Driving up Wages: The Effects ofRoad Construction in Great Britain ∗

Rosa Sanchis-Guarner† (LSE, SERC, Grantham)

September 2013

Abstract

Reductions in travel time between locations have direct consequences on the effectivesize of labour markets. However, there is little evidence on how road improvementsaffect labour market performance. This paper estimates the effects of road constructionon individual outcomes using micro data from Great Britain. To capture the effects, Iuse a measure of accessibility to employment through the road network. To address theendogeneity of new roads placement, I compare workers located close to the schemesand identify the effect from different exposure to accessibility changes. I further use workand home location specific individual fixed-effects to separate the effects due to mobilityfrom those due to accessibility changes in-situ. For stable work-home locations, I finda positive impact of increases in accessibility from work location on wages and hoursworked, but no effect of accessibility from home on either outcome.

Keywords: Job accessibility, Labour markets, Roads, Spatial sorting

JEL codes: J31, 018, R12

∗The work in this paper is based on data from BSD, ASHE and NSPD produced by the Office for NationalStatistics (ONS) and supplied by the Secure Data Service (SDS) at the UK Data Archive and by the Virtual Mi-crodata Laboratory (VML). The data is Crown copyright and reproduced with the permission of the controllerof HMSO and Queen’s Printer for Scotland. The use of the ONS statistical data in this work does not implythe endorsement of the ONS in relation to the interpretation or analysis of the statistical data. This work usesresearch datasets which may not exactly reproduce National Statistics aggregates. All the results have beengranted final clearance by the staff of the VML and by the staff of the SDS. Additionally, any interpretations oropinions expressed in this presentation are those of the authors and do not necessarily reflect the views of theDepartment for Transport who also provided data to the study.

†I thank Henry Overman, Steve Gibbons, Teemu Lyytikainen, Olmo Silva, Thierry Mayer, Guy Michaels,Gilles Duranton, Gabriel Ahlfeldt, Steve Ross, Ferdinand Rauch and Felix Weinhardt for comments and sug-gestions. I am also grateful to conference and seminar participants at the LSE Geography Economics Clusterseminar, the Urban Economics Association Meeting 2011, the LSE Labour Markets Workshop, the Finnish Gov-ernment Institute for Economic Research (VATT), the University of Valencia, the Society of Labour EconomicsMeeting 2012, the 2nd IEB Workshop in Urban Economics 2012, the 2nd European-UEA Meeting 2012 andthe IEB seminar February 2013 for comments. Richard Welpton and the staff at the SDS kindly helped withthe data. All errors and opinions are my own. Comments and suggestions are welcomed. I acknowledge theBank of Spain, the LSE and the ERSC grant ES/J007382/1 “Transport investments and spatial economic per-formance” for financial support. Contact: Spatial Economics Research Centre, London School of Economics,Houghton Street, WC2A 2AE, London, United Kingdom. m.r.sanchis-guarner-herrero@lse.ac.uk

1 Introduction

Road network is a hugely important part of infrastructure in all countries. In Great Britainroads are the most important transportation mode both for goods and passengers. Over90% of passenger transport is done by road. In 2001 over 80% of the commutes in the UKwere done using motor vehicles, and this percentage increases in city size 1. New transportprojects require large amounts of public investment. Between 2001 and 2008, investment inroad infrastructure increased in Great Britain around 55% in nominal terms and around 40%in per capita terms. In 2008/09 over £16,000 million of public expenditure was invested intransport infrastructure, and around 40% of it on national and local roads 2. During thisperiod, the major roads network was extended by around 175 kilometres.

Better transport infrastructure brings places and people closer together. This has twoeffects on the actual size of markets. Firstly, for a given location of firms and workers, ef-fective density increases, as it becomes easier to reach other locations using the improvedtransportation network. Secondly, new infrastructure changes the attractiveness of locations,which may boost spatial concentration if firms and workers relocate. These effects may re-inforce each other and create positive agglomeration spillovers. Disentangling the empiricalrelationship between infrastructure construction and economic outcomes is essential to thedesign of transport policy, and, given the importance of road transportation for the move-ment of people and goods, the correct identification of the impact of road investments isalso important for economic policy as a whole. Yet, the number of empirical studies thathave tried to quantitatively assess these effects is still limited.

The objective of this paper is to provide causal estimates of the effects of transport im-provements on individual labour markets outcomes. I provide evidence for Great Britainbetween 2002–2008. To measure the local impact of road construction, I combine data onroad projects with micro data on firms. For the outcome variables I use data on individuallabour market outcomes for a panel of workers. I test the effect that reductions in traveltimes induced by road construction have their wages and labour supply. My results providethe first causal evidence on the effects of changes in proximity on individual labour marketperformance.

Building-up on my previous work in Gibbons et al. (2012), the effect of road constructionis captured using an index of accessibility to employment (effective density). This measure issimilar in nature to “market-access” indices. It quantifies the amount of employment whichis reachable using the road network from a given location, inversely weighed by the traveltime to reach other locations. Previous related studies rely on area variation to identify theeffects of road infrastructure on economic outcomes. They use measures such as density ofroads within an area or area connectivity to the network 3. These studies estimate the effectsof road by comparing road endowment of different locations over time.

The accessibility index has important advantages over these measures. Firstly, it is notconstrained to artificial geographical boundaries. It captures how a location is affected by

1Except for cities over with over 1 million inhabitants, where there is larger availability of public transport.Source: UK Census 2001.

2Source: Transport Statistics Great Britain and HM Treasury.3Some examples are Baum-Snow (2007a); Michaels (2008); Duranton & Turner (2011); Faber (2012)

1

road construction taking place over the whole geography. Secondly, accessibility captureshow locations are affected by changes in the road network even if these changes are small.Regardless of a location being previously connected to the network, when new links reducethe travel time necessary to reach other locations, they affect relative “connectivity”. Theextent of the effect would depend on the economic size of close-by locations and on theposition of the area in the road network relative to the rest of the country.

The geographical unit used as the basis for the calculation of the accessibility index is theelectoral ward. Wards (as defined in 1998) are quite small areas and there are over 10,500in Great Britain 4. Having such a large number of units implies that there is a fair amountof variation in the values of accessibility changes due to road construction and that theychange almost continuously over space. I exploit the fine geographical detail of the data toimplement an empirical strategy capable of identifying causal effects.

I estimate the impact of accessibility changes on individual labour market outcomes(mainly on wages and hours worked). I have information on workers home and job loc-ations over time, so in order to investigate the different channels through which accessib-ility can impact individual labour market performance, I estimate both the effects of ac-cessibility from home and from work locations, conditional on predicted commuting traveltime. To causally identify the effects I follow an empirical strategy which is based on theuse of individual micro datasets and small geography. The potential sources of bias are ad-dressed in three steps. Firstly, cross-sectional estimates of the effect of accessibility on labourmarket outcomes could be biased if the model does not capture underlying time-invariantfactors (such as individual specific productive advantages) that affect both effective densityand economic outcomes. I use a fixed-effects estimation method to address this problem.I exploit the longitudinal nature of the data and control for individual heterogeneity. Thesecond identification challenge arises because over time accessibility in locations varies dueto employment (economic size) and travel times (proximity) changes. If local employmentchanges as a result of road construction 5, I would not be able to separate the effect of ef-fective density changes on labour markets arising from changes in local employment fromthat due to changes in proximity. Furthermore, accessibility changes due to relocation of em-ployment may be partly driven by the outcome variable studied or be correlated with thesame unobserved shocks. To address these concerns, I construct the measure of accessibilityfixing location employment to pre-period levels (2001), so it only changes over time due tovariations in travel times between locations. This way, I avoid using spurious variation oneffective density that arises due to endogenous changes in employment across space in re-sponse to the new transport schemes. Additionally, by using this definition of accessibility,I focus on the variation in accessibility over time stemming from changes in the channel ofinterest, e.g. reductions in travel times between locations due to construction of new roadlinks.

Finally, a major challenge to the identification of the effects is the fact that new roadlinks are not randomly allocated, but targeted specifically at locations to improve traffic

4The average area of British wards is 21 square kilometres (they are smaller in England, 15 km2, and muchbigger in Scotland, 62 km2) and the average population estimate in 2001 (for England and Wales) is around6,000 people.

5A result that Gibbons et al. (2012) suggests.

2

flows, decrease congestion or improve connectivity. For this reason, to identify the effect ofroads on individual labour market outcomes, we cannot simply compare workers locatedin areas that were “treated” with workers in areas “non-treated”. The standard approach toaddress this issue is the use of instrumental variables, normally based on historical routesplacements or national infrastructure plans (Baum-Snow, 2007a; Michaels, 2008; Duranton& Turner, 2011, for example). Instead, I exploit the geographical detail of the data and I onlycompare workers which are located close to the road schemes 6. When new links are addedto the network, optimal travel times decrease and employment accessibility increases, butby different amounts according to where a place is in relation to the existing road network,the characteristics of the new road links and the location of major centres of employment.All the workers within this distance band are “treated” by the accessibility changes, but todifferent extents and at different points in time, depending on their location. I identify theimpact of road construction by exploiting the treatment intensities, conditional on a largeset controls. This strategy is furthermore supported by the fact that the road links are aimedat larger areas than these narrow distance band and aimed at connecting distant places andnot at improving the local economy (Highways Agency, 2009). It is quite unlikely that thenew links are aimed at specific individuals in wards within those narrowly defined distancebands, especially after controlling for different growth trends around the schemes.

By including individual fixed-effect, I estimate the coefficients using the within-individualchanges in accessibility. Over time, workers are exposed to different levels of employmentaccessibility depending on their work and job locations. Within individual variation in ac-cessibility could be due to changes in accessibility for given locations (so stemming solelyfrom changes in travel times) or due to workers relocation over space (potentially as a res-ult of new road links or due to reasons correlated with them). As I follow workers workand home locations over time, different sets of fixed-effects (individual, individual-work,individual-home, individual-work-home) can be used. These control for individual and areatime-invariant unobservables and allow me to investigate the impact that work and homemobility has on the estimates. The changes in the coefficient of work and home accessibil-ity when I allow for different degrees of mobility informs about the role that spatial sort-ing plays on identifying the impact of accessibility on individual labour market outcomes.When use I individual-home-job fixed-effects (so I exploit the variation in accessibility foran individual while he is keeping his work-home location pair fixed) I identify the effectof accessibility only from changes in travel times for a given location. As mobility mightbe a result of road construction, this strategy isolates the effect of changes in accessibility“in-situ” (pure effective density effects) from the effect due to spatial sorting.

Allowing for spatial mobility, I find positive significant effects of accessibility from bothwork and home on wages and hours worked, conditional on predicted commuting time.The effects could be working via increased access in-situ or work or home relocation. WhenI use individual-work-home fixed effects, I find positive significant effects of accessibilityfrom work on earnings and hours worked but I find no evidence that accessibility fromhome affects any of the outcomes. This suggests that residential mobility is the source ofvariation from which I identify the impact of accessibility from home on labour market out-

6This follows from Gibbons et al. (2012)

3

comes. When I use the changes in accessibility for a given location, only accessibility fromwork location has an effect on outcomes. Potential explanations of this finding could beincreased spatial competition for jobs (which results in higher individual performance) oragglomeration externalities.

My research contributes to the existing evidence in three ways. First, I use very richworker microdata and a detailed dataset on road projects, which allows detailed study ofthe relationship between transport improvements and individual labour market outcomesat a very small spatial scale and to study several inter-related outcomes. To my knowledge,there exists almost no previous evidence on the effect of road construction on individuallabour market outcomes. Secondly, given the quality of the data used, I can adopt a carefulempirical strategy to tackle several identification issues which could undermine the causalinterpretation of the results. This improves the validity of the results in order to extractpolicy recommendations. Finally, I provide robust empirical evidence on some strands ofthe existing theoretical urban labour economics literature, as the paper provides empiricalevidence on the spatial mismatch theory and the effect of effective density on labour pro-ductivity.

The remainder of the paper is structured as follows. Section 2 sets out the channelsthrough which transport investments can affect labour market outcomes and reviews therelated theoretical and empirical literature. The empirical strategy is explained in detail inSection 3. The data sources and the construction of the variables used in the empirical ana-lysis are also described in this section. The results are fully explained in section 4, and ro-bustness checks and an interpretation of the coefficients is also provided. Finally, section 5concludes and explains the future steps of this research.

2 Theoretical framework and related literature

The role of transportation in the spatial distribution of economic activity and economic per-formance has become of increased interest to researchers in the last years. Decreasing trans-port costs are considered to be a central driver of economic integration and of the rise of ag-glomeration externalities, but solid empirical evidence on the channels through which theseeffects operate is scarce. Even if some authors have explicitly included the role of transporta-tion into spatial economic analysis (Combes & Lafourcade, 2001; Puga, 2002; Venables, 2007),there is still need to empirically establish the causal link from transportation infrastructure tospatial economic performance, especially with regard to the effects on individual outcomes.

While most of trade and location theory regards transport investments as having an effecton (goods) transport costs (Michaels, 2008, for example), urban labour theories (Zenou, 2009)consider transport policy as having an effect on commuting costs and therefore mostly oper-ating through the labour market. Glaeser & Kohlhase (2003) document the declining role ofgoods-transportation costs in developed countries and highlight the increasingly importantrole of the mobility of workers in modern economies.

Transport improvements affects labour markets through multiple channels (for a reviewsee Gibbons & Machin, 2006). First of all, new transport links reduce commuting time costs.

4

The “spatial mismatch hypothesis” (SMH) predicts that larger physical distance 7 betweenresidential location and job location has detrimental effects on the labour market outcomesof those living further away from employment centres. A reduction in commuting time andcosts associated with transport improvements enables people to increase the scale of their jobsearch and could also encourage potential workers to participate in the labour market (Vick-erman, 2002; Jiwattanakulpaisarn et al., 2009). For unemployed or inactive workers reducedcommuting costs decrease search costs and reservation wages, and in this way, they can helpmitigate frictional unemployment and increase the employment probabilities of those whoare jobless (Inhanfeldt, 2006). By affecting optimal labour market choices, shorter commutesalso impact observed wages and unemployment rates (see Phillips, 2012, for some evidenceon these channels). If employers require workers with specific characteristics or skills, em-ployees can bargain to improve their labour market conditions (wages, hours worked, occu-pation) to compensate for longer commutes. However, if labour markets are thin (Manning,2006), i.e. workers have access to a limited number of potential employers, longer commutescould not be fully capitalised into nominal wages. Longer commutes can also have an effecton wages through their effect on productivity, if shorter commutes are related to healthier ormore motivated workers (van Ommeren & Gutierrez-i-Puigarnau, 2011). Evidence suggeststhat commuting costs have effects on workers labour supply (Gutierrez-i-Puigarnau & vanOmmeren, 2010), absenteeism (van Ommeren & Gutierrez-i-Puigarnau, 2011) and wages(Manning, 2003; Mulalic et al., 2010).

Secondly, faster connections bring employers and workers closer together and henceaffect the effective size of the labour markets. Accessibility affects the tightness of the la-bour market, i.e. the ratio of unemployed relative to the number of job vacancies (Detang-Dessendre & Gaigne, 2009). If labour markets are more accessible to unemployed workersresiding outside the labour market area, or the labour market becomes effectively biggerbecause it is better connected, for a given number of vacancies the number of potential can-didates would increase. This increases competition for jobs and might have two effects. Onthe one hand, unemployed workers living further away from job might become employeddue to the increase in accessibility 8. Manning & Petrongolo (2011) use British data to studyhow “local” labour markets are and find that workers tend to look for and to accept jobswhich are relatively close to them, so labour markets are very local. Reduced commutingtime due to road construction can have an effect on the actual size of these local labour mar-kets and affect the number of competitors for a give search area. On the other hand, dueto increased competition for jobs and workers, we could see an increase in the quality ofjob matches which could be translated into higher productivity and wages. Better matchescan occur because for a given travel time workers have access to a larger pool of vacancies(Harmon, 2013).

Changes in effective density can also affect labour supply, e.g participation and hoursworked. It can weaken barriers to participation in the labour market, encouraging the entryof disadvantages groups (female or low-skilled) into job search. It can affect hours worked

7Distance includes geographical distance but could also include bad public transport provision or pooraccess to private transport (car).

8It is beyond the scope of this paper to study the effect of accessibility on employment. I use data for a panelof employees and this focus on labour market outcomes for employed workers.

5

if wages increase due to transport improvements, making work more appealing; or evenif increased competition in denser areas induce young professionals to behave in a morerivalrous manner (the “rat race” argument in Rosenthal & Strange, 2008).

Thirdly, reductions in travel times would increase the geographical scope of the agglom-eration economies, as for a given physical distance employers and employees are nearer toeach other 9. If agents are closer, there is more potential for interaction. Ahlfeldt & Wend-land (2011), using the construction of the railway in Berlin at the beginning of last century,show how distance becomes less important and the scope of spatial interaction increases astransport infrastructure improves. The emergence agglomeration externalities might havepositive effects on workers earnings if urbanisation economies 10 and proximity to markets(Krugman, 1991; Krugman & Venables, 1995) give raise to productivity gains which are cap-italised into nominal wages.

Finally, transport investments have an effect on location decisions. If better connectivityis perceived as a locational amenity it would be taken into account when workers decidewhere to reside. When transport investments are capitalised into housing prices (Gibbons &Machin, 2005; Ahlfeldt, 2011), they in turn affect real household incomes, and thus residen-tial choices and labour market behaviour.

In short, better transport connections can improve the efficiency of labour markets andhave positive impacts on workers performance. Nevertheless, given the multiplicity of chan-nels, the size and direction of the effects of transport improvements on labour market out-comes remains mainly an empirical question (Gibbons & Machin, 2006).

A number of recent papers have studied, with emphasis on the estimation of causal ef-fects, the impact of roads on a variety economic outcomes. Most of these papers use datafor the US. Using the 1947 planned Interstate Highway System as an exogenous source ofvariation, Baum-Snow (2007b) studies the effect of highways on the process of suburban-isation of American cities since the 50s, and on the changes in commuting patterns sincethe 60s (Baum-Snow, 2010). Michaels (2008) uses a similar source of exogenous variation toestimate the effect of reduced trade barriers on the demand for skills. Hymel (2009) uses across section of US metropolitan areas to assess the impact of traffic congestion on aggregateemployment growth. Duranton & Turner (2011) examine the effect of road construction onvehicle-kilometres traveled (VKT) in US cities. These same authors, using a similar identific-ation strategy, also estimate the effects of highway construction on urban growth (Duranton& Turner, 2012) and on trade (Duranton et al., 2013). A few papers (Donaldson, 2013; Faber,2012) have focused on developing countries (railroads in colonial India and highways inChina) to study the effect of the reduction of transport costs, due to transport network de-velopment, on trade integration and resulting economic development.

