THE REGIONAL DISTRIBUTION OF SPANISH UNEMPLOYMENT. A SPATIAL ANALYSIS. * E. LOPEZ-BAZO T. DEL BARRIO M. ARTIS

Transcription

1 THE REGIONAL DISTRIBUTION OF SPANISH UNEMPLOYMENT. A SPATIAL ANALYSIS. * E. LOPEZ-BAZO T. DEL BARRIO M. ARTIS Address: Research Group Anàls Quanttatva Regonal Dept of Econometrcs, Statstcs and Spansh Economy Unversty of Barcelona Avda. Dagonal 690, Barcelona, Span Phone: Fax: Emal: elopez@eco.ub.es * The authors acknowledge fnancal support by the CICYT, SEC

2 Abstract: Ths paper proposes a set of tools for analysng the regonal dstrbuton of unemployment. As we were nterested n the characterstcs of the dstrbuton as a whole, results from a tradtonal regresson analyss were complemented wth those obtaned by estmatng ts external shape before and after beng condtoned to factors underlyng regonal unemployment. In addton, the paper specfcally consders the spatal characterstcs of the dstrbuton, and the emprcal model developed n order to determne explanatory factors ncludes spatal effects. Ths framework s appled to the study of the provncal dstrbuton of unemployment rates n Span. Results pont to ncreasng spatal dependence n the dstrbuton of regonal unemployment rates, and a change n the factors causng regonal dfferentals over the last decade. Keywords: geography of unemployment, spatal analyss, Spansh regons JEL codes: C21, E24, R12, R23

3 Span s cookng can be consdered by way of the dfferent autonomous regons, although there are smlartes between neghbourng areas Ibera Arlne Magazne (December 1999) 1. INTRODUCTION Hgh unemployment rates have typfed some European countres n recent decades. Most studes pont to structural condtons and rgdtes of the labour market, together wth the system of unemployment beneft n those European economes, as the major causes of such hgh fgures (Bean, 1994). In addton to the queston of natonwde aggregate unemployment, another nterestng, but less studed, aspect s the geographcal dstrbuton of unemployment. There s, however, evdence as to the relevance of spatal dfferentals wth respect to unemployment rates n Europe, Canada and the US. Asde from the fact that labour markets reman essentally regonal, there are reasons for consderng unemployment from a regonal pont of vew. Elhorst (2000) proposes three: the magntude of regonal dfferences between regons wthn countres; the absence of explanatons for the exstence of regonal unemployment dspartes n macroeconomcs; and the neffcency created by such dspartes n the economy as a whole. In ths regard, most prevous contrbutons have amed to analyse the determnants of regonal unemployment by usng a regresson analyss, n whch unemployment n regons of a gven economy s explaned by a set of explanatory varables that nclude characterstcs of the regonal labour market, of the populaton, the ndustral mx, natonwde unemployment, etc (Marston, 1985; Elhorst, 1995; Partrdge and Rckman, 1997; Taylor and Bradley, 1997) 1. Such analyses provde an estmate of the effect that each factor has on the unemployment rate of an average or representatve regon n the sample beng analysed. Quah (1993, 1996) ntally rased ths pont wth respect to growth regressons. He suggested studyng the effect on the whole dstrbuton of the economc varable under analyss by complementng the 1 Taylor and Bradley (1997) and Elhorst (2000) dscuss the determnants of regonal unemployment and 1

4 tradtonal analyss wth alternatve technques. Ths approach has recently been appled to the analyss of the dynamcs of regonal unemployment rates (Overman and Puga, 2002; López- Bazo et al, 2000). In ths paper, we develop t further by combnng the results of a regresson analyss wth the estmaton of the shape of the regonal dstrbuton of unemployment, condtonal to some of the above-mentoned factors. Comparng the real observed dstrbuton wth that n whch the mpact of the explanatory varables has been removed allows ther effect on the characterstcs of the dstrbuton as a whole to be determned. Our results provde nterestng nsghts nto, for example, the formaton of groups of regons wth separate unemployment rates. We beleve, also, that analyses of regonal unemployment should specfcally consder the spatal characterstcs of the dstrbuton, and emprcal models developed n order to determne explanatory factors should nclude the possblty of spatal effects. Spatal nteractons across regonal labour markets may be the result of workers n a regon beng wllng to fll vacances n other regons and frms lookng for workers outsde the regons n whch they are located. Burda and Proft (1996), for local labour markets n the Czech Republc, and Burgess and Proft (2001), for the travel-to-work areas n Brtan, have provded evdence for the exstence of such spatal nteractons. More generally, the outcome of the labour market n a regon could be nfluenced by the crcumstances of other regons n the system. In ths regard, Bronars and Jansen (1987) and Molho (1995) report the sgnfcance of spatal spllovers n the process by whch unemployment dfferentals adjust to local shocks n the UK and the US, respectvely. Accordngly, our study ncludes an explct spatal econometrc analyss of the regonal dstrbuton of unemployment and, therefore, s consstent wth the work of Rey and Montoury (1999), who reconsdered the queston of regonal economc growth from a spatal econometrc perspectve. Ther paper provded new nsghts nto the geographcal dynamcs summarse results from prevous research. 2

5 of US regonal ncome growth patterns by applyng methods of exploratory spatal data analyss and ncludng spatal effects n the econometrc models used to study regonal ncome convergence. In our paper, we apply the analyss of the regonal dstrbuton of unemployment rates, ncludng spatal effects, to Spansh unemployment. Several studes have tred to explan why unemployment n Span has behaved the way t has, and also, why t has followed a dfferent pattern to that experenced n other countres (Bentolla and Blanchard, 1990; Blanchard and Jmeno, 1995; Dolado and Jmeno, 1997; Marmon and Zlbott, 1998). However, the regonal dstrbuton of unemployment rates n Span has attracted less attenton. Yet, as wll be shown below, the Spansh case s somewhat extreme n ths regard as well. The dstrbuton of unemployment rates s charactersed by szeable dfferences between regons and a remarkable stablty n ther rankng. Thus, the Spansh provnces (NUTS III regons n Span 2 ) wth the hghest unemployment n the late nnetes have rates that are almost double the Spansh average for those years. They were also among the regons wth the hghest rates n prevous decades. In contrast, some other provnces had rates that were actually below the EU average. Indeed, n recent decades they have consstently been among the most favoured provnces n Span, wth rates never above half the Spansh average. Our analyss s focused on the dstrbuton of unemployment n the50 Spansh provnces for two partcular years, 1985 and It s nterestng to study changes n the dstrbuton over a perod n whch the Spansh economy underwent mportant economc reforms as a result of the processes of market lberalsaton, openness and ntegraton nto the European Unon. In addton, labour market reforms n that perod were amed at ncreasng flexblty and deregulaton 3. It s lkely that Spansh regons dd not all react n the same way to these 2 Most labour commutng takes place wthn these terrtoral unts, so they can be taken to approxmately defne ntegrated labour markets. The sze of the average provnce, as measured by the labour force, was 571,654 workers n 1985 and 644,879 n There have been varous reforms n labour market legslaton n Span over the last two decades (1984, 1992, 3

6 reforms. In addton, the determnants of unemployment dfferences across provnces may have changed durng that perod, and ths s n fact confrmed by our results. The frst year n our analyss comes at the end of a decade of crss and ndustral restructurng, and was the year before Span joned the European Communty. At that tme, unemployment fgures reached ther hghest levels ever. Twelve years later, the Spansh economy had undergone a perod of notable growth and a fall n unemployment rates (late eghtes and early nnetes), followed by some years of deceleraton and a rse n unemployment to prevous levels. Thus, the two ponts n tme that we are consderng encompass a complete cycle and, therefore, the analyss s not contamnated by separate regonal responses to the dfferent phases of the busness cycle. In addton, smlar aggregate unemployment rates for Span n both years mean that the analyses of relatve or absolute devatons do not dffer greatly (see Martn, 1997 for a dscusson of regonal unemployment dspartes n terms of relatvtes or dfferentals). The rest of the paper s organsed as follows: a prelmnary spatal exploratory analyss of the dstrbuton of unemployment rates n the Spansh provnces s presented n secton 2, where the technques used throughout the paper are concsely descrbed; secton 3 brefly summarses the explanatory varables of regonal unemployment ncluded n our study and presents the emprcal model used n secton 4. Ths secton descrbes the results of the explanatory analyss. It ncludes the regresson results and analyss of the mpact of the varables nfluencng unemployment n the provncal dstrbuton. The paper s fnal secton offers some concludng comments. 1994, 1997). These ntroduced new types of contracts (part tme, tranng, fxed duraton), decreased the cost of frng workers, and redefned the system of unemployment beneft. However, doubts have been rased about the effectveness of such measures, whle none of the reforms addressed the problem of heavly centralsed labour market barganng (see Segura, 2001 for further detals). 4

7 2. EXPLORATORY ANALYSIS OF THE SPATIAL DISTRIBUTION OF UNEMPLOYMENT RATES Changes over tme n the aggregate Spansh unemployment fgures have been wdely reported and the causes behnd ther hgh level n recent decades have been analysed n prevous contrbutons (Blanchard and Jmeno,1995; Marmon and Zlbott,1998). Durng the sxtes, the average rate remaned stable at around 2-3%. It clmbed moderately throughout the next decade, reachng a fgure of around 10% by the begnnng of the eghtes. The unemployment rate then doubled n a fve-year perod so that more than 20 out of every 100 workers were unemployed (Table 1). Later, unemployment rates moved n parallel wth the busness cycle, yet always wthn a range far above those n other Western economes -t was around 20% n In ths same perod, the standard devaton, as a raw measure of unemployment dfferentals n the Spansh provnces, ncreased markedly up to the md-eghtes. Afterwards, t remaned generally stable 4. A comparson of unemployment rates n those provnces whch each year report extreme values provdes a clear pcture of the magntude of the spatal dfferences. The last row of Table 1 shows the dfference n unemployment rates between the provnces wth the hghest and lowest rates n 1985 and Usng unemployment rates as a rough measure of the probablty of beng unemployed, these fgures reveal that workers n certan provnces were much more lkely to be unemployed than those n some other provnces. Furthermore, ths probablty may be ncreasng. Indeed, more recent fgures seem to ndcate that certan provnces n north-east Span are close to full employment, whle at a dstance of a few hundred klometres rates reman above 20%. The provncal dstrbuton of unemployment, 4 These fgures, as well as the ones used throughout the paper on labour market varables, come from the Labour Force Survey (EPA) carred out by the Spansh Statstcal Offce (INE) followng the homogeneous EU-wde methodology of EUROSTAT. The survey defnes an unemployed person as someone aged 16 or over who has not been employed that week, but who s avalable for work and s actvely seekng a job. Another major source of unemployment data n Span s the unemployment records of the Natonal Employment Offce (INEM). We have dscarded ths latter source as only part of the unemployed are regstered n the INEM. 5

8 however, seems to be charactersed by strong, though not perfect, persstence, as the correlaton coeffcent for unemployment dfferentals n both perods s Wth the am of provdng further nsghts nto the regonal pattern of unemployment rates n Span, we estmated the densty functon assocated wth the dstrbuton of unemployment n 1985 and Ths functon proxes the shape of the dstrbuton, and actually gves more nformaton than the sngle measures of poston and dsperson do. The densty functon s estmated non-parametrcally by the kernel method. The kernel densty estmator replaces the boxes n a hstogram by smooth bumps (Slverman 1986). Smoothng s done by puttng less weght on observatons that are further from the pont beng evaluated. More techncally, the kernel densty estmate of a seres X at a pont x s estmated by f n 1 x X (1) K Nh = 1 h ( x) = where N s the number of observatons, h s the bandwdth (or smoothng parameter) and K( ) s a kernel functon that ntegrates to one. The kernel functon s a weghtng functon that determnes the shape of the bumps. We have used the Gaussan kernel n our estmates: exp u 2π 2 (2) where u s the argument of the kernel functon. The bandwdth, h, controls the smoothness of the densty estmate; the larger the bandwdth, the smoother the estmate. Bandwdth selecton s of crucal mportance n densty estmaton, and varous methods have been suggested n the lterature. In ths paper we have used the data-based automatc bandwdth suggested by Slverman (1986, equaton 3.31): 1 5 h = 0.9N mn { s, R /134} (3) 5 The coeffcent of a smple regresson between unemployment dfferentals n 1997 and those n 1985 s 0.82, wth an R 2 of 63.7%. 6

9 where s s the standard devaton and R the nterquartle range of the seres. The external shape of two or more dstrbutons can be compared by means of the estmated densty functons. More specfcally, the change n shape of the dstrbuton over the perod under analyss can be assessed by comparng the densty functon for provncal unemployment rates n 1985 and However, ths method comes up aganst one of the man drawbacks of ths type of analyss, namely, how to test the equalty of the dstrbutons from the estmated denstes. We have addressed ths by applyng an overlappng coeffcent (OVL). Bradley (1985) and Inman and Bradley (1989) promote the use of OVL as an ntutve measure of substantve smlarty between two probablty dstrbutons. The closer the OVL s to 1, the more smlar the dstrbutons beng compared. Confdence ntervals can be computed by bootstrap technques n order to test that samples of unemployment rates n both years were ndeed drawn from the same theoretcal dstrbuton. Addtonally, the OVL can be splt nto the overlap assocated wth three ranges of unemployment rates: low, medum and hgh. Further detals of ths coeffcent are provded n the Appendx. Fgure 1 plots the estmated denstes for the dfference between the unemployment rate n each provnce and the average rate for the Spansh economy n the years under analyss. In addton to the hgh degree of dsperson - already llustrated by the data n Table 1 - the fgure would seem to show that the shape of the dstrbuton dd not undergo mportant changes. However, a closer look at both denstes reveals a tendency towards the concentraton of the mass of probablty n partcular unemployment rate ntervals. The most strkng feature s the consoldaton of a peak at very hgh postve dfferentals n 1997, whle another peak may be formng to the left of the dstrbuton. Ths s confrmed by the OVL whch, for the whole range of unemployment rate dfferentals, has a value of 0.873, below the crtcal value, and thus the hypothess that both dstrbutons are smlar s rejected. Coeffcents for the three ntervals ndcate that dfferences n the dstrbuton are due to the 7

10 range of low (0.804) and, especally, hgh (0.763) unemployment rate dfferentals, whle smlarty n the ntermedate nterval cannot be rejected. Summng up, changes n provncal unemployment rates over the perod under analyss may have caused the formaton of two clusters of provnces. The clearest s the one n the range of unemployment rates far above the Spansh average, whle the other, perhaps stll beng formed, s charactersed by low relatve unemployment. The above analyss does not, however, consder the partcular spatal locaton of the provnces. Thus, the mpact of geography on the dsperson of the dstrbuton and on the process of cluster formaton over the perod, detected by means of the estmated denstes, cannot be assessed. A smlar pont has recently been rased n studes dealng wth the regonal dstrbuton of producton, and specfc tools have been appled n such cases n order to detect the type and ntensty of spatal assocaton (Rey and Montour, 1999; López-Bazo et al, 1999). The type and ntensty of spatal assocaton n the regonal dstrbuton of unemployment rates can be easly depcted by an X-Y plot n whch the standardsed value for each regon s represented on one axs and the standardsed value n the neghbourng regons (spatal lag) on the other a Moran scatterplot, as suggested n Anseln (1996). In addton, the degree of spatal assocaton can be summarsed by means of what s known as Moran's I statstc (Moran, 1948). It s defned as: I N N w ( x j j = N = 1 ( x x)( x x) 2 j x) (4) where x and x j are the observatons for regon and j of the varable under analyss; x s the average of that varable n the sample of regons; and w j s the -j element of a rowstandardsed matrx of weghts, W. Ths s an NxN matrx of spatal weghts whose characterstc element, w j, summarses the nteracton between regons and j. Dfferent 8

11 defntons of nteractons cause dfferent W matrces. Here, we adopted the smplest, but probably also the most popular, defnton: the bnary contguty matrx, whereby the element -j of the weght matrx, ω j =1 --before beng row-standardsed-- f regons and j share a border, and ω j = 0 otherwse 6. Therefore, the spatal lag s smply the average of the unemployment rate n the neghbourng provnces. The top panel of Fgure 2 shows the Moran scatterplot for unemployment rates n 1985, whle the plot for the fnal year, 1997, s shown at the bottom. Results for the Moran's I statstc n each year are also shown n the fgure. The poston of the provnces n quadrants I and III n the Moran scatterplot correspondng to 1985 ndcates that provnces wth hgh unemployment rates have neghbours wth the same characterstc, whle low-unemployment provnces are more lkely to be surrounded by provnces wth low values. The postve spatal relatonshp seems to be even stronger n Accordngly, the value of Moran's I s sgnfcant n both cases and hgher for the fnal year under analyss 7. Therefore, we can conclude that the regonal dstrbuton of unemployment rates n Span s charactersed by ntense spatal dependence. Furthermore, t seems to have ncreased over the last two decades. In order to shed lght on the effect whch the observed spatal dependence could have on the characterstcs of the dstrbuton detected above, we compared the shape of the dstrbuton of provncal unemployment rates, relatve to the average rate n Span, wth that for the dfference between the rate n each provnce and the average rate n the neghbourng provnces, that s, the spatal lag of unemployment rates. If some of the dsperson n the dstrbuton s lnked to spatal dependence, then we would expect the latter dstrbuton to be more concentrated. Smlarly, f cluster formaton s, at least partly, a geographcal phenomenon, the dstrbuton of unemployment rates n each provnce mnus the rate n the 6 It should be stressed that the man results n ths paper were not affected by the use of a dstance weght matrx. On the contrary, the role of spatal dependence was even larger n that case. 7 Spatal dependence s observed n each one of the years between 1985 and 1997, wth contnuous ncrease. These results are not reported n order to save space; they wll be provded upon request. 9

12 neghbourng ones should not show the mass of probablty at the very hgh and low unemployment rates. Gven that we prevously detected some changes n the shape of the dstrbuton, and n the degree of spatal autocorrelaton between 1985 and 1997, we made the comparson for both years (Fgure 3). It can be seen how the dstrbuton n 1985 shfts to the rght when the neghbourng effect s removed, beng the mode located now around zero. It s also moderately more concentrated than the orgnal dstrbuton, although the mass of probablty remans at the large postve dfferentals. The OVL clearly ndcates that dstrbutons are dfferent, partcularly n the case of the low unemployment range. The same exercse for 1997 reveals that the neghbourng effect could be responsble for most of the characterstcs of that year s dstrbuton. Not only s the dstrbuton now less dspersed, but also the clusters detected n the orgnal dstrbuton completely dsappear when the densty functon s estmated for devatons wth respect to neghbourng provnces. Once agan, conclusons from vsual nspecton are confrmed by the OVL coeffcents. Summng up, a smple descrptve analyss shows how the regonal dstrbuton of unemployment s Span s largely dspersed, and that there s a trend toward the formaton of clusters of extreme values. Furthermore, the spatal dstrbuton s far from random or homogeneous. On the contrary, the unemployment rate n a provnce s ncreasngly related to the one n the surroundng provnces, and ths phenomenon could be responsble for the majorty of the dstrbuton s characterstcs. Ths s partcularly so n 1997, where t seems to account for the above-mentoned clusters of provnces. 3. EMPIRICAL MODEL OF REGIONAL UNEMPLOYMENT Causes of regonal unemployment have been dscussed n detal n the lterature (Marston, 1985; Partrdge and Rckman, 1997; Martn, 1997; Taylor and Bradley, 1997). Elhorst (2000) 10

13 has recently produced a comprehensve revew that ncludes a lst of the explanatory varables suggested as havng an nfluence on regonal unemployment. Among the factors on the lst are the natural change n the labour force, the partcpaton rate, net n-mgraton and commutng, wages, employment growth, the ndustral mx, the educatonal attanment of the populaton, market potental, and other characterstcs of the labour market such as the degree of unonsaton. Although we do not ntend to descrbe n detal the effects whch those varables have on regonal unemployment rate dfferentals, we should pont out that we used the above-mentoned papers n selectng the varables for our emprcal model. The process of selecton was also nfluenced by studes provdng partcular evdence about factors whch affect Spansh unemployment (e.g. Rodríguez-Pose, 1996 and 1998; Marmon and Zlbott, 1998), and the avalablty of relable data at the provncal level. In ths regard, we were not able to nclude factors such as long-term unemployment and unemployment benefts due to the lack of spatally dsaggregated data for those varables. However, the omsson of these factors should not alter the man results f they have a homogeneous mpact on all provnces, gven that we are focusng on unemployment dfferentals. Furthermore, some of the varables already ncluded may capture at least some of ther effects. The factors fnally selected n our analyss were as follows: 8 Employment growth (EMP): It s expected that addtonal jobs decrease the unemployment rate, and most of the studes whch have consdered ths varable support that negatve effect. However, the sgn of the nfluence can be reversed, as ponted out by Harrs and Todaro (1970), through nduced urban-rural mgraton. Net mgraton (M): The effect of net mgraton on regonal unemployment rates s not straghtforward, as t may ncrease labour supply and demand over a long tme perod. Accordngly, emprcal evdence has produced mxed results. In the case of Span t 8 The precse defnton of varables and sources can be found n the data appendx. 11

14 should be stressed that nternatonal as well as nterregonal mgraton flows were an mportant mechansm n balancng the labour market up to the eghtes, though they fall to lower levels n the last two decades. Unt labour costs (ULC): We assumed that frms are concerned wth wages n relaton to labour productvty, snce wage dfferences across regons accommodate to productvty dfferences. In so dong, we are bascally consderng the postve nfluence of labour costs on unemployment through the effect on labour demand. The mpact through labour supply would requre the use of data on real wages. As far as we know, seres on provncal prces are not avalable to correct nomnal labour costs for dfferences n purchasng power. Nomnal labour costs were always non-sgnfcant when they were ntroduced n the analyss. Industral mx: We controlled for the share of agrculture (%AGR) and manufacturng (%MANU) n employment. Regons specalsed n declnng ndustres are expected to exhbt hgher unemployment rates than those based around growng actvtes. Industral restructurng n the seventes and eghtes was partcularly severe n Span. As a result, employment n agrculture and manufacturng fell markedly. Consequently, a negatve relatonshp between employment share n those ndustres and unemployment rates would be expected. However, Elhorst (2000), consderng the possblty of ndustral msmatch and some drawbacks n the use of employment shares, ponts out that t s not clear what the sgn of these varables should be. Ths s confrmed by the dversty of results obtaned by emprcal studes whch have ncluded these varables (e.g. Elhorst, 1995; Partrdge and Rckman, 1995 and 1997;Taylor and Bradley, 1997). Human captal (H): For a number of reasons (hgher demand for sklls, lower probablty of lay off, nfluence on mgraton decsons, etc) the educatonal attanment of workers s expected to be negatvely related to unemployment rates. Unemployment rates for 12

15 workers wth hgher level studes have been reported to be lower than for workers who leave educaton wth few or no qualfcatons (Nckell and Bell, 1996). There has been a constant ncrease n the level of educaton of the Spansh populaton over recent decades, but regonal dfferences n these levels reman great (Rodríguez-Pose, 1996). If the average human captal of the labour force n the Spansh provnces dffers, ths mght explan some of the nequalty n the geographcal dstrbuton of unemployment. We have proxed ths factor by the percentage of the labour force that has at least started secondary schoolng. Demography and partcpaton: The structure of the populaton has an obvous effect on labour supply. Unemployment rates have been notably hgher for people aged In the Spansh economy, 36 out of 100 workers aged under 25 were unemployed n well above the 19% average for the EU as a whole. Furthermore, dfferences across Spansh regons are notable: above 40% n those wth hgher youth unemployment and below 25% where the problem s less ntense. Therefore, our model ncludes the share of workng age populaton aged 16 to 25 (YOU). As regards partcpaton, there s a controversy about the effect of partcpaton rates on unemployment, as several opposte mechansms mght be at work smultaneously (Elhorts, 2000). To allow for the possblty that these mechansms exerted a separate nfluence on male and female partcpaton decsons, we ncluded both partcpaton rates (MALE, FEMALE) as explanatory varables n our model. As a result the model to be estmated can be expressed as: U t = β 0 + β 1 EMP t + β 2 M t + β 3 ULC t + β 4 %AGR t + β 5 %MANU t + (5) β 6 H t + β 7 YOU t + β 8 MALE t + β 9 FEMALE t +e t where U t s the vector of dfferences between the unemployment rate n each provnce n year t (=1985,1997) and the average unemployment rate n Span. The explanatory varables, as 13

16 defned above, are all expressed as devatons from the Spansh average as well. Fnally, ε s a random perturbance. The unknown coeffcents were estmated by ordnary least squares (OLS) usng the observatons from the 50 Spansh provnces. Gven that the effect on unemployment rates of the explanatory varable may have changed over the perod under analyss, we dd not mpose equalty restrctons on the coeffcents across equatons for each one of the years. However, we dd check for spatal dependence n the resduals of the regressons for each one of the years. Three tests of spatal dependence were computed: the resduals Moran's I, and the robust Lagrange multpler tests for spatal lag and spatal error autocorrelaton. Whle the Moran test s not able to dstngush the two types of spatal autocorrelaton, the robust tests have been shown to have good power aganst a specfc alternatve (Anseln et al, 1996), and thus can be used to formulate the approprate spatal model (Florax and Folmer, 1992). More specfcally, the spatal error model consders the followng structure for the perturbance of (5): e = δwe + x (6) where e s the perturbance vector, W the matrx of spatal weghts defned n the prevous secton, δ the spatal error coeffcent, and x ~N(0,σ 2 I). The spatal lag model ncludes the spatal lag of the unemployment rates (WU) n the lst of regressors: U t = β 0 + β 1 EMP t + β 2 M t + β 3 ULC t + β 4 %AGR t + β 5 %MANU t + (7) β 6 H t + β 7 YOU t + β 8 MALE t + β 9 FEMALE t + γ WU t + e t where γ s the spatal autoregressve parameter. 4. EXPLANATORY ANALYSIS OF THE SPATIAL DISTRIBUTION OF UNEMPLOYMENT RATES 14

17 In ths secton we study the nfluence whch the factors outlned n the prevous secton have on the man characterstcs of the dstrbuton of regonal unemployment dfferentals n the two years under analyss, and partcularly on the dsperson and clusterng descrbed n the exploratory analyss. Several varables proposed n the lterature as affectng the level of regonal unemployment are consdered. These are factors wthn each provnce that may nfluence the performance of the labour market n general, and the rate of unemployment n partcular. Gven that we have already shown spatal dependence to be an mportant characterstc of the provncal dstrbuton of unemployment, we also consder the lkely exstence of nteractons across provnces whch may help n understandng unemployment rates. As our nterest was not only focused on a representatve or average provnce, we estmated the effect of those factors on the whole dstrbuton of unemployment rates. Therefore, we began wth a tradtonal regresson analyss n whch estmates of the parameters should provde evdence about the effect whch the dfferent varables have on the unemployment dfferentals of an average Spansh provnce for each of the two years beng analysed 9. Then, usng the tools descrbed n secton 2, we complemented that analyss by comparng the orgnal dstrbuton wth that condtonal to the factors under analyss. In so dong, we were able to assess ther mpact on the whole range of unemployment rate dfferentals, ncludng, for example, ther contrbuton to the formaton of clusters n the dstrbuton Regresson results. 9 Poolng observatons for both years would allow unobservable regonal effects n unemployment dfferentals to be accounted for. However, ths would be at the cost of mposng equalty constrants on the effects of the varables under analyss across tme. Ths hypothess was clearly rejected by standard tests. 15

18 We appled OLS to the lnear specfcaton gven by (5), although the dependent varable was restrcted to the nterval {-u NAT, 1-u NAT }, where u NAT s the natonwde unemployment rate 10. Ths s a common problem n emprcal analyses of unemployment rates, and only a few studes have appled the logstc transformaton n order to address ths (see the summary of the collecton of studes n Table 1 n Elhorst, 2000). When the focus of the analyss s the regonal unemployment rate, the dependent varable ranges wthn the nterval {0,1}, and can be taken to be the probablty of an average worker n a regon beng unemployed. Thus, the model proposed for analysng regonal unemployment rates s based on proportons data, and so the logstc transformaton s approprate. Unfortunately, n our case, such a transformaton could not be appled as regonal unemployment rate dfferentals may be negatve. Therefore, we contnued to estmate the coeffcents based on the lnear model, but reported the standard errors from the Whte (1980) heteroskedastcty consstent estmator of the covarance matrx for the parameter estmates. In so dong, we sought to account for the heteroskedastc perturbance of a model of proportons data (see Greene, 1993 for further detals). The OLS estmates of (5) for 1985 and 1997 are summarsed n Table 2. Before dscussng the sgn and magntude of the estmated coeffcents, t should be stressed that the overall ft acheved by the factors ncluded n the specfcaton for both years s qute hgh. Furthermore, the degree of collnearty among the regressors, as summarsed by the condton number, s surprsngly moderate, takng nto account the cross nfluence of the dfferent factors. Ths enables us to be more confdent n the estmates of sngle coeffcents. However, the spatal dependence tests pont to the presence of spatal autocorrelaton n the resduals of the equatons for both years. In accordance wth the results n the exploratory secton, ths phenomenon seems to be more ntense n As spatal autocorrelaton would nvaldate conclusons based on the msspecfed model, we have not commented on the value of the parameters from the OLS estmates. Instead, we have estmated the model whch best 10 We thank an anonymous referee for pontng ths out to us. 16

19 accounts for spatal dependence. In ths regard, the values of the robust tests clearly pont to the spatal lag model as the preferred specfcaton. However, the OLS s nconsstent n ths case due to smultanety nduced by the spatal lag (Anseln, 1988). Instrumental varables and maxmum lkelhood estmators have been suggested to provde consstent estmates. Table 3 presents the maxmum lkelhood estmates of the spatal lag model (7), where reported standard errors come from the heteroskedastcty consstent estmator of the covarance matrx of the maxmum lkelhood parameters, as suