Light Rail Transit, Local Labour Market Outcomes and Household Sorting [early draft]

Transcription

1 Light Rail Transit, Local Labour Market Outcomes and Household Sorting [early draft] Justin Tyndall March 16, 2018 Sauder School of Business, University of British Columbia Abstract High urban transportation costs are often considered as a potential barrier for workers seeking employment. Light rail transit has become a common policy intervention to enhance access to job opportunities and revive lagging urban areas. The expansion of transit may improve labour market outcomes through reduced commuting costs. Identifying the effect of transit on individual labour market outcomes is hindered by the possibility of endogenous household sorting. The effect of light rail is analysed for four US metros between 2000 and Analysis is restricted to neighbourhoods that may have received light rail exogenously, independent of local economic conditions. Results of a neighbourhood level propensity score weighting exercise show that light rail caused both a significant rise in public transit mode share among commuters and improved local labour market outcomes. A decomposition exercise reveals that induced demographic shifts cannot explain the transportation mode switching effect or the local labour market gains. A structural model is subsequently proposed to explicitly disentangle the effect of sorting from the potential individual causal effects of increased job access. Models are estimated with detailed commuter flow information combined with novel Google navigation data. Results suggest that light rail transit increases transit commuting, expands labour force participation and reduces unemployment among local residents. For presentation at the 59th Annual Transportation Research Forum

2 1 Introduction Ameliorating poor labour market outcomes in US cities is a perennial topic of acute policy interest. Labour market outcomes display strong spatial heterogeneity in US cities. Spatial mismatch theory proffers a cause for concentrated unemployment and poverty, suggesting that high unemployment neighbourhoods are limited by their spatial access to employment opportunities (Kain, 1968). The expansion of access to labour market opportunities through improved transportation systems is a natural policy response to observed pockets of low employment. In recent years, the introduction of light rail transit (LRT) has been adopted by many US cities as a means to revitalize urban areas. The premise that LRT connects workers to firms is a common contention. Consider the following two examples. The original environmental impact assessment for Seattle s LRT system advised: One result of the improved access and improved travel times is improved access to employment opportunities (Sound Transit, 1999). Similarly a chief political advocate for LRT in Minneapolis proffered: We re trying to reconnect people, particularly people with high levels of unemployment, to the job market of today. 1 This study will search for evidence that LRT improves labour market connections. The first contribution of this paper will be to employ panel census tract data to estimate the neighbourhood change effects of LRT stations. The rapid proliferation of LRT construction and proposals in US cities inspire an urgency for such estimates. Analysis reveals that LRT stations lead to a significant rise in the share of local commuters using public transit. Concurrently, LRT is found to cause substantial improvements in local labour market outcomes. The quasi-random introduction of LRT transit systems in four US cities over the period is used to estimate causal neighbourhood effects. Estimating unbiased 1 Quotes by Peter McLaughlin, Hennepin County Commissioner, Minnesota, from The Train Line That Brought the Twin Cities Back Together, by E. Trickey, Politico Magazine, March 16,

3 neighbourhood effects provides limited inference on individual level effects due to the fact that households may endogenously relocate. The second contribution of this paper will be to disentangle the two elements that potentially contribute to LRT s positive effect on neighbourhood outcomes, namely: 1. The beneficial expansion of an individual s accessible labour market opportunities (overcoming spatial mismatch). 2. Endogenous neighbourhood sorting in response to local transit infrastructure. A Blinder-Oaxaca style decomposition exercise will be undertaken to estimate the share of neighbourhood effects explainable by observable demographic sorting (Blinder, 1973; Oaxaca, 1973). LRT induced sorting cannot explain the rise in public transit use, and only explains a small share of the improved local labour outcomes. This paper proposes a model of spatial sorting among workers to provide inference on the distribution of welfare effects. Locational choices are reconciled with labour market activities and spatial access through transportation amenities. The study will proceed as follows. Section 2 reviews relevant literature on spatial mismatch and LRT. Section 3 proposes an instrument to identify instances where LRT allocation was exogenous to neighbourhood conditions. Section 4 will summarize the LRT projects under analysis. Section 5 introduces data sources. Section 6 provides neighbourhood level results. Section 7 provides evidence of LRT induced shifts in commuter flows. Section 8 proposes a structural model to estimate welfare effects and section 9 concludes. 2

