commuter velocity leads to a decreasein job accessibility (Figure 13). This back of the envelope calculation is suggestive evidencethat the effect of Ω on job access is roughly zero, with the sprawl effects offsetting themobility effects.Table 15: Impact of Commuter Velocity (Ω) on Mean Localised Job Density (ζ)5km 10km 20km 30km 40km 50km 60km(1) (2) (3) (4) (5) (6) (7)Outcome: Mean Localised Job DensityEstimation Method: IVCommuter -12.114 -8.695 -5.724∗ -3.744∗ -2.482∗ -1.723∗ -1.255∗Velocity (7.081) (4.763) (2.671) (1.711) (1.166) (.826) (.610)Estimation Method: OLSCommuter -2.231∗∗ -1.349∗∗ -.527∗∗ -.255∗∗ -.137∗∗ -.080∗ -.051∗Velocity (.332) (.226) (.130) (.081) (.050) (.032) (.022)Obs. 3585 3585 3585 3585 3585 3585 3585Significance levels: ∗ : 5% ∗∗ : 1%. Standard errors shown in parenthesis are clustered at the CBSAlevel. Each regression includes standard controls and fixed effects for CBSA and year.As a robustness check, a second metric for job access is tested. I calculate the averagedistance between a random worker and a random job for each metro. To calculate thismetric, a geodesic distance matrix is constructed for census tract pairs within each CBSA.Within CBSAs in 2014, the average distance between a randomly selected worker and arandomly selected job was 27.4 km. Repeating the IV regression procedure on this outcomevariable indicates a 1 km/hr increase in commuter velocity results in a 0.29 km increase inthe average worker-job distance, significant at the 5% level (Table 16). The correspondingOLS estimate is 0.167 km, significant at the 1% level. These results provide furtherevidence that increases in Ω cause a dispersion of jobs relative to workers.Table 16: Impact of Commuter Velocity (Ω) on Average Distance to a JobIV OLS(1) (2)Commuter Velocity .289∗ .167∗∗(.145) (.046)Obs. 3585 3585Significance levels: ∗ : 5% ∗∗ : 1%. Standard errors shown in parenthesis are clustered at the CBSAlevel. Each regression includes standard controls and fixed effects for CBSA and year.714.7.2 Labour Market Impacts of Commuter MobilityThis section will test for an effect of changing urban mobility on three metropolitanlabour market outcomes: the employment rate, the labour force participation rate, and theaverage working hours per week amongst employed workers. Differential subgroup effectswill be investigated.Success in the labour market is related to the accessibility of jobs. Improvements inmobility could improve labour market outcomes through its effect on easing job search(Zenou, 2009) or causing a reduction in commuting costs (Hu, 2015). The previoussubsection suggested the accessibility benefits of mobility may be counteracted by inducedurban sprawl.Table 17, panel A presents regression results of the effect of commuter velocity on themetropolitan employment rate. Under OLS, the effect of a 1 km/hr increase in Ω is areduction in the employment rate of 0.2 percentage points. For reasons discussed above,OLS may suffer from endogeneity and attenuation bias. Under instrumentation of Ω, theeffect is estimated to be -0.9 percentage points, though the result is statisticallyinsignificant.Adverse labour market outcomes across a metropolitan area are consistent withspatial and transportation mismatch literature which suggests that the expansion of highmobility policy leads to a decline in the ability of some workers to access and secureemployment. Prior research suggests the groups most likely to suffer adverse labour marketoutcomes include nonwhite workers, youth, low skill workers, and workers without access toa private vehicle (Ihlanfeldt and Sjoquist, 1990; Kain, 1968; Kawabata and Shen, 2007;Tyndall, 2017; O’Regan and Quigley, 1998).Panel A, columns 2-6 estimate employment effects by demographic subgroup. Theestimated partial effect of Ω on the employment rate of the non-college educated, and youthare lower than the aggregate estimate. I find that the negative effects of commuter mobilityon white workers, the college educated and workers with access to a private vehicle are allhigher than the aggregate estimate. In fact, I find the employment rate of those with acollege education increases by 0.9 percentage points per unit increase in Ω. Youth workersstand out for having particularly adverse reactions to increases in mobility conditions.Table 17, panel B shows the results for labour force participation. An IV regression forthe full sample suggests a statistically insignificant 0.5 percentage point increase in labourforce participation per km/hr increase in Ω. However, division of the labour force intosubgroups reveals interesting heterogeneity. White workers, the college educated and thosewith access to a private vehicle experience larger increases in labour force participationthan the aggregate estimate. Noncollege educated workers and youth are more