With the aim of estimating causal impacts, my research is related to this recent strand ofthe literature. However, the identification strategy is different to the studies above. Most ofthese papers use road network proxies at the metropolitan area level as their regressor of

9Rosenthal & Strange (2003, 2008) and Jofre-Monseny (2009) are some of the few studies which investigatethe geographical scope of agglomeration economies. They however examine the geographical scope of theagglomeration effect, not test the effect of agglomeration via changes in proximity.

10Sharing, matching and learning in Duranton & Puga (2004) terminology; input-output linkages, labourmarket pooling and knowledge spillovers in the classical Marshall (1890) terminology.

6

interest. In my case, accessibility is computed at very small area level, so the treatment ofspace is quasi-continuous. Additionally, the accessibility index changes smoothly over spaceeven in areas where road construction did not take place. Finally, my estimation methodo-logy does not rely on the use of instrumental variables but exploits the geographical detailof the data by only comparing individuals affected by road construction.

My findings add to existing evidence on the effects of agglomeration and market accesson nominal wages (for example Mion & Naticchioni, 2005; Combes et al., 2008). Most of thesestudies identify the impact from local variations on employment density, but little attentionhas been given to the effect of agglomeration via changes in proximity between locations.New road links decrease travel time between locations and, for given economic sizes, have adirect effect on effective density. In contrast with these papers, my paper provides evidenceon how changes in “proximity” can give rise to agglomeration externalities. If there are anypotential productivity benefits to workers arising from changes in accessibility, these couldbe induced through road constructions. Therefore, my estimates inform not only on the over-all effect of accessibility on worker economic outcomes but also on the channel throughwhich policy can potentially impact these outcomes. Additionally, my findings also informsome of the predictions of the spatial mismatch literature (see Inhanfeldt, 2006, for a review)and of the thin labour markets theory (Manning, 2003).

My main results explore the impact of road construction on workers pay, but I also ex-amine the effects on labour supply and investigate how spatial mobility affects the results.By looking at several outcomes, this paper provided a wider analysis on the relationshipbetween transport investments and individual labour market outcomes than the existingevidence.

3 Research design

3.1 Measuring road construction

3.1.1 Accessibility measures

The aim of this paper is to estimate the causal effect of road construction on economic out-comes 11. Therefore, the first challenge is to find a measure that captures changes in the roadnetwork. A variety of measures have been used in the literature 12. As the current road net-work in Great Britain is very dense and the length of the new links constructed is relativelyshort, using measures like changes in connectivity or density of roads to capture the effectsof road construction would not be appropriate. Instead, I use a measure of accessibility toemployment “through the road network” (or effective-density) from each location at everypoint in time.

11The methodology described in this section builds from previous work I did with co-authors in Gibbonset al. (2012). When necessary, the methodology was adapted to the specific context and research questions ofthe current paper.

12Some measures are connectivity to the network (Faber, 2012), kilometres of roads within a given area (Meloet al., 2010; Duranton et al., 2013), distance to closest highway (Baum-Snow, 2007a), number of rays crossing agiven area (Baum-Snow, 2007a, 2010), presence of highways in a given location in a particular year (Chandra &Thompson, 2000; Michaels, 2008), “lowest-cost route effective distance” (Donaldson, 2013) or amount of publicexpenditure on road infrastructure in a given area (Fernald, 1999).

7

Accessibility to employment measures the amount of employment which is reachableusing the road network from a given location, inversely weighed by the travel time to reachthese other locations. One advantage of using this measure is that it is not constrained toartificial geographical boundaries like some of the alternative measures. Moreover, it allowsus to use variation due to road construction which affects optimal travel times, even if thelocation was previously connected to the network. Additionally, it captures the effects oftransport improvements over the whole geography. This measure is, thus, appropriate in asetting where road density was already high at the beginning of the period of study andwhere one aims to study the effect of additions to the existing network.

Formally, accessibility to employment Art from a given location r at time t is defined as:

Art = ∑Rj 6=r

[a(crjt

)∗ econ sizejt

](1)

where a (.) is the transport cost function, crjt are the transport costs between locations r and jat time t and econ sizejt measures the economic size of the location at time t. Finally, t corres-ponds to years 2002 to 2008 13. I use electoral wards (as defined in 1998) as the geographicalunit.

This index is a measure of the economic mass accessible to a firm or a worker in a par-ticular location, given the local transport network. At a given origin location r at time t,accessibility Art is a weighted sum of economic mass in all destinations j that can be reachedfrom origin r by incurring a transport cost crjt along some specified route between r and j(for example straight line distance or minimum cost route along a transport network – meas-ured in travel time or in distance). The function a (.) determines how the weights enter inthe calculation of Art. The economic size of locations is measured using total ward employ-ment and travel time between wards is used as measure of transport costs (further details onthe construction of ward-to-ward travel times and ward employment is provided in SectionA.1.). Assuming inverse cost weights 14 and a cost decay equal to one 15, the final expressionfor accessibility to employment becomes:

Art = ∑Rj 6=r

[(1/travel timerjt

)∗ employmentjt

](2)

The accessibility index as defined in (2) is similar in structure to market potential meas-ures used in economic geography (e.g. Harris, 1954; Krugman, 1991), and to the access-

13As explained in Section A.1.2., road networks and travel times are dated at the beginning of the year. As theward employment data is dated in April of each year, in practice I combine travel time data at the beginning ofyear t (for example 2002) with employment data in April of the previous year t− 1 (April 2001). This impliesthat in the empirical specifications of section 3.2, accessibility measures are in fact lagged one year with respectto the outcome variables.

14In the definition of Art the value of the weight a (.) attached to any destination r is a decreasing functionof the cost of reaching destination j from origin r. α is the cost decay. Potential weighting schemes include:“cumulative opportunities” weights a

(crjt

)= 1 if j is within a specified distance of r, zero otherwise; “ex-

ponential weights” a(crjt

)= exp

(αcrjt

); “logistic weights” a

(crjt

)= [1 + exp

(−αcrjt

)]−1 or “inverse cost

weights” a(crjt

)= c−α

rjt . See Graham et al. (2009) for further discussion of these indices.15Graham et al. (2009), using the inverse cost weighting scheme, estimate the parameter of distance decay

functions for several sectors using similar British data to ours and find values between 1.8 and 1, dependingon the sector. The estimated alpha for the whole economy is around 1.6. In practice, I check the robustness ofthe results to different distance decays.

8

ibility indices used more generally in the transport literature (e.g. Vickerman et al., 1999;El-Geneidy & Levinson, 2006). The calculation of this index requires the construction of anorigin–destination (O–D) matrix whose components are travel times between the locations.When computing the O–D matrix I apply a limit of 75 minutes drive time (1.25 hours). Thislimit facilitates O–D matrix computation but hardly affects the value of the accessibility in-dex because wards beyond 75 minutes have negligible weights in the calculation of Art.Moreover, as shown in Table A.1, more than 99% of commutes in the Great Britain are below90 minutes. I also exclude location r from the calculation of the accessibility measures tolessen the potential reverse causality problem. This problem arises because the dependentvariables include labour market outcomes the individuals located in the ward and it is likelythat these outcomes and the economic size of a given location are be jointly determined. Asexplained below, I use an alternative definition of accessibility to address this issue, so theinclusion of the own economic-size in the calculation of Art would not be a major issue.

Accessibility Art changes in a given origin r are driven both by changes in travel timesbetween wards (stemming from road construction) and by changes in the employment oforigin ward r and in the different wards j around r. This may lead to endogeneity problemsin the estimation of the effect of accessibility, if the employment changes near the origin arecausally linked with changes in the economic outcomes in the origin or driven by the sameunobserved factors. Moreover, as I focus on the effect of road construction, the examina-tion of Art would be limited as it would be impossible to differentiate between the changesdriven by employment relocation and the changes driven by changes in travel time betweenlocations. It is thus useful to construct an alternative accessibility measure Art that focuseson the changes in accessibility that stem only from road construction:

Art = ∑Rj 6=r

[(1/travel timerjt

)∗ employmentjt0

](3)

where t0 is some fixed period of time at the beginning of the period of analysis. In the em-pirical analysis below t0 corresponds to 2001. Fixing employment to its t0 level ensures thatchanges in the accessibility index (3) over time occur only as a result of changes in thecosts crjt (e.g. travel time) and not as a results of changes in wards employment. By de-fault, changes in Art are always positive as employment is fixed over time and changes inthe denominator are always reductions in optimal travel times (see more details on the con-struction of these in the Appendix). In the empirical work, I use Art in order to use only thevariation in accessibility which stems from the road construction. In the robustness checksI also instrument Art with Art, and the results remain unchanged. Section A.1. in the ap-pendix contains full details on the construction of the ward employment and the travel timebetween locations used in the calculation of the accessibility measures.

The empirical methodology uses the changes in the accessibility index at each locationr to estimate the extent to which individuals in location r are “potentially” affected by theconstruction of new roads close to them. The construction of Art is based on predicted op-timal travel times which change due to new routing after the construction of new links inthe road network. I do not use information on how individuals (works and firms) actuallyuse roads and which are the actual travel times between locations. In this sense, the measure(changes in accessibility) captures the “intention-to-treat” of road construction. Neverthe-

9

less, it is road construction (and not actual observed travel times) that is the instrumentused by transport policy to improve locations connectivity and alleviate traffic congestion.In this sense, using changes in predicted travel times is the appropriate measure to evaluatethe effect to this transport policy.

Over all possible improvements through which transport investments can affect roadnetworks I focus on construction for two reasons. First of all, among all the different type ofinterventions (e.g. dualling, resurfacing, improving lighting and signalling, etc), construc-tion of new links is a clear source of variation which is easily identifiable and measurable.Construction of new road links affects optimal routes along the road networks and it con-stitutes the source which is likely to have the larger impact on the variation in travel timesbetween locations. It also is a well-defined tool which can be used by policy to influenceproximity between locations. The second reason is data availability. The collection of reli-able data on transport improvements and projects over a length of time is demanding andtheir measurement can be problematic. The methodology I use to generate variation in op-timal travel times needs a clear definition of changes in the road network and addition newlinks to the network provides a reasonably source of variation (see Sections A.1.2., A.1.3.and A.1.4. in the Appendix for more information). Nevertheless, one must bear in mind thatother type transport improvements might also affect optimal travel times between locationand the analysis is limited from that perspective.

3.1.2 Descriptive statistics

Table 1 provides summary statistics on wards’ log accessibility. The top panel displays thestatistics for the change log accessibility between 2002 and 2008 (growth rate) 16. The bottompanel displays values for the statistics for the annual log accessibility, which is the variation Iuse in the empirical estimations. The table shows statistics for the accessibility indices for allthe wards in Great Britain (10,540) and for wards which are situated within 5, 10, 20 and 30kilometres of the road schemes carried out during the period of analysis 17. In the main es-timations I use observations from workers located in wards within 30 kilometres of the newlinks, but here I show summary statistics for smaller distance bands in order to illustratehow accessibility changes increase the closer we get to the new links. The accessibility meas-ure is calculated fixing employment at 2001 levels (Art), so its variation stems from changesin predicted travel times (denominator) and not from changes in ward employment. For thisreason its range of variation is much lower than standard market-access measures.

[INSERT TABLE 1 HERE]

The upper panel of Table 1 shows that, for the whole Great Britain, between the first andlast year of the period of analysis employment accessibility induced by road constructionwas on average small, only 0.3% (which is very similar to the decrease in travel times ob-served in Table A.12). However, average accessibility change increases significantly whenwe focus on wards closer to projects. For example, within the 5-kilometre distance band the

16I calculate the growth rate between the first and last period of the analysis for descriptive purposes, but Ido not use it in the empirical analysis.

17Along the road scheme. When schemes are close to each other, these bands overlap and are combined intoa single area. See figure 1(c) for more detail.

10

mean change is 3.5% and the 90th percentile is 8.5%. As we expand the sample away fromthe schemes changes in accessibility tend to fall. Within 10 kilometres, mean accessibilitychange is around 2% and the 90th percentile is almost 5%. Within 20 kilometres these valuesdecrease to around 1% and 2%, and within 30 kilometres of the schemes mean accessibilitychange is 0.75% and the 90th percentile is 1.5%. It is worth noting that the standard vari-ation of the changes is relatively large, suggesting a substantial amount of spatial variationof changes in accessibility over space.

In the estimation of the empirical results, I exploit the annual variation in road access-ibility. For this reason, the bottom panel of Table 1 provides summary statistics on the levelof log accessibility in every year. Focusing on the statistics for the 30-kilometre band we seethat the even if the overall standard deviation is not very large, there is a quite a large rangeof variation between the minimum value and the maximum value of log accessibility. Infact, in the panel estimates of section 4, I exploit the within-variation of the variable. For thisreason, the between and within standard variation of log accessibility is also provided in thetable. We notice that most of the overall variation stems from the differences across wards,and this increases the further away we move from the schemes. However, as we see in theempirical results, the amount of within variation of the variable within the 30-kilometreband is sufficient to yield precise estimates.

Figures 1(a) and 1(b) illustrate the spatial relationship between the location of road schemesand resulting accessibility increases. Figure 1(b) shows the changes in log accessibility between2002 and 2008 which stem only from road construction (Art). The biggest changes in access-ibility are around the schemes plotted in Figure 1(a), but there is substantial spatial variationacross the country. The amount of spatial variation is more evident in Figure 1(c). It repres-ents the same values as Figure 1(b) but focuses on the Manchester-Bradford-Leeds area. Thisfigure shows the small scale of the spatial data and also the substantial amount of variationin accessibility values close to the new links. It also allows us to illustrate the identifica-tion strategy (which is explained in Section 3.2). The location of the new links constructedbetween 2002 and 2007 are indicated by bold white lines, with the name of the road labeled.The dark grey lines are ward boundaries. The thick black lines delimits the wards whichare included within 30 kilometres of the road projects within the area. The map illustratesthat the effects of road construction on accessibility vary considerably across wards in thevicinity of the same scheme. In the identification strategy I argue that these differences inaccessibility changes across wards are coincidental and can be treated as exogenous, espe-cially when controlling for differential time trends near different schemes. Figure A.1 showsthe geographical extent of these distance bands.

[INSERT FIGURES 1(a), 1(b) and 1(c) HERE]

3.2 Empirical specification and identification strategy

The Annual Survey of Hours and Earnings (ASHE) is used to obtain results on individualearnings and hours worked (more details in Section 4). I use a panel of workers to estimatethe effect of accessibility to employment, both from work and from home, on individuallabour market outcomes. The main regressions use data for years 2002 to 2008. In the main

11

results I focus on the effect of accessibility on wages, as most of the literature on the effectof effective density on labour market outcomes has focused on this variable. In section 4.3.1I provide evidence on alternative outcomes in order to investigate the multiple channelsthrough which road construction might affect labour markets. Without loss of generality,the notation below is refers to wages

The estimated relationship is:

yihwt = β1Aiht + β2Aiwt + β3cihwt + θXit + δZht + λWwt + µi + ξt + εihwt (4)

The wage of a worker i living in ward h and working in ward w at time t is denoted withyihwt. Given that accessibility is measured at the ward level, I initially ignore changes ofhome or job within the wards 18. At each point in time, workers work and live in spe-cific wards. Over time, workers location can change, if they change jobs, change home orchange both. Initially, I use individual variation in accessibility and wages which could alsobe driven by spatial relocations (more on this below). Aiht denotes accessibility to employ-ment from home ward h at time t, Aiwt denotes accessibility to employment from workward w at time t 19, cihwt denotes the commuting costs (travel time) between work and homeat time t, Xit is a vector of personal and job characteristics, Zht is a matrix of home wardcharacteristics and Wwt is a matrix of work ward characteristics. ξt are year-industry dum-mies that control for year-industry specific shocks affecting all wards in a given year. εihwt

is the idiosyncratic error. Both the labour market outcomes and the accessibility indices aretransformed to natural logarithms so we can interpret their coefficients as elasticities. Ac-cessibility and commuting time are lagged with respect to the outcome variable, as labourmarket outcomes measured in April of year t, travel times are measured in January of yeart and the economic size of locations is fixed to April 2001 employment.

The ASHE dataset does not report information on travel time or distance travelled toworkplace, but provides detailed information on the work and home location at every pointin time. I use the road networks between 2002 and 2008 (created as explained in the ap-pendix) to calculate optimal travel time between ward of home and ward of work along theroad network. We do not have any information on the travel mode of the workers. It couldbe the case that they are not commuting by road or not commuting using the optimal routepredicted by the GIS software. However, using optimal travel time to proxy for commut-ing costs has the advantage of getting rid of potential measurement error on self-reportedtravel time 20. In practice, I estimate a reduced-form effect of commuting costs, in which the

18For feasibility, accessibility is calculated using a very large but limited number of geographical units (over10,000). This geographical division is a good approximation of the continuity of space as most home and jobrelocations induces a ward relocation. However, workers can move houses or jobs within a given ward. AsI have information on the home postcode and the work plant (as a combination of firm-sector-postcode), insection 4.3 I also check the robustness of the result to allowing or restricting home and plant movements withinwards.

19For generality, I define accessibility as in expression (2), therefore allowing for changes in both ward em-ployment and in travel times over time. As explained below, in the formulation of accessibility used in theempirical application I keep ward employment fixed to 2001 values, focusing on the variation stemming onlyfrom changes in predicted travel times.

20Which is common, as noted for example by Gutierrez-i-Puigarnau & van Ommeren (2010). Indeed, when ifwe tabulate the answers on commute travel times using data from the UK Labour Force Survey –non reported–,responses are disproportionably cumulated in values such 5, 10, 15, 20 and 30 minutes, likely due to rounding-

12

measure of commuting cost is the optimal travel time through the road network. Moreover,as this travel time measure changes over time due to the road construction, it does not dropout when I include individual-home ward-work ward fixed-effects below, as opposed tostraight geodesic distance. I include this information in the estimated specification (4) inorder to capture the effect of commuting costs on labour market outcomes. This way I canestimate the effect of accessibility from work and home conditional on commuting costs.This helps to interpret the results given the numerous theoretical channels through whichtransport construction can affect labour market outcomes, as discussed above. For example,some of the effects of transport policy on wages could be due to employers compensatingworkers for longer commutes and some could come through increased spatial competitionor agglomeration externalities. By controlling for commuting costs in equation (4) we canbe more certain that the effects of accessibility are not due to compensation for longer com-mutes as this explicitly controlled for. Moreover, the interpretation of the coefficient β3 alsoinforms us about the relationship between varying commuting costs and labour market out-comes.

I am interested in the consistent estimation of parameters β1 and β2. There might be (timeinvariant) unobservable individual characteristics that affect both individual labour marketoutcomes and accessibility indices (in levels) at the same time, and that are not included inXit. For example, more able individuals may live or work in areas where accessibility andwages are higher. I include worker fixed effects µi to control for this (which in practice isequivalent to estimate the demeaned model). By introducing the individual fixed effects Iestimate coefficients β1 and β2 using the variation of accessibility over time with respect tothe average individual accessibility in the period of observation. At each point in time theworker may hold different jobs in different work locations (w) and live in different places (h).Therefore the level of employment accessibility to which the worker is exposed at home andat work varies over time for three reasons: when work-home location changes, when wardemployment changes and when road construction takes place (changing predicted traveltimes).

If workers are sorting spatially in order to take advantage of the changes in accessibilitywe would not be able to identify the separate effects on labour market outcomes which stemfrom spatial sorting and those which are due to changes in accessibility for a given location(for example agglomeration externalities). Sorting could be an outcome of accessibility orcould be due to other unobservable reasons correlated with individual labour market out-comes. For example, if workers with higher ability move to areas where accessibility andwages are growing faster, then the correlation between the changes in accessibility and the(demeaned) error term could be different from zero. The same could occur if more ableworkers choose jobs in areas where wages and accessibility are growing pro-cyclically. Forthese reasons, even after controlling for unobservable time invariant characteristics of theindividuals, there would still be reasons to think the estimates of β1 and β2 could be biased.

To investigate this issue I define individual home-ward-work-ward specific fixed effects,µihw and exploit the individual variation in accessibility for a given work-home (ward) loc-ation. This strategy is useful for two reasons. Firstly, it isolates the effect of accessibility on

up.

13

labour markets for a “given-location”. When I use within-individual variation, I am identi-fying the effects of accessibility as a combination of changes due to spatial relocations andchanges in effective density in-situ. Apart from the endogeneity concerns this type of vari-ation raises (if movements are correlated with labour market performance and changes ineffective density), most of the theoretical predictions of the effect of accessibility on labourmarkets apply to individuals who do not move over space, and thus this strategy allows meto link the results with the theoretical explanations in a more direct way. Secondly, compar-ing the estimates when I allow for different degrees of spatial mobility informs about therole that spatial mobility plays on identifying the impact of accessibility on individual la-bour market outcomes. Finally, individual-location specific fixed-effect additionally controlfor location-specific unobservables which might be correlated with labour market perform-ance and changes in predicted travel time, for example unobservable ward characteristicswhich might have determined the location of new road links on those specific locations.

Replacing the individual fixed effects in (4) with µihw gives:

yihwt = β1Aiht + β2Aiwt + β3cihwt + θXit + δZht + λWwt + µihw + ξt + εihwt (5)

Rewriting equation (5) in demeaned terms by subtracting the individual-fixed location meansacross time (focusing only on the accessibility measures):

(yihwt − ωi.) = β1 (Aiht − Ai.) + β2 (Aiwt − Ai.) + (εihwt − εi.) (6)

The effects of accessibility on individual labour market outcomes in (5) are obtained by ex-ploiting the changes of accessibility over time for an individual while keeping their workand home locations constant. When including µihw, coefficients β1 and β2 are not identifyout of variation stemming from spatial relocations. This reduces the endogeneity bias thatmight arise if the relocations are endogenously determined after the changes in accessibil-ity take place (that could be caused by increased accessibility or other unobserved reasons).The comparison of the estimates obtained by estimating (4) and (5) inform us about howresidential and work location sorting affects the estimation of the accessibility effects.

To implement this, each individual in the panel is allocated a different individual-locationfixed-effect depending on where he works and where he lives. The same individual can getseveral fixed effects depending on his work-home location. In the data, we can identify twotypes of individuals: those that never move work-home location pair (while observed in thedata) and those that eventually change ward of work, ward of residence or both. Individu-als in this last group may be spatially relocating (sorting) due to changes in accessibility, soin contrast to (4), specification (5) does not use this variation for the estimation of the ef-fects of accessibility but includes both types of individuals in the estimation sample. Usingindividual-location specific fixed effects is different in nature from restricting the sample toindividuals that do not move locations over time, as in the second case we are selecting oursample (and therefore might introduce sample selection bias) while in the first case we usethe whole sample but exploit variations over time around mean accessibility values for in-dividuals while their keep the same work-home locations. By not restricting the sample tonon-movers, for the same individual I exploit both the within variation for a given location

14

and the between variation for different individual-location observations, so for the sameworker I am able to exploit a larger range of variation to identify the effects. Restricting thesample to non-movers drops the information for individuals that move and thus the estim-ation of the effects relies only in a sub-sample of workers which might rise precision andsample selection problems 21.

However, there are still other possible sources of bias in the estimation of the effect of ac-cessibility using (5). Even when keeping the work-home location fixed, accessibility changesaround the individual means due to changes in employment (numerator) and in travel timesdue to road construction (denominator). If changes in labour market outcomes, for examplewages, affect accessibility changes by means of attracting workers to wards around the onesin which the worker lives or works, a reverse causality problem could challenge the validityof the estimates. There could also exist unobservable trends which affect both the location ofemployment and the labour market outcomes which could bias the estimates. As explainedin Section 3.1.1, to overcome this issue, I calculate Art, accessibility in t, as in equation (3),where ward employment is fixed to 2001 values and travel times correspond to predictedroad optimal values in January of t. Over time, the variation in this measure only occurswhen new road links are constructed and predicted travel times between wards change. Inpractice, this is similar to instrumenting accessibility Art with Art, but using a reduced-formapproach. Using the reduced form approach allows me to focus on the variation in accesswhich I want to study (that stemming from road construction) and, as results in Section4 show, does not provides very different estimates from a two-step instrumental variablesestimation.

Possibly the most important threat for the causal interpretation of the estimates is thefact that new links are not randomly allocated to locations, but specifically aimed at certainplaces in order to affect their economic performance. It could be argued that the approachis not valid if the construction of new links is aimed at areas which are experiencing un-observable shock which are correlated with individual labour market outcomes. Transportinvestments may be taking place in areas in which workers would have done better anyway.Most of the literature on the evaluation of transport projects has devoted a large discussionabout how to tackle the endogenous placement of transport investment and most authorshave dealt with this problem using instrumental variables estimation (Baum-Snow, 2007a;Michaels, 2008; Duranton & Turner, 2011, are good examples.). Instead, I rely in two thingsin order to tackle this identification issue. Firstly, according to the Department for Transportdocumentation about the schemes, improving very local economic outcomes is not one ofthe key objectives of the road projects carried out by the Government 22. Transport projectsare generally aimed at a large spatial scales and designed to improve safety or reduce con-gestion within a wider area 23. As projects are aimed at connecting areas (for reasons related

21For robustness, in section 4.2 I obtain estimates of equation (4) restricting the sample to individuals thatdo not move locations over time and, even if the results are qualitatively similar to restricting the sample tonon-movers, I avoid introducing a potential sample selection bias in the estimates. Table A.3 in the Appendixreport some summary statistics (mean, standard deviation) for the movers and non-movers samples for initialvalues of some variables.

22More details on British Transport Policy context are provided in Sanchis-Guarner (2012).23For example, the objectives of scheme “A6 Clapham Bypass” in Table A.13 were to “improve road safety”,

“relieve congestion” and “provide the opportunity for environmental improvement in Clapham by removing

15

or not to economic performance), I exclude observations in wards situated at the beginningand at the end of the road construction schemes as these locations might have been specific-ally aimed by transport policy (similarly to Faber, 2012).

However, even in these non targeted areas, it could still be the case that some individuals“treated” by new road links and others are not because of unobserved reasons. One cannotsimply compare individuals which are treated to those who are not. I rely on the smallgeography and micro data nature of the data and only compare workers who are close tothe new road links and that experienced changes in accessibility to different extents andat different points in time, depending on the size of the closest road scheme, its locationwithin the network and when it was completed. I define a 30-kilometre wide distance band24 around the 23 schemes undertaken during the study period and only compare individualslocated in wards within this band. I identify the impact of road construction by exploitingthe treatment intensities within this distance band, conditional on controls and individual-location fixed effects. Individuals placed within a given distance band are more likely to beexposed to similar shocks and road projects are quite unlikely to be aimed at specific indi-viduals within this narrowly defined distance band. This strategy implies that changes inaccessibility (predicted travel times), conditional on controls, can be considered exogenousto individual workers located close to the new links. Moreover, even if the placement of thenew roads would be endogenous to the individual labour market outcomes, its exact pla-cing within the road network, and thus how the accessibility of the wards is affected by thenew link, is likely to be exogenous to individual workers. Figure A.1 in the Appendix showsthe extent and location of the wards within this band. Distance bands are defined for bothwork and home location, but I experiment with the overlap of these bands in the robustnesssection (4.3) by restricting the commute distance between work and home.

In order to ensure that the accessibility index Art is uncorrelated with the underlyingarea trends, I further control for differential trends in the vicinity of road schemes. I controlfor the initial level of accessibility interacted with a linear trend. I include a set of nearestscheme dummies (23 schemes) interacted with year in equation (5) and I also control for thedistance to the closest scheme within the distance bands and a dummy which indicates ifthe scheme has been opened in the year of observation. These three sets of controls (closestscheme trend, distance to closest scheme trends and opened scheme trend) capture any un-observable trends specific to the scheme (shape, length, opening date) which could be cor-related with the placement of the scheme and initial employment values and future labourmarket outcomes around the vicinity of the road project. I also control for differential trendsof the wards (work and home) based on 2001 characteristics. I used CENSUS data (providedby CASWEB) at the ward level to calculate the share of population aged 15-64 (workforce)with higher education, mean age of population, share of population living on social hous-ing, the rate of unemployment, proportion of workers commuting using motor vehicles andthe average distance traveled to work. I also calculated a residential density measure, usingaddress counts data in 2001 from the National Statistics Postcode Directory (NSPD) and thearea of the wards in square kilometres, obtained from EDINA-UKBORDERS. Finally, I con-

through traffic”. See http://www.highways.gov.uk/roads/projects/6006.aspx for more informa-tion and the evaluation report.

24Narrower bands are used in the robustness checks.

16

trol for “travel-behaviour” in these wards controlling for the proportion of commuters thatuse motor vehicles and the average commuting distance. As they are time invariant, I inter-act the 2001 characteristics with a linear trend. These trends control for differential growthin labour market outcomes, e.g. wages, depending on the level of these ward characterist-ics in 2001, before the period of analysis which starts in 2002. All the specifications includeyear-industry fixed effects to control for year industry-specific shocks. In the estimation of(4) I introduce these ward characteristics in levels.

I furthermore control for individual personal and job characteristics. Some of these char-acteristics, for example full time status or occupation, could be regarded as “bad controls”because they could be outcomes of the transport policy. To help address this issue, I definethe level of the characteristics at the beginning of the period (the first time the individual isobserved within each of the two panel definitions) and I interact that level with a time-trend.By doing this I control for differential trends in the evolution of the labour outcome depend-ing on the initial level of the job and personal characteristics. I use occupation (9 categoriesdefined below), age group (10 year groups, from 16 to 65), full-time status, plant-type status(private sector, propietor/partner, public sector, non-profit sector), firm regulated by collect-ive agreement dummy and gender trends; and firm size (number of employees in the firm).Finally, I also control by commute distance dummies (5 kms bands) interacted with a lineartrend. These control for different trends on individual labour market outcomes around theschemes depending on how long apart are work and home locations. Without loss of gen-erality, to define the commute distance dummies I use straight-line distance between workand home postcodes. In the estimation of specification (4) I introduce these individual char-acteristics in levels, and in the estimation of (5), I introduce them interacted with a lineartrend (except for firm size).

Table 2 reports some balancing tests to check the validity of the empirical approach. Ionly report the tests for work ward, but the results for home ward are very similar. I test if ameasure of economic performance of a given ward at the beginning of the period of analysis(2001 or 2002 depending on the availability of the data) is systematically correlated with thedistance to the closest road scheme, within the 30-kilometre distance band (excluding thewards at the beginning and the end of the schemes). I include 2001 ward census controlsand scheme dummies, and I cluster the standard errors at the ward level. I test the effect ofdistance within the band on initial ward accessibility, average weekly wages, average hourlyearnings, average total hours worked, average commute distance (as crow-fly) and averagecommute travel time (as predicted by the network given work and home ward locations).The coefficients for distance to closest scheme are insignificant in all cases (only weaklysignificant for commute travel time) and close to zero. This indicates that, conditional onscheme dummies and census controls, there is not systematic relationship between howclose the wards is to a given scheme within the distance band, so the variation of accessibilitywithin the distance band is “as-good-as-random”.

[INSERT TABLE 2 HERE]

Finally, in the estimation of the empirical results I cluster the standard errors to correctfor arbitrary within-group correlation of the individual shocks in two distinct non-nestedcategories. In the case of the estimation of specification (4), I cluster at the individual level. In

17

the case of the estimation of (5), I implement 2-way clustering at the home-ward and work-ward level. By doing this, I allow the errors of the workers to be correlated at the treatmentlevel (wards). In addition, the standard errors are also robust to arbitrary heteroskedasticity.

There are a few limitations to this approach. The network data is simplified due to datalimitations, and some assumptions on the impact of new links additions on travel times aremade in order to be able to create a clear source of variation on travel times (full details areprovided in the Appendix). I focus on additions of new links to the network and disreg-ard other road improvements like dualling and resurfacing, which might also have positiveeffect on minimum travel times. Finally, the level of traffic congestion is fixed in the con-struction of predicted travel times to be able to separate the policy instrument (more roadkilometres) from the policy objective (reduce congestion). Given these limitations, there islikely attenuation bias due to measurement error in the calculation of the accessibility index,so my estimates would be a lower-bound of the real impacts.

One limitation of the paper is the lack of substantial evidence on how accessibility dir-ectly affects work and home spatial sorting. Using individual-work-home fixed-effects ex-ploits the variation stemming only from changes in accessibility for a given work-homelocation pair choice. Comparing the results using this approach and a standard individualfixed-effects estimation (which allows variation in accessibility to arise from job and homerelocations) sheds some light on the effect that spatial mobility has on the source of variationof accessibility when identifying the coefficients. But the causal relationship between access-ibility and work and home location is relevant on its own and requires further investigation.

Finally, my approach does not take into account that there might be other externalitiesstemming from the road construction, for example changes in congestion or pollution. Evid-ence (Duranton & Turner, 2011) suggests the elasticity of traffic flows with respect to newhighways is close to one, so it is quite likely that congestion was substantially relieved afterthe road construction. Nevertheless more evidence on the impact of road construction ontraffic flows is needed in order to assess the whole impact of the new schemes.

3.3 Data

The individual labour market outcomes data comes from the Annual Survey of Hours andEarnings (henceforth ASHE), which provides information about employees on an annualbasis. The ASHE is an annual survey of the earnings of employees in Great Britain, whichfrom 2004 replaced the New Earnings Survey (NES). Its primary purpose is to obtain in-formation about the levels, distribution and make-up of earnings, and for the collectiveagreements that cover them. It is designed to represent all categories of employees in busi-nesses of all kinds and sizes. The questionnaire is directed to the employer, who completesit on the basis of payroll records for the employee. The earnings, hours of work and otherinformation relate to a specified week in April of each year.

ASHE is based on a survey of a 1% sample of employees on the Inland Revenue PAYEregister (Pay As You Earn). The information is provided by the employer. The sample con-sists of employees whose National Insurance numbers end with two specific digits. It coversapproximately 160,000 individuals a year. The survey is designed as a panel of workers, inwhich the same workers are observed for multiple years. The sample is replenished as work-

18

ers leave the PAYE system (e.g. to self employment, retirement, overseas or death) and newworkers enter it (e.g. from school, self-employment, immigration).

ASHE contains information on the make-up of weekly earnings and hours worked (ba-sic, gross and overtime), occupation (using Standard Occupational Classification – SOC),industrial sector (using Standard Industry Classification – SIC), collective agreement status,whether the job is private or public sector, age, gender, postcode of workplace, and from2002, postcode of residence. In order to clean the data I drop the 0.5% top and bottom ex-treme values of the labour market outcome variables (wages and hours) and of the commut-ing times. I define total pay consistently over the whole period 2002–2008 25. I also removedobservations with negative values of the variables and individuals which show inconsist-ency in their age or gender over time. I only keep main jobs for those individuals that havemore than one job in the same year. Finally, I drop the individuals for which earnings wereaffected by absence (loss of pay) and those paid at trainee/junior rates.

ASHE provides information on the occupation level of the individuals using SOC 2000codes from 2002. I define broader occupation codes using the first digit of the code 26. I alsodefine five broader industrial categories based on 2-digit codes from the SIC 2003 classifica-tion 27.

The great advantage of this data is the good quality of the earnings and hours inform-ation and the detailed information on the geographical location of both the workplace andthe place of residence. Furthermore, its panel structure allows us to control for unobservabletime invariant characteristics of the workers which might be correlated with the variable ofinterest. However, the survey contains information only on workers who are employed, soI am unable to observe unemployment spells. The responses to the survey are providedby employers, rather than the employees themselves, so personal and household informa-tion about the employees is very limited (essentially gender and age). However, most of thehousehold characteristics change slowly over time so the use of individual and individual-location fixed effect control for time-invariant heterogeneity captures most of these charac-teristics.