4 2 Related Literature 2.1 Spatial Mismatch This study will be concerned with identifying the local labour market consequences of spatial connectivity. The analysis of black populations in Chicago and Detroit in Kain (1968) provided early evidence that the spatial isolation of marginalized populations may be in part responsible for lagging labour market outcomes. Numerous subsequent studies have extended Kain s notion of spatial mismatch to explain spatial heterogeneity in urban labour market outcomes and particularly to explain the lagging outcomes of racial minorities and youth. A significant attempt at empirically estimating the effect of proximity to jobs was undertaken in Immergluck (1998). Results suggested a highly significant connection between household proximity to jobs and the local unemployment rate. Stoll (1999) provided evidence from Los Angeles that black and Hispanic workers undertook more extensive spatial job search than white residents. This behaviour appeared to be a compensating response to inferior locations relative to jobs among minority groups. Overall, the unemployed and poor tend to live in places that are isolated from relevant job opportunities (Holzer, 1991; Sanchez et al., 2004). However, observed correlations can not elucidate whether this pattern is the result of heterogeneous workers self-selecting into particular neighbourhoods, or if there is a causal effect of neighbourhood connectivity on individual employment outcomes. Distance from employment mechanically increases transportation costs. But perhaps more importantly it acts as a barrier to information diffusion, as price ignorance increases through space. Stigler (1961) provides the seminal analysis of information search in a spatial setting. Zenou (2013) modelled social networks, embedding network analysis into a spatial employment model. Zenou (2013) demonstrated that if black workers are relegated to spatially isolated areas, and information on employment opportunities is propagated through social weak ties that are formed in the central business district, then black 3

5 workers will experience heightened unemployment in equilibrium. In an empirical paper, Holzer and Reaser (2000) found that black workers are less likely to obtain employment in suburban areas, primarily due to a lack of information on suburban opportunities and heightened costs to applying. A thorough summary of the underlying theory of spatial mismatch and related theoretical works can be found in Gobillon et al. (2007). 2.2 The Role of Transit The role of transit in neighbourhood sorting and neighbourhood change has been the subject of some prior work. LeRoy and Sonstelie (1983) provided a dynamic model of transportation induced urban change, noting that when the wealthy enjoy proprietary access to a rapid transportation technology they will optimally locate in the suburbs. However, when rich and poor residents have access to the same transportation technology, the rich will locate in the city center due to their increased time value of money. Glaeser et al. (2008) provided an analysis of why the poor concentrate in central cities. The presence of public transportation was examined as the likely culprit that holds poor residents in central cities. The current study extends analysis of heterogeneity in transport mode demand and its role in neighbourhood choice. Baum-Snow and Kahn (2000) catalogue rail transit investment in five US cities during the 1980s. Results point towards localized increases in home values and increased transit use. Baum-Snow and Kahn (2000) also demonstrated that rail investment during this era was preferentially directed towards suburban areas. Kahn (2007) stressed the endogeneity of route and station site selection, finding new transit stations were resisted in affluent areas, and that stations were cited in tracts that were poorer and had higher black population shares. Estimated home price effects were somewhat disparate between cities in Kahn (2007). In general, new walk and ride stations were found to cause some gentrification in the area immediately surrounding the station, on average. Walk and ride stations were found to increase the share of the local population with a college degree by 4

6 5.1 percentage points, and increase local household income by 2%. Tyndall (2017) made use of exogenous variation in transit access brought about by subway outages in New York City induced by a 2012 hurricane event. The sudden and spatially heterogeneous change in neighbourhood transit access provided an opportunity for identification. Results showed a rise in the local unemployment rate of 1.4 percentage points in neighbourhoods that lost subway access to the CBD. Results were robust to controlling for changes in observable neighbourhood demographics. Negative labour market effects were found to be strongest amongst those without access to private vehicles as well as among the Hispanic population. Heterogeneity in commute costs and mode choice give rise to heterogeneity in neighbourhood composition. Wasmer and Zenou (2002) outlined a bid-rent model wherein employed and unemployed cohorts face differing commuting costs. Wasmer and Zenou (2002) imposed dichotomous linear bid-rent curves which leads to only two possible urban configurations: a city where the unemployed live adjacent to the CBD and the employed live on the periphery, or the converse. Given that employed workers have a more frequent need to commute suggests they have higher transportation costs and therefore choose to locate close to the CBD. As will be explored in the empirical structural of the current project, the introduction of mode choice heterogeneity in a polycentric city implies non-linear bid-rent curves that can generate more complex patterns of neighbourhood sorting. 2.3 Light Rail Studies A few studies have attempted to estimate the neighbourhood effects of LRT construction. Cao and Schoner (2014) implemented propensity score matching to study the ridership effects of a Minneapolis LRT line. Methodologically, Cao and Schoner (2014) discussed the importance of both spatially endogenous infrastructure investment and household location choice in estimating causal effects. Cao and Schoner (2014) found that 5