adversely72Table 17: Impact of Commuter Velocity (Ω) on Labour Market OutcomesAll White College No Car YouthCollege(1) (2) (3) (4) (5) (6)A. Outcome: Employment RateEstimation Method: IVCommuter Velocity -.009 -.005 .009 -.013 -.004 -.045∗(.006) (.005) (.006) (.007) (.004) (.020)Estimation Method: OLSCommuter Velocity -.002∗∗ -.002∗∗ -.001∗ -.002∗∗ -.001∗∗ -.002(.0003) (.0004) (.0005) (.0003) (.0002) (.001)Employment 0.560 0.567 0.701 0.521 0.593 0.515Rate (mean)B. Outcome: Labour Force Participation RateEstimation Method: IVCommuter Velocity .005 .007 .015∗ .003 .010∗ -.022(.005) (.005) (.008) (.005) (.005) (.012)Estimation Method: OLSCommuter Velocity -.001∗∗ -.001∗∗ -.001 -.001∗∗ -.0004∗ -.001(.0002) (.0002) (.0004) (.0003) (.0002) (.0005)Labour Force 0.605 0.608 0.724 0.573 0.638 0.611Partic. (mean)C. Outcome: Weekly Hours WorkedEstimation Method: IVCommuter Velocity -.375∗ -.392∗ -.194 -.451∗ -.263 -1.171∗(.187) (.191) (.180) (.217) (.146) (.530)Estimation Method: OLSCommuter Velocity -.043∗∗ -.053∗∗ -.008 -.054∗∗ -.039∗∗ -.050∗∗(.009) (.010) (.011) (.010) (.006) (.019)Hours Worked 37.99 38.16 40.63 36.99 38.30 30.63(mean)Obs. 3585 3585 3585 3585 3585 3585Significance levels: ∗ : 5% ∗∗ : 1%. Standard errors shown in parenthesis are clustered at the CBSAlevel. Each regression includes standard controls and fixed effects for CBSA and year.affected. Youth cohorts actually experience a decrease in labour force participation as aresult of rising commuter velocity.Examining the effect of commuter velocity on hours worked amongst employedworkers allows for the estimation of mobility’s effect on the intensive margin of laboursupply. Variable spatial access to place of work may impact the number of hours suppliedper week, through the effect on commuting costs. Table 17, panel C provides estimatedeffects. Amongst the full sample IV regression (column 1), a 1 km/hr increase in Ωcorresponds to a significant reduction in labour supplied of 23 minutes per week.Examination by subgroup reveals familiar patterns. Youth are once again73disproportionately negatively affected. The large effect amongst youth may be reflective ofa high share of youth holding part time jobs with adjustable hours. A pronounced effect onyouth is consistent with the findings of O’Regan and Quigley (1991, 1996, 1998).90.1% of ACS respondents have access to a private vehicle. Those with a privatevehicle fare better in response to an increase in mobility. Findings are suggestive of anoverarching mechanism: expansion of high mobility policy is harmful to individuals whoare not able to fully participate in the upside of high mobility infrastructure, while beingexposed to the downside burden of employment sprawl.4.8 ConclusionUS cities differ widely in how quickly workers can move through urban space. Theimpact of mobility on employment accessibility can be considered through two channels.First, rapid movement extends the area that is accessible to commuters, improving access.Second, mobility enables sprawl, which reduces the spatial density of jobs. This paperprovides a direct estimate of the causal effect of mobility on sprawl and metropolitanlabour market outcomes, unique to prior literature. I present significant evidence linkingincreased mobility to urban sprawl. I find no evidence that mobility increases overalllabour market outcomes, but I do find evidence that increased mobility exacerbatesinequalities in labour market outcomes, with downside effects concentrated amongstnonwhite workers, less educated workers and youth. Policy and infrastructure investmentthat seek to aid labour markets through reducing vehicle congestion or otherwise increasingthe mobility of commuters may be self defeating due to endogenous sprawl.Spatial and transportation mismatch literature provide viable theory to understandwhy high urban mobility may bring about negative job market consequences for particularpopulations. Future research should investigate heterogeneous effects more deeply throughuse of individual level micro data.The near future prospect of autonomous vehicles may enable a further increase incommuter mobility. Compounding the effect, the mobility benefits of autonomous vehiclesmay initially accrue to an affluent minority while resultant urban spatial dispersion wouldaffect the wider population. Understanding how increases in mobility generate urbansprawl is therefore likely a topic of continued interest.745 ConclusionThe amount of time people spend commuting per day has been remarkably constantover human history, a phenomenon known as Marchetti’s Constant. Ancient Greekvillagers and modern American urbanites both travel roughly one hour per day in commuteon average, despite remarkable differences in transportation technologies (Marchetti, 1994).This is possible because improvements in travel speed can translate into increases in thedistance between people’s homes and work places, rather than time savings. Thismechanism is salient for much of the above work, and is a vital consideration in evaluatingtransportation policy and infrastructure. When it is assumed that the location decisions ofworkers and firms are exogenous to transportation investments, evaluating infrastructure’simpact appears deceivingly simple. Transportation infrastructure is built in areas withparticular economic characteristics. Furthermore, worker location decisions are highlyendogenous to the provided transportation networks. The evaluation of transportationsystems must be cognisant of these sources of confounding endogeneity.This thesis has provided some new approaches to dealing with empirical identificationissues. The above papers have suggested several sources of random variation to estimatetransportation infrastructure’s impact, including a natural disaster (Chapter 2), regionalplanning constraints (Chapter 3) and randomness in the political process that fundsprojects (Chapter 4). Results suggest transportation amenities improve average economicoutcomes in the local area (Chapters 2 and 3). Understanding how these neighbourhoodeffects translate into a distribution of welfare effects is complicated by the endogenoussorting of workers. Chapter 3 develops structural estimation methodology that is able toaccount for the presence of endogenous worker choices. The application of structuralapproaches to evaluating spatial urban policies provides a promising direction for futureacademic work in the field of urban economics.Despite enormous spending and policy concern directed towards urban transportation,policy makers rarely consider the general equilibrium consequences of transportationinvestments. Failure to account for general equilibrium effects will result in projects thatfail to direct benefits to the intended groups (for example due to displacement) or fail toachieve the intended accessibility benefits (for example due to sprawl). Increasing thesophistication of analysis used to evaluate transportation projects could significantlyimprove the efficacy of investment.75BibliographyAhlfeldt, G. M., Redding, S. J., Sturm, D. M., and Wolf, N. (2015). The economics ofdensity: Evidence from the berlin wall. Econometrica, 83(6):2127–2189.Alonso, W. et al. (1964). Location and land use. toward a general theory of land rent.Location and land use. Toward a general theory of land rent.Altshuler, A. A. and Luberoff, D. E. (2004). Mega-projects: The changing politics of urbanpublic investment. Brookings Institution Press.American Automobile Association (2007). Your driving costs.Anas, A. (2011). Discovering the efficiency of urban sprawl. In Handbook of UrbanEconomics and Planning, pages 123–149.Anas, A. (2015). Why are urban travel times so stable? Journal of Regional Science,55(2):230–261.Andersson, F., Haltiwanger, J. C., Kutzbach, M. J., Pollakowski, H. O., and Weinberg,D. H. (2018). Job displacement and the duration of joblessness: The role of spatialmismatch. Review of Economics and Statistics, 100(2):203–218.Angel, S. and Hyman, G. M. (1970). Urban velocity fields. Environment and Planning A,2(2):211–224.Angrist, J. D. (2001). Estimation of limited dependent variable models with dummyendogenous regressors. Journal of business & economic statistics, 19(1).Angrist, J. D. and Pischke, J.-S. (2008). Mostly harmless econometrics: An empiricist’scompanion. Princeton university press.Atkinson-Palombo, C. (2010). Comparing the capitalisation benefits of light-rail transitand overlay zoning for single-family houses and condos by neighbourhood type inmetropolitan phoenix, arizona. Urban Studies, 47(11):2409–2426.Baum, C. L. (2009). The effects of vehicle ownership on employment. Journal of UrbanEconomics, 66(3):151–163.Baum-Snow, N. (2007). Did highways cause suburbanization? The Quarterly Journal ofEconomics, pages 775–805.76Baum-Snow, N. and Kahn, M. E. (2000). The effects of new public projects to expandurban rail transit. Journal of Public Economics, 77(2):241–263.Bay Area Rapid Transit District (1987). 