Table A.2 provides summary statistics for the main variables of analysis for both paneldefinitions. The table shows the mean, standard deviation, maximum value, minimum valuefor the overall, between and within dimensions of the panels for basic weekly pay (wages),basic weekly hours worked, accessibility from workplace, accessibility from home and pre-dicted road travel time between work and home, all in natural logarithms. The top panel dis-plays statistics for the panel defined using individual fixed-effects for the estimation sample(320,105 observations). We can see there are 78,305 individuals which appear an average

25Different stratifications of ASHE define gross pay differently. I use the most recent definition: total (gross)pay=basic pay + incentive pay + shift and premium payments + overtime pay + other pay.

26They correspond to: 1 Managers and Senior Officials; 2 Professional Occupations; 3 Associate Professionaland Technical Occupations; 4 Administrative and Secretarial Occupations; 5 Skilled Trades Occupations; 6Personal Service Occupations; 7 Sales and Customer Service Occupations; 8 Process, Plant and Machine Oper-atives; 9 Elementary Occupations.

27I define 10 broad categories based on 2-digit SIC codes: 1 Primary Activities - agriculture, fishing, forestry,mining - (SICs 1-14), 2 Manufacturing (SICs 15-37), 3 Energy & Construction (SICs 40-45), 4 Wholesale &Retail (SICs 50-52), 5 Hotels & Restaurants (SIC 55), 6 Transport & Communications (SICs 60-64), 7 FinancialIntermediation (SICs 65-67), 8 Real Estate & Producer Services (SICs 70-74), 9 Public Administration, Education& Health (SICs 75-85), 10 Other Activities (SICs 90 and over).

19

of 4.1 times. For all variables we can see that most of the variance is between groups. Thebottom panel displays the summary statistics for the panel defined using individual-work-home fixed-effects. Even if the number of observations is smaller 28, the number of groupsis larger (79,868), and groups appear fewer times in the panel (3.35 times). In this setup ofthe panel the concentration of the variance in the between dimension is even more evident,especially for the accessibility and the travel time variables.

[INSERT TABLE A.2 HERE]

Apart from ASHE, other datasets are used to construct the accessibility measures (moreinformation is available in section A.1. in the Appendix). I use the National Statistics Post-code Directory (NSPD), which provides a look-up between postcodes and higher UK geo-graphies, to allocate the individuals to work and home wards. This dataset also providesinformation about the number of residential addresses in each postcode, which is used tocalculate population-based effective density measures. I also use data from the 2001 Brit-ish Census to calculate socio-economic controls at the ward level, which are included in theestimation of equation (5).

4 Results

4.1 Individual fixed effects

Table 3 presents the results for the estimates of equation (4) on log wages (log of basic weeklypay). In this section I focus on the results on wages, while in sections 4.3.2 I explore the ef-fects on other labour market outcomes (mainly on labour supply). These results are obtainedusing the 30 kilometre band both from work-ward and home-ward, and excluding the wardswhere the nearest scheme begins and ends. The table reports the estimated coefficients forlog accessibility from both work ward, log accessibility from home ward and log of pre-dicted travel time between work and home wards. The variables are constructed as definedabove. The table reports the parameters and the standard errors in square brackets. Theseare clustered at the individual level to allow for arbitrary correlation of individual shocksover time 29. The different columns report results for different specifications where differentsets of controls are introduced. These are indicated in the last rows of the table. The samplesize is 320,105 observations, for 78,305 individuals observed an average of around 4 periodsof time (see table A.2). All observations include sic-by-year fixed-effect to control for nation-wide industry-specific shocks. The star symbols next to the estimated coefficient indicate thelevel at which the coefficients are statistically different from zero.

[INSERT TABLE 3 HERE]28If an individual now keeps a specific work-home combination only for one period, it drops from the panel

estimation. This is why the number of observations is smaller.29Alternatively I could clustered the s.e. at ward levels, but as individuals move over space, these levels are

not nested. However, when we allow for spatial mobility the relevant clustering level might be the individual,not the area, as he is choosing the “level of treatment” by his moving decisions. In practice, the main resultsrely on fixed-location estimates from equation (5), where I am able to cluster at the area treatment level (thepair work ward-home ward location).

20

Column 1 reports the Ordinary Least Square results (OLS) where only sic-by-year dum-mies are included. I use a standard definition of accessibility based on Art in columns (1)and (2). Given the large number of observations, these partial correlations are precisely es-timated. I find a positive and significant relationship between accessibility from work onwages, a negative relationship between accessibility from home and wages and a positivecoefficient on the effect travel time on pay. The positive effect of accessibility from work onwages and hours could be due to agglomeration externalities or to the fact that professionals,who earn more and work full time, are concentrated in work locations where accessibility ishigher. At the same time, workers in low paid jobs such as basic services could also be livingin these locations and this could explain the negative coefficient of accessibility from homeon earnings and hours. Finally, these results suggest that longer commutes are capitalisedinto higher wages as suggested for example by Manning (2003).

In column 2 I introduce individual fixed-effects to control for unobserved individual het-erogeneity that might be correlated with accessibility levels and wage levels. I now estimatethe effects using the within-individual variation and exploiting the changes in accessibilityover individual means over time. The coefficient for accessibility from workplace is reducedsubstantially, and that of accessibility from home becomes positive. This suggests that indi-viduals which have unobservable characteristics correlated negatively with wages are loc-ated in wards in which accessibility from home is higher. This is in line with the previousargument that less skilled workers live in denser areas in which they can access more jobseasily, as predicted for example by the spatial mismatch hypothesis.

In column 3 I use Aiht, the measure of accessibility that keeps ward employment fixed to2001 levels. I do this in order to avoid using the variation which comes from spatial reloca-tion of employment, as these changes and individual wages are probably influenced by thesame unobservable factors. The coefficient barely moves, which suggests that both variablesare highly correlated and that the instrument is very strong. In column 4 I control for ini-tial accessibility (both for work and home) and initial job and personal characteristics, andthe coefficients remain unchanged. In column 5 I add (closest) scheme dummies, distance tothe scheme and opened scheme dummies to control for specific unobservables around thescheme. In column 6 I add 2001 census controls and commute band dummies. This is mypreferred specification as it includes all the individual, job and area level controls. The coef-ficient of accessibility from workplace is 0.038, that of accessibility from home is 0.021 andthe effect of predicted road commute time is 0.014. The are all significantly different fromzero at 1% level. These elasticities are fairly small, suggesting that doubling accessibility atworkplace (what, given the summary statistics in 1 would require large amounts of newlinks) would result in an increase in wages of around 4%, and increasing accessibility fromworkplace of around 2%. Even if small, the estimates of the effect of accessibility from workare similar in magnitude to the estimates of the effect of market potential on wages providedfor example by Combes et al. (2008) or Mion & Naticchioni (2009).

In columns 7 and 8 I provide additional estimates for robustness. In column 7 I estimatethe same specification as in 6 but using instrumental variables, and the coefficients changelittle. The F-statistic of the first-stage (Kleibergen-Paap rk Wald F statistic) indicates thatthe instrument is very strong. This is not surprising for two reasons; the first one is thatAiht and Aiht share one source of variation, changes in predicted travel times, so there is a

21

mechanical relationship between the two. The second one is that, as shown in Gibbons et al.(2012), new roads attract employment to areas around where they are placed, so the changesin travel times are good predictors of changes in local employment. The second additionalresult, presented in column 8, uses the controls no only in levels (initial values) but alsointeracted with a linear trend (when possible). The results barely change with respect to themain results in column 6.

As discussed above, the identification of the three coefficient relies on the individualvariation stemming from spatial relocations. This implies that individuals might be chan-ging jobs or home as a result of changes in accessibility or due to unobservables correlatedwith these changes. In the next section I investigate if using the variation of accessibilityfor individuals while they keep their location fixed has any impact on the estimates of theeffects.

4.2 Exploring the role of spatial mobility

This section explores the impact that restricting spatial mobility has on the estimates of theeffect of accessibility changes on wages. I do this using two methods. The first one is re-stricting the sample used in the estimates of table 3 to individuals that do not move homeward, work ward or both during the period of observation. This implies dropping observa-tions from the sample, and running the results on individuals that meet the stable locationconditions. The second approach uses individual-location fixed effects. As explained above,this has the advantage of not “dropping” individuals but it allocates different individualfixed-effects depending on their work-home location 30. Individuals might change locationsover time but, conditional on having at least two observations for a given location choice,we use the information for the whole spell in which in which the worker is observed. Table4 presents these results.

[INSERT TABLE 4 HERE]

The dependent variable in this table is log of basic weekly pay. All the columns include allcontrols either in dummies or in trends, depending on the specification. Column 1 replicatesthe main result of table 3 for comparison. In this specification individuals are allowed tochange work and home wards at any point in time. In columns 2-4 I restrict the sampleto workers which keep stable locations (either work or home, or both) over the period ofstudy. Column 2 drops the individuals that change homes at least once during the periodof analysis. Almost 90,000 observations drop. The coefficient of the effect of accessibilityfrom work changes a little with respect to column 1, and the same for the effect of traveltime. However, the coefficient of the effect of accessibility from home become insignificant.This is not due to changes in the size of the coefficient (in fact, this increases), but to thefact that standard errors increase dramatically when we restrict the sample to non-home-movers. This suggests that most of the identification of the effect of accessibility from homeon wages in column 1 relies on variation stemming from spatial home relocation. Even if

30Of course, the number of observations in both panels is different because we need individuals to keepa stable location for at least two periods in order to be able exploit the within individual-location variation.When we use individual fixed-effects individuals need to appear more than one but they can be located indifferent wards.

22

this strategy does not directly test if workers move as a consequence of road construction, itsuggests that they move to areas in which accessibility is increasing due to road construction31. This means that residential sorting is a consequence of new links.

In column 3 I restrict the sample to workers that do not change work ward during theperiod of study, but they can change home location. The sample is even smaller than be-fore, around 190,000 observations as compared to 320,000. The coefficient of accessibility forworkplace remains significant at 1%, but both the size of the parameter and the standarderrors increase. The estimate become less precise because there is less variation from whichto estimate the effect, given that workers keep stable work wards in this sample. The coeffi-cient of predicted travel time on wages is also significant and positive. In column 4, I restrictthe sample to individuals that do not move either work or home ward during the period ofanalysis. The sample in which I estimate this results is substantially smaller than the originalone, around 150,000 individuals. The coefficients are very similar to those of column 3. How-ever, the coefficient of the effect of travel time collapse to zero which suggests that workersget compensated for their commute when they change jobs, and for a given job, changes intravel time due to road construction do not have a direct effect on pay. The standard errorsfor all 3 variables are larger given that I have less variation in both accessibility and traveltimes when work and home locations are stable over time. In both columns 3 and 4 onlyaccessibility from workplace is significant, an the elasticity is around 2.5%.

The difference in size of the coefficient of the effect of accessibility from workplace ob-tained in column 1 and those obtained in column 3 or 4 is quite substantial. Specifically, thereduced-form estimates of column 4 are almost one order of magnitude larger than the onesusing individual fixed-effects. The difference in the size of the coefficient originates fromkeeping work ward locations for individuals fixed over time, regardless of home mobility.This would suggest that individuals capitalise any benefits that firms draw from increasesin effective density to a larger extent when they stay in the same job-area for a period oftime.

For simplicity, we can separate the expected effect of accessibility from workplace onwages in the effect due to job mobility (better job matches) and that due to changes in densityat current workplace (related to agglomeration economies in-situ). If the channels throughwhich increases in effective density are positively affecting wages are linked to area specificagglomeration economies (for example knowledge spillovers or labour market pooling), it issensible to assume that workers would increasingly capitalise these benefits over time whilestaying at the same work location. In these specification we are keeping work ward fixed, soworkers could still change jobs within their ward (this is explored in section 4.3.1). Recentevidence suggests that workers benefit from better job matches in larger labour markets(Harmon, 2013), but my finding suggests that the effect of changes in the size of labourmarkets on wages works mostly via location-specific channels, potentially agglomerationeconomies. Moreover, recent evidence from De la Roca & Puga (2012) suggests that workersin more agglomerated areas do not necessarily have higher initial ability, but the effect ofdensity in wages increases with workers’ experience in agglomerated areas.

31Recall that accessibility changes are non-negative as in the calculation of Aiht ward employment is fixedto 2001 levels and changes in optimal travel times between locations are non-negative. See Appendix for moredetail.

23

The difference in the coefficients between column 1 and columns 4 could also be due tosample selection, as I drop the observations which do not meet the stable location criteriain the different columns. To check the extent to which this is an issue, in table A.3 I providesome summary statistics on the initial values for some variables for the whole sample andfor individuals that move any of the two dimensions (movers) or those that keep stablework-home locations over time (stayers). We can observe that these differences are minorand not statistically significant, especially if we compare the stayers and the whole sample.Therefore, the difference in the size of the coefficients is likely due to the effect that restrictingmobility has on the source of variation for the accessibility and travel time variables.

In columns 5 to 7 of Table 4 I use individual-location specific fixed-effects. By includingworker-ward specific fixed-effects I use a larger number of observations and exploit thevariation for the same individual for different locations. The coefficients are very similarto those of columns 2 to 4 due to the absence of sample selection bias in the estimates asdiscussed above. In my preferred specification, that of column 7, the elasticity of accessibilityfrom workplace on wages is around 3.1%, and there is no effect of accessibility from homenor predicted commute travel time. This coefficient implies that doubling accessibility fromwork, for a given work-home location, would increase weekly wages around 3%. This effectis not negligible but it is fairly small. Even if this coefficient is very similar to that of column4, I prefer this specification because it uses a larger range of variation and a larger number ofobservations. Therefore, the robustness checks of next section are performed with referenceto these coefficients.

4.3 Additional results and robustness

4.3.1 Robustness checks

The main result from the previous section is that, once we control for spatial mobility, onlyaccessibility from workplace has a positive significant effect on workers’ wages. In this sec-tion I test the robustness of these findings. Only the coefficients for work accessibility, homeaccessibility and predicted travel time are reported in the tables. Unless specified, all theresults are obtained using the 30-kilometre band (excluding the extremes) and include allthe controls and trends. The main results are reproduced in the tables to ease comparison.

In order to link the empirical results to the theoretical predictions, in equation (5) I in-clude accessibility from both work and home locations and predicted commuting time. Butthese three variables might be very correlated, threatening the correct identification of theparameters. For this reason, in columns 2 to 5 in table 5 I test the robustness of the findings tochanges in the specification. We can see that the effect of work accessibility remains signific-ant even if when I exclude travel time or home accessibility for the specification, and that byitself, home accessibility remains insignificant. This rules out the possibility that, due to col-linearity, work accessibility would be capturing all the effect of effective density for workerswhich have work and home locations close to each other. In column 6 I exclude observationsfor which work and home wards is the same (when both accessibility measures would beoverlapping). The main result remains mostly unchanged. In column 7 I use one lag of theaccessibility measures and still find a positive significant impact of accessibility from work

24

on wages. Finally, in column 8 I use the accessibility indices as defined in Art and instrumentthem with Aiht. The coefficient is very similar to that of column 1.

[INSERT TABLE 5 HERE]

In the main results, the individual-location fixed effects are defined using work and homeward locations. Wards are very small units and the incidence of house and job changeswithin these small units is very small and unlikely to affect the main results. However,changes in the size of the coefficient when we add additional mobility restrictions pointstowards the channels through which accessibility is affecting wages. Table 6 explores this.Workers’ home postcode are available in ASHE and, given that in Great Britain postcodesrefer roughly to house blocks, we can track individuals almost at their address level. For joblocations, ASHE does not provide information on plant identifiers, but it provides informa-tion on work location postcodes, firm sector and firm identifier. I use a combination of thesethree variables to define specific plant locations. In column 2 of table 6 I use individual-homepostcode-plant location effects, which adds additional mobility restrictions with respect tothe main specification of column 1. The qualitative results remain unchanged: only accessib-ility from workplace has an effect on wages and the elasticity is now equal to 2.5%. Columns3 and 4 investigate if this change in the coefficient is due to house or plant moves within thewards. In column 3 I use individual-home postcode-work ward fixed-effects: this impliesthat individuals keep the same house and work wards but are allowed to change plantswithin the ward. The coefficient is larger than in column 2 and more significant, but smallerthan that of column 1. In column 4 I use individual-home ward-plant specific fixed-effects,so I exploit variation in accessibility can still stem from changes in home postcodes withinthe wards. The result is very similar to that of column 3. The fact that the coefficient is lar-ger when we allow for in-ward job movements suggests that one of the channels throughwhich effective density is being capitalised into wages is by very local plant moves. Thereare two plausible explanations for this: firms located in wards which experience changes inaccessibility are experiencing productivity shocks and are able to pay more, and workerscould take advantage of this by moving to other plants close-by and capitalising these gainsvia higher wages. Another explanation would be that labour markets are effectively largerdue to faster road connections now and firms located close to the schemes start competingfor specialised labour with a larger pool of firms. In order to keep their workers, they bidtheir wages up. This would imply that employed workers market-power increases due tothe road construction. These explanations are consistent with the findings in Gibbons et al.(2012).

[INSERT TABLE 6 HERE]

A potential concern for the validity of the results would be that the is a lot of overlapbetween the work 30-kilometre band and the home 30-kilometre band, so we are unable toseparate the effects of work and home accessibility and that is why I only find significantimpacts from one of the dimensions (work). A first attempt to control for this is to includecommute distance trends in the estimations, as I do in the main results. I explore this furtherin table 7. I restrict the sample to individuals for whom work and home locations are locatedat different distances.

25

Given that the bands around the schemes are defined using crow-fly distance, I do thesame here to restrict for commuting distances so it is a better control for the overlap of thework and home distance bands. I use restrictions for 30, 20, 15 and 10 kilometres 32. Whenthese distances are small the overlap between the work and home distance bands would belarge, while when they are larger, there would be less overlap between the work and homeareas. Column 1 reproduces the main results. In columns 2 to 5 I use different work-homedistance restrictions and the main conclusions remain unchanged. The coefficients changea bit with respect to column 1, but they are very similar and still significantly positive. Incolumns 6 to 10 I restrict the observations to workers for which the closest scheme is thesame for work and home locations. This is not necessarily the case in the previous results,as depending on the location of the schemes and the work and home wards workers mightbe closer to different schemes in the two dimensions. The results in columns 6 to 10 are verysimilar to before.

[INSERT TABLE 7 HERE]

Tables A.4 and A.5 check the robustness of the findings to excluding different wardsand using different size of the distance bands around work and home locations. Column2 replicates the main results. The coefficient on accessibility from workplace changes verylittle when I do not exclude any or when I exclude more wards at the beginning and end ofthe schemes. Column 6 is a bit different. I exclude not only the wards at the extreme of theschemes but also the wards along which the scheme goes through. This strategy is similarto that used by Faber (2012). The coefficient is smaller but less significant (only at 10%). Thissuggests that most of the effect is drawn from locations very close to the schemes but notnecessarily aimed at them. This indicates that the effects are very localised. In table A.5 Iexpand (columns 1 to 4) and reduce (columns 6 to 10) the size of the work and home dis-tance bands. The coefficients are very similar across the columns only becoming insignificantwhen the size of the bands is very small and the number of observations very reduced.

In table A.6 I use alternative definitions of the accessibility index Aiht. Column 1 repro-duces the main results where I use ward employment in 2001 as the measure of economicsize of the wards. Columns 2 and 3 use different measures: 2001 number of plants and 2001residential address counts (a proxy for population). The coefficients are very similar to themain results. Columns 4 to 6 use ward 2001 employment but change the form of the costfunction and/or the degree of the distance decay. As expected the size of coefficients is dif-ferent, as the weights given to the different wards in the computation of the accessibilityindices change. But the qualitative results remain unchanged.

In table A.7 I exclude observations of workers located in London 33. In column 1 I rep-licate the main results. In column 2 I exclude individuals working in London at least onceduring the panel, in column 3 individuals living in London and in column 4 individualsliving or working in London at least once while observed. The results remain very similarto the main findings.

Finally, In table A.8 I explore the effect of using individual-location fixed-effects further.

32The average distance between work and home locations in the individual-location panel is 9 kilometres,the 25th percentile is 2.3 kilometres, the median is 5.3 kilometres and the 75th percentile is 11.7 kilometres.

33Defined using the local labour market definition of London as in Gibbons et al. (2010)

26

The main results the observations for workers irrespectively of their work-home location, Ijust allocate a new fixed-effect when a new location-pair is chosen. Of this sample, there willbe some individual that always keep the same location (in both or one of the dimensions) sothe individual-location fixed-effects would be equivalent to individual fixed-effects. Otherindividuals would have more than one spell, meaning the get different sets of individual-location fixed-effects over time. In columns 2 to 3 I restrict the sample to those who keepthe same work-home location over time (column 2), keep the same work location (column3) or keep the same home location (column 4). The results remain very similar to the mainfindings. In columns 5 to 7 I focus on individuals that get more than one individual-locationfixed effect. The samples are quite small, and yet I still get very similar results, especially forcolumn 5.

4.3.2 Other labour market outcomes

In this section I explore the impact of changes in accessibility in other labour market out-comes. In particular, I study the effect of changes in effective density due to road construc-tion on labour supply (hours and full-time status) and hourly earnings. I use the same iden-tification strategy than in the previous sections.

Changes in effective density might affect hours worked via several channels. If wages arereacting as a consequence of road construction, workers’ might have an incentive to worklonger hours in order to gain more. Alternatively, if labour markets become more competit-ive because better connections are making then tighter (Detang-Dessendre & Gaigne, 2009),for a given commuting cost workers might work more in order to keep their jobs or to es-tablish themselves in the labour markets (Rosenthal & Strange, 2008). However, number ofhours worked per week might not be a flexible outcome to adjust given a contractual setup.For this reason, I explore if workers are more likely to work part-time or full-time as a con-sequence of the road investments. As male and female labour supply might react differently,I also study the differences by gender for this variables.

Table 8 presents the results for specification (5) using different labour market outcomesas dependent variables. The outcomes can be divided into basic and gross/total (includingoverpay and overtime). Most of the outcomes are weekly, but I also test the effect for totalannual pay which might include additional extras from for example bonuses. Columns 1 to 3display the results for the wage variables; using gross wages or annual wages does not causelarge changes in the coefficients. Columns 4 and 5 test the effect of changes in accessibility onweekly hours worked, both basic and total. As for pay, only accessibility from work has aneffect on labour supply, and the coefficient is slightly larger for total hours (even thought notstatistically significant from the one on basic hours). Columns 6 and 7 study the impact onhourly earnings, which are defined as the weekly pay over weekly hours. Focusing on theeffects on basic outcomes, the elasticity of wages with respect to work accessibility is largerthan the one of hours. However, when I test the effect on the ratio of both variables, basichourly earnings, the effect is weaker but still significant and positive. The elasticity is lowerthan in the case of weekly wages, around 1.1%, which is consistent with the elasticity of paybeing larger than that of hours. For total hourly earnings, the coefficient is non-significant.

[INSERT TABLE 8 HERE]

27

We would expect that increasing access to jobs would lower barriers to participation andhave an effect especially on women, for example if they commute shorter distances thanmen (as documented for example by Rosenthal & Strange, 2012). I cannot directly test thisin my data, but I can explore if there are gender differences on labour supply for employedwomen. I do this in table 9. In column 1, I test if accessibility has an effect on workingfull-time (versus working-part time). This result is reported in column 1; I find a positiveeffect of accessibility from work on working-full time, but negative significant effects ofaccessibility from home and commute travel time on this outcome. As discussed above, asit is mainly full-time workers who benefit from the accessibility-induce wage increases, thiscould results in workers moving from part time to full time status while remaining in thesame work-home locations. Commute distance and labour supply can be positively relatedbecause, the longer the commute is, the more worthy is to work longer hours in order tocompensate for the time spent travelling to work. There is no straightforward explanationwhy better home accessibility could increase the probability of working full time, given thatthis estimates are obtained for fixed work-home locations.

It could be that there are substantial gender differences with respect to this margin ofadjustment when road construction takes place. I explore this in the remaining columns oftable 9. Columns 2 and 3 test the effect of accessibility and travel time on full time statusseparately for male and female workers. I only find effects for female workers, and only foraccessibility from workplace (positive) and commute time (negative). The impact of workaccessibility is also only significant for women when I test the impact on basic hours worked(columns 4 and 5). Only female workers seem to be working longer as a result in increasesin work accessibility and reductions in commute travel time (as suggested by Black et al.,2013). This suggests that women are more likely to adjust their labour supply when wagesincrease as a result of changes in work accessibility. As I report in the previous section, theaccessibility wage premium is larger for women, which could explain the result.

[INSERT TABLE 9 HERE]

Section A.2.1. in the Appendix explores the heterogeneity of the effects.

4.4 Discussion

Section 2 discusses the multiple channels through which changes in effective density viathe construction of new road links can affect individual labour market outcomes, especiallywages.

For individuals working and living in different locations, I test the impact of accessibilityfrom workplace and accessibility from home, and also the impact of changes in predictedtravel time between work and home on wages. We could explain the impact from work-place as a result of agglomeration economies. Firms might become more productive due tothe transport investments (as suggested by Gibbons et al., 2012) and workers might be ableto capitalise these gains, especially if they have market power over firms when specialisedlabour is scarce. Changes in accessibility from home could have an impact on wages if work-ers get better labour market matches as a result of increased density. Finally, lower commutetimes after road construction could be also capitalised into higher wages if workers become

28

more productive when they have to spend less time traveling to work (van Ommeren &Gutierrez-i-Puigarnau, 2011).

Results in table 3 provide the estimates when I use individual fixed-effects and I use thevariation in the three variables for an individual irrespective of his location. All three coeffi-cients are significant and positive. Given the summary statistics reported in table 1, between2002 and 2008 the average increase in wages due to increases in accessibility from workplacewould be 0.012%, and that from home would be 0.0064%. When we get close to the site ofthe new links, these effects increase (up to 0.077% and 0.043% within 10 kilometres), but theyare still small. The size of the coefficient of work accessibility is similar in nature to the es-timates of the impact of “market-access” (or market potential) on wages (some examples areMion & Naticchioni, 2005; Fingleton, 2006; Combes et al., 2008; Amiti & Cameron, 2007). Intheir review of the literature, Graham & Melo (2009) find that these estimates are normallybetween 0.02 and 0.2. Even if my estimate are not directly comparable as my definition ofmarket-access only relies on variation due changes in optimal travel times, the coefficientfor work accessibility falls within that range. In fact, in column 7 of table 3, when I also usethe variation in access stemming from changes in local employment, the coefficients are onlymarginally larger.

The estimates using individual fixed-effects rely on variation stemming from spatial re-location of workers across work and home wards. As these moves could be due to road con-struction, or to unobservables related to them, in table 4 I use two methods to explore theimpact of mobility on the estimates. I either restrict the samples using mobility restrictions orI use individual-location specific fixed-effects. The coefficient using both methods are verysimilar. The results change substantially with respect to table 3; the coefficient of accessib-ility from home becomes insignificant while the coefficient of accessibility from workplaceremains significant and increases significantly. At the same time, the coefficient of home ac-cessibility becomes insignificant when using individual variation for a given location choice.This indicates that the variation in home accessibility stemming from home mobility is es-sential to be able to identify the effects of accessibility from home on wages. This suggeststhat home sorting is an outcome of the reduction of travel times. Given the construction ofthe travel times, the changes I exploit are always non-negative, which implies that workersare moving to locations where accessibility is increasing.

Once individuals have chosen their residential location, access to more jobs from homedoes not seem to have an effect on their wages neither on hours, conditional on predictedcommuting time. The theoretical channel through which better accessibility from homewould have an effect on wages once controlling for sorting is unclear. The spatial mismatchhypothesis predicts positive effects of better access to jobs from home on labour market out-comes, especially on the probability of becoming employed or on the length of the unem-ployment spells. Better home access can boost wages if workers can find better job-matcheswhen changing jobs. However, from the results obtained using the sample, in which all theindividuals are already working and keep their work-home location fixed, I do not find anyevidence in favour of the spatial mismatch hypothesis. This could be due to householdshaving already optimised their residential location in order to have access to better jobs andshorter commutes so further increases in accessibility from home do not have any impact ontheir wages or hours worked.

29

When I use individual-work-home fixed-effects the coefficient of accessibility from workon wages increases largely, almost tenfold. Spatial sorting could be explaining part of thedifference in size between my estimates and the coefficients found in previous evidence.The identification strategy I follow substantially helps to reduce the bias caused by the spa-tial sorting of workers. As explained above, I control for sorting at the individual level (bothwork and residential sorting) by using the individual-work-home fixed effects. This way,for the identification of the effects I exploit the changes in accessibility over time for anindividual while staying a given location combination. Following the discussion above, se-lection of workers does not seem a major issue, so the difference in the estimates when Irestrict spatial mobility are likely due to the difference source of variation used when I ex-ploit the variations in accessibility in-situ. This way, the effects of agglomeration on wageswould work “while workers stay in the same location”. Moreover, as outlined above, theinstrumental variables strategy (using Art) helps to eliminate the bias induced by spatialrelocation of workers across space which might be driven by the changes in accessibility. Inother words, I tackle the sorting of workers at the ward level and avoid using these endo-genous variation in the estimation of the effects 34.

The positive effect of accessibility from work on wages and hours, given that in tables3 and 4 I control for commuting travel time, could be working through spatial competitionor agglomeration externalities. If firms in which the workers are employed can access alarger pool of workers, then employees might behave more competitively and work harder(especially high skilled workers, as suggested by Rosenthal & Strange, 2008). Workers couldbecome more productive in these denser and better connected areas, as has been suggestedand empirically verified in the agglomeration economies literature (see Duranton & Puga,2004, for a review) or in the new economic geography literature (Krugman & Venables, 1995;Redding & Venables, 2004, for example). Given that the effects are obtained for a given work-home location, the channels is work-area or firm-area specific. Faster travel times increasethe effective size of the labour markets (Harmon, 2013). If firms compete for specialisedworkers in the local area and competition for these workers increases with effective densitychanges, firms might have to pay workers higher wages in order to keep them in the firm.

In my preferred specification, when I use individual-work-home specific fixed-effects,the coefficient of work accessibility is equal to 0.309. Given the average growth in accessib-ility between 2002 and 2008, almost 0.1% of the growth in wages could be attributed to roadconstruction. This is a small effect but significantly different from zero. As we get closerto the new links (10 kms) it amounts to almost 0.63%. This effect is larger around schemeswhich had larger impacts on predicted travel times. Within 10 kilometres of the A5 NesscliffeBypass, opened in 2003 and which was almost 21.5 kms long (see table A.13), the averagechange in accessibility between 2002 and 2008 was 3.62%, which implies a wage growth of

34In their investigation of the determinants of individual wages using a large sample of French workers,Combes et al. (2008) find that differences in the skill composition of the labour force account for 40 to 50%of aggregate spatial wage disparities. They conclude that workers with better labour market characteristicstend to agglomerate in the larger, denser and more skilled local labour market. The intuition is that sortingof workers across space has an effect on individual wages because more skilled workers sort in specific areas.My approach is different as I explicitly control for the work and home sorting of the individuals. Furthermore,I focus on accessibility changes stemming from road construction and do not use the variation coming fromspatial changes in employment. This should help to reduce sorting issues in the same line as these authors.

30

around 1.12%. Around the A1(M) Ferrybridge to Hook Moor (19.2 kms and opened in 2006),the effect on wages is almost 1%. Given that the average growth in basic weekly pay for thewhole Great Britain between 2002 and 2008 was around 25%, the magnitude of the impact ofroads on wages is rather limited. For the whole Great Britain the size of the coefficients im-plies that road construction explains around 0.4% of total wage growth during this period.As we get close to the schemes, the effects become more important.

One of the main contributions of the paper is to provide robust estimates on the impactmarket potential on wages using a different strategy from most of the existing evidenceand focusing on a different channel. The papers mentioned above use a general marketpotential definition (Harris, 1954), which is similar in structure to (2) but uses geodesic timeinvariant distance as the measure of proximity between locations. As the distance betweenlocations (weights) is fixed over time, the effects of market potential on wages is estimatedusing changes in employment in the different locations. The endogeneity of these changesis dealt in different ways, for example constructing instruments with historical or geologicaldata (see for example Ciccone & Hall, 1996; Mion & Naticchioni, 2009; Combes et al., 2011).In the current paper, the main results are obtained using a variation of (2) which keepsemployment constant and changes over time only due to reductions on optimal travel timesbetween locations. This strategy focuses on changes induced by reduction on travel timesafter new links are added to the road network. As discussed above, it also avoids usingendogenous variation in Art from spatial changes in ward employment.

5 Conclusion

In this paper I investigate the effect of changes in accessibility, induced by road construction,on individual labour market outcomes. I make use of rich individual datasets and smallgeographical scale to be able to infer causality on the estimates.

I provide evidence of the effect of accessibility both from work and home wards on indi-vidual wages, and also on labour supply. Controlling for commuting time allows us to learnabout the potential theoretical channels through which accessibility might be impacting onlabour market outcomes. The methodology used overcomes several endogeneity problemscommon in the identification of effects of access to jobs: unobserved individual heterogen-eity, reverse causation between accessibility and spatial distribution of local employmentand endogenous placement of the new road links. Moreover, I explore the impact that spa-tial mobility (sorting) has on the significance and size of the estimates. When I exploit thechanges in accessibility for a given work-home location, I find positive effects of accessibil-ity from work on earnings and total hours. These results could be driven by agglomerationexternalities which are capitalised into higher nominal wages and by increased spatial com-petition which might make employees work longer hours. Workers seem also to be respond-ing to higher nominal wages by switching from part time to full time jobs. These results arevery robust to different specifications and across different groups of workers.

Two main conclusion can be drawn from these results. Firstly, my identification strategyseparates the effect of spatial sorting from other channels through which accessibility mightbe impacting labour market outcomes. As discussed in the text, once I focus on the variation

31

of accessibility for a given work-home location combination, there is no effect of accessibilityfrom home on wages or hours. This suggest that workers sort residential location in responseto changes in accessibility. On the other hand, the effect of accessibility from workplace issignificant even once controlling for sorting, and the size of the coefficient increases substan-tially. This result stresses the importance of accounting for sorting to be able to identify thechannels through which accessibility is affecting wages. My findings add to the evidence ofthe importance of learning on the determination density wage premium (see Baum-Snow& Pavan, 2012; De la Roca & Puga, 2012, for recent evidence). Given that I exploit the vari-ation on effective density only via changes in travel times (keeping initial ward employmentfixed) and for a given work-home location choice, my estimates do not capture the effect thatspatial sorting of workers (both at the aggregated and the individual level) has on the effectof density on wages (Combes et al., 2008). Moreover, the results suggest that learning effectmight be not only location-specific but even job-specific.

The main results estimate the effect of accessibility changes on wages, but I also explorethe impact on other labour market outcomes, namely hourly earnings and hours worked.I find effect of accessibility from workplace on all three variables, suggesting that workerswage gains are a combination of longer working weeks and increased hourly pay.

The findings in this paper provide new evidence on the effect of agglomeration external-ities and proximity to markets on individual wages and hours worked using a novel strategythat carefully tackles multiple endogeneity issues and a detailed dataset on road construc-tion. I use changes in optimal travel times between location stemming from road construc-tion as the source of changes in accessibility. Road construction can be directly used by policymakers to influence travel times and therefore affect effective density and proximity betweenlocations. Transport policy is a substantial part of economic policy and the estimates of thispaper help to shed light on the economic impacts that transport infrastructure investmentscan have on individual labour market outcomes.

32

Tables and figures

CHANGE LOG ACCESSIBILITY BETWEEN 2002 AND 2008

Wards Mean Standard Minimum Median 90th Maximum Proportiondeviation value percentile value of zeroes

All 10540 0.305% 1.634% 0.000% 0.004% 0.572% 52.368% 42.884%30 kms 4130 0.749% 2.546% 0.000% 0.122% 1.570% 52.368% 10.823%20 kms 2751 1.042% 3.069% 0.000% 0.223% 2.209% 52.368% 10.978%10 kms 1119 2.024% 4.592% 0.000% 0.556% 4.708% 52.368% 9.294%5 kms 539 3.445% 6.208% 0.000% 1.260% 8.504% 52.368% 5.937%ANNUAL LOG ACCESSIBILITY 2002–2008

Wards Mean Overall Between Within Minimum Median 90th Maximumstd. dev. std. dev. std. dev. value percentile value

All 73780 14.705 1.172 1.172 0.007 5.538 14.875 16.010 19.31630 kms 28910 15.234 0.807 0.807 0.011 10.312 15.324 16.164 18.19320 kms 19257 15.166 0.739 0.739 0.013 11.505 15.290 15.952 18.16610 kms 7833 15.030 0.750 0.750 0.020 11.505 15.195 15.778 17.6215kms 3773 14.970 0.755 0.755 0.028 12.075 15.156 15.705 17.082Source: Department of Transport, BSD and own author’s own calculations. Index calculated using 2001 ward employment and2002-2008 predicted travel times. Top panel in percentage values.

Table 1: Summary statistics ward log of employment accessibility. Based on Gibbons et al. (2012)

Balancing tests: 2001 work ward values 30 kms band

DEPENDENT VARIABLE:Log of Log of Log of Log of Log of Log of

employment weekly hourly total commute commuteaccessibility wages rate hours distance travel time

Log distance to closest 0.005 0.014 0.005 0.002 -0.01 -0.031*scheme [0.010] [0.009] [0.004] [0.005] [0.020] [0.017]Observations 4,037 3,627 3,624 3,624 3,761 3,749Scheme dummies Y Y Y Y Y YCensus controls Y Y Y Y Y YNotes: S.e. clustered at ward level. Wards up to 30 kms to the new schemes, exclude those located in the extremes of theschemes. The regressions include 2001 census ward controls (mean age, unemployment rate, proportion of workforce withhigher education, proportion of population living in social housing, ward address density, proportion of commutes by motorvehicles, average commuting distance) and closest scheme dummies. Commute distance and travel time values correspondto 2002, the rest to 2001 values. Results for home ward very similar. Sources: ONS, DfT, CASWEB and author’s own calcu-lations. * p<0.1, ** p<0.05, *** p<0.01.

Table 2: Balancing tests: work ward values for 2001 (30 kms band). Based on Gibbons et al. (2012)

33

!

!

!

!

!

!

!

!

!

!

!

! !

!

!

!

!!

A5

A66

A6

M6(T)

A43

A47

A1(M)A6

Leeds

OxfordLondonBristol

Glasgow

Cardiff

Bradford

Aberdeen

Newcastle

Liverpool

Cambridge

Notingha m

Birmingham

Manchester

2008 Major Road Network, type and location of new projectsNew routes Faster routes

(a) Major road network 2008 and new roads projects

!

!

!

!

!

!

!

!

!

!

!

! !

!

!

!

!!

Leeds

OxfordLondonBristol

Glasgow

Cardiff

Bradford

Aberdeen

Newcastle

Liverpool

LeicesterCambridge

Notingha m

Birmingham

Change in log accessibility between 2002 and 20080.00% - 0.05% 0.0501% - 0.50% 0.501% - 5.00% 5.01% - 52.37%

(b) Change in log accessibility (2001 employment)

!

!

! A63

A1(M

)

A650

A1(M)

LeedsBradford

Manchester

Change in log accessibility between 2002 and 20080.00% - 0.05% 0.051% - 0.50% 0.501% - 5.00% 5.001% - 52.37%

0 10 20 30 405Kil ometers

(c) Change in accessibility in Manchester-Leeds area (2001 employment)

Figure 1: Location of projects and changes in log accessibility 2002–08. Based on Gibbons et al. (2012)

34

DEPVAR: Log of basic weekly pay1 2 3 4 5 6 7 8

Log of accessibility 0.206*** 0.071*** 0.073*** 0.073*** 0.065*** 0.038*** 0.044*** 0.034***from work ward [0.005] [0.007] [0.007] [0.007] [0.007] [0.008] [0.008] [0.008]Log of accessibility -0.065*** 0.039*** 0.041*** 0.041*** 0.031*** 0.021*** 0.024*** 0.016**from home ward [0.006] [0.007] [0.007] [0.007] [0.007] [0.008] [0.008] [0.008]Log of travel time 0.130*** 0.029*** 0.029*** 0.029*** 0.029*** 0.014*** 0.028*** 0.012***between work & home [0.002] [0.002] [0.002] [0.002] [0.002] [0.002] [0.002] [0.002]Observations 320105 320105 320105 320105 320105 320105 320105 320105Individual FX N Y Y Y Y Y Y YReduced form access N N Y Y Y Y Y YInitial access N N N Y Y Y Y YJob-perso characteristics N N N Y Y Y Y YScheme dummies N N N N Y Y Y YDistance/open scheme N N N N Y Y Y Y2001 Census dummies N N N N N Y Y YCommute bands dummies N N N N N Y Y YInstrumental variables N N N N N N Y NAdd controls in trends N N N N N N N YNotes: DEPVAR stands for dependent variable. Clustered standard errors shown in brackets. Distance band 30 kms (excluding extremewards). Personal and job characteristics include log of firm size (employment) and female, initial 1-dig occupation, initial age group, andfull-time dummies interacted with a linear trend. Ward 2001 attributes include mean age of the population, proportion of household livingin social housing, proportion of household with higher qualifications, unemployment rate, address density, average distance traveled toplace of work and proportion of employees going to work by motor vehicles. All specification include sector-year dummies (10 categories).F-stat first stage in column 7 is equal to 297,528. “Ind” stands for individual. S.e. clustering (individual). * p<0.1, ** p<0.05, *** p<0.01.Source: ONS, DfT, CASWEB and author’s own calculations.

Table 3: Individual fixed effects results, log of basic weekly pay

DEPVAR: Log of basic weekly payIndiv FX Indiv FX + restrictions Indiv-ward FX

1 2 3 4 5 6 7Log of accessibility 0.038*** 0.045*** 0.244*** 0.261*** 0.039*** 0.246*** 0.309***from work ward [0.008] [0.010] [0.087] [0.080] [0.008] [0.090] [0.096]Log of accessibility 0.021*** 0.048 -0.008 -0.002 0.098 -0.004 -0.058from home ward [0.008] [0.111] [0.009] [0.086] [0.101] [0.007] [0.080]Log of travel time 0.014*** 0.017*** 0.000 -0.006 0.017*** -0.002 -0.002between work & home [0.002] [0.004] [0.003] [0.033] [0.003] [0.002] [0.021]Observations 320105 230630 189644 151398 297840 282961 267018Fixed-effects Ind Ind Ind Ind Ind-hw Ind-ww Ind-hw-wwStable home ward N Y N Y N N NStable work ward N N Y Y N N NAll controls in dummies Y Y Y Y N N NAll controls in trends N N N N Y Y YNotes: DEPVAR stands for dependent variable. Clustered standard errors shown in brackets. Distance band 30 kms (excluding ex-treme wards). Personal and job characteristics include log of firm size (employment) and female, initial 1-dig occupation, initialage group, type of plant, collective agreement and full-time dummies interacted with a linear trend. Ward 2001 attributes includemean age of the population, proportion of household living in social housing, proportion of workforce with higher education, un-employment rate, address density, average distance traveled to place of work and proportion of employees going to work by motorvehicles. All specification include sector-year dummies (10 categories), commute ring, scheme, distance to scheme and openedscheme dummies or trends. “Ind” stands for individual, “hw” stands for home ward, “ww”stands for work ward. S.e. clustering(individual/ward levels). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table 4: Mobility restrictions and ward fixed effects, log of basic weekly pay

35

DEPVAR: Log of basic weekly pay1 2 3 4 5 6 7 8

Log of accessibility 0.309*** 0.310*** 0.280*** 0.332*** 0.187* 0.299***from work ward [0.096] [0.099] [0.099] [0.088] [0.098] [0.093]Log of accessibility -0.058 -0.057 0.067 -0.042 -0.075 -0.050from home ward [0.080] [0.084] [0.079] [0.081] [0.079] [0.075]Log of travel time -0.002 -0.019 0.002 -0.012 0.011 -0.006 -0.001between work & home [0.021] [0.028] [0.023] [0.025] [0.026] [0.025] [0.021]Observations 267018 267018 267018 267018 267018 235087 268758 267018

Robustness MainNo travel Only travel Only work Only home Excl. same Lagged Instrumental

time time access access ward access variablesAll controls Y Y Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls. Distance band 30 kms (excluding ex-tremes). F-stat first stage in column 8 is equal to 1,015. * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table 5: Robustness: specification changes, log of basic weekly pay

DEPVAR: Log of basic weekly payIndiv-ward FX Indiv-house-plant FX

1 2 3 4Log of accessibility 0.309*** 0.250** 0.264*** 0.275***from work ward [0.096] [0.100] [0.099] [0.099]Log of accessibility -0.058 0.022 -0.022 -0.018from home ward [0.080] [0.083] [0.080] [0.082]Log of travel time -0.002 0.008 0.003 0.002between work & home [0.021] [0.019] [0.021] [0.019]Observations 267018 241624 261007 246818Fixed-effects Indiv-hw-ww Indiv-hp-pl Indiv-hp-ww Indiv-hw-plAlways same hward N N N NAlways same wward N N N NAll controls Y Y Y YNotes: Clustered standard errors in brackets. “Ind” stands for individual, “hw” stands for home ward,“ww”stands for work ward, “hp” stands for home postcode, “pl” stands for plant (work postcode-firm-sector combination). S.e. clustering (individual/ward levels). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS,DfT, CASWEB and author’s own calculations.

Table 6: Robustness: additional mobility restrictions and FX, log of basic weekly pay

DEPVAR: Log of basic weekly pay1 2 3 4 5 6 7 8 9 10

Log of accessibility 0.309*** 0.314*** 0.343*** 0.338*** 0.236* 0.349*** 0.354*** 0.392*** 0.396*** 0.263**from work ward [0.096] [0.099] [0.100] [0.117] [0.121] [0.102] [0.101] [0.099] [0.116] [0.125]Log of accessibility -0.058 -0.059 -0.118 -0.084 -0.023 -0.081 -0.078 -0.138 -0.109 -0.029from home ward [0.080] [0.087] [0.097] [0.097] [0.136] [0.093] [0.094] [0.103] [0.103] [0.139]Log of travel time -0.002 -0.003 -0.011 -0.006 -0.017 -0.003 -0.003 -0.012 -0.008 -0.019between work & home [0.021] [0.022] [0.022] [0.025] [0.029] [0.023] [0.023] [0.022] [0.026] [0.029]Observations 267018 253154 236864 218972 187466 230633 225837 215579 202248 176158Work-home distance All 30 kms 20 kms 15 kms 10 kms All 30 kms 20 kms 15 kms 10 kmsSame scheme N N N N N Y Y Y Y YAll controls Y Y Y Y Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls. Distance band 30 kms (excludingextremes). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table 7: Robustness: work-home distance restrictions, log of basic weekly pay

36

DEPVAR: Other labour market outcomes1 2 3 4 5 6 7

Log of accessibility 0.309*** 0.267*** 0.237** 0.196*** 0.199** 0.113* 0.067from work ward [0.096] [0.097] [0.093] [0.063] [0.078] [0.068] [0.056]Log of accessibility -0.058 -0.04 -0.044 -0.021 -0.022 -0.037 -0.015from home ward [0.080] [0.068] [0.118] [0.052] [0.050] [0.063] [0.055]Log of travel time -0.002 -0.011 -0.071** -0.025 -0.032 0.023 0.021between work & home [0.021] [0.022] [0.030] [0.023] [0.025] [0.017] [0.019]Observations 267018 264543 266795 267018 267018 267018 266795Type Basic Gross Gross Basic Total Basic TotalOutcome Pay Pay Pay Hours Hours Earnings EarningsPeriod Weekly Weekly Annual Weekly Weekly Hourly HourlyAll controls Y Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls. Distance band30 kms (excluding extremes). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table 8: Other labour market outcomes: wages and hours

DEPVAR: Labour supply by gender1 2 3 4 5

Log of accessibility 0.180** 0.109 0.297** 0.037 0.400***from work ward [0.071] [0.068] [0.143] [0.065] [0.099]Log of accessibility -0.163** -0.062 -0.277 0.057 -0.135from home ward [0.077] [0.053] [0.180] [0.038] [0.118]Log of travel time -0.079** -0.027 -0.108** -0.007 -0.038between work & home [0.031] [0.021] [0.048] [0.019] [0.039]Observations 267018 131163 135855 131163 135855Type Full time Full time Full time Weekly WeeklyOutcome Status Status Status Hours HoursGroup All Males Females Males FemalesAll controls Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes allcontrols. Distance band 30 kms (excluding extremes). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS,DfT, CASWEB and author’s own calculations.

Table 9: Labour supply: gender differences

37

References

Ahlfeldt, G. (2011). If We Build, Will They Pay? Predicting Property Price Effects of TransportInnovations. SERC Discussion Papers, 0075.

Ahlfeldt, G. & Wendland, N. (2011). Fifty Years of Urban Accessibility: the Impact of theUrban Railway Network on the Land Gradient in Berlin 1890-1936. Regional Science andUrban Economics, 41(2), 77–88.

Amiti, M. & Cameron, L. (2007). Economic Geography and Wages. The Review of Economicsand Statistics, 89(1), 15–29.

Baum-Snow, N. (2007a). Did Highways Cause Suburbanization? The Quarterly Journal ofEconomics, 122(2), 775–805.

Baum-Snow, N. (2007b). Suburbanization and transportation in the monocentric model.Journal of Urban Economics, 62(3), 405–423.

Baum-Snow, N. (2010). Changes in Transportation Infrastructure and Commuting Patternsin US Metropolitan Areas, 1960-2000. American Economic Review, 100(2), 378–82.

Baum-Snow, N. & Pavan, R. (2012). Understanding the city size wage gap. Review of EconomicStudies, 79(1), 88–127.

Black, D. A., Kolesnikova, N., & Taylor, L. J. (2013). Why do so few women work in NewYork (and so many in Minneapolis)? Labor supply of married women across US cities.Journal of Urban Economics.

Chandra, A. & Thompson, E. (2000). Does public infrastructure affect economic activity?:Evidence from the rural interstate highway system. Regional Science and Urban Economics,30(4), 457–490.

Ciccone, A. & Hall, R. E. (1996). Productivity and the density of economic activity. AmericanEconomic Review, 86(1), 54–70.

Combes, P.-P., Duranton, G., & Gobillon, L. (2008). Spatial wage disparities: Sorting matters!Journal of Urban Economics, 63(2), 723–742.

Combes, P.-P., Duranton, G., & Gobillon, L. (2011). The identification of agglomeration eco-nomies. Journal of Economic Geography, 11(2), 253–266.

Combes, P.-P. & Lafourcade, M. (2001). Transport Cost Decline and Regional Inequalities: Evid-ence from France. CEPR Discussion Papers 2894, Centre for Economic and Policy Research.

De la Roca, J. & Puga, D. (2012). Learning by working in big cities. Working paper.

Department of Transport (1998a). A new Deal for Transport: Better for Everyone. White paper -Department of Transport.

Department of Transport (1998b). A New Deal for Trunk Roads in England: Understanding theNew Approach. Report - Department of Transport.

Detang-Dessendre, C. & Gaigne, C. (2009). Unemployment duration, city size, and the tight-ness of the labor market. Regional Science and Urban Economics, 39(3), 266–276.

38

Donaldson, D. (2010). Railroads of the Raj: Estimating the Impact of Transportation Infra-structure. NBER Working Papers, 16487.

Donaldson, D. (2013). Railroads of the Raj: Estimating the Impact of Transportation Infra-structure. American Economic Review, forthcoming.

Duranton, G., Morrow, P., & Turner, M. A. (2013). Roads and trade: Evidence from the US.Technical report. Mimeo, Wharton School, University of Pennsylvania.

Duranton, G. & Puga, D. (2004). Micro-foundations of urban agglomeration economies. InJ. V. Henderson & J. F. Thisse (Eds.), Handbook of Regional and Urban Economics, volume 4of Handbook of Regional and Urban Economics chapter 48, (pp. 2063–2117). Elsevier.

Duranton, G. & Turner, M. A. (2011). The fundamental law of road congestion: Evidencefrom US cities. American Economic Review, 101(6), 2616–52.

Duranton, G. & Turner, M. A. (2012). Urban growth and transportation. Review of EconomicStudies, 79(4), 1407–1440.

El-Geneidy, A. M. & Levinson, D. M. (2006). Access to destinations: Development of accessibilitymeasures. Report, Research Services Section. Mn/DOT 2006-16, University of Minnesota.Dept. of Civil Engineering and Minnesota. Dept. of Transportation.

Faber, B. (2012). Trade Integration, Market Size, and Industrialization: Evidence fromChina’s National Trunk Highway System. London School of Economics, mimeo.

Fernald, J. G. (1999). Roads to prosperity? Assessing the link between public capital andproductivity. American Economic Review, 89(3), 619–638.

Fingleton, B. (2006). The new economic geography versus urban economics: an evaluationusing local wage rates in great britain. Oxford Economic Papers, 58(3), 501–530.

Gibbons, S., Lyytikainen, T., Overman, H., & Sanchis-Guarner, R. (2012). New road infra-structure: the effects on firms. SERC Discussion Papers, (0117).

Gibbons, S. & Machin, S. (2005). Valuing rail access using transport innovations. Journal ofUrban Economics, 57(1), 148–169.

Gibbons, S. & Machin, S. (2006). Transport and Labour Market Linkages: empirical evidence,implications for policy and scope for further research. Background paper for the Eddingtonreport to the Department of Transport.

Gibbons, S., Overman, H. G., & Pelkonen, P. (2010). Wage Disparities in Britain: People orPlace? SERC Discussion Papers, (0060).

Glaeser, E. L. & Kohlhase, J. (2003). Cities, regions and the decline of transport costs. Papersin Regional Science, 83(1), 197–228.

Graham, D. J., Gibbons, S., & Martin, R. (2009). Transport investments and the distance decay ofagglomeration benefits. Draft – Department of Transport.

Graham, D. J. & Melo, P. C. (2009). Agglomeration economies and labour productivity:Evidence from longitudinal worker data for gbs travel-to-work areas. SERC DiscussionPapers, (0031).

39

Gutierrez-i-Puigarnau, E. & van Ommeren, J. N. (2010). Labour supply and commuting.Journal of Urban Economics, 68(1), 82–89.

Harmon, N. A. (February 2013). Are Workers Better Matched in Large Labor Markets? Jobmarket paper, Princeton University.

Harris, C. D. (1954). The Market as a Factor in the Localization of Industry in the UnitedStates. Annals of the association of American geographers, 44(4), 315–348.

Highways Agency (2009). The Highways Agency: Framework Document. Highways Agency -Department of Transport.

Hymel, K. (2009). Does traffic congestion reduce employment growth? Journal of UrbanEconomics, 65(2), 127–135.

Inhanfeldt, K. R. (2006). A primer on spatial mismatch within urban labor markets. In Acompanion to urban economics, Blackwell companions to contemporary economics (pp. 404– 417). Wiley-Blackwell.

Jiwattanakulpaisarn, P., Noland, R. B., Graham, D. J., & Polak, J. W. (2009). Highway infra-structure and state-level employment: A causal spatial analysis. Papers in Regional Science,88(1), 133–159.

Jofre-Monseny, J. (2009). The scope of agglomeration economies: Evidence from catalonia.Papers in Regional Science, 88(3), 575–590.

Krugman, P. & Venables, A. J. (1995). Globalization and the inequality of nations. TheQuarterly Journal of Economics, (pp. 857–880).

Krugman, P. R. (1991). Increasing returns and economic geography. Journal of Political Eco-nomy, 99(3), 483–99.

Manning, A. (2003). The real thin theory: monopsony in modern labour markets. LabourEconomics, 10(2), 105–131.

Manning, A. (2006). A generalised model of monopsony. Economic Journal, 116(508), 84–100.

Manning, A. & Petrongolo, B. (2011). How local are labor markets? Evidence from a spatial jobsearch model. CEPR Discussion Papers 8686, Centre for Economic and Policy Research.

Marshall, A. (1890). Principles of Economics. London: Macmillan.

Melo, P. C., Graham, D. J., & Noland, R. B. (2010). Impact of transport infrastructure on firmformation: Evidence from the portuguese municipalities. Transportation Research Record,2163, 133–143.

Michaels, G. (2008). The effect of trade on the demand for skill: Evidence from the interstatehighway system. The Review of Economics and Statistics, 90(4), 683–701.

Mion, G. & Naticchioni, P. (2005). Urbanization Externalities, Market Potential and Spatial Sort-ing of Skills and Firms. CEPR Discussion Papers 5172, Centre for Economic and PolicyResearch.

Mion, G. & Naticchioni, P. (2009). The spatial sorting and matching of skills and firms.Canadian Journal of Economics, 42(1), 28–55.

40

Mulalic, I., van Ommeren, J. N., & Pilegaard, N. (2010). Wages and commuting: Quasi-natural experiments’ evidence from firms that relocate. Tinbergen Institute Discussion Pa-pers, (10-093/3).

Phillips, D. (2012). Getting to Work: Experimental Evidence on Job Search and TransportationCosts. Technical report. TECHREPORT, Department of Economics, Hope College MI.

Phimister, E. (2005). Urban effects on participation and wages: Are there gender differences?Journal of Urban Economics, 58(3), 513–536.

Puga, D. (2002). European regional policies in light of recent location theories. Journal ofEconomic Geography, 2(4), 373–406.

Redding, S. J. & Venables, A. J. (2004). Economic geography and international inequality.Journal of International Economics, 62(1), 53–82.

Rosenthal, S. S. & Strange, W. C. (2003). Geography, industrial organization, and agglomer-ation. The Review of Economics and Statistics, 85(2), 377–393.

Rosenthal, S. S. & Strange, W. C. (2008). Agglomeration and hours worked. The Review ofEconomics and Statistics, 90(1), 105–118.

Rosenthal, S. S. & Strange, W. C. (2012). Female Entrepreneurship, Agglomeration, and aNew Spatial Mismatch. The Review of Economics and Statistics, 94(3), 764–788.

Sanchis-Guarner, R. (2012). Essays on Urban and Spatial Economics. PhD Theses. The LondonSchool of Economics and Political Science.

van Ommeren, J. N. & Gutierrez-i-Puigarnau, E. (2011). Are workers with a long commuteless productive? An empirical analysis of absenteeism. Regional Science and Urban Econom-ics, 41(1), 1–8.

Venables, A. J. (2007). Evaluating urban transport improvements: Cost-benefit analysis inthe presence of agglomeration and income taxation. Journal of Transport Economics andPolicy, 41(2), 173–188.

Vickerman, R. (2002). Transport and Economic Development. Transport and economic devel-opment round table 119, European Conference of Ministers of Transport, OECD.

Vickerman, R., Spiekermann, K., & Wegener, M. (1999). Accessibility and economic devel-opment in europe. Regional Studies, 33(1), 1–15.

Zenou, Y. (2009). Urban Labor Economics. Cambridge University Press.

41

Appendix

A.1. Further details on data construction

The data described in this section follows closely Gibbons et al. (2012). When necessary, thedata construction was adapted to the setup of the empirical analysis of this paper.

A.1.1. Economic size of the locations

The accessibility indices Art and Art of section 3.1.1 are calculated combining the economicsize of locations and ward-to-ward travel times in years 2002–2008 using the expressions (2)and (3). I use total employment in the ward as measure of economic size of the locations. Theemployment figures at the ward level were obtained using data from the Business StructureDatabase (BSD). The BSD is a yearly snapshot of the Inter-Departmental Business Register(IDBR), which is accessible to researchers through the Secure Data Service (SDS) deliveredby the UK Data Archive (UKDA). I use data from 2001 to 2008.

This dataset is maintained by the Office of National Statistics (ONS) and contains a yearlyupdated register of the universe of businesses in the United Kingdom. It is drawn from ad-ministrative registers. It covers about 98% of business activity (by turnover) in Great Britain.The smallest unit of observation is the establishment or plant (“local unit”), but there is alsoinformation on the firm to which the plant belongs (“reporting unit”) and the enterpriseand enterprise group of the firm. The dataset provides detailed information on the location(postcode), the sector of production (up to 5 digits) and employment of the plant. This levelof detail allows us to calculate employment at any geographical level aggregating up frompostcodes.

Alternatively indices (2) and (3) can be defined using different measures of economicsize. Specifically, I use number of residential addresses in the ward obtained from the Na-tional Statistics Postcode Directory (NSPD) and the number of establishments at the wardlevel (local unit counts), calculated using the BSD. I use these alternative indices in the ro-bustness checks carried out in Section 4.3.

A.1.2. The calculation of travel times

The second component in the accessibility index is then an origin-destination (O–D) mat-rix containing the costs crjt (journey time) between each origin and destination (the wardcentroid). This matrix is required for all the years of the sample. To calculate travel times Iconstruct GIS-networks for every year between 2002 and 2008 (dated at the beginning of theyear). I do this by combining the dataset on road schemes described in Section A.1.3. andtwo GIS-networks (for years 2003 and 2008) for Great Britain provided by the Departmentof Transport (DfT).

The 2008 GIS-network contains all the major road links existing at the beginning of 2008.It includes all major roads that, according to the DfT, cover roughly 65% of vehicle kilo-metres. The network includes information on several characteristics of the road links: thecount point code (CP) of each road section (which helps is to identify the links and refers

42

to the point where the traffic is counted), the grid reference for the traffic count point, aunique reference for the local and national transport authorities which manage the link, theroad to which the link belongs to (number and type), the maximum permitted speed, thetotal length of network road link in kilometres and the traffic total flows 35. Total flow isdefined as the “Annual Average Daily Flow” (AADF) and it is measured in terms of numberof vehicles. It corresponds to the average over a full year of the number of vehicles passinga point in the road network each day.

I use road construction as the source of variation in travel times over time. I geo-locateall the road links belonging to each of the 23 schemes listed in Table A.13 and I match themto the 2008 road network based on their CP code. Starting from the 2008 network, in everyyear I remove the new links opened in that year in order to reconstruct the network as itwas in years prior to 2008. I construct a network at the beginning of each year of the period2002–2008. The road network consists of roughly 17,000 road links annually. These networkscontain the links existing every year. In order to construct Art and Art I need to calculate thecost of crossing those links crjt, which I define based on travel times.

In order to calculate travel times, I use data on traffic speeds from the 2003 generalisedprimary road GIS-network provided by DfT. Traffic speeds are modeled from traffic flowcensus data using the Road Capacity and Costs Model (FORGE - Fitting On of RegionalGrowth and Elasticities) component of the National Transport Model (NTM). FORGE is thehighway supply module of NTM 36. The National Transport Model provides “a means ofcomparing the national consequences of alternative national transport policies or widely-applied local transport policies, against a range of background scenarios which take intoaccount the major factors affecting future patterns of travel”. It is used to produce forecastson traffic flows in order to design transport policies 37. The Road Capacity and Costs Modelis one of the three sub-models included in the NTM and it corresponds to the highwaysupply module.

The Road Capacity and Costs Model (FORGE) is used to show the impact of road schemesand other road-based policies 38. As explained in the documentation 39: ”The inputs to theRoad Capacity and Costs Model are car traffic growth (based on growth in car driver trips)and growth in vehicle-miles from other vehicle types. This traffic growth is applied to a data-base of base year traffic levels to give future “demand” traffic flows. These are compared tothe capacity on each link, and resulting traffic speeds are calculated from speed/flow rela-tionships (which link traffic volumes, road capacity and speed) for each of 19 time periodsthrough a typical week”. One of the outputs of FORGE is therefore vehicle speeds by roadtype, and this is what I use in the calculation of travel times between wards.

I use journey times, obtained from FORGE, in the non-busy direction averaged over alltime periods between Monday-Friday 08:00 and 18:00. I focus on non-busy travel directionsbecause the busy travel directions are, in principle, more sensitive to changes in conges-tion induced by new travel links. In practice, there are only minor differences between the

35More information in http://www.dft.gov.uk/matrix/estimates.aspx.36See http://www.rudi.net/files/FORGE.pdf for more information.37See http://www2.dft.gov.uk/pgr/economics/ntm/ for more information.38See http://www.rudi.net/files/FORGE.pdf for more information.39See http://www2.dft.gov.uk/pgr/economics/ntm/pdfntmoverview.pdf for more information

43

modeled journey times in the busy and non-busy directions, or between averages over theworking day and full 24-hour periods (see Gibbons et al., 2012, for more details). Due todata limitations, I use journey times in 2003 for the whole period 2002–2008. These speedsare based on traffic flows for 2003, and I keep these constant for all years in the sample,except for the links opened after 2003. For links opened after 2003 I use estimated journeytimes from a regression model using a dataset of over 17,000 links in the 2003 network. Iregressed link speeds from the 2003 FORGE network on speed limit dummies, traffic flows,traffic flows squared, road category dummies (six categories) and local authority dummies40. The regression predicts speeds from the FORGE reasonably well (R-squared =0.76). Ithen used the results to predict travel times for links opened after 2003 for which no FORGEspeed is available. To obtain these predictions I use the link characteristics and 2008 trafficflows. The 2008 flows are the only ones available for all the links constructed between 2003and 2007. For some of the links, the prediction exceeded travel time implied by the speedlimit. I replaced predicted speed with the speed limit for these links.

Using traffic flows for a single year (2003 when possible, and 2008 for new links) alsohelps to avoid using endogenous changes in traffic flows that could be induced by roadconstruction and that would affect travel times. I am interested in using only the variationin travel times which stems from road construction, which is the policy variable I aim toevaluate. After the construction of the new road links, actual travel times between locationswould be the result of the use of new links for optimal routing and the changes in trafficcongestion, which are endogenous to the new routes. Using traffic flows fixed over time al-lows me to focus on changes in predicted travel times which stem from new optimal routingafter the addition of new road links.

Some of the road schemes in the Highways Agency data are bypasses around villagesand small towns (Faster Routes in Table A.13). Typically, before the bypass was opened therewas a primary road through the village or town but after the introduction of the bypass theold road was downgraded. The downgrading of the old road implies that it is not presentin the 2008 primary road network. Hence, using the method of deleting links based on theiropening years would create an artificial break in the primary road network, when it comes tobypasses. Therefore, I keep the bypasses in the network in the pre-opening years and assumethat travel time on the bypass before opening year was twice the post opening travel time.Scheme evaluation reports available to us support the assumption of significantly longertravel time through the village/town before the bypass is opened.

After constructing the generalised traffic networks of each year, I use the network ana-lysis algorithms in ESRI ArcGIS R© to compute least-cost (minimum journey time) routesbetween each origin ward j and destination ward k in years 2002–2008. When computingthe O-D matrix I apply a limit of 75 minute drive time. This limit facilitates O-D matrixcomputation but does not affect the value of accessibility index because wards beyond 75minutes would have negligible weights in the calculation of Art. Table A.12 contains sum-mary statistics for the travel times between wards for year 2002–2008. The number of O–Dcrossings is over 9,000,000 in every year. Over the period, the average travel time betweentwo wards decreased 0.38%. Given the assumptions in the construction of the road networks

40The results of this regression are available on request.

44

over time, the optimal travel time between two given locations is either unchanged (if thenew link is not used in the new optimal route) or reduced (if the new link is used) afterthe addition of a new road link. This is a simplification but clearly identifies the source ofvariation in the construction of the accessibility measures.

It should be noted that the network is highly generalised. Journeys via the minor roadnetwork are not modeled. Forbidden turns and one way systems are not modeled. All linkintersections are treated as junctions. The changes in accessibility must therefore be regardedas approximate. This might have implications in the estimation of the effect of accessibilityon economic outcomes if it induces measurement error in the accessibility indices calcula-tion. Hence, the estimates could be too small due to attenuation bias. In this case, I shouldinterpret any effect as a lower bound of the “real” effect.

A.1.3. The road schemes

In order to calculate travel times between location in every point of time, I need to knowwhich road links existed every year. I collected information on completed road projects forthe British major roads network by combining information provided by the Departmentof Transport and other data sources41. I collected data on around 75 projects which werecompleted between 2002 and 2007 42.

The nature of these projects is diverse. They cover construction of new junctions, du-alling, widening, upgrades and construction of new roads. I focus on construction of newroads and, as the analysis covers the period 2002–2008, I retain 23 road schemes, which arelisted in Table A.13. The length of these schemes is also provided in Table A.13. As clearlystated in the policy documentation (see for example Department of Transport, 1998b; High-ways Agency, 2009, for more information), road interventions undertaken by the local andnational authorities 43 aim to improve traffic flows and road security, and indirectly affectthe environment and the economy by reducing traffic congestion. This supports the empir-ical strategy as wards located very close to the schemes were not specifically targeted toimprove their economic performance.

I define two types of projects. Type 2 (Faster Routes) correspond to roads for which therewas an alternative route before, but the road was a minor road (not existing in the majorroad network) and an upgrade (which involves improvement and the construction of newlanes) was carried out so the road becomes part of the major road network. They mainlycorrespond to bypasses which relieve traffic congestion from villages and usually flow inparallel to an existing alternative minor road. Type 1 (New Routes) corresponds to “genu-inely” new roads, i.e. roads for which I do not have an alternative minor road flowing in

41Mainly The Highways Agency, the Motorway Archive, Transport Scotland and Wikipedia.42More details in Gibbons et al. (2012). For details on the policy context see Sanchis-Guarner (2012).43Due to rising concerns about the increase in traffic flows and the limited capacity of existing networks, in

1997 the Labour Government carried out a reform of the management of major roads. Major roads comprisetrunk roads, motorways and principal roads (A roads). The aim of this reform was to “radically change trans-port policy” (Department of Transport, 1998a). This reform involved the transfer of parts of the English TrunkNetwork to local authorities, while the Department of Transport kept control of the most strategic roads (thoseconnecting major population centers, ports and airports, key cross-border links and the Trans-European RoadNetwork). The management of these roads was transferred to the Highways Agency in England (which is partof the Department of Transport), to Transport Scotland and to the Welsh Assembly Government.

45

parallel. In practice, as detailed in Section A.1.2., I defined these two types of improvementsin order to be able to calculate travel times given the characteristics of the road network datathat I use. They are not substantially different, both involve road construction, except for thetreatment I give them in them in the calculation of the travel O–D matrices.

Figure 1(a) displays the location of these projects and the major road network at theend of the period of analysis (2008). Projects are scattered all over Britain. I focus on newconstruction because these improvements are the ones I can expect to have a substantialeffect on travel times between wards. Table A.14 summarises the length of the links by typeand displays the percentage of kilometres they represent with respect to the total networklength. Total new links represent 0.49% of the network length, and the majority of them(61.9%) correspond to type 2 links, i.e. faster routes.

A.1.4. Accessibility changes arising from road construction

Accessibility indices Art and Art can be applied to study the economic effects arising fromchanging accessibility by road when the costs crjt in (2) and (3) are calculated using routingalong the transport network. This works because road construction changes the structureof costs crjt along the transport network and the structure of costs along routes from r topotential destinations j. This in turn changes the accessibility index.

For example, consider a transport improvement that involves a journey time reductionon a road link between two nodes p and q. This scheme has a first order effect on the costsof the least-cost route between r and j if:(a) the least-cost route between r and j passes along the link p–q in both the pre and post-

improvement periods, such that the transport improvement reduces the cost of the jour-ney along p–q and brings employment at destination j “closer” to origin r in cost terms.

(b) the least-cost route between r and j bypasses link p–q in the pre-improvement period,but switches to use the link p–q in the post-improvement period because of the reductionin costs; again this brings employment at destination j “closer” to origin r in cost terms.

There are also “second order” effects arising when:(c) the least cost route between r and j bypasses link p–q in the pre-improvement and in

the post-improvement periods. However, journeys between other origin and destinationpairs have switched to using the link p–q, which reduces congestion on the alternativelinks in the network used by the routing between r and j; again this brings employmentat destination j “closer” to origin r in cost terms.

In the empirical work below I rely only on the first order effects of type (a) and (b) arisingfrom new transport infrastructure. I have to ignore second order effects of type (c) because,as explained in the previous section, the road transport network data does not allow usto observe changes in travel time induced by changes in congestion occurring as a result ofroad construction (I have no information on traffic flows observed prior to the construction).

Changes in cost of all these types imply changes in the accessibility indices (i.e. a changein effective density). The amount of change in the accessibility index at a location j dependson the likelihood that a route between r and j uses the improved link p–q, and on economicmass in j.

46

A.2. Further results

A.2.1. Heterogeneous effects

In this section I explore if the effects of accessibility from workplace are different for differentgroups of workers. I use the same specification as in column 7 of table 4, and divide thesample in sub-groups according to different criteria. To explore the heterogeneity of theeffects I focus on the impact on wages.

First, I test if the wage effects are different by gender and age. The results are reportedin table A.9. Columns 1 and 2 test the effects for male and female workers separately; theeffect for women is almost double as that for men. This result is in line with the findings ofPhimister (2005), which finds that urban wage premium is larger for women. Columns 3 and4 checks what happens to the coefficient if I focus on primary working-age population (aged25-55), for both genders and solely for male. Many labour papers use one of these sub-groupsto construct their main samples, as workers that just entered the labour market (below 25) orare close to retirement (over 55), and women, could behave different in the labour markets asother groups. The size of the coefficients is different from the main findings, but qualitativethe results remain unchanged.

In table A.10 I divide the sample depending on the type of firm in which the workeris employed. Only workers in the private sector (column 1) benefit from the changes inaccessibility, and the coefficients for the other groups are non-significant and less precisegiven the sample sizes.

In table A.11 I test the effect by skill level. Given that ASHE does not provide informationon education levels, the skill levels are defined using occupation codes 44, and I use the levelof occupation the first time an individual is observed to classify workers, as changes in occu-pation could be endogenous to the accessibility changes. Column 1 includes the less skilledworkers (in elementary occupations), and group 4 includes the more skilled (professionalsand managers). We see that the effect of accessibility from workplace increases with skilllevel, and that is non-significant for the less skilled (the number of observations is smallerin this sub-sample, so the coefficients are also less precise). The effect for the most skilledworkers is almost double than for category 2.

In columns 5 and 6 of the same table I test the effects for individuals which work full-timeand those who work part-time. Only individuals that work full-time capitalise the impactof higher effective density into higher pay. This could become an incentive for individualsto switch from part-time to full-time jobs in order to earn higher wages. I explore this in thefollowing section.

44Based on the 2000 Standard Occupation codes, see page 6 of volume 1.

47

Commuting timeTravel mode

Rail/tube Motor Walk/Cycle Commuters % times15 mins or less 0.25% 78.13% 21.63% 25,298 49.27%16-30 mins 2.12% 87.65% 10.23% 15,666 30.51%31-45 mins 8.07% 87.85% 4.08% 5,489 10.69%45-60 mins 21.63% 76.71% 1.66% 3,315 6.46%61-90 mins 33.74% 65.78% 0.48% 1,242 2.42%91-120 mins 30.21% 69.79% 0.00% 288 0.56%120 mins or more 31.11% 68.89% 0.00% 45 0.09%% modes 4.04% 81.63% 14.33% 100% 51,343Notes: Percentages with respect to commuters. Source: British Household Panel Survey, years 2002–2008.

Table A.1: Commuting times by travel mode

!

!

!

!

!

!

!

!

!

!

!

! !

!

!

!

!!

A5

A66

A6

M6(T)

A43

A47

A1(M)A6

Leeds

OxfordLondonBristol

Glasgow

Cardiff

Bradford

Aberdeen

Newcastle

Liverpool

Cambridge

Notingha m

Birmingham

Manchester

Distance bands from the new road projects0 - 10 kms 10 - 20 kms 20 - 30 kms Beyond 30 kms

Figure A.1: Wards within 10-20-30 kms distance bands around the project sites - 2002–2008

48

INDIVIDUAL FX SAMPLE

Variable Mean Std. Dev. Min Max ObservationsLog of basic weekly pay Overall 5.730 0.709 2.028 7.654 N = 320105

Between 0.693 2.698 7.621 n = 78305Within 0.241 2.115 8.004 T-bar = 4.088

Log of basic weekly hours Overall 3.445 0.403 0.728 4.094 N = 320105Between 0.385 0.833 4.077 n = 78305Within 0.186 1.144 5.305 T-bar = 4.088

Log of accessibility from Overall 15.562 0.713 10.312 18.504 N = 320105workplace Between 0.705 10.364 18.412 n = 78305

Within 0.144 12.372 18.434 T-bar = 4.088Log of accessibility from Overall 15.473 0.655 10.312 18.300 N = 320105home Between 0.651 10.364 18.110 n = 78305

Within 0.112 12.003 18.191 T-bar = 4.088Log of predicted travel time Overall -1.941 1.323 -12.480 0.223 N = 320105between work and home Between 1.211 -12.480 0.222 n = 78305

Within 0.555 -10.241 5.651 T-bar = 4.088INDIVIDUAL-WORK-HOME FX SAMPLE

Variable Mean Std. Dev. Min Max ObservationsLog of basic weekly pay Overall 5.745 0.702 2.028 7.654 N = 267018

Between 0.698 2.667 7.627 n = 79868Within 0.171 2.130 8.019 T-bar = 3.343

Log of basic weekly hours Overall 3.448 0.393 0.728 4.094 N = 267018Between 0.385 0.833 4.077 n = 79868Within 0.139 1.454 5.308 T-bar = 3.343

Log of accessibility from Overall 15.544 0.710 10.312 18.504 N = 267018workplace Between 0.714 10.364 18.463 n = 79868

Within 0.037 15.154 16.057 T-bar = 3.343Log of accessibility from Overall 15.458 0.655 10.312 18.300 N = 267018home Between 0.653 10.364 18.205 n = 79868

Within 0.038 15.042 15.971 T-bar = 3.343Log of predicted travel time Overall -1.977 1.328 -12.480 0.223 N = 267018between work and home Between 1.3228 -12.4802 0.2229 n = 79868

Within 0.0171 -2.9013 -0.3546 T-bar = 3.343Sources: Author’s own calculations using data from ASHE 2002-2008. N denote total number of observations, n denotes number ofobservation Within the group and T-bar denotes the average number of observed period Within the group. Groups correspond toindividuals (top panel) and individual-work-home location spells (bottom panel). All variables are in logs.

Table A.2: Summary statistics ASHE, both samples, 2002-2008

49

BALANCING TESTS: Initial values for movers and stayersMovers Stayers All

Mean Std. Dev No Obs Mean Std. Dev No Obs Mean Std. Dev No ObsBasic pay 339.942 224.708 166624 340.085 237.539 153481 340.011 230.949 320105Basic hours 33.310 9.172 166624 32.849 9.389 153481 33.089 9.279 320105Basic earnings 9.984 6.009 166624 10.087 6.342 153481 10.033 6.171 320105Occupation code 4857.264 2544.243 166624 5006.678 2604.991 153481 4928.904 2574.627 320105Full time 0.774 0.418 166624 0.742 0.437 153481 0.759 0.428 320105Age 35.961 10.752 166624 40.769 11.140 153481 38.266 11.200 320105Female 0.513 0.500 166624 0.506 0.500 153481 0.510 0.500 320105Log w access 15.547 0.709 166624 15.434 0.713 153481 15.493 0.713 320105Log h access 15.451 0.639 166624 15.358 0.667 153481 15.406 0.654 320105Work access 7287830 6355985 166624 6431514 5281399 153481 6877252 5880943 320105Home access 6167741 3980881 166624 5672440 3721551 153481 5930259 3866635 320105Travel time 0.274 0.251 166624 0.232 0.232 153481 0.254 0.243 320105Commute distance 9.846 11.241 166624 8.254 9.870 153481 9.083 10.636 320105Observations correspond to sample average for the values the first time the individuals are observed. Movers refer to individuals who changeswork or home wards at least once while in the sample, and stayers to individuals that keep the same work-home wards combination during thewhole panel. Observations correspond to years 2002-2008. Pay variables are in £, occupations correspond to SOC2000 classification, travel timeis in hours, and commute distance is crow-fly distance in kilometres. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.3: Balancing tests: individual FX sample, movers and stayers

DEPVAR: Log of basic weekly pay1 2 3 4 5 6

Log of accessibility 0.265*** 0.309*** 0.305*** 0.294*** 0.360** 0.224*from work ward [0.103] [0.096] [0.102] [0.109] [0.141] [0.129]Log of accessibility -0.039 -0.058 -0.083 -0.073 -0.023 -0.041from home ward [0.078] [0.080] [0.089] [0.089] [0.116] [0.105]Log of travel time 0.01 -0.002 -0.011 0.07 0.115 0.047between work & home [0.021] [0.021] [0.020] [0.067] [0.100] [0.063]Observations 271962 267018 261615 243683 224096 262273Excluded extremes N 0 kms 1 kms 3 kms 5 kms 0 kmsExclude along scheme N N N N N 0 kmsAll controls Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all con-trols. Distance band 30 kms (excluding extremes). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT,CASWEB and author’s own calculations.

Table A.4: Robustness: exclude extremes, log of basic weekly pay

50

DEPVAR: Log of basic weekly pay1 2 3 4 5 6 7 8 9 10

Log of accessibility 0.293*** 0.284*** 0.286*** 0.287*** 0.309*** 0.314*** 0.308*** 0.345*** 0.222** 0.209from work ward [0.096] [0.097] [0.098] [0.099] [0.096] [0.095] [0.100] [0.103] [0.108] [0.165]Log of accessibility -0.055 -0.058 -0.044 -0.036 -0.058 -0.065 -0.073 -0.067 0.032 -0.018from home ward [0.073] [0.075] [0.075] [0.075] [0.080] [0.080] [0.082] [0.089] [0.116] [0.143]Log of travel time 0.002 0.000 0.000 0.000 -0.002 -0.001 -0.001 0.004 0.012 -0.001between work & home [0.022] [0.022] [0.022] [0.022] [0.021] [0.022] [0.021] [0.022] [0.023] [0.022]Observations 414646 388726 349624 310192 267018 217169 164456 100362 46980 17453Distance band 50 kms 45 kms 40 kms 35 kms 30 kms 25 kms 20 kms 15 kms 10 kms 5 kmsAll controls Y Y Y Y Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls. Distance band 30 kms (excludingextremes). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.5: Robustness: distance bands, log of basic weekly pay

DEPVAR: Log of basic weekly pay1 2 3 4 5 6

Log of accessibility 0.309*** 0.296*** 0.307*** 0.712*** 0.153** 1.690***from work ward [0.096] [0.096] [0.098] [0.196] [0.061] [0.455]Log of accessibility -0.058 -0.055 -0.054 -0.145 -0.02 -0.355from home ward [0.080] [0.079] [0.080] [0.166] [0.044] [0.389]Log of travel time -0.002 -0.002 -0.002 -0.002 -0.002 -0.002between work & home [0.021] [0.022] [0.021] [0.022] [0.022] [0.021]Observations 267018 267018 267018 267018 267018 267018Economic size Empl No plants Addr cts Empl Empl EmplDistance function Inverse Inverse Inverse Inverse Inverse ExponDistance decay 1 1 1 0.5 1.5 0.2All controls Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls.Distance band 30 kms (excluding extremes). “Empl” stands for employment, “No plants” stands for num-ber of plants and “Addr cts” stands for number of residential addresses (proxy for population counts). *p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.6: Robustness: accessibility measures, log of basic weekly pay

DEPVAR: Log of basic weekly pay1 2 3 4

Log of accessibility 0.309*** 0.304*** 0.322*** 0.321***from work ward [0.096] [0.094] [0.093] [0.093]Log of accessibility -0.058 -0.056 -0.063 -0.067from home ward [0.080] [0.078] [0.079] [0.080]Log of travel time -0.002 0.001 0.002 0.002between work & home [0.021] [0.022] [0.022] [0.022]Observations 267018 216448 212088 209638

Sample AllNo work No home Neither

in London in London in LondonAll controls Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. In-cludes all controls. Distance band 30 kms (excluding extremes). * p<0.1, ** p<0.05, ***p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.7: Robustness: exclude London, log of basic weekly pay

51

DEPVAR: Log of basic weekly pay1 2 3 4 5 6 7

Log of accessibility 0.309*** 0.290*** 0.320*** 0.293*** 0.461*** 0.209 0.629***from work ward [0.096] [0.108] [0.105] [0.101] [0.169] [0.222] [0.226]Log of accessibility -0.058 -0.05 -0.063 -0.043 -0.155 -0.041 -0.361from home ward [0.080] [0.086] [0.086] [0.080] [0.195] [0.213] [0.255]Log of travel time -0.002 -0.004 0.001 -0.004 0.013 -0.025 0.077between work & home [0.021] [0.023] [0.023] [0.022] [0.052] [0.047] [0.112]Observations 267018 210953 228774 241117 56065 38244 25901Individual moves Y N N N Y Y YDimension Both Neither Work w Home w Both Work w Home wAll controls Y Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls. Distanceband 30 kms (excluding extremes). “Work w” stands for work ward and “Home w” stands for home ward. * p<0.1,** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.8: Robustness: individual-location FX and mobility, log of basic weekly pay

DEPVAR: Log of basic weekly pay1 2 3 4

Log of accessibility 0.226** 0.442*** 0.276*** 0.283**from work ward [0.115] [0.121] [0.072] [0.118]Log of accessibility 0.097* -0.234 0.008 0.114from home ward [0.058] [0.163] [0.093] [0.072]Log of travel time 0.038* -0.028 0.004 0.013between work & home [0.022] [0.033] [0.025] [0.022]Observations 131163 135855 209930 102045Age group 16-65 16-65 25-55 25-55Gender Males Females Both MalesAll controls Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and homeward. Includes all controls. Distance band 30 kms (excluding extremes). * p<0.1, **p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.9: By groups: gender and age, log of basic weekly pay

DEPVAR: Log of basic weekly pay1 2 3 4

Log of accessibility 0.341*** 0.395 0.317 0.521from work ward [0.094] [0.378] [0.315] [0.381]Log of accessibility -0.063 0.109 -0.127 -0.179from home ward [0.087] [0.282] [0.149] [0.359]Log of travel time 0.029 -0.105 -0.077** 0.159**between work & home [0.031] [0.226] [0.031] [0.075]Observations 155211 13408 76772 19336Type of firm Private Owner Public Non-profitAll controls Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward.Includes all controls. Distance band 30 kms (excluding extremes). * p<0.1, ** p<0.05,*** p<0.01. Source: ONS, DfT, CASWEB and author’s own calculations.

Table A.10: By groups: firm type, log of basic weekly pay

52

DEPVAR: Log of basic weekly pay1 2 3 4 5 6

Log of accessibility -0.263 0.280** 0.372** 0.432** 0.202*** 0.113from work ward [0.389] [0.112] [0.146] [0.204] [0.057] [0.226]Log of accessibility 0.196 -0.034 -0.089 -0.095 0.027 0.023from home ward [0.480] [0.136] [0.113] [0.110] [0.048] [0.297]Log of travel time -0.081 -0.01 0.021 0.026 0.022 0.100between work & home [0.078] [0.040] [0.029] [0.055] [0.014] [0.078]Observations 33053 113468 61755 58742 199992 57710Skill group (initial) Lower 2 3 Higher All AllStatus All All All All Full time Part timeAll controls Y Y Y Y Y YNotes: Clustered standard errors in brackets, 2-way clustering work and home ward. Includes all controls.Distance band 30 kms (excluding extremes). * p<0.1, ** p<0.05, *** p<0.01. Source: ONS, DfT, CASWEB andauthor’s own calculations.

Table A.11: By groups: initial skill/full-time, log of basic weekly pay

YearNumber of

MeanStandard 1st 10th 90th 99th

observations deviation percentile percentile percentile percentile2002 9,170,886 49.745 17.515 6.681 24.104 70.872 74.5932003 9,170,886 49.735 17.512 6.679 24.099 70.859 74.5912004 9,170,886 49.613 17.476 6.674 24.049 70.724 74.5732005 9,170,886 49.597 17.471 6.672 24.044 70.704 74.5712006 9,170,886 49.591 17.469 6.672 24.043 70.696 74.5692007 9,170,886 49.575 17.463 6.672 24.037 70.675 74.5662008 9,170,886 49.554 17.457 6.669 24.025 70.645 74.560Growth 2002–2008 -0.38% -0.33% -0.18% -0.33% -0.32% -0.04%Notes: Times in minutes. Min=0 and Max=75 mins. Times measured at the beginning of the year. Source: Department of Transportand own authors’ calculation.

Table A.12: Summary statistics O–D matrices of travel times

53

OpeningType Road Scheme

Lengthyear in kms2002 New Route A27 A27 Polegate Bypass 3.22002 Faster Route A43 A43 Silverstone Bypass 14.22002 Faster Route A6 A6 Clapham Bypass 14.62002 Faster Route A66 A66 Stainburn and Great Clifton Bypass 4.12003 Faster Route A41 A41 Aston Clinton Bypass 7.32003 Faster Route A5 A5 Nesscliffe Bypass 21.52003 Faster Route A500 A500 Basford, Hough, Shavington Bypass 7.72003 Faster Route A6 A6 Alvaston Improvement 4.72003 Faster Route A6 A6 Great Glen Bypass 6.82003 Faster Route A6 A6 Rothwell to Desborough Bypass 8.42003 New Route A6 A6 Rushden and Higham Ferrers Bypass 5.42003 Faster Route A650 A650 Bingley Relief Road 4.42003 New Route M6(T) M6 Toll. Birmingham Northern Relief Road 29.72004 Faster Route A10 A10 Wadesmill to Colliers End Bypass 7.02004 New Route A63 A63 Selby Bypass 9.52005 New Route A1(M) A1(M) Wetherby to Walshford 8.12005 Faster Route A21 A21 Lamberhurst Bypass 2.42005 Faster Route A47 A47 Thorney Bypass 10.72005 New Route M77 M77 Replaces A77 from Glasgow Road 18.22006 New Route A1(M) A1(M) Ferrybridge to Hook Moor 19.22006 Faster Route A421 A421 Great Barford Bypass 7.62007 Faster Route A2 A2 / A282 Dartford Improvement 4.2

2007 Faster Route A66A66 Temple Sowerby Bypass

26.2and Improvements at WinderwathTOTAL 245.1

Source: Own authors calculations using information from the Department for Transport, the Highways Agency, theMotorway Archive, Transport Scotland and Wikipedia.

Table A.13: Road schemes – 2002–2007

Length PercentageMAJOR ROADS NETWORK IN 2008 50,093.7All new links 2002–2007 (both types) 245.1 kms 0.49%of which: New Routes 93.35 kms 38.1%of which: Faster Routes 151.78 kms 61.9%Notes: The right column reports the percentage of total new links over the whole network(2nd row) and the percentage of the new links by type over total improvement length(3rd and 4th rows). Source: Own authors calculations using information from the Depart-ment for Transport, the Highways Agency, the Motorway Archive, Transport Scotlandand Wikipedia.

Table A.14: Length of schemes – 2002–2007

54