7 residents moving towards new transit were less likely to use LRT than the original residents; potentially because incoming residents were less dependent on transit. Analysing a wider set of cities, the current paper will make the congruent finding that induced household sorting cannot explain LRT s propensity to increase public transit use among commuters. Baker and Lee (2017) investigated LRT projects and related gentrification in fourteen US cities in the 1980s and 1990s. The author s find no evidence of neighbourhood change in response to LRT construction, suggesting LRT does not trigger significant household sorting. Chatman et al. (2012) undertook a repeat sales index analysis of the home price effect attributable to the construction of a LRT system in New Jersey and found little evidence of price appreciation. The LRT project analysed in Chatman et al. (2012) is somewhat unique in that it was built to facilitate intercity trips. 3 Spatially Endogenous Transit Placement Policy makers potentially direct rail investment towards neighbourhoods that have already achieved a reasonably high level of economic development, or are primed for future growth. Therefore, the comparison of neighbourhood change between tracts gaining and not gaining transit may yield biased results. This section proposes deriving quasi-randomness in rail investment by focussing on a common transit project type: lines connecting the CBD to the metro s primary airport. In 1975, of the current 25 largest US metros, only the Boston area had a direct rail link from the CBD to the largest regional airport, today the majority of these metros have such a link. Figure 1 plots this progression. Several additional US cities are currently planning LRT extensions to their airports; including Buffalo, Pittsburgh and Sacramento. The economic development motivation for constructing rail links from city centres to airports is based on very little economic literature. Case studies have generally been unable 6

8 to provide compelling arguments in favour of such projects (Stubbs and Jegede, 1998; Widmer and Hidber, 2000). Contrastingly, the political motivations for constructing such mega-projects appear to be strong. The origins of these large rail projects are often attributed to state or regional governments who are promoting broad economic development goals and are unlikely to be apprised of neighbourhood level transit demand. Figure 1: Share of large US metros with a direct rail links from downtown to the largest airport. This study will use the rise in airport rail connections as a source of randomness in the spatial allocation of rail infrastructure. The identifying assumption will be that neighbourhoods that happened to be en route to the airport were arbitrarily provided with rail service, while otherwise similar neighbourhoods in other areas of the city were not. Efforts will be taken to control for the possibility that being located on the way to the airport may have had additional time varying impacts on neighbourhood change. 4 Light Rail Investment in Four US Cities LRT transit has become a popular transportation and economic development strategy across the US. Figure 2 shows the national growth in LRT stations. Between 2000 and 2016 the number of LRT stations in the US grew by 60%. The empirics of this study will focus on four metropolitan areas: Minneapolis, Minnesota; Portland, Oregon; Salt Lake City, Utah; and Seattle, Washington. These four metropolitan areas all completed 7

9 substantial LRT construction over the period of study and included an extension to the metro s largest airport. The metros range in population from 1.1 million (Salt Lake City) to 3.6 million (Seattle). Table 1 displays metropolitan level characteristics as contrasted with the full sample of US metropolitan residents. Figure 2: The Proliferation of LRT Stations Station counts obtained from the annual American Public Transportation Association Fact Book. Table 1: Summary Statistics, Metropolitan Areas Minneapolis Portland Salt Lake City Seattle All US Metros Metro 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 54,516 Public 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) LRT Stations (2013) Public transit comprises only a small share of total commutes in these metros. Seattle had the highest rate of commuting by transit in 2013, with 8.4% of commuters using public transit. Salt Lake City had the lowest public transit mode share in the sample at 3.6%. Public transit mode share increased in all four metros during the period. Seattle experienced the largest increase, expanding public transit mode share among commuters by 27%. 8

10 5 Data 5.1 US Census Data This study will use census tract level data from the 2000 US Decennial Census as well as the 2015 American Community Survey, five-year estimates. Census data from 2000 are crosswalked to 2010 boundaries using the Missouri Census Data Center s Geographic Correspondence Engine. Metropolitan areas will be bounded according to 2013 Bureau of Labor Statistics core-based statistical area boundaries. 5.2 Commuter Flow Data Job flow data is obtained from the Longitudinal Employer-Household Dynamics, Origin-Destination Employment Statistics (LODES) data set. LODES provides linked workplace and residence data at the census block level. The data is compiled by the US Census Bureau. LEHD data coverage extends to 95% of wage and salary employment nationally, exceptions include self-employed individuals and US military personnel (Graham et al., 2014) data is used for post-treatment commute flows, and 2002 data is used for pre-treatment commute flows. Data from 2000 is not available. Estimating the accessibility impact of rail transit changes will be aided by detailed trip routing data from the Google Maps Application Programming Interface (API). A full matrix of potential tract to tract commute routes was constructed for each metro, resulting in 1,412,602 origin-destination pairs. Subsequently, all routes were queried through the Google API for travel instructions for both driving and public transit. Queries are made for an 8 am departure on a Wednesday. The API provides estimated trip times and distances. Step-by-step navigation instructions are collected for public transit routes. A list was created containing the name of every LRT station built between 2000 and 2013, as it is identified within the API. A text search program is then run to identify all of the origin-destination pairs that make use of the new LRT infrastructure. Commuter flow 9