1987 five year plan, volume 1, operations andservice plan.Bayer, P., Ferreira, F., and McMillan, R. (2007). A unified framework for measuringpreferences for schools and neighborhoods. Journal of political economy, 115(4):588–638.Bayer, P. and McMillan, R. (2012). Tiebout sorting and neighborhood stratification.Journal of Public Economics, 96(11-12):1129–1143.Bayer, P., McMillan, R., and Rueben, K. (2004). An equilibrium model of sorting in anurban housing market. Technical report, National Bureau of Economic Research.Bayer, P. and Timmins, C. (2005). On the equilibrium properties of locational sortingmodels. Journal of Urban Economics, 57(3):462–477.Boarnet, M. G. (1997). Infrastructure services and the productivity of public capital: thecase of streets and highways. National Tax Journal, pages 39–57.Bogin, A. N., Doerner, W. M., Larson, W. D., et al. (2016). Missing the mark: House priceindex accuracy and mortgage credit modeling. Technical report.Bollinger, C. R. and Ihlanfeldt, K. R. (1997). The impact of rapid rail transit on economicdevelopment: the case of atlanta’s marta. Journal of Urban Economics, 42(2):179–204.Bollinger, C. R. and Ihlanfeldt, K. R. (2003). The intraurban spatial distribution ofemployment: which government interventions make a difference? Journal of UrbanEconomics, 53(3):396–412.Brueckner, J. K. and Zenou, Y. (2003). Space and unemployment: The labor-marketeffects of spatial mismatch. Journal of Labor Economics, 21(1):242–262.Cao, X. J. and Schoner, J. (2014). The influence of light rail transit on transit use: Anexploration of station area residents along the hiawatha line in minneapolis.Transportation Research Part A: Policy and Practice, 59:134–143.Cervero, R. (1997). Paradigm shift: from automobility to accessibility planning. UrbanFutures (Canberra), (22):9.Cervero, R. (2001). Efficient urbanisation: economic performance and the shape of themetropolis. Urban Studies, 38(10):1651–1671.77Cervero, R. and Kockelman, K. (1997). Travel demand and the 3Ds: density, diversity, anddesign. Transportation Research Part D: Transport and Environment, 2(3):199–219.Clark, G. (2014). The son also rises: Surnames and the history of social mobility,volume 49. Princeton University Press.Cooper, D. and Griffith, J. (2012). Highway robbery: How congress put politics before needin federal highway and transit funding. Technical report, Center for American Progress.Coulson, N. E., Laing, D., and Wang, P. (2001). Spatial mismatch in search equilibrium.Journal of Labor Economics, 19(4):949–972.Couture, V., Duranton, G., and Turner, M. A. (2018). Speed. Review of Economics andStatistics, 100(4):725–739.Couture, V. and Handbury, J. (2017). Urban revival in america, 2000 to 2010. Technicalreport, National Bureau of Economic Research.Crane, R. (2000). The influence of urban form on travel: an interpretive review. Journal ofPlanning Literature, 15(1):3–23.Credit, K. (2018). Transit-oriented economic development: The impact of light rail on newbusiness starts in the Phoenix, Az region, USA. Urban Studies, 55(13):2838–2862.Davis, M. A. and Ortalo-Magne´, F. (2011). Household expenditures, wages, rents. Reviewof Economic Dynamics, 14(2):248–261.Duranton, G. and Turner, M. A. (2011). The fundamental law of road congestion:Evidence from US cities. The American Economic Review, pages 2616–2652.Elliott, J. R. (2001). Referral hiring and ethnically homogeneous jobs: How prevalent is theconnection and for whom? Social Science Research, 30(3):401–425.Epple, D. and Sieg, H. (1999). Estimating equilibrium models of local jurisdictions.Journal of political economy, 107(4):645–681.Farley, J. E. (1982). Black male unemployment in u.s. metropolitan areas: The role of blackcentral city segregation and job decentralization. Journal of Urban Affairs, 4(3):19–34.Ferreyra, M. M. (2007). Estimating the effects of private school vouchers in multidistricteconomies. The American Economic Review, 97(3):789–817.78Fujita, M. and Ogawa, H. (1982). Multiple equilibria and structural transition ofnon-monocentric urban configurations. Regional science and urban economics,12(2):161–196.Giacomin, D. J. and Levinson, D. M. (2015). Road network circuity in metropolitan areas.Environment and Planning B: Planning and Design, 42.Giuliano, G. and Dargay, J. (2006). Car ownership, travel and land use: a comparison ofthe US and Great Britain. Transportation Research Part A: Policy and Practice,40(2):106–124.Glaeser, E. L., Kahn, M. E., and Rappaport, J. (2008). Why do the poor live in cities? therole of public transportation. Journal of urban Economics, 63(1):1–24.Gobillon, L., Selod, H., and Zenou, Y. (2007). The mechanisms of spatial mismatch. Urbanstudies, 44(12):2401–2427.Gomez-Ibanez, J. A. (1996). Big-city transit rider snip, deficits, and politics: Avoidingreality in boston. Journal of the American Planning Association, 62(1):30–50.Gordon, P., Kumar, A., and Richardson, H. W. (1989). The spatial mismatch hypothesis:Some new evidence. Urban Studies, 26(3):315–326.Graham, M. R., Kutzbach, M. J., McKenzie, B., et al. (2014). Design comparison of lodesand acs commuting data products. Technical report.Grengs, J., Levine, J., Shen, Q., and Shen, Q. (2010). Intermetropolitan comparison oftransportation accessibility: Sorting out mobility and proximity in San Francisco andWashington, DC. Journal of Planning Education and Research, 29(4):427–443.Ham, J. C., Swenson, C., I˙mrohorog˘lu, A., and Song, H. (2011). Government programs canimprove local labor markets: Evidence from state enterprise zones, federal empowermentzones and federal enterprise community. Journal of Public Economics, 95(7-8):779–797.Hanson, A. (2009). Local employment, poverty, and property value effects ofgeographically-targeted tax incentives: an instrumental variables approach. RegionalScience and Urban Economics, 39(6):721–731.Harrison, B. (1972). The intrametropolitan distribution of minority economic welfare.Journal of Regional Science, 12(1):23–43.79Holzer, H. J. (1991). The spatial mismatch hypothesis: What has the evidence shown?Urban Studies, 28(1):105–122.Holzer, H. J., Quigley, J. M., and Raphael, S. (2003). Public transit and the spatialdistribution of minority employment: Evidence from a natural experiment. Journal ofPolicy Analysis and Management, 22(3):415–441.Hu, L. (2015). Job accessibility of the poor in Los Angeles. Journal of the AmericanPlanning Association, 81(1).Hughes, M. A. and Madden, J. F. (1991). Residential segregation and the economic statusof black workers: New evidence for an old debate. Journal of Urban Economics,29(1):28–49.Ihlanfeldt, K. R. and Sjoquist, D. L. (1990). Job accessibility and racial differences inyouth employment rates. The American Economic Review, 80(1):267–276.Ihlanfeldt, K. R. and Sjoquist, D. L. (1998). The spatial mismatch hypothesis: a review ofrecent studies and their implications for welfare reform. Housing policy debate,9(4):849–892.Imbens, G. W. and Kolesar, M. (2016). Robust standard errors in small samples: Somepractical advice. Review of Economics and Statistics, 98(4):701–712.Immergluck, D. (1998). Job proximity and the urban employment problem: do suitablenearby jobs improve neighbourhood employment rates? Urban Studies, 35(1):7–23.Jacobs, J. (1961). The death and life of great american cities. New-York, NY: Vintage.Kahn, M. E. (2007). Gentrification trends in new transit-oriented communities: Evidencefrom 14 cities that expanded and built rail transit systems. Real Estate Economics,35(2):155–182.Kain, J. F. (1968). Housing segregation, negro employment, and metropolitandecentralization. The Quarterly Journal of Economics, 82(2):175–197.Kain, J. F. (1992). The spatial mismatch hypothesis: Three decades later. Housing PolicyDebate, 3(2):371–460.Kasarda, J. D. (1989). Urban industrial transition and the underclass. Annals of theAmerican Academy of Political and Social Science, 501:26–47.80Kawabata, M. (2003). Job access and employment among low-skilled autoless workers in usmetropolitan areas. Environment and Planning A, 35(9):1651–1668.Kawabata, M. and Shen, Q. (2007). Commuting inequality between cars and public transit:The case of the San Francisco bay area, 1990-2000. Urban Studies, 44(9):1759–1780.Kline, P. (2010). Place based policies, heterogeneity, and agglomeration. AmericanEconomic Review, 100(2):383–87.Kline, P. and Moretti, E. (2014). People, places, and public policy: Some simple welfareeconomics of local economic development programs. Annu. Rev. Econ., 6(1):629–662.Knight, B. (2002). Endogenous federal grants and crowd-out of state government spending:Theory and evidence from the federal highway aid program. The American EconomicReview, 92(1):71–92.Knight, R. L. and Trygg, L. L. (1977). Evidence of land use impacts of rapid transitsystems. Transportation, 6(3):231–247.Kou, W., Chen, X., Yu, L., Qi, Y., and Wang, Y. (2017). Urban commuters’ valuation oftravel time reliability based on stated preference survey: A case study of beijing.Transportation Research Part A: Policy and Practice, 95:372–380.Krizek, K. J. and El-Geneidy, A. (2007). Segmenting preferences and habits of transit usersand non-users. Journal of public transportation, 10(3):5.LeRoy, S. F. and Sonstelie, J. (1983). Paradise lost and regained: Transportationinnovation, income, and residential location. Journal of Urban Economics, 13(1):67–89.Levine, J., Grengs, J., Shen, Q., and Shen, Q. (2012). Does accessibility require density orspeed? A comparison of fast versus close in getting where you want to go in USmetropolitan regions. Journal of the American Planning Association, 78(2):157–172.Levinson, D. and El-Geneidy, A. (2009). The minimum circuity frontier and the journey towork. Regional Science and Urban Economics, 39(6):732–738.Lindley, J. A. (1987). Urban freeway congestion: quantification of the problem andeffectiveness of potential solutions. ITE journal, 57(1):27–32.Marchetti, C. (1994). Anthropological invariants in travel behavior. Technologicalforecasting and social change, 47(1):75–88.81McKenzie, B. and Rapino, M. (2011). Commuting in the United States: 2009. USDepartment of Commerce, Economics and Statistics Administration, US Census Bureau.Melo, P. C., Graham, D. J., and Brage-Ardao, R. (2013). The productivity of transportinfrastructure investment: A meta-analysis of empirical evidence. Regional Science andUrban Economics, 43(5):695–706.Meyer, J. R. (1959). The economics of competition in the transportation industries.Number 107. Harvard University Press.Mills, E. S. (1967). An aggregative model of resource allocation in a metropolitan area.The American Economic Review, 57(2):197–210.Morrison, G. M. and Lawell, C.-Y. C. L. (2016). Does employment growth increase traveltime to work?: An empirical analysis using military troop movements. Regional Scienceand Urban Economics, 60:180–197.Muth, R. F. (1969). Cities and housing; the spatial pattern of urban residential land use.Nelson, P., Baglino, A., Harrington, W., Safirova, E., Lipman, A., et al. (2007). Transit inwashington, dc: Current benefits and optimal level of provision. Journal of urbanEconomics, 62(2):231–251.Neumark, D. and Kolko, J. (2010). Do enterprise zones create jobs? evidence fromCalifornia’s enterprise zone program. Journal of Urban Economics, 68(1):1–19.Ong, P. M. and Miller, D. (2005). Spatial and transportation mismatch in los angeles.Journal of Planning Education and Research, 25(1):43–56.O’Regan, K. M. (1993). The effect of social networks and concentrated poverty on blackand hispanic youth unemployment. The annals of regional Science, 27(4):327–342.O’Regan, K. M. and Quigley, J. M. (1991). Labor market access and labor marketoutcomes for urban youth. Regional Science and Urban Economics, 21(2):277–293.O’Regan, K. M. and Quigley, J. M. (1996). Teenage employment and the spatial isolationof minority and poverty households. The Journal of Human Resources, 31(3):692–702.O’Regan, K. M. and Quigley, J. M. (1998). Where youth live: economic effects of urbanspace on employment prospects. Urban Studies, 35(7):1187–1205.Prud’homme, R. and Lee, C.-W. (1999). Size, sprawl, speed and the efficiency of cities.Urban Studies, 36(11):1849–1858.82Raphael, S. (1998). The spatial mismatch hypothesis and black youth joblessness: evidencefrom the san francisco bay area. Journal of Urban Economics, 43(1):79–111.Raphael, S., Stoll, M. A., Small, K. A., and Winston, C. (2001). Can boosting minoritycar-ownership rates narrow inter-racial employment gaps?[with comments].Brookings-Wharton Papers on Urban Affairs, pages 99–145.Rogers, C. L. (1997). Job search and unemployment duration: Implications for the spatialmismatch hypothesis. Journal of urban Economics, 42(1):109–132.Ruggles, S., Alexander, J. T., Genadek, K., Goeken, R., Schroeder, M. B., and Sobek, M.(2014). Integrated public use microdata series: Version 5.0 [machine-readable database].Minneapolis: University of Minnesota.Sanchez, T. W. (1999). The connection between public transit and employment: The casesof Portland and Atlanta. Journal of the American Planning Association, 65(3):284–296.Sanchez, T. W., Shen, Q., and Peng, Z.-R. (2004). Transit mobility, jobs access andlow-income labour participation in us metropolitan areas. Urban Studies,41(7):1313–1331.Schrank, D., Eisele, B., Lomax, T., and Bak, J. (2015). 2015 urban mobility scorecard.Schwanen, T. and Mokhtarian, P. L. (2005). What affects commute mode choice:neighborhood physical structure or preferences toward neighborhoods? Journal ofTransport Geography, 13(1):83–99.Severen, C. (2018). Commuting, labor, and housing market effects of mass transportation:Welfare and identification.Sieg, H., Smith, V. K., Banzhaf, H. S., and Walsh, R. (2004). Estimating the generalequilibrium benefits of large changes in spatially delineated public goods. InternationalEconomic Review, 45(4):1047–1077.Small, K. A., Winston, C., and Yan, J. (2005). Uncovering the distribution of motorists’preferences for travel time and reliability. Econometrica, 73(4):1367–1382.Smith, T. E. and Zenou, Y. (2003). Spatial mismatch, search effort, and urban spatialstructure. Journal of Urban Economics, 54(1):129–156.Sound Transit (1999). Central Link Light Rail Transit Project, Seattle, Tukwila andSeatac: Environmental Impact Statement, Volume 1. pages 4–21.83Stock, J. H. and Yogo, M. (2005). Testing for weak instruments in linear iv regression.Identification and Inference for Econometric Models: Essays in Honor of ThomasRothenberg, page 80.Stoll, M. A. (1999). Spatial job search, spatial mismatch, and the employment and wagesof racial and ethnic groups in los angeles. Journal of Urban Economics, 46(1):129–155.Stubbs, J. and Jegede, F. (1998). The integration of rail and air transport in britain.Journal of transport Geography, 6(1):53–67.Taylor, B. D. and Ong, P. M. (1995). Spatial mismatch or automobile mismatch? anexamination of race, residence and commuting in us metropolitan areas. Urban studies,32(9):1453–1473.Tiebout, C. M. (1956). A pure theory of local expenditures. Journal of political economy,64(5):416–424.Tsivanidis, N. (2018). The aggregate and distributional effects of urban transitinfrastructure: Evidence from bogota’s transmilenio. Working Paper, University ofChicago Booth School of Business.Tyndall, J. (2017). Waiting for the r train: Public transportation and employment. UrbanStudies, 54(2):520–537.Tyndall, J. (2018). Bus quality improvements and local commuter mode share.Transportation Research Part A: Policy and Practice, 113:173–183.Walters, A. A. (1961). The theory and measurement of private and social cost of highwaycongestion. Econometrica: Journal of the Econometric Society, pages 676–699.Wasmer, E. and Zenou, Y. (2002). Does city structure affect job search and welfare?Journal of urban Economics, 51(3):515–541.Wasmer, E. and Zenou, Y. (2006). Equilibrium search unemployment with explicit spatialfrictions. Labour Economics, 13(2):143–165.Widmer, J.-P. and Hidber, C. (2000). Effects of rail stations at airports in europe.Transportation Research Record: Journal of the Transportation Research Board,(1703):90–97.Wilson, W. J. (2011). When work disappears: The world of the new urban poor. Vintage.84Yang, J. (2008). Policy implications of excess commuting: Examining the impacts ofchanges in US metropolitan spatial structure. Urban Studies, 45(2):391–405.Zax, J. S. and Kain, J. F. (1996). Moving to the suburbs: do relocating companies leavetheir black employees behind? Journal of Labor Economics, 14(3):472–504.Zenou, Y. (2009). Search in cities. European Economic Review, 53(6):607–624.85AppendicesA Google API DataA1. Distribution of Google API Generated Trip TimesFigure A1 displays the distribution of travel times for both driving and public transitcommuting for the full matrix of home and work locations. Across the four metros, 66% ofpublic transit commutes take over 90 minutes and 43% take over 2 hours. Only 0.6% ofdrive times exceed 90 minutes.86A2. Correlation Between Google Travel Times and Geodesic DistanceDriving Times Public Transit TimesEach dot represents one potential commute. Constructing the full commuting matricesrequired extensive data collection. A more easily constructed alternative to the Google APItrip data would be to use a matrix of straight line travel distances and assume that thesecorrelate with actual trip times. Analysis reveals that straight line distances may be areasonable proxy for drive times, but are a poor proxy for public transit durations. Thecircuity of a public transit commute is often high, as transit infrastructure funnelstravellers along indirect routes. Across the 944,085 origin-destination pairs that areconnected through public transit, straight line distance can explain 89% of the variation inprivate vehicle trip duration, but only 38% of the variation in public transit trip duration.87B Structural Results - Change in Potential IncomeFigure 14: Structural Results, Change in Potential Income (Skill)A. Minneapolis B. PortlandC. Salt Lake City D. Seattle< −4% (−4%,−2%) (−2%, 0%) + (0%, 2%) + (2%, 4%) + > 4%- New rail station - Pre 2000 rail station(“+” symbols are redundant with colours, but enable interpretation of figures if viewed in grayscale.)Local LRT stations generate a rise in the average job market abilities of residents, as employedresidents sort towards neighbourhoods with new stations. Potential income is the averagewage a worker would expect to receive in the labour market.88C Policy Extension: Local Transit Pass RequirementIn this appendix I propose a novel policy instrument to increase the societal benefits ofLRT. When proposing LRT projects, policy makers normally state the joint objectives ofimproving labour market outcomes and increasing public transit use. As shown in thispaper, LRT induced neighbourhood demand may crowd out the population most likely touse transit for commuting, undercutting policy goals.Consider the following program: the local government imposes a levy on every residentliving in a tract that received a new LRT station. The levy is set to be exactly the cost of alocal transit pass (t). In exchange, the local government provides a transit pass to everyresident who pays the levy. The program essentially mandates the purchase of a transitpass within the neighbourhoods gaining a LRT station. The policy causes LRTneighbourhoods to become less desirable for workers who wish to live in a LRTneighbourhood but commute by private vehicle. Such workers are charged for a transitpass that they do not use.I rerun the model with the preference parameters identified in the originalspecification, but apply the new policy in addition to LRT. Figure 15.A displays thecombined effect of the policies on public transit mode share. While LRT alone increasedtransit commuting in the metro by .43 percentage points, LRT combined with themandatory transit pass program increases the share of commuters using transit by .56percentage points.The policy is also effective at reducing the negative aggregate employment effects ofLRT. Figure 15.B displays employment effects. Compared to LRT alone, the addition of thelocal mandatory transit pass program cuts the negative employment effect roughly in half.The program provides money to the government in two ways. First, a majority ofresidents in the tracts gaining LRT continue to drive to work, consistent with low overalltransit uptake in the metros. These workers pay for transit passes but do not use transitfor commuting, transferring revenue to the government. Second, more employed workersmeans that fewer workers require government transfers. In the model, transfers come fromoutside of the metropolitan economy, which is consistent with the reality that social welfareprograms are mainly supplied by state and federal governments. However, depending onthe relationships between levels of government, these savings may also be salient.If the money raised from unused passes is rebated to local residents through a uniformtransfer, welfare effects are approximately the same as without the mandatory transitprogram (+.16%). If the money saved from reduced transfers to the jobless is also rebatedthrough a uniform transfer, the welfare gains are higher (+.18%). When all benefits areconsidered, the mandatory transit pass program improves the welfare benefits of LRT.89Figure 15: Transit Pass Structural Results, Distribution Across Potential Income Per-centilesA. Public Transit Mode Share B. Employment RateAverage Effects Average EffectsLRT: +0.43 pp LRT: −0.10 ppLRT and Transit Pass Requirement: +0.56 pp LRT and Transit Pass Requirement: −0.05 ppResults are scaled to represent the effect of ten LRT stations in a metro of average size (1.25 millionworkers).The program is effective because it reduces the appeal of neighbourhoods gaining LRTamong rich car users, as they do not stand to gain from the provided transit passes. Themechanism preserves accessible housing for those who do use transit.90D Policy Extension: “Bus Rapid Transit”Many cities have begun implementing Bus Rapid Transit (BRT) to improve urbantransportation. BRT aims to provide mobility benefits similar to LRT, but rather thanusing LRT infrastructure, implements the program through buses. BRT service includesseveral improvements over traditional bus service, for example, more frequent buses,dedicated bus lanes, traffic signal priority, greater distances between stops, and ticketingimprovements.As BRT relies on traditional buses, it may lack appeal among higher income residents.Additionally, as BRT does not require large, permanent capital investment, it is unlikely tohold the same economic development properties of LRT. In this appendix I model a BRTsystem by using a blunt assumption: that BRT provides mobility improvements identicalto LRT, but does not lead to any local consumption benefits. I estimate the paper’s mainstructural model but “turn off” the local consumption multiplier (ρj = 1). Local rents maystill rise on account of BRT, but only through the capitalization of commuting mobilitybenefits.I find that, unlike LRT, BRT leads to an increase in aggregate metropolitanemployment. While LRT reduced aggregate employment by 0.10 percentage points (per 10stations), the BRT treatment raises aggregate employment by 0.11 percentage points(Figure 16.B). BRT does not raise local property values to the extent LRT does, andtherefore lower earning, transit dependent populations are not repelled from the area to thesame extent. I find that the BRT treatment leads to a slightly larger rise in the share ofthe metropolitan population using transit for commuting. BRT raises overall transit use by.52 percentage points while LRT only raised transit use by .43 percentage points. Lowerearning workers are more likely to switch from private vehicles to transit under BRTcompared to LRT, while the highest earning workers are slightly less likely to switch totransit (Figure 16.A). In terms of welfare effects, I find that while 10 LRT stationsincreased the average individual’s welfare by 0.16%, 10 BRT stations increased welfare byan average of only 0.05%. However, BRT systems are a fraction of the cost of LRTsystems, suggesting the benefit per dollar of the BRT system may be higher than LRT.Workers at the 90th percentile of the income distribution receive roughly zero welfareimprovement under BRT, while they enjoyed substantial benefits under LRT.91Figure 16: BRT Structural Results, Distribution Across Potential Income PercentilesA. Public Transit Mode Share B. Employment RateAverage effects Average effectsLRT: +0.43 pp LRT: −0.10 ppBRT: +0.52 pp BRT: +0.11 ppResults are scaled to represent the effect of ten LRT (or BRT) stations in a metro of average size (1.25million workers).92

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
https://iiif.library.ubc.ca/presentation/dsp.24.1-0379885/manifest

Comment

Related Items