the accessibility to employment offices in the spanish labour market

The accessibility to employment offices in the Spanishlabour market*

Patricia Suárez1, Matías Mayor1, Begoña Cueto1

1 Department of Applied Economics, University of Oviedo, Avda. del Cristo s/n Oviedo, Asturias 33006 Spain(e-mail: [email protected], [email protected], [email protected])

Received: 1 October 2010 / Accepted: 29 January 2012

Abstract. This paper focuses on the differences in the levels of accessibility to publicemployment offices in the Spanish municipalities. Hence the main purpose is to evaluate the roleof the public employment services in local labour markets by considering the physical distanceto employment offices and the spatial structure of their catchment areas. First, we proposean accessibility measure and, second, we estimate a spatial model and test whether a higheraccessibility to employment offices could contribute to reduce local unemployment rates.The results suggest that policy-makers should strive to improve the accessibility to employ-ment offices so that adequate assistance to find suitable employment may be ensured to everyjobseeker.

JEL classification: J68, J60, C21, R12

Key words: Employment offices, unemployment, accessibility, spatial econometrics

1 Introduction

After a 15 year period of sustained reduction in the Spanish unemployment rates and conver-gence with most EU countries, the economic downturn has sent Spain back to the top ofcountries with higher unemployment rates. More than two million jobs were destroyed in 2008and 2009, and the unemployment rate is currently over 20 per cent (Labour Force Survey (LFS),INE). This has led to the passing of the Labour Market Reform Act 2010 (Law 35/2010), whichaims to improve labour mediation. Specifically it regulates the activities of private employmentagencies and sets a commitment for the improvement and promotion of public employment

* Part of this work was done while Patricia Suárez and Matías Mayor were visiting research fellows at the Institutefor Economic Research (University of Lugano), whose support is gratefully acknowledged. Special thanks are due tothree anonymous referees as well as Roberto Patuelli, Enrique López-Bazo, Andrew Clark, Uwe Blien and RonMcQuaid for helpful observations and comments. Patricia Suárez would also like to thank the financial support by theMinisterio de Ciencia e Innovación of Spain (FPU Grant) and Fundación Banco Herrero. A previous version of this paperwas published as working paper no 610/2011 of the Fundación de las Cajas de Ahorros (FUNCAS). The usual disclaimerapplies.

doi:10.1111/j.1435-5957.2012.00425.x

© 2012 the author(s). Papers in Regional Science © 2012 RSAI. Published by Blackwell Publishing, 9600 Garsington Road,Oxford OX4 2DQ, UK and 350 Main Street, Malden MA 02148, USA.

Papers in Regional Science, Volume •• Number •• •• 2012.

services (PESs) in the autonomous communities (NUTS 2),1 providing them with additionalhuman and material resources. Since the PES in the autonomous communities are responsiblefor implementing active labour market policies (ALMPs), the spatial analysis of the accessibilityof the unemployed to employment offices becomes quite relevant, as well as its differencesacross communities.

Placement services are located in space, hence analyses of the accessibility to employmentoffices require spatially explicit tools. Also, since PESs are essentially aimed at meeting thedemand for job search assistance, any improvements in accessibility would translate into betterPES performance. Consequently, we need to discuss whether the accessibility to employmentoffices is really equitable regardless of place of residence. In fact, according to the EuropeanCommission, the main task of any PES is to ensure that no job-seeker is marginalized by a lackof adequate assistance to find suitable employment. Furthermore, the LFS lists 12 active jobsearch methods, among them registration at a public employment office. They are not mutuallyexclusive so that an unemployed individual may use multiple job search methods simulta-neously. In 2009, registration at a public employment office was used as a job search method by67.9 per cent of the country’s jobless population.

This paper gives a clear picture of PES in Spain in terms of access to employment officesand provides regional policy-makers with important information about the lack of opportunitiesto go to such offices in some places. The aim of the paper is twofold. First we introduce a newaccessibility measure and tackle differences in access at municipal level. Second we propose anempirical model which includes the existence of spatial dependence in its specification so as toaddress the relation between local unemployment rates and accessibility to employment offices.

Before proceeding any further, let us review some of the transformations that PESs haveundergone in Spain since the 1990s. In 1998, the Spanish government started to decentralizethe PESs to the autonomous communities, which were granted complete authority on ALMPs.2

However, to ensure standards of service provision regardless of place of residence, PESs inthe autonomous communities have remained integrated in the national employment system. Thedecentralization of ALMPs was undertaken so that each region adopted a needs-based approachwhich could bring in better management of the available resources, and adapted employmentand training programmes to the features of its labour market and unemployed population profile.

The ALMPs, which also cover the PESs, have been hardly evaluated, so there is littleinformation available about their performance. Whenever figures of registered vacancies areconsidered, the Spanish PES efficiency is regarded as low. This poor performance may be partlyexplained by the small number of job counsellors. In 2006 before the start of the economic crisis,there were 1.837 million unemployed and just 7,996 employees at PES offices (ConsejoEconómico y Social 2009). Consequently, each counsellor saw about 230 job-seekers – one of thehighest records in the EU – and the unemployed were likely to compete for time with theircounsellor.3

In Spain there are no studies of employment offices and their catchment areas at the locallevel and, as in other countries (Fertig et al. 2006), we do not know how public funding isdistributed among the offices. Studies on the efficiency of PES offices at local level have beendone in Germany (Hagen 2003), Switzerland (Sheldon 2003) and Sweden (Althin and Behrenz

1 Spain is made up of two autonomous cities (Ceuta and Melilla) and 17 autonomous communities, each with its ownheritage, government and PES.

2 This decentralization process has been culminated with the devolution of this power to the Basque regionalgovernment in 2011 (see Suárez 2011 and the references given therein).

3 The special plan for job counseling, professional training and work placement, estimates that 1,500 new jobcounsellors – approx. two counsellors per office – would render a coefficient of 3.5 beneficiaries per counsellor and day.Even though the hiring of 1,500 job counsellors has led to a significant increase in staff since 2008, current staff numbersare far from meeting the counselling and mediation needs of the unemployed, especially at employment offices that haveto attend to a high number of jobless.

2 P. Suárez et al.


2004). However, these studies have not analysed whether the spatial distribution of employmentoffices ensures homogeneous access to such offices. Blien et al. (2010) make a useful contri-bution to the literature by developing a classification of employment office areas so as to showthe diversity of the problems that labour market policy has to deal with. Joassart-Marcelli andGiordano (2006) use a geographic information system to look into the location of one-stopcentres in Southern California and their level of accessibility. Our work differs in that weconsider the problems of endogeneity and heteroscedasticity using context appropriate estima-tion methods. As far as we know, this is the first paper on the spatial distribution of employmentoffices and their levels of accessibility in Spain. All in all, our paper attempts to bridge this gapin Spain by combining the methodology of spatial economics with new accessibility measuresthat take into account the size of an employment office catchment area.

The outline of the paper is as follows: Section 2 describes the data and examines basicfeatures of the unemployed and employment offices in Spain. It also introduces the accessibilitymeasures proposed. Section 3 presents comparative evidence of the spatial distributions of theunemployed and employment offices across the Spanish municipalities. In Section 4 we estimatethe spatial unemployment rate model which includes the accessibility to employment offices asexplanatory variable. Section 5 concludes with some policy recommendations.

2 Data and methodology

2.1 Data

In Spain, high regional unemployment rates have been endemic and persistance (for a moredetailed discussion, see e.g. Jimeno and Bentolila 1998; Bande et al. 2008; Garcia-del-Barrioand Gil-Alana 2009). This stylized fact leads to an increasing importance of the local level inthe design and implementation of ALMPs which calls for a more disaggregated analysis of thelabour market, namely, beyond the provincial level (NUTS 3).

Since employment offices operate at local level, we have regarded the municipality as theappropriate spatial unit for our analysis. It is essential to establish clusters of unemployed peopleat the local level, since active job-seeking policies and the modernization of PESs should bemore intense in such municipalities.

Unemployment data in the following pages have been taken from the official unemploymentstatistics, which are published monthly by the SEPE (Servicio Público de Empleo Estatal).4 Datareferring to the local employment offices and their catchment areas have been taken from theregional employment authorities websites and the SEPE website.

Since there are no data on the economically active population at municipal level, localunemployment rates have been calculated by dividing the number of unemployed workersregistered at PES offices by the number of people of working age (i.e., population aged 16–64)on the 2009 municipal register. Even if we had had access to local economically activepopulation data in the LFS, they could hardly have been used as a reference for registeredunemployment, since the active population in the LFS is calculated by adding the number ofpeople in employment to the number of jobless, and the latter is calculated in accordance withinternational statistical standards, which need not correspond to those in administrative-basedstatistics (Toharia 2005). Alonso-Villar and Río (2008) and Alonso-Villar et al. (2009) also relyon this definition to obtain unemployment rates at municipal level.5

4 The SEPE basically manages the unemployment benefits system, co-ordinates regional PES and integrates availablestatistical information about the labour market.

5 One of the main features of the Spanish labour market is its low female participation rate. Consequently, the femaleunemployment rate diminishes when the working-age group stands for the whole population. The male rate is ratheraccurate, though.

3The accessibility to employment offices


Figure 1 shows the concentration of the high local unemployment rate, and especially upperoutliers (1,238), in the south and along the Mediterranean coast. Broadly speaking, the map alsoshows a concentration of low local unemployment rates in Castile and Leon, as well as Aragon(except Zaragoza), Cuenca and Guadalajara.

The current location pattern of public employment offices in Spain stems from politicaldecisions over the last 30 years. More precisely, employment offices are administrativeunits established long before the autonomous communities took over ALMPs. The question iswhether this location pattern is the most adequate and, if not, how it could be possibly improved.For example, in 2008, the Government of Navarre opened a new employment office in Tudelaas part of their Plan for the Modernization of the Employment Service of Navarre. Since thenthe office has provided service to 21 municipalities, as well as the municipality of Tudela itself.Besides alleviating the workload of the employment offices which had to attend to these joblessup to 2008, the office represents a step forward in the autonomy the autonomous communitieshave been conferred to modernize the PES and improve job counselling and work placementservices.

Table 1 and Figure 2 show the spatial distribution of employment offices in Spain. Clearly,its most striking feature is the large number of municipalities lacking employment offices –7,524 out of 8,109.

The many municipalities with no employment offices are predominantly concentrated inCastile and Leon, whereas those with are in the south and the south-east, Madrid and Barcelona.Notwithstanding that, the data shows employment offices in every municipality with over4,000 jobless, except Paterna and Milasta (Valencia metropolitan area), San Vicent del Raspeig(Alicante metropolitan area), Mijas (Malaga) and Los Realejos (Tenerife). When the urban-nonurban perspective is considered, we find that the 73 per cent of the employment offices are

Fig. 1. Box map of local unemployment rate (2009)

4 P. Suárez et al.


located in urban areas. However, the distribution of employment offices differs across autono-mous communities. Thus, Madrid, Catalonia and the Valencian Community have 98 per cent, 74per cent and 64 per cent of their employment offices located in municipalities within large urbanareas, respectively. Other autonomous communities such as Extremadura, Castile-La Manchaand Castile and Leon have about 50 per cent of their employment offices in municipalities whichdo not belong to either large or small urban areas.

Figure 3 shows the existence of steep differences between the Spanish autonomous com-munities in the number of unemployed workers per employment office. The number of employ-ment offices seems to be far below the number of jobless they have to attend to, especially inMadrid, the Canary Islands, the Valencian Community and Catalonia, so differences in acces-sibility may be expected.

Table 1. Distribution of the employment offices

Number of municipalities Employment offices

7,524 0526 135 210 36 42 52 62 81 111 19

Total: 8,109 718

Fig. 2. Employment office location



2.2 Measuring accessibility

One of the aims of this paper is to assess whether the accessibility to employment officesis equitable in Spain. The core issue we have to address is the measure of accessibility itself.Several authors from different perspectives have analysed the concept of accessibility within theframework of urban and regional economies. For instance, Krugman (1991) and Fujita et al.(1999) study the importance of accessibility in economic development from a regional perspec-tive. Most existing studies on accessibility belong to the field of transportation economy.Gutiérrez (2001) and Holl (2007) analyse accessibility improvements in Spain. From a theoreti-cal perspective, Geurs and Van Wee (2004) review is remarkable for its analysis of the usefulnessof accessibility measures in the evaluation of changes in transportation infrastructures and its useby researchers and policy-makers alike. With respect to labour markets, accessibility measuresare given consideration in few works. For instance, Van Wee et al. (2001) develop a conceptof accessibility to analyse whether jobs are accessible for employees. Détang-Dessendre andGaigné (2009) study the impact of the place of residence on unemployment duration. They relyon an accessibility measure to convey workers’ competition for jobs and subsequently tacklelabour market tightness.

It is currently intended that ALMPs become an asset in the fight against unemployment sothat assurance of equal access to employment offices is essential. We may begin by stating that,even though employment offices are administrative units that were created long ago, their spatialdistribution is by no means random. However, regardless of the fact that it does follow a pattern,such distribution may cause either equity or inequity of access to the offices. Accessibilityconditions should be the same regardless of the autonomous community of residence – whosegovernment, in turn, is responsible for the administration of the employment offices. In otherwords, every unemployed worker should be equally treated, no matter where they may live. Forus, spatial equity is just equal access to employment offices. That leads us not only to calculatethe accessibility to employment offices but also to analyse their spatial distribution.

Fig. 3. Average number of unemployed workers per employment office. NUTS 2 (2009)

6 P. Suárez et al.


The simplest measure to analyse job-seeker accessibility to employment offices consistsin counting the existing employment offices within a given area (see Figure 2). However, thesemeasures, based on the count of employment offices per municipality, do not take into accountother spatial interactions such as the inverse relation existing between the size of an employmentoffice catchment area and its level of accessibility. Catchment areas are set by counting theunemployed assigned to an employment office, namely, by adding up the number of joblessin the municipalities serviced by a given employment office. In the case of Spain, everyunemployed worker is assigned an employment office by the National Employment Authoritydepending on their place of residence. Consequently, we need further accessibility measures,similar to those that transcend the mere count of employment offices.

Next, we will consider two more types of accessibility measures. The first type of measureonly takes into account the number of employment offices for each regional labour office andthe distance to the municipality in which the corresponding employment office is located. Thelimited scope of this measure leads us to propose a new accessibility measure which also takesinto account the unemployed within each employment office catchment area. We would like tohave had access to the number of job counsellors and/or counselling sessions per unemployedworker, but access to this information is not provided at local level.

The first type of measure (denoted by the superscript I) is based on the number of employ-ment offices in the same regional labour office, adjusted for the distance between a municipalityi and its corresponding employment office.

A EO eiIa

jdij= ( )−λ

(1)

where Ai is a measure of the municipal accessibility to the employment offices in the regionallabour office j, EOj is the number of employment offices in the regional labour office j, dij is thedistance between the municipality i and that in which is located the employment office theunemployed living in i have to go to, measured as the Euclidian distance between the munici-palities’ centroids. Finally, l is a parameter of the distance-decay function and determines thedegree of interaction between the place of residence of the jobless and the employment officethey have to go to, that is, the accessibility quality decreasing as distance to the office increases.We have no data on trips to employment offices, so we have been unable to set the parametersof the distance-decay function. Even though several values were used for this parameter whiledoing this paper, the performance of a sensitivity analysis led us to the results presented here,which were eventually obtained using the following values: l = 0.10 and l = 0.25.6 Neverthe-less, it should also be noted that results do not vary significantly when we use either parameter,especially when we analyse the spatial distribution of accessibility, as will be shown later.

The study of the internal accessibility or ‘self-potential’ of employment offices presentsfurther problems, since there are no data on the exact distance to the office when job-seekersare assigned an office within their municipality of residence. Even though this issue has beenstudied by some authors (Bröcker 1989; Bruinsma and Rietveld 1993; Frost and Spence 1995),it remains largely unsolved in the literature. We may proceed by estimating internal distance.Hence, following Condeço-Melhorado et al. (2011), we used the formula by Rich (1975) toestimate the municipalities’ internal distance:

darea km

ii = 1

2

2( ).

π

6 Joassart-Marcelli and Giordano (2006) establishes l = 0.25.



Another possibility is to assign the same distance value to such municipalities, since theyshare many similarities (73% of the municipalities lacking employment offices are locatedin urban areas). Accordingly, once the distribution of dij had been considered, we imputed avalue of 1 km. for these municipalities (7.2% out of total). Gutiérrez-i-Puigarnau et al. (2010)applies a similar solution to that of ours in the assignment of daily commuting distances forworkers who commute to workplace locations within their municipality of residence.

Accessibility levels have also been calculated using the gravity potential measure so that wecould use other well-known expressions.

AEO

diIb j

ij

= α (2)

In this case, a is a parameter of the distance-decay function. The higher the value of theparameter, the greater will be the resulting differential between near and distant municipalities.This value crucially depends on the type of activity involved (Holl 2007). Higher values areusually assigned to accessibility measures of public services. In this study, the parameter hasbeen set to a value of 2 and 1.5.7

We have refined these measures by including the number of employment offices togetherwith the distance and size of their catchment areas. Consequently, the proposed accessibilitymeasure is more empirically adequate, since some employment offices serve approximately20,000 jobless, for example, Fuenlabrada (Madrid), while others serve just 1,000, for example,Caudete (Albacete). The accessibility to employment services is determined by this fact and thatcannot be overlooked.

The second type of measure (denoted by the superscript II) is based on the number ofemployment offices per unemployed worker within a catchment area, adjusted for the distancebetween the municipality i and its corresponding employment office.

AEO

ue A p ei

IIa j

ii j

diIIa

jdij ij=

⎡

⎣

⎢⎢⎢

⎤

⎦

⎥⎥⎥

→ = ⎡⎣ ⎤⎦∈

− −

∑( ) ( )λ λ

(3)

where Ai is the municipality accessibility, pj is the number of employment offices (EOj) per

employment office catchment area (∑∈i j

iu ), measured as the number of unemployed workers

in the municipalities i within a single catchment area. Finally, dij is the distance between amunicipality i and its corresponding employment office, and l is a parameter of the distance-decay function.

Similarly, the index AiIb has been modified so that we know which municipalities imple-

ment active labour market policies more extensively. The higher the value of pj, the greater is thepotential of the employment office for providing better service.

AEO

ud A p di

IIb j

ii j

ij iIIb

j ij=⎡

⎣

⎢⎢⎢

⎤

⎦

⎥⎥⎥

→ = [ ]∈∑

α α (4)

Again, a is a parameter of the distance-decay function and adopts the values set before.

7 Bruinsma and Rietveld (1993), Gutiérrez (2001) and Holl (2007) assume a = 1 in their respective analyses of theaccessibility to economic activity.

8 P. Suárez et al.


One concern that may arise is the inclusion in the accessibility measure of both the numberof job-seekers and offers per catchment area. That is not essential, for PESs are not aimed atproviding job vacancies directly to the unemployed, but adequate job search assistance. Besides,the number of unemployed workers is highly correlated with the economic situation and thenumber of job offers.

Information about the Spanish municipalities (surface, distance to employment offices andaccessibility indicator) is provided to ensure the understanding of the results. Table 2 shows theaverage of these variables in each autonomous community.

There may be observed differences in accessibility, surface and distance to employmentoffices across the Spanish regions. In view of the results, we may not establish a clear relationbetween the size of a municipality and its accessibility. That is hardly surprising becausefirst there are no pre-existing criteria for the definition of a municipality as such, but the localadministrative division arises from historical reasons. Second the location pattern of employ-ment offices stems from political/electoralist decisions, although it is fair to say that they aremostly located in higher unemployment areas (see Suárez 2011).

Once the accessibility measures have been defined, the following step consists of analysingtheir spatial distribution.

3 Accessibility measure clustering

The methodology used in this paper to analyse the geographical differences in access toemployment services relies upon the exploratory spatial data analysis (ESDA). This type ofanalysis allows us to identify the main clusters of municipalities with higher numbers ofunemployed and test whether the level of accessibility to employment offices is also higher inthem. Talen and Anselin (1998) point out the advantages of using local indicators of spatialassociation (LISA) and focus on the fact that this facilitates the detection of relevant patternsof local spatial association. Tsou et al. (2005) also recommend a spatial analytical perspective toevaluate suitability of urban public facilities in assessing whether or not, or to what degree, thedistribution of urban public facilities is equitable.

Table 2. Average municipality surface, accessibility and distance

Regions Surface Accessibility AiIIa Distance to

employment offices

Madrid 44.33 0.03 21.51Valencian Com. 42.93 0.07 14.30Balearic Islands 74.97 0.07 11.41Canary Islands 85.62 0.07 10.61Catalonia 34.02 0.08 14.32Cantabria 51.64 0.10 14.18Basque Country 28.35 0.11 10.33Murcia 251.36 0.12 6.41Galicia 94.19 0.15 12.54Aragon 65.12 0.15 28.14Navarre 30.82 0.16 16.11Extremadura 108.82 0.16 19.02Castile and Leon 40.82 0.18 23.52Andalusia 113.63 0.21 12.80Castile-La Mancha 86.22 0.21 22.96La Rioja 28.81 0.29 14.63Asturias 136.04 0.31 11.21



Using the information available, we intend to compare the spatial distribution of theunemployed and the existing clusters of unemployed workers with the distribution of offices sothat we may establish the degree of matching between the two distributions. Notwithstandingthat, not only is the spatial pattern of the offices relevant, but more complex aspects mustalso be taken into account, such as those relating to the accessibility indices calculated. Ideally,accessibility to employment offices should be kept at an adequate level even in high localunemployment rate contexts – in other words, there should be no clusters of municipalities withlow accessibility levels.

This section examines global and local spatial autocorrelations in local unemploymentrates, employment offices and accessibility measures. First, we analyse the existence of spatialautocorrelations using Moran’s I and the randomization approximation (Table 3). Since thestatistics are significant, all the variables show positive spatial autocorrelation, which suggeststhe existence of spillovers across municipalities. That is, the spatial structure of these variablesis clear so that none is scattered randomly or independently in space.

3.1 Spatial distribution of the unemployed and employment offices

Once the null hypothesis of spatial randomness has been rejected, two additional questions areraised: where are the clusters and what is their spatial extent (Fischer and Getis 2010). We haveused the LISA statistic (Anselin 1995) which describes the degree of similarity or dissimilaritybetween values in spatially close areas. The local version of Moran’s I for each municipality iscomputed as follows:

Iz

z nw z z x xi

i

ii

ij j i ij= = −

∑ ∑2; (5)

where a positive value for Ii indicates spatial clustering of similar values (high or low), whereasa negative value indicates spatial clustering of dissimilar values between a municipality and itsneighbours.

Table 3. Measure of global spatial autocorrelation (Moran’s I)

Variables I Z

Unemployed people 0.147 24.334Local unemployment rate 0.574 85.300Employment offices* 0.119 18.214Ai

Ia (l = 0.1) 0.618 91.427(l = 0.25) 0.505 75.891

AiIIa (l = 0.1) 0.625 92.272

(l = 0.25) 0.624 91.711Ai

Ib (a = 2) 0.142 21.562(a = 1.5) 0.165 25.126

AiIIb (a = 2) 0.057 8.891

(a = 1.5) 0.076 11.738

Notes: All statistics are significant at the 1% level. The expected value forMoran’s I is -1.234e-04. We also applied Moran’s I to the square root trans-formed employment offices variable due to the large number of municipalitieswithout employment offices (I = 0.137; Z = 20.230). The conclusion is the samewhen BB joint-count statistics and empirical Bayes test are computed (EB,Assunção and Reis 1999); the p-value is 0.001 and 0.016 respectively.

10 P. Suárez et al.


Figure 4 shows the LISA map for unemployed people in Spain for 2009. Most significanthigh-high (HH) municipalities are located in southern Andalusia, Murcia, central Asturias,Madrid, Barcelona, Tenerife and Las Palmas de Gran Canaria, among others.

The map also points to the existence of clusters of low-low (LL) municipalities, that is,municipalities where the total number of unemployed people are significantly below average –which, in turn, are surrounded by municipalities with similar values. Most significant LLmunicipalities are located in Castile and Leon and Aragon.

As regards the LISA map for local unemployment rates (Figure 5), it seems that Spanishmunicipalities are characterized by positive spatial autocorrelation, as in the case of thelevelled variable. In this case, the clusters (HH and LL) are made up of a greater number ofmunicipalities, and two areas stand out very clearly: HH in the south and LL in the north-east.The map also reveals the existence of some atypical municipalities, characterized by negativespatial autocorrelation.

Below we analyse the clusters of municipalities according to the number of employmentoffices within their territory. Our purpose is to establish a relation between these clusters andthose of unemployed workers and test whether the employment offices are located in munici-palities belonging to HH clusters.8

There is a clear pattern of HH spatial clustering in the south (Seville, Cordoba and Cadiz),Valencia, Alicante, Murcia, Barcelona, Madrid, central Asturias and Extremadura. The existenceof HL municipalities and the non-existence of LL clusters are good in terms of equity, forit ensures the existence of an employment office near any municipality. In other words, thereare no big clusters of municipalities lacking employment offices. Figures 4 and 6 confirm the

8 Since this variable does not conform to a normal distribution (7,254 out of 8,109 municipalities have no employ-ment offices), we have also transformed it by calculating its square root (Talen and Anselin 1998).

Fig. 4. LISA map for unemployed people (2009)



Fig. 5. LISA map for local unemployment rates (2009)

Fig. 6. LISA map for employment offices



overlapping between HH spatial clusters of unemployed workers and employment offices.Employment offices are located around municipalities with high numbers of jobless. Thisdistribution may be deemed efficient, but it is not equitable. Nevertheless, limiting the measureto the number of employment offices is far too simple, since we need to consider some otherissues which also have an effect on employment office accessibility.

3.2 Results based on the comparison of accessibility indices

This subsection compares the accessibility measures proposed by resorting again to Moran’slocal indicators. Generally speaking, decreases in accessibility should be expected as we proceedfurther away from major towns.

Figure 7 shows the LISA maps for the accessibility indices AiIa (upper half of figure) and

AiIIa (lower half of figure). When we examine the LISA for the Ai

Ia index, in which only thenumber of employment offices and the distance have been considered, we notice the presenceof clusters that are coincidental with those in Figure 4 (815 out of 1,120 municipalities show thesame type of association in both variables).

However, when the number of employment offices per catchment area (pj) is taken intoaccount, the Ai

IIa index shows a different spatial distribution (only 351 out of 1,185 munici-palities show the same type of association in both variables). In this case, the HH and LL clustersare not coincidental with those detected using the Ai

Ia index, but they include municipalitiesnot necessarily linked to major cities. Therefore, previously detected HH accessibility clustersdisappear in Madrid, Barcelona and their surrounding municipalities, as well as Extremadura,Cadiz, the Balearic Islands and the Canary Islands.

The most interesting results are obtained when we analyse the LL clusters detectedespecially in Madrid, along the Mediterranean coast, as well as Toledo, Zaragoza, the BalearicIslands and the Canary Islands. This is due to the fact that, even though most employment officesare concentrated in urban areas, these offices are not sufficient to attend to the high number ofjobless from the city itself and the surrounding municipalities.

Figure 8 shows the LISA maps for the accessibility indices AiIb (upper half of figure)

and AiIIb (lower half of figure). When we examine the LISA for the Ai

Ib index we may noticethat, in general, some HH clusters are coincidental with those detected using the accessibilityindex Ai

Ia and, therefore, with those in Figures 4 and 6. In the case of the accessibility indexAi

IIb, its spatial pattern does not differ greatly from that of the index AiIb. Even though it is true

that some HH clusters in Figures 4 and 6 disappear (i.e. Madrid, Barcelona and Valencia) whenwe take into account the number of unemployed people, the HH cluster in Cadiz remains andwestern Asturias becomes an area with high levels of accessibility to employment offices. Asregards the LL clusters, there is no relevant difference between the Ai

IIb and the AiIb indices. The

results show the influence of the proposed functional form, in which there are greater decreasesin accessibility as distance increases. Therefore, it is less sensitive to the variations in the otherterms of the expression, namely pj.

To sum it up, both Figures 7 and 8 reveal differences in the levels of accessibility to publicemployment services. The detected LL clusters are especially worrying, even more so if theseare coincidental with HH clusters of unemployed people or local unemployment rates. It isfor this reason that, on the basis of the research carried out, it is deemed more adequate to usean accessibility measure based on exponential expressions and take into account the size of theemployment office catchment area in order to include a competition factor. Job-seekers have toqueue at some employment offices due to the high number of unemployed people and, conse-quently, office performance gets compromised. Although, most companies use Internet selfservice solutions, physical presence is still necessary for the unemployed, especially when theyrequest job mediation, counselling and training.



Fig

.7.

LIS

Am

aps

for

indi

ces

A iIaan

dA iII

a



Fig

.8.

LIS

Am

aps

for

indi

ces

A iIban

dA iII

b



Finally, we may also consider how the results would vary if the scope of the analysis wasexpanded to include other years besides 2009. It should be taken into account that on the one hand,the spatial distribution of offices remains quite stable over time, but on the other hand, the sizeof their catchment areas is very variable. To analyse the sensitivity of the measure to the currenteconomic situation, we calculated the accessibility index Ai

IIa for 2006 (a ‘good’ year for theSpanish economy). The 2006–2009 correlation coefficients obtained show the persistence ofaccessibility levels regardless of the economic cycle moment.9 Broadly speaking, the 2006 acce-ssibility index is higher than that of 2009, but its corresponding spatial distribution is quite alike.

4 Local unemployment rates and access level to employment offices

4.1 Theoretical framework

Finally, we will consider in this section whether the accessibility to employment offices haseffect on local unemployment rates. Within the field of labour market studies, several contri-butions have taken into account the spatial dimension of regional labour markets and pointedout the high degree of interdependence of local labour markets (Molho 1995; Lopéz-Bazoet al. 2002; Overman and Puga 2002). Furthermore, Patacchini and Zenou (2007) analyse thereasons for the spatial dependence in local unemployment rates. This spatial autocorrelation ismainly due to the fact that the unemployed may seek and find work in different areas, so spatialinteractions result from the mobility of the unemployed.10 This paper adds consideration ofspatial dependence in local unemployment rates to the diverse influences exerted by publicemployment services across different levels of accessibility.

Recent studies on spatial job search have shown that distance to jobs may reduce thelikelihood of leaving unemployment (e.g. Détang-Dessendre and Gaigné 2009). Ihlanfeldt(1997) asserts that labour market information acquisition is considered a type of investmentbehaviour. At present, theory suggests that the unemployed will go to placement offices in searchof information or job-broking services when benefits are greater than costs. The unemployed mayrefuse to go to a placement office because travelling expenses are too costly and, in some cases,they have to queue at the office.

From a political perspective, insofar as the relation between unemployment rates andaccessibility to employment offices remains negative, investment in bettering accessibilitywill be regarded as meaningful. Joassart-Marcelli and Giordano (2006) point out that ‘one-stops’ are well positioned to serve the unemployed and that access to them does help to reducelocal unemployment rates. The accessibility measure in our study, unlike that in Joassart-Marcelli and Giordano’s, takes into consideration that whenever a job-seeker finds work,the unemployment rate in their municipality of residence is reduced, accessibility levels (pj)grow in municipalities within the regional labour office and, consequently, the performance ofthe employment services gets improved. When we refer to employment services, we mean notonly job-seeking mediation but also career counselling, which allows the identification anddevelopment of each individual’s talent (Servicio Público de Empleo Estatal 2008). A com-prehensive study on the impact of the accessibility to placement offices on job accessibility isstill pending, but that is beyond the scope of this paper.

9 This analysis is included on the suggestion of one of the referees. The 2006–2009 correlation coefficients obtainedwere 0.920 (Andalusia), 0.898 (Aragon), 0.998 (Asturias), 0.978 (Balearic Islands), 0.910 (Basque Country), 0.921(Canary Islands), 0.997 (Cantabria), 0.958 (Castile and Leon), 0.978 (Castile-La Mancha), 0.949 (Catalonia), 0.968(Extremadura), 0.982 (Galicia), 0.968 (La Rioja), 0.937 (Madrid), 0.957 (Murcia), 0.935 (Navarre) and 0.921 (ValencianCom.).

10 Longhi and Nijkamp (2007) show that spatial models improve the forecasting performance of non-spatial models,provided that the data available are not correspondent with a well-defined local labour market area.



Regional unemployment differentials have been widely analysed theoretically and empiri-cally. Elhorst (2003) has reviewed the papers on regional and labour economics publishedsince 1985 and distinguished four models of regional unemployment, namely the singleequation model, the implicit model, the accounting identity and the simultaneous equation model.Notwithstanding that, Elhorst (2003, p. 724) also asserts that “Whichever model is used . . . theyall result in the same reduced form equation of the regional unemployment rate”. In this case, theunavailability of economic information for every Spanish municipality has led us to proposea single equation model. One of the problems arising from this model is the placement of signexpectations on the explanatory variables for local unemployment rates, which is hindered by theexistence of conflicting theories and the uncertain overall effect of the said variables, since theymay affect supply, demand and/or wages. Be that as it may, the effects of the explanatory variablesmust be interpreted with caution and ensuring they are consistent with economic theory.

We should also carefully select the variables to incorporate in the empirical model. Ideallylabour supply, labour demand and wage setting factors should be incorporated as explanatoryvariables. In this case, the need to use municipal level data for the reasons described above causesa trade-off between the high level of spatial disaggregation and the economic data available.

The model in this paper includes as explanatory variables the rates of foreign populationand males and females of working age, the educational attainment of the population, industries’employment shares (so that disparities in the local economy structure may be accounted for)and two dummy variables, one for municipalities within HH clusters and the other for munici-palities within LL ones. The local accessibility level to placement offices is also included. All thevariable related information is in Table 4. The basic specification is:

log u log A X ei iIIa

i i( ) = ( ) + +η β (6)

where ui is the unemployment rate of each municipality, AiIIa is the selected accessibility

measure and the X matrix collects the explanatory variables described above. In previoussections, we have established the existence of spatial dependence in unemployment rates, sospatial models must be considered in our specification.

Table 4. Summary statistics

Variable Mean SD Definition Data source

Local unemploymentrate

0.087 0.043 Unemployed population / Total populationof working age (16–64)

SPEE and MunicipalRegister

ILLI 0.024 0.028 % Illiteracy Population CensusPRI* 0.324 0.149 % Complete primary education Population CensusSEC 0.396 0.138 % Secondary education or vocational training Population CensusUNI 0.079 0.048 % Higher education Population CensusHH – – HH cluster Own elaborationLL – – LL cluster Own elaborationAi

IIa with l = 0.10 0.155 0.204 Accessibility measure Own elaborationAi

IIa with l = 0.25 0.088 0.148 Accessibility measure Own elaborationFLF 0.571 0.103 Female population 16–64 / Total female population Municipal RegisterMLF 0.644 0.074 Male population 16–64 / Total male population Municipal RegisterFOR 0.088 0.093 Foreign population (16–64) / Total population

of working age (16–64)Municipal Register

WI 0.191 0.119 Share of employment in industry Population CensusWB 0.115 0.078 Share of employment in construction Population CensusWS* 0.628 0.221 Share of employment in services Population Census

Note: * The percentage of population with incomplete primary education and the share of employment in agriculturehave been omitted so as to avoid multicollinearity.



4.2 Empirical model

First, the model has been estimated by means of ordinary least squares (OLS) (Table 5).Both local unemployment rates and accessibility measures have been considered in logarithmicform, but it should be stressed that the use of these variables in levels makes no considerabledifference. The accessibility measure Ai

IIa with l = 0.10 is included in models I and II, whereasl = 0.25 in models III and IV.

The effect of the accessibility to placement offices is significant and negative (-0.062 modelI and -0.026 model III). In model I, the unemployment rate decreases by 0.062 per cent whenaccessibility rises by 1 per cent. This estimated elasticity diminishes when the accessibilitymeasure Ai

IIa with l = 0.25 is included in the model.The models above have been estimated with the correction by Condeço-Melhorado et al.

(2011) so as to test the robustness of the accessibility measure in relation to the self-potentialissue. The results have been quite similar, for example, when l = 0.1 (model II), the estimatedcoefficient is -0.044 (0.004) and when l = 0.25 (model IV) with -0.011 (0.004).

The coefficients’ signs are as expected and in accordance with previous theoretical andempirical studies. With the exception of the percentages of males of working age (MLF) andpopulation with complete primary education (PRI) in models II and IV, all the coefficientsare significant. Additionally, we may conclude that the effect of the percentage of females ofworking age (FLF) is higher than that of the males.11

Standard tests have been carried out so as to assess the adequacy of the regressions. TheBreusch-Pagan test for homoscedasticity of the error terms points to heteroscedascity, which inturn is related to the different sizes of the municipalities considered. In any case, since spatial

11 Cracolici et al. (2007) reach the same conclusion for the Italian provinces.

Table 5. OLS regression of local unemployment rate

Model I Model II Model III Model IV

Intercept -2.947 (0.062)*** -3.316 (0.123)*** -2.939 (0.062)*** -3.376 (0.124)***Ai

IIa with l = 0.10 -0.062 (0.006)*** -0.058 (0.005)*** – –Ai

IIa with l = 0.25 – – -0.026 (0.005)*** -0.026 (0.004)***FLF 1.067 (0.114)*** 0.890 (0.130)*** 1.148 (0.115)*** 0.995 (0.130)***MLF -0.295 (0.092)** -0.066 (0.166) -0.293 (0.092)** -0.094 (0.167)HH 0.333 (0.025)*** 0.282 (0.013)*** 0.350 (0.025)*** 0.295 (0.013)***LL -0.278 (0.015)*** -0.115 (0.016)*** -0.291 (0.015)*** -0.124 (0.016)***ILLI 4.571 (0.232)*** 3.686 (0.268)*** 4.549 (0.234)*** 3.684 (0.271)***PRI -0.126 (0.046)** -0.055 (0.550) -0.128 (0.046)** -0.049 (0.050)SEC -0.127 (0.055)** -0.276 (0.057)*** -0.100 (0.056) -0.254 (0.057)***UNI -2.070 (0.134)*** -2.065 (0.118)*** -2.023 (0.134)*** -2.031 (0.120)***FOR -0.176 (0.060)*** -0.494 (0.051)*** -0.141 (0.060)** -0.458 (0.052)***WI – 0.415 (0.105)*** – 0.484 (0.106)***WB – 0.704 (0.107)*** – 0.770 (0.108)***WS – 0.494 (0.100)*** – 0.562 (0.101)***

Breusch–Pagan 614.61*** 236.41*** 563.12*** 226.87***Kolmogorov-Smirnov 0.207*** 0.248*** 0.208*** 0.248***R2 (adj.) 0.271 0.268 0.262 0.257Number of obs. 7,754 6,292 7,754 6,292Log-likelihood -5,164.954 -2,457.991 -5,211.458 -2,505.280AIC 10,351.91 4,943.98 10,444.92 5,038.56SBC 10,428.42 5,038.26 10,521.43 5,132.84

Note: ***, ** and *denote significant at 1%, 5% and 10%, respectively.



dependence may cause this heteroscedasticity (McMillen 1992), the result has been interpretedwith caution. Similarly, Anselin and Bera (1998, p. 250) assert that “every type of spatiallydependent error process induces heteroscedasticity as well as spatially autocorrelated errors,which will greatly complicate specification testing in practice”. We may also note that theKolmogorov-Smirnov test has led us to reject the assumption of normality of the residuals.12

Another issue is whether the accessibility variable is endogenous. The measure considersthe size of the catchment area in terms of the number of unemployed people. Accordingly, anyunemployment increase in a municipality may cause, ceteris paribus, a decrease in the acces-sibility of the catchment area. Also, it may increase the local unemployment rate, which in turncould lead to reverse causality endogeneity. Notwithstanding that, it should be borne in mindthat most catchment areas are made up of more than one municipality and, consequently, anunemployment increase in a municipality does not necessarily decrease its accessibility, sincethat depends on all the municipalities within the catchment area.

Be that as it may, the endogeneity of the variable has been tested taking into account theexistence of heteroscedasticity. The scarcity of data at municipal level limited the selectionof instrumental variables to geographic and demographic features. More specifically, we haveconsidered three potential instruments for the accessibility variable: den – population densityin each municipality; rpopofi – number of people assigned to each public employment office; andselfdistance – internal distance of each municipality, as shown in Rich (1975).13 We used theKleibergen-Paap rank F-test14 on the excluded instruments to test the relevance of the potentialinstruments. Since the value is considerably larger (247.90), the instruments may be consideredstrong.

Once the relevance of the instruments had been tested, a robustified Durbin-Wu-Hausman(DHW) test and the bootstrap Hausman test were computed to test the null hypothesis that theaccessibility measure is exogenous. In the presence of heteroscedastic errors, the first optioncan be implemented provided that we use robust variance whereas the second overrides the needto assume that one of the estimators is fully efficient under the null hypothesis. There is noevidence to reject the null hypothesis that accessibility is exogenous in either case. In the firstoption, the value of the statistic is F6197

1 0 587= . , whereas in the second option the value of theWald statistic associated with the potentially endogenous variable is χ1

2 0 129= . .Put in context, the result is not surprising because, in accordance with our own theoretical

proposal, the accessibility to employment offices of the unemployed in a municipality dependson the total number of unemployed workers in the municipalities within the same catchmentarea, the total number of offices and the distance between the said municipality and its corre-sponding office, which make the reverse causality argument harder to sustain.

Second, we analyse the existence of spatial autocorrelation. Moran’s I is widely used to detectspatial dependencies based on OLS residuals. The resulting statistic standard deviation is 41.462in model II and 35.748 in model IV. Here we have used a row-standardized rook contiguity matrixso that w w wij

sij j ij= ∑ when i � j and wij

s = 0 when i = j. At this point, we could consider thatthe accessibility-related variable fully tackles the spatial dependence in the dependent variable,as Martin and Gråsjö (2009) show in their study. However, we should also bear in mind that theaccessibility measure in this paper only covers some of the spatial interactions within locallabour markets, namely, those related to the activity of public employment services.

Once spatial autocorrelation has been detected, we may proceed to incorporate it into theproposed model. In spatial econometrics, spatial autocorrelation is modelled by means of the

12 The Kolmogorov-Smirnov, Cramer-vol Mises and Anderson-Darling tests are recommended when N is large.13 The instruments den and rpopofi were computed considering the population levels in 1950 and the conclusions of

the exogeneity test were the same.14 The Kleibergen-Paap rank F-test is a generalization of the Cragg-Donald (Cragg and Donald 1993) statistic that

allows for heteroscedastic errors.



relation between the dependent variable Y or the error term and its associated spatial lag, Wyfor a spatially lagged dependent variable (spatial lag model) and We for the spatially lagged errorterm (spatial error model) respectively. The general form for the spatial lag model is:

log u log A Wlog u X N IIIa( ) = ( ) + ( ) + + ≈ ( )η ρ β ε ε σε; ,0 2 (7)

where Wlog(u) is the spatially lagged dependent variable for the weight matrix W, r is the spatialautoregressive parameter, h is the accessibility coefficient and b is a vector of regressionparameters.

The most common specification for the spatial error model is:

log u log A X e

e We N I

IIa( ) = ( ) + +

= + ≈ ( )η β

θ ε ε σε; ,0 2(8)

where q is the spatial autoregressive coefficient for the error lag We.Only a few papers deal with how to specify a spatial econometric model. Then the problem

is how to best identify the structure of the underlying spatial dependences in a given data set.This paper relies on a widely used strategy (specific to general), which is based on the Lagrangemultiplier (LM) test and its robust version for local misspecifications (Anselin et al. 1996).

In the classical approach, the Lagrange multiplier for error dependence (LMERR) andthe Lagrange multiplier for spatially lagged dependent variable (LMLAG) are compared. If theLMERR is lower than the LMLAG, the spatial lag model should be specified. If not, the spatialerror model is to be specified. Florax et al. (2003) have developed a hybrid approach based onthe robust version of these tests. Mur and Angulo (2009), however, point out that the robust andthe classical approaches render identical results.

These tests have been computed on OLS residuals of the previously estimated models.We have also considered different criteria to build the spatial weight matrices that allowed us toanalyse the sensitivity of the results. As regards the structure of the spatial effects, three criteriaare usually considered in the creation of a spatial weight matrix: contiguity, k-nearest anddistance. First, we define a rook contiguity matrix, where wij = 1 if municipalities i and j sharea common edge and wij = 0 otherwise. Second, we apply a k-nearest neighbours’ criterion(k = 3, 4 ¥ 5). Then, we obtain a distance-based matrix, where wij = 1 if dij � d and wij = 0if i = j or d > dij (d = 10, 20, 30 and 40 km).15 Table 6 shows the values of the test statistics. Thep-value is included only if it is greater than 0.01.

For the matrices Wq, WK3, WK4, WK5, W10 and W20, the robust version of these tests(R-LMLAG and R-LMERR) renders the same results and LMLAG>LMERR, so the spatial lagmodel is appropriate. However, when these tests are computed with W30 and W40, the spatial

15 Other options for defining the spatial weight matrix are also available in the literature, for example, the inversedistance matrix. This option has been excluded to guarantee non-overlapping between the accessibility measure and thespatial weight matrix.

Table 6. Spatial dependence statistics by alternative spatial weight matrices

Wq WK3 WK4 WK5 W10 W20 W30 W40

LMERR 1,707.659 1,716.667 1,733.624 1,988.605 1,399.577 3,924.033 6,431.013 8,552.674LMLAG 1,931.005 2,168.793 2,186.351 2,495.631 1,770.705 4,226.901 5,492.595 5,872.466RLMERR 65.739 9.812 9.888 2.650 (0.103) 10.881 315.179 1,805.844 3,422.358RLMLAG 289.085 461.937 462.614 509.676 382.010 618.047 665.619 742.150SARMA 1,996.744 2,178.605 2,196.239 2,498.281 1,781.587 4,542.081 7,096.633 9,294.824



error model becomes more appropriate. Notwithstanding this, our research is based on localdata and, from an economic perspective, any attempt at considering long distance is difficult tojustify. Anselin (2002) describes the problems associated with the criterion of distance. A highdegree of heterogeneity in the municipalities makes it difficult to find a suitable critical distance.On the one hand, a small distance renders many municipalities isolated from the rest, but on theother hand, a distance large enough to ensure that every municipality has at least one neighbourmay result in small municipalities having large numbers of neighbours when the matricesW30 and W40 are considered.16 In fact, when these matrices are used, we lose the advantageof working with a high level of spatial disaggregation. Additionally, a spatial Hausman testhas been applied to detect the presence of omitted variables. The null hypothesis17 (statistic value= 410.925) may be rejected in this case and hence a model with a spatial lag of the dependentvariable is more plausible than a spatial error model.

Consequently, a spatial lag specification has been chosen and, more specifically, one basedon both the economic theoretical framework and the results of the specification test. Similarly,LeSage and Pace (2009) assert that spatial lag models have been used in contexts where thereis a theoretical motivation for Y to be dependent on neighbouring values of Y. In this sense,Molho (1995) and Patacchini and Zenou (2007) provide theoretical explanation for the spatialcorrelation between unemployment rates.

Maximum likelihood (ML) is the most conventional estimation method for a standard spatialautoregressive model (SAR) where the error terms are assumed to follow a normal distribution.The computational complexities of the Jacobian term (|I - rW| in the SAR model and |I - qW|in the SEM model) represent the main problem of this method. This computational problem issorted out by means of the simplification solution proposed by Ord (1975) or the approximationoption developed by Smirnov and Anselin (2001).

The use of the spatially lagged dependent variable Wy as explanatory variable may beunderstood as a form of endogeneity or simultaneity leading to the instrumental variable approach(IV) / two stage least squares (2SLS). Anselin (1988) considers this method more appropriatewhen the error terms are not normally distributed, but some recent studies have pointed out theinefficiency of 2SLS estimators (2SLSE), especially when compared to the maximum-likelihoodestimator (MLE). Furthermore, 2SLSE will be inconsistent when the exogenous regressors areirrelevant (Lin and Lee 2010).

The generalized method of moments estimator (GMME) for the autoregressive parameterin a spatial model, proposed by Kelejian and Prucha (1999), also allows us to solve the problemspreviously described. They prove that the GMME is consistent without the assumption ofnormality. More recently, Lin and Lee (2010) have shown the robustness of the GMME underunknown heteroscedasticity – a context in which the MLE is usually inconsistent.

The local unemployment rate equation has been estimated by means of ML, 2SLS andgeneralized methods of moments (GMM) (Table 7). We have also considered some spatialweight matrices based on either geographic contiguity (municipalities sharing boundaries) ordistance between municipalities, but we present only the results obtained with a k-nearestneighbour matrix k = 5.

Generally speaking, it should be noted that the results are qualitatively similar across thedifferent methods. Also, they are quantitatively the same when 2SLS and GMM are compared.The first column in Table 7 shows the estimation results of the model by ML. The coefficient ofthe spatial lag term is 0.54 and is highly significant, but according to the LM test for residualautocorrelation, uncontrolled spatial effects remain in the residuals.

16 For example, in the matrices W30 and W40 there are municipalities with 147 and 201 neighbours, respectively.17 The Hausman test statistic follows a chi-squared distribution with degrees of freedom equal to the number of

explanatory variables.



According to the results obtained, the unemployment rates in the Spanish municipalitiesshow strong spatial autocorrelation, with an estimated spatial coefficient of around 0.84 (2SLSand GMM). Thus, about 84 per cent of the changes in the unemployment rates of neighbouringmunicipalities will be absorbed by a municipality’s own unemployment rate. A possible expla-nation for the smaller value of the estimated spatial coefficient when the model is estimated byML could lie in the non-normality of the error term and the aforementioned heteroscedasticityproblem.18 Therefore, 2SLS and GMM are more adequate.

In accordance with our hypotheses, local unemployment rates seem to be inversely related tothe accessibility measure. Its coefficient is significant and negative, but it is constrained to -0.029(ML), -0.009 (2SLS) and -0.010 (GMM). If we were analysing two linear regression models, itcould be possible to conclude that the accessibility elasticity is lower when spatial autocorrela-tion is included explicitly into the model. However, the interpretation of the parameters is morecomplicated in models containing the spatial lag of the dependent variable. In the spatial lagmodel, any change in the dependent variable for a single region may affect the dependent variablein all the other regions. Thus, a change in one explanatory variable in the municipality i will notonly exert a direct effect on its own unemployment rate, but also an indirect effect on theunemployment rates of other municipalities. As consequence, the impact that a change in oneof the explanatory variables has on the dependent variable of a region is not usually equal to itsestimated coefficient.

LeSage and Pace (2009) propose new measures to collect all these interactions betweenregions so that we may reach a correct interpretation of the spatial models and distinguish

18 Lin and Lee (2010) show that the MLE estimator is generally inconsistent with unknown heteroscedasticity if theSAR model were estimated as if the disturbances were independent and identically distributed (i.i.d.).

Table 7. Estimation results for spatial models

ML 2SLS GMM GMM–H

Intercept -1.602 (0.067)*** -0.843 (0.224)*** -0.808 (-0.091)*** -0.568 (0.071)***Ai

IIa with l = -0.10 -0.029 (0.004)*** -0.009 (0.006)** -0.010 (0.004)*** -0.007 (0.002)**FLF 0.439 (0.092)*** 0.189 (0.120) 0.182 (0.093)** 0.148 (0.078)*MLF 0.189 (0.113) 0.297 (0.143)** 0.326 (0.112)*** 0.142 (0.096)HH 0.157 (0.016)*** 0.077 (0.017)*** 0.082 (0.017)*** 0.067 (0.009)***LL -0.034 (0.012)*** 0.014 (0.018) 0.014 (0.013) -0.008 (0.009)ILLI 1.901 (0.171)*** 1.034 (0.289)*** 0.097 (0.181)*** 0.736 (0.151)***PRI 0.024 (0.037) 0.084 (0.043)* 0.081 (0.042)** 0.041 (0.029)SEC -0.058 (0.045) 0.055 (0.054) 0.059 (0.045) 0.038 (0.034)UNI -1.260 (0.102)*** -0.824 (0.137)*** -0.823 (0.106)*** -0.597 (0.074)***FOR -0.286 (0.042)*** -0.123 (0.508)** -0.148 (0.043)*** 0.112 (0.031)***WI 0.115 (0.040)*** 0.191 (0.092)** 0.141 (0.040)*** 0.107 (0.030)***WB 0.267 (0.039)*** 0.256 (0.100)** 0.209 (0.039)*** 0.159 (0.033)***WS 0.076 (0.033)** 0.081 (0.093) 0.029 (0.032) 0.066 (0.027)*r 0.541 (0.011)*** 0.848 (0.064)*** 0.847 (0.026)*** 0.873 (0.019)***l – – – -0.639 (0.024)***

Breusch–Pagan 266.722*** 398.432*** 259.929*** 299.643***Kolmogorov–Smirnov 0.250*** 0.278*** 0.231*** 0.276***R2 0.453 0.519 0.514 0.507LM test 464.3*** – – –Number of obs. 6,292 6,292 6,292 6,292Sigma^2 0.089 0.085 0.086 0.087

Notes: In the 2SLS model, the spatial lag of the explanatory variables is included as instrumental variables (WX). ***,** and * denote significant at 1%, 5% and 10%, respectively.



between the direct and the indirect impact. The direct impact shows the average response of thedependent variable to independent variables, including feedback influences that arise fromimpacts passing through neighbours and back to the municipality itself.19 The indirect impacttackles the effect that any change in a region has on others and how changes in all regions affecta region. Table 8 shows the estimated direct and indirect impacts by means of 2SLS and GMM.

The accessibility to placement offices has a slightly higher (and significant) direct effectthan the coefficient estimate. This difference is caused by impacts passing through neighbouringregions and back to the region itself. Consequently, a positive feedback effect is obtained.Even more interesting is the estimation result of the indirect impact, which is significant andhigher than the coefficient estimate in the models, showing an influence of the accessibility toplacement offices across the spatial dependences between municipalities. The total impacts are-0.0673 and -0.0639 for GMM and 2SLS respectively. This means that if accessibility increasesby 1 per cent, the unemployment rate decreases by 0.0673 per cent/0.0639 per cent. Someconclusions for policy-makers may be drawn from this result.

Just to give you an example, in the municipalities of Valencia and Zaragoza, the creation ofa new employment office would translate into, ceteris paribus, a decrease of 0.75 per cent and1 per cent in the province’s unemployment rate, respectively. The creation of an office in amunicipality previously lacking one, results in higher increases in accessibility. Looking furtherinto the example provided on page 4, the opening of an employment office in the municipalityof Tudela and the subsequent reorganization of the catchment area, according to the estimatedmodel, the creation of the office entailed a decrease of 2.3 per cent in the local unemploymentrate. Be that as it may, the magnitude is in accordance with the expected, since not all registeredjobless have the same chance of leaving unemployment and finding a job. Some find it verydifficult to access employment, which limits the impact on unemployment rates.

All coefficients of the independent variables – except PRI, SEC and WS – are statisticallysignificant (Table 7). In addition to this, there is evidence in support of the geographicalperspective hypothesis on persistent unemployment. The coefficient of the dummy variable HHis positive, which means that a municipality belonging to an HH cluster is strongly constrainedby this spatial pattern. However, the variable LL is not significant in 2SLS and GMM models.It is significant and negative in the ML estimation, which means that a municipality belongingto a LL cluster receives a positive influence in terms of unemployment rates. Regarding theestimation results of the educational attainment variables, the percentage of university graduatesis significant and negative, whereas the percentage of illiterates is significant and positive.

Finally, the residuals of the spatial lag model have been analysed to check whether thespatial autocorrelation had been fully removed. The result of the LM test is significant to rejectthe null hypothesis of no spatial correlation in the residual errors. However, as we explainedabove, the heteroscedasticity problem points to the specification of a model in which suchunknown heteroscedasticity in the error term may be controlled.

19 The main diagonal of higher order spatial weight matrices is non-zero, which allows us to collect these feedbackeffects.

Table 8. Direct, indirect and total impact estimations by means of 2SLS and GMM

Direct Indirect Total

Accessibility AiIIa (2SLS) -0.0142 (-2.24)*** -0.0497 (-2.15)*** -0.0639 (-2.20)***

Accessibility AiIIa (GMM) -0.0149 (-2.39)** -0.0523 (-2.35)** -0.0673 (-2.41)**

Accessibility AiIIa (GMM–H) -0.0112 (-2.46)** -0.0489 (-2.49)** -0.0504 (-2.52)**

Notes: Z–statistics in parentheses are based on 2000 simulated draws of the parameters and ***, ** and * denotesignificant at 1%, 5% and 10%, respectively.



As we described above, the GMM estimation is a good choice when normality cannot beverified. Recently, Kelejian and Prucha (2010) and Arraiz et al. (2010) have extended the GMMapproach to a spatial autoregressive disturbance process with heteroscedasticity innovations.It should also be noted that this specification allows for heteroscedasticity of unknown form. Thegeneral form is:

log( ) log( ) log( ) ; .u W u A X e e W eIIa= + + + = +ρ η β θ ε1 2 (9)

In this case, heteroscedasticity of unknown form is permitted with E(ei) = 0 and E i i( )ε σ2 2= . Thefourth column in Table 7 shows the estimation results of the model by GMM with heterosce-dasticity innovations.20 The estimated coefficient of the accessibility measure is negative andstatistically significant. We have obtained a strong spatial dependence between local unemploy-ment rates with a significant spatial effect. The direct and indirect effects of the accessibilityvariable have been calculated (Table 8), as was done in the estimations above and their valuesare slightly smaller than the others. Thus, the presence of heteroscedasticity has no main impacton the coefficient estimates of this empirical model when 2SLS and GMM methods arecompared.

5 Conclusions and policy recommendations

We have obtained that there are spatial differences across the local employment offices in Spain.We have computed the number of unemployed persons per employment office and found that,in some autonomous communities, the number of placement offices is far below the unemployedthey have to attend to (especially Madrid, the Canary Islands, the Valencian Community andCatalonia). Even though employment offices are located around municipalities with highnumbers of jobless, it may be concluded that, in terms of the accessibility to employment offices,this distribution is synonymous with spatial inequity.

In addition to that, we have improved the precision of the measure of the level of accessi-bility of a municipality to its corresponding employment office by including the size ofthe employment office catchment area in the accessibility measure. Also, we have detected themain clusters of municipalities with low accessibility to employment offices and those with highunemployment in 2009. The results suggest that policy-makers should strive to improve theaccessibility to placement offices, especially in the municipalities with low accessibility levelsso that adequate assistance to find suitable employment may be ensured to every job-seeker.

Using ML, 2SLS and GMM methods, we have shown a strong spatial correlation betweenunemployment rates, namely, neighbourhood influences are very important in labour markets.This view is consistent with other empirical studies such as Molho (1995) and Patacchini andZenou (2007) and, therefore, the spatial perspective cannot be ignored in the analysis of theSpanish labour market. Furthermore, in accordance with our hypotheses, unemployment ratesappear to be inversely related to the accessibility measure. In addition to that, when we computethe direct and indirect impacts of the accessibility measure on unemployment rates, the indirectimpact is shown to be higher than the coefficient estimate in 2SLS and GMM models. This,in turn, shows a positive influence on the reduction of unemployment rates across the spatialinteractions between municipalities. Also, it does strengthen our conclusions about the impor-tance of catchment areas and impact of accessibility measures on unemployment rates.

As a consequence of the decentralization process over the last 20 years in Spain, theautonomous communities have taken over the ALMPs. We recommend the creation and/or

20 The estimation has been done using the R package sphet (Piras 2010).



reorganization of the employment offices so that the distance and size of catchment areas maybe reduced. Support to inter office collaboration is also recommended, especially when it maylead to higher diffusion of job openings. However, we need to learn more about the efficiencyof the several types of services provided at each employment office. Therefore, greater infor-mation transparency should be encouraged so that researchers may evaluate the ALMPs imple-mented by each employment office.

References

Alonso-Villar O, Río C Del (2008) Geographical concentration of unemployment: A male-female comparison in Spain.Regional Studies 42: 401–412

Alonso-Villar O, Río C Del, Toharia L (2009) Un análisis espacial del desempleo por municipios. Revista de EconomíaAplicada 49: 47–80

Althin R, Behrenz L (2004) An efficiency analysis of Swedish Employment Offices. International Review of AppliedEconomics 18: 471–482

Arraiz I, Drukker DM, Kelejian HH, Prucha IR (2010) A spatial Cliff-Ord-type model with heteroscedastic innovations:Small and large sample results. Journal of Regional Science 50: 592–614

Anselin L (1988) Spatial econometrics methods and models. Kluwer Academic Publishers, Dordrecht.Anselin L (1995) Local indicators of spatial association: LISA. Geographical Analysis 27: 93–115Anselin L (2002) Under the hood: Issues in the specification and interpretation of spatial regression models. Agricultural

Economics 27: 247–267Anselin L, Bera A (1998) Spatial dependence in linear regression models with an introduction to spatial econometrics.

In: Ullah A, Giles ED (eds) Handbook of applied economic statistics. Marcel Dekker, New YorkAnselin L, Bera A, Florax R, Yoon M (1996) Simple diagnostic tests for spatial dependence. Regional Science and

Urban Economics 26: 77–104Assunção RM, Reis EA (1999) A new proposal to adjust Moran’s I for population density. Statistics in medicine 18:

2147–2162Bande R, Fernández M, Montuenga V (2008) Regional unemployment in Spain: Disparities, business cycle and wage

setting. Labour Economics 15: 885–914Blien U, Hirschenauer F, Hong Van P (2010) Classification of regional labour markets for purposes of labour market

policy. Papers in Regional Science 89: 859–880Bröcker J (1989) How to eliminate certain defects of the potential formula. Environment and Planning A 21: 817–830Bruinsma F, Rietveld P (1993) Urban agglomerations in European infrastructure networks. Urban Studies 30: 919–934Condeço-Melhorado A, Gutierrez J, Garcia-Palomares JC (2011) Spatial impacts of road pricing: Accessibility, regional

spillovers and territorial cohesion. Transportation Research Part A 45: 185–203Consejo Económico y Social (2009) Memoria sobre la situación socioeconómica y laboral en España. CES, MadridCracolici MF, Cuffaro M, Nijkamp P (2007) Geographical distribution of unemployment: An analysis of provincial

differences in Italy. Growth and Change 38: 649–670Cragg JG, Donald SG (1993) Testing identificability and specification in instrumental variable models. Econometric

Theory 9: 222–240Détang-Dessendre C, Gaigné C (2009) Unemployment duration, city size, and the tightness of the labour market.

Regional Science and Urban Economics 39: 266–276Elhorst JP (2003) The mystery of regional unemployment differentials: Theoretical and empirical explanations. Journal

of Economic Surveys 17: 709–748Fertig M, Schmidt CM, Schneider H (2006) Active labour market policy in Germany: Is there a successful policy

strategy. Regional Science and Urban Economics 36: 399–430Fischer MM, Getis A (2010) Handbook of applied spatial analysis. Springer, BerlinFlorax R, Folmer H, Rey S (2003) Specification searches in spatial econometrics: The relevance of Hendry’s method-

ology. Regional Science and Urban Economics 33: 557–579Frost ME, Spence NA (1995) The rediscovery of accessibility and economic potencial: The critical issue of self-

potential. Environment and Planning A 27: 1833–1848Fujita M, Krugman P, Venables AJ (1999) The spatial economy: Cities, regions and international trade. MIT Press,

Cambridge, MAGarcia-del-Barrio P, Gil-Alana PL (2009) New revelations about unemployment persistence in Spain: Time series and

panel data approaches using regional data. Applied Economics 41: 219–236Geurs KT, Van Wee B (2004) Accessibility evaluation of land-use and transport strategies: Review and research

directions. Journal of Transport Geography 12: 127–140



Gutiérrez J (2001) Location, economic potential and daily accessibility: An analysis of the accessibility impact of thehigh-speed line Madrid-Barcelona-French border. Journal of Transport Geography 9: 229–242

Gutiérrez-i-Puigarnau E, Van Ommeren JN (2010) Labour supply and commuting. Journal of Urban Economics 68:82–89

Hagen T (2003) Three approaches to the evaluation of active labour market policy in East Germany using regional data.ZEW Discussion Paper 03-27, ZEW-Mannheim

Holl A (2007) Twenty years of accessibility improvements: The case of the Spanish motorway building programme.Journal of Transport Geography 15: 286–297

Ihlanfeldt KR (1997) Information on the spatial distribution of job opportunities within metropolitan areas. Journal ofUrban Economics 41: 218–242

Jimeno JF, Bentolila S (1998) Regional unemployment persistence (Spain 1976–1994). Labour Economics 5: 25–51Joassart-Marcelli P, Giordano A (2006) Does local access to employment services reduce unemployment? A GYS

analysis of One-Stop Career Centres. Policy Sciences 39: 335–359Kelejian HH, Prucha IR (1999) A generalized moments estimator for the autoregressive parameter in a spatial model.

International Economic Review 40: 509–533Kelejian HH, Prucha IR (2010) Specification and estimation of spatial autoregressive models with autoregressive and

heteroscedastic disturbances. Journal of Econometrics 157: 53–67Krugman P (1991) Increasing returns and economic geography. Journal of Political Economy 99: 483–499LeSage J, Pace RK (2009) Introduction to spatial econometrics. CRC Press, Boca Raton, FLLin X, Lee LF (2010) GMM estimation of spatial autoregressive models with unknown heteroscedasticity. Journal of

Econometrics 157: 34–52Longhi S, Nijkamp P (2007) Forecasting regional labour market developments under spatial autocorrelation.

International Regional Science Review 30: 100–119López-Bazo E, del Barrio T, Artis M (2002) The regional distribution of Spanish unemployment: A spatial analysis.

Papers in Regional Science 81: 365–389Martin A, Gråsjö U (2009) Spatial dependence and the representation of space in empirical models. The Annals of

Regional Science 43: 159–180McMillen DP (1992) Probit with spatial autocorrelation. Journal of Regional Science 32: 335–348Molho I (1995) Spatial autocorrelation in British unemployment. Journal of Regional Science 35: 641–658Mur J, Angulo A (2009) Model selection strategies in a spatial setting: Some additional results. Regional Science and

Urban Economics 39: 200–213Ord JK (1975) Estimation methods for models of spatial interaction. Journal of the American Statistical Association 70:

120–126Overman HG, Puga D (2002) Unemployment clusters across euro regions and countries. Economic Policy 17: 115–148Patacchini E, Zenou Y (2007) Spatial dependence in local unemployment rates. Journal of Economic Geography 7:

169–191Piras G (2010) sphet: Spatial models with heteroscedastic innovations in R. Journal of Statistical Software 35: 1–21Rich D (1975) Accessibility and economic activity: A study of locational disadvantage in Scotland. PhD Thesis,

University of CambridgeServicio Público de Empleo Estatal (2008) Memoria del Servicio Público de Empleo Estatal 2008. Ministerio de Trabajo

e Inmigración, MadridSheldon GM (2003) The efficiency of public employment services: A nonparametric matching function analysis for

Switzerland. Journal of Productivity Analysis 20: 49–70Smirnov O, Anselin L (2001) Fast maximum likelihood estimation of very large spatial autoregressive models: A

characteristic polynomial approach. Computational Statistics & Data Analysis 35: 301–319Suárez P (2011) El servicio público de empleo en España: Ensayos desde una perspectiva regional. Ph.D. Thesis,

Universidad de OviedoTalen E, Anselin L (1998) Assessing spatial equity: An evaluation of measures of accessibility to public playgrounds.

Environment and Planning A 30: 595–613Toharia L (2005) El desempleo en España. In: Navarro V (ed) La situación social en España. Fundación Largo

Caballero y Biblioteca Nueva, MadridTsou KW, Hung YT, Chang YL (2005) An accessibility-based integrated measure of relative spatial equity in urban

public facilities. Cities 22: 424–435Van Wee B, Hagoort M, Annema JA (2001) Accessibility measures with competition. Journal of Transport Geography

9: 199–208



the accessibility to employment offices in the spanish labour market

Documents