11 analysis will look for evidence that the new LRT infrastructure triggered shifts in the propensity of a worker to work along an LRT route. 6 Light Rail and Neighbourhood Change In order to identify the partial effect of gaining a local LRT station on neighbourhood level outcomes, this section will propose and implement a propensity score weighting framework. 6.1 Identifying Treatment and Control Neighbourhoods A census tract will be considered to be treated by LRT if a new station was built within the tract, or within 1 km of the tract s centroid, between the pre and post treatment periods ( ). Selection of treated neighbourhoods will be further limited by focusing on tracts that received LRT due to their location within an airport corridor. As discussed in section 3, the assignment of LRT to these tracts is plausibly exogenous to prior demographic levels or trends. Tracts within the CBD are excluded from analysis, proxied as any tract that has a centroid within one km of city hall. Tracts within a km of the airport are also excluded from analysis to avoid trends that might be specific to airports. Any tract that possessed a LRT station prior to 2000 is also omitted from analysis. There are 39 treatment tracts identified across the four metros. A control group is formed through propensity score weighting. Initially, a logit regression is run to estimate the probability that any tract would receive treatment based on its pretreatment characteristics. Subsequently, a synthentic control group is created by weighting untreated observations by their relative probability of receiving treatment ex ante. Any observation with less that a 0.5% chance of receiving treatment is omitted from analysis. The methodology is based on the pioneering work of Rosenbaum and Rubin (1983). The 10

12 proposed methodology amounts to a causal estimator of the neighbourhood level effect of LRT. The method is preferable to a standard OLS approach as it establishes orthogonality between treatment status and pretreatment characteristics. In order for the causal estimates to be valid the propensity score weighting must generate a synthetic control group with attributes comparable to the treatment group. Table 2 displays the matching variables selected. Pretreatment values from 2000 and 1990 are included to establish orthogonality of both levels and trends to treatment status. Of the 18 matching variables selected in Table 2, only one is significantly different between treatment and control observations at the 5% level. Treatment observations are somewhat closer to the city center on average than the controls. Mode share, race, education and labour market variables are all well balanced between the treatment and control groups, suggesting that treatment was orthogonal to ex ante neighbourhood conditions. Table 2: Matching Balance Test Variable Year Treatment Group Control Group P-value Mean Std. Dev. Mean Std. Dev. Public transit mode share White population share College degree share Unemployment rate Labour force participation rate Median household income Share of homes: detached Log distance to city hall CBSA: Minneapolis CBSA: Portland CBSA: Salt Lake City N Significance levels: : 5% : 1%. All estimates are weighted by the inverse of the propensity score. Figure 3 maps the location of all treated and control tracts used. Control tracts that 11

13 were omitted from analysis or had less than a 1% chance of receiving treatment according to the logit regression are shown in grey. Such observations have very little weight in the causal estimates. Figure 3: Propensity Score Matching, Treatment and Control Tracts A. Minneapolis B. Portland C. Salt Lake City D. Seattle - Treatment tract - Control tract - New rail station 12 - Pre 2000 rail station

14 6.2 Neighbourhood Change Results Propensity score weighted estimates can provide information on how the average characteristics of neighbourhoods were changed as a result of new LRT amenities. All estimates in this section correspond to the so-called average treatment effect (ATE). Standard errors are calculated according to the methodology proposed in Abadie and Imbens (2006). All estimates are executed in first differences to focus analysis on changes in neighbourhood characteristics between 2000 and A frequently stated policy goal of LRT transit investment is to increase the use of public transit and decrease reliance on privately owned vehicles. Table 3 estimates the partial effect of gaining a local station on the transportation mode share of local commuters. Column 1 indicates that LRT caused the share of local commuters using public transit to increase by 3.3 percentage points. The effect is highly statistically significant. Given that the (unweighted) average treated tract had 14.5% of commuters choosing public transit, the causal effect amounts to a 23% rise in public transit mode share. The effect is generally consistent with that estimated in Baum-Snow and Kahn (2000), that found a 1.42 percentage point increase in transit mode share when the closest rail station moved from 3 to 1 km away. The estimated effect of LRT on the share of commuters using a private vehicle is negative and significant. The introduction of LRT is found to reduce private vehicle mode share by 2.8 percentage points among commuters. The effect is small when compared to the average pre-treatment private vehicle mode share in treated tracts of 76.0%. Column 3 estimates the change in the average commute duration of a local resident. LRT treatment caused a 1.13 minute rise in the average commute. Public transit use can include walking to and from stations and generally is associated with longer commutes. Table 4 estimates the change in neighbourhood level labour market outcomes. Labour market outcomes are found to respond positively to the introduction of LRT. Table 4 reveals that the local employment rate rose by 3.0 percentage points, labour force participation rose by 2.0 percentage points and the unemployment rate fell by

15 Table 3: Commuter Transportation Effects Public Transit Private Vehicle Commute Mode Share Mode Share Time (min) (1) (2) (3) Gained Rail Access (.006) (.006) (.507) Mean 2000 value (treated obs) Significance levels: : 5% : 1%. Robust standard errors in parenthesis. percentage points. All three effects are statistically significant. The effect of log median income is estimated in column 4 but no significant effect is found. Viewed as a type of place-based policy, LRT appears to be a powerful tool to advance local labour market outcomes. However, the improvement of labour market outcomes amongst the neighbourhood s original residents is only one possible explanation for the finding. If local demographics were to shift such that residents with superior labour market abilities endogenously moved into the treated neighbourhoods, much of the observed labour market improvements might be explained by neighbourhood sorting. Table 4: Labour Market Effects Employment Labour Force Unemployment Log Median Rate Participation Rate Rate Income (1) (2) (3) (4) Gained Rail Access (.008) (.010) (.008) (.046) Mean 2000 value (treated obs) Significance levels: : 5% : 1%. Robust standard errors in parenthesis. Prior research has generally found transit amenities to have a positive effect on home values, with some work finding a negative effect directly adjacent to new stations. Kahn (2007) found new walk-and-ride transit stations increased local home values by 5.4% 10 years after station construction and 10.8% after 20 years. Table 5 estimates the effect of treatment on local home values and rents. The point estimates suggest LRT increased local home values by 3.4% and local rents by 2.3% in the mean tract, but the results are not 14

16 statistically significant. Note that local housing price and rent data are not adjusted for potentially changing housing quality. LRT allocation is often accompanied by local real estate development. Column 3 tests for whether LRT treatment is associated with an expansion in local housing units. Tracts that received LRT saw an average increase in local housing supply of 95 units. Table 5: Real Estate Effects Log Home Log Housing Value Rent Units (1) (2) (3) Gained Rail Access (.039) (.027) (45.811) Mean 2000 value (treated obs) Significance levels: : 5% : 1%. Robust standard errors in parenthesis. Table 6 and 7 test for neighbourhood demographic sorting in response to treatment. Increased home prices may reduce the appeal of a neighbourhood to lower income residents. However, higher valuation of public transit amongst low income residents suggests they may preferentially sort into LRT neighbourhoods. The effect of LRT on neighbourhood composition is therefore ambiguous in theory. The issue of race has strong relevance to both spatial mismatch and gentrification concerns. Table 6 displays the estimated effect of obtaining a local rail station on neighbourhood racial composition. There is no clear evidence of LRT induced shifts in racial composition. Point estimates suggest LRT led to an increase in the local black population share of 2.4 percentage points. Table 6: Racial and Ethnic Composition Effects White Black Asian Hispanic (1) (2) (3) (4) Gained Rail Access (.022) (.015) (.010) (.013) Mean 2000 value (treated obs) Significance levels: : 5% : 1%. Robust standard errors in parenthesis. 15

17 Improved transportation may conceivably improve local educational outcomes by supplying better access to educational institutions for local residents. However, an alternative mechanism would be through endogenous household sorting. Table 7 suggests LRT led to an increase in local education levels. The share of the adult population with a high school diploma, college degree and graduate degree all increased by 1.3%, 13.2% and 9.7% respectively. The three effects are not statistically significant. Table 7: Education Level Effects High School College Graduate Diploma Degree Degree (1) (2) (3) Gained Rail Access (.006) (.023) (.013) Mean 2000 value (treated obs) Significance levels: : 5% : 1%. Robust standard errors in parenthesis. This section has shown that LRT has a clear and positive effect on public transit mode share and local labour market outcomes. There is some weak evidence of LRT induced endogenous demographic sorting and property value uplift. The remainder of this paper will attempt to more thoroughly decompose the positive labour market effects between endogenous sorting and the facilitation of overcoming spatial mismatch. 6.3 Statistical Decomposition of Neighbourhood Effect LRT stations are associated with contemporaneous local increases in public transit mode share and improved labour market outcomes. This section will implement a decomposition exercise to estimate the share of neighbourhood effects that can be explained by endogenous sorting on observables. The methodology will follow the Blinder-Oaxaca framework proposed in Blinder (1973) and Oaxaca (1973). Blinder-Oaxaca is a superior methodology to simply controlling for observables in an OLS setting as it avoids the linearity assumptions of pooled OLS regressions by estimating separate models 16

18 for separate treatment groups. The following observable shifts in demographics are included as explanatory variables: white, black, Asian and Hispanic population shares; high school, college and graduate school education rates; the share of the population under 30 years of age and the share of the population over 65 years of age. A Blinder-Oaxaca decomposition produces the following results. Point estimates suggest that local demographics shifted slightly towards populations who were more likely to use public transit and less likely to use a private vehicle for commuting. However, these demographic shifts only explain 7.2% of the overall rise in public transit mode share and 13.1% of the fall in private vehicle share. The large majority of mode shift effects appear to be due to either LRT having a causal effect on the mode choice of individuals, or due to sorting on characteristics that are orthogonal to race, ethnicity, education or age. The Blinder-Oaxaca decomposition of labour market effects suggest that endogenous sorting explained a portion of the improved labour market outcomes. Once again, the effects that can be attributed to sorting are much less that the portion that remains unexplained. 15.6% of the unemployment reduction can be tied to sorting and 24.2% of the labour force participation rise is due to sorting. Figure 4 allows for the comparison of total neighbourhood causal effects to the share explained by sorting. For all four decompositions, the hypothesis that sorting played no role in the neighbourhood effects cannot be rejected. 7 Effect of Light Rail on Commute Flows This section will search for evidence of whether transit dependent commuters shift towards commute routes provided by new LRT infrastructure. Google routing data is used to identify all home-work pairs for which the fastest transit route involves a segment of newly added LRT infrastructure. Of 1,412,602 possible home-work pairs, 84,063 (6.0%) 17

19 Figure 4: Blinder-Oaxaca Decomposition Results Total effects correspond to propensity score weighted neighbourhood estimates. were connected through new LRT infrastructure. These newly connected routes covered 6.2% of 2013 jobs. Figure 5 maps 2013 observed commute flows across the four cities. Commute routes that are populated by at least one commuter are captured on the map, with darker lines indicating more commuters. The left column of Figure 5 displays all commutes in The right column displays only those commutes which directly benefit from the LRT service added between 2000 and In Salt Lake City, the large expansion of LRT resulted in 26.5% of 2013 home-work commute pairs including new LRT infrastructure as a component of the quickest transit route. In Portland, 9.8% of home-work commutes became connected through LRT. Relative to metro size, LRT 18

20 construction in Minneapolis and Seattle were less expansive. In Minneapolis, 2.4% of commute routes are covered by new LRT infrastructure and in Seattle the figure is 2.0%. Analysis will compare the propensity of a commute to be executed along a LRT treated route before and after the LRT infrastructure was built. A home-work route is considered to be treated if the fastest transit link between home and work includes an LRT station built between 2000 and Table 8 regresses the share of a tract s workers who commute along a LRT treated route against a dummy variable for post treatment status, as shown in equation 1. Fixed effects are included for the tract as well as the year. Table 8, column 1 shows the fixed effects estimation of the effect of a LRT connection on commute share. Results show a negligible increase in the overall share of the local workforce commuting along LRT serviced routes. In the average tract in 2000, 7.0% of workers were commuting along a route that would be treated with LRT, in 2013 the fraction had risen to 7.2%. The relatively modest public transit use across the three cities (5.3% of commuters in 2013) is consistent with LRT having only a minor effect on overall commute flows. LRT Route Share jt = β 0 + β 1 [Post LRT] jt + Φ j + Ψ t + ε jl (1) j indexes home tract. t indicates year; either 2000 or Φ is a home census tract fixed effect. Ψ is a year fixed effect. β 1 is the parameter of interest. LRT Route Share is the share of the tract s workforce that commutes along a route that received new LRT service. Limiting the analysis to populations who are more likely to be reliant on public transit yields results that are much larger, suggesting LRT may in fact be successfully expanding the labour market prospects of particular populations. Column 2 limits analysis to jobs that pay less than $15,000 annually likely capturing part time and low wage employment and shows an effect three times higher than the aggregate, with the share of low income jobs traversing LRT treated routes increasing from an average of 6.7% to 7.1%. Column 3 tests for an effect among jobs that pay over $40,000 annually, but finds no effect. The presence of a commuting effect within only low income jobs is consistent with the higher rates of transit dependence amongst the poor. Column 4 tests for an effect amongst 19

21 Figure 5: Mapping Commuter Flows A. Minneapolis A1. All Commutes A2. New LRT Commutes B1. All Commutes B. Portland B2. New LRT Commutes C. Salt Lake City C1. All Commutes C2. New LRT Commutes D1. All Commutes D. Seattle D2. New LRT Commutes - Commute Flow 20 - LRT Station

22 workers under 30 years of age. Workers under 30 had a response more than twice as large as the aggregate. The strong effect on employment of young people is consistent with prior literature that has found youth to be more sensitive to the spatial proximity of jobs, for example O Regan and Quigley (1998). Table 8: Change in Commuter Flows All Jobs Low Income Mid-High Income Under Age 30 (1) (2) (3) (4) Treatment Effect (.0007) (.0010) (.0010) (.0010) Home Tract FE? Y Y Y Y Year FE? Y Y Y Y Mean 2000 Flow Share, Treated Routes Significance levels: : 5% : 1%. Robust standard errors in parenthesis. N = 1,296,326 8 Urban Structural Estimation in the Presence of Sorting The practice of estimating structural neighbourhood choice models is becoming increasingly popular due to advances in methodology as well as increased ubiquity of computational power. Structural estimation provides an important advantage in its ability to recover the parameter estimates that account for complex and endogenous choice. In regards to the current research question, the ability of new rail infrastructure to advance neighbourhood development is of general interest. However, a more fundamental question is how these investments translate into societal welfare changes. The contributions of Alonso et al. (1964), Muth (1969) and Mills (1967) provide a basis for modelling urban spatial structure. From this early work it was clear that the rational decisions of utility maximizing agents who face commuting costs give rise to spatial heterogeneity in the characteristics of residents. Epple and Sieg (1999) provided a pioneering methodological description of estimating 21

23 structural neighbourhood choice models. The focus of Epple and Sieg (1999) and related earlier work was primarily on reconciling observed data with the predictions of Tiebout (1956). Bayer et al. (2004) further develops a framework of discrete neighbourhood choice. Sieg et al. (2004) applied a neighbourhood choice model to estimate the welfare effects of a change in air quality in southern California. Estimation results demonstrated that reduced form methods were not able to capture the distribution of welfare effects due to neighbourhood sorting. The exploration of parental schooling decisions and neighbourhood choice has become an interesting application for neighbourhood sorting structural models. Ferreyra (2007) and Bayer et al. (2007) estimated neighbourhood choice models to determine the welfare implications of school quality. Ahlfeldt et al. (2015) investigated agglomeration and amenity forces in a time variable urban environment. Ahlfeldt et al. (2015) used the fall of the Berlin wall as a potentially exogenous instrument that affected connectivity in the city of Berlin. The common challenge shared by these papers and the current task is to estimate the benefits of a spatially delineated public good in the presence of sorting. 8.1 A Neighbourhood Choice Model This project seeks to develop a model of spatial sorting among workers and firms. The neighbourhood effects estimated in the initial stage of this project can be used to calibrate sorting simulations. By assigning workers a utility function, the observed changes in the location of LRT stations and the corresponding changes in the commute time matrix can be reconciled with individual decisions of labour force participation, neighbourhood choice and mode choice. The resultant parametrized model can answer whether LRT is successful at catalysing welfare improving labour market linkages. The model of worker choice will take the following general framework. The utility of a worker is represented by a Cobb-Douglas style function (equation 2). U ijkv = C γ H (1 γ) + ξ ijk (2) 22

24 Workers derive utility from numeraire consumption (C) and the consumption of generic units of housing (H). Workers have a distribution of idiosyncratic preferences across home and work locations (ξ ijk ). i indexes the worker, j indexes home tract, k indexes work tract and v indexes a binary vehicle ownership decision. Utility maximization is subject to a budget constraint (equation 3). w i = Hp H j + C + c ijkv (3) w i is worker income, p H j is the location specific price of a unit of housing and c ijkv is the transportation cost incurred from commuting. Commuting costs differ based on the endogenous choice of the worker whether to own a vehicle. A full matrix of commute distances by either car or transit are recovered from Google routing data. Equations 2 and 3 result in an indirect utility function (equation 4). V ijkv = (w i c ijkv )γ γ ( 1 γ ) 1 γ + ξ p H ijk (4) j LRT influences worker decisions through reduced commute costs and by the local amenity value associated with LRTs ability to facilitate non-work trips and potential effect on local development. LRT amenity values are captured in the model by a rise in worker preference for the neighbourhood. Workers in the model make decisions of where to live, where to work and what mode of transportation to use. Firms are modelled to include labour as an input. Computationally, local rents and wages are adjusted using contraction mapping until the system of equations reaches an equilibrium wherein every tract contains the number of employees and residents dictated in the data. The model will be solved to recover parameters of worker preference for commuting by LRT, that can reconcile observed neighbourhood change effects. In the model, low income residents are endogenously more likely to forgo vehicle ownership and be reliant on public transit. Therefore, lower income residents disproportionately benefit from lowered transit commute costs introduced by LRT. Simultaneously, low income residents are also more likely to be harmed or displaced by 23

25 local rent increases. Therefore, the welfare effect of LRT on original residents is ambiguous prior to model estimation. If transit dependent populations are displaced to transit poor environments, LRT may ultimately degrade the accessibility of some workers. Solving the model will provide the necessary parameters to conduct interesting counterfactual exercises. For example, the household displacement effect of LRT stations could be mitigated by laws that limit local rent increases. Dawkins and Moeckel (2016) suggested this may improve the overall welfare gains of LRT. The model could be employed to test such a hypothesis. Work on structural modelling and estimation is ongoing. Model details and results are forthcoming in future drafts. 9 Conclusion This study analyses recent LRT transport investment in four US cities. The arrival of LRT infrastructure to a neighbourhood is found to increase the share of the local workforce who commute via public transit and improves local labour market outcomes. Analysis of the effect of LRT on property values reveals only small and insignificant increases. Local demographics are not found to shift substantially in response to LRT. A decomposition exercise suggests that the mode share and labour market effects are not the result of neighbourhood sorting. Novel use of commuter flow data combined with Google routing data allows for the estimation of LRT s effect of commute flows. Among young and low income populations, LRT is linked to an increase in the share of workers commuting along LRT routes. Future work seeks to isolate the role of overcoming spatial mismatch in LRT s propensity to improve local labour market outcomes. The proposed structural estimation methodology aims to provide a more nuanced understanding of the mechanics that connect LRT construction to labour market outcomes. 24

26 References Abadie, A. and Imbens, G. W. (2006). Large sample properties of matching estimators for average treatment effects. Econometrica, 74(1): Ahlfeldt, G. M., Redding, S. J., Sturm, D. M., and Wolf, N. (2015). The economics of density: Evidence from the berlin wall. Econometrica, 83(6): 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. Baker, D. M. and Lee, B. (2017). How does light rail transit (lrt) impact gentrification? evidence from fourteen us urbanized areas. Journal of Planning Education and Research, page X Baum-Snow, N. and Kahn, M. E. (2000). The effects of new public projects to expand urban rail transit. Journal of Public Economics, 77(2): Bayer, P., Ferreira, F., and McMillan, R. (2007). A unified framework for measuring preferences for schools and neighborhoods. Journal of political economy, 115(4): Bayer, P., McMillan, R., and Rueben, K. (2004). An equilibrium model of sorting in an urban housing market. Technical report, National Bureau of Economic Research. Blinder, A. S. (1973). Wage discrimination: reduced form and structural estimates. Journal of Human resources, pages Cao, X. J. and Schoner, J. (2014). The influence of light rail transit on transit use: An exploration of station area residents along the hiawatha line in minneapolis. Transportation Research Part A: Policy and Practice, 59: Chatman, D. G., Tulach, N. K., and Kim, K. (2012). Evaluating the economic impacts of light rail by measuring home appreciation: A first look at new jersey s river line. Urban Studies, 49(3): Dawkins, C. and Moeckel, R. (2016). Transit-induced gentrification: Who will stay, and who will go? Housing Policy Debate, 26(4-5): Epple, D. and Sieg, H. (1999). Estimating equilibrium models of local jurisdictions. Journal of political economy, 107(4): Ferreyra, M. M. (2007). Estimating the effects of private school vouchers in multidistrict economies. The American Economic Review, 97(3): Glaeser, E. L., Kahn, M. E., and Rappaport, J. (2008). Why do the poor live in cities? the role of public transportation. Journal of urban Economics, 63(1):1 24. Gobillon, L., Selod, H., and Zenou, Y. (2007). The mechanisms of spatial mismatch. Urban studies, 44(12):

27 Graham, M. R., Kutzbach, M. J., McKenzie, B., et al. (2014). Design comparison of lodes and acs commuting data products. Technical report. Holzer, H. J. (1991). The spatial mismatch hypothesis: What has the evidence shown? Urban Studies, 28(1): Holzer, H. J. and Reaser, J. (2000). Black applicants, black employees, and urban labor market policy. Journal of Urban Economics, 48(3): Immergluck, D. (1998). Job proximity and the urban employment problem: do suitable nearby jobs improve neighbourhood employment rates? Urban Studies, 35(1):7 23. Kahn, M. E. (2007). Gentrification trends in new transit-oriented communities: Evidence from 14 cities that expanded and built rail transit systems. Real Estate Economics, 35(2): Kain, J. F. (1968). Housing segregation, negro employment, and metropolitan decentralization. The Quarterly Journal of Economics, 82(2): LeRoy, S. F. and Sonstelie, J. (1983). Paradise lost and regained: Transportation innovation, income, and residential location. Journal of Urban Economics, 13(1): Mills, E. S. (1967). An aggregative model of resource allocation in a metropolitan area. The American Economic Review, 57(2): Muth, R. F. (1969). Cities and housing; the spatial pattern of urban residential land use. Oaxaca, R. (1973). Male-female wage differentials in urban labor markets. International economic review, pages O Regan, K. M. and Quigley, J. M. (1998). Where youth live: economic effects of urban space on employment prospects. Urban Studies, 35(7): Rosenbaum, P. R. and Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1): Sanchez, T. W., Shen, Q., and Peng, Z.-R. (2004). Transit mobility, jobs access and low-income labour participation in us metropolitan areas. Urban Studies, 41(7): Sieg, H., Smith, V. K., Banzhaf, H. S., and Walsh, R. (2004). Estimating the general equilibrium benefits of large changes in spatially delineated public goods. International Economic Review, 45(4): Sound Transit (1999). Central Link Light Rail Transit Project, Seattle, Tukwila and Seatac: Environmental Impact Statement, Volume 1. pages Stigler, G. J. (1961). The economics of information. Journal of political economy, 69(3):

28 Stoll, M. A. (1999). Spatial job search, spatial mismatch, and the employment and wages of racial and ethnic groups in los angeles. Journal of Urban Economics, 46(1): Stubbs, J. and Jegede, F. (1998). The integration of rail and air transport in britain. Journal of transport Geography, 6(1): Tiebout, C. M. (1956). A pure theory of local expenditures. Journal of political economy, 64(5): Tyndall, J. (2017). Waiting for the r train: Public transportation and employment. Urban Studies, 54(2): Wasmer, E. and Zenou, Y. (2002). Does city structure affect job search and welfare? Journal of urban Economics, 51(3): 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): Zenou, Y. (2013). Spatial versus social mismatch. Journal of Urban Economics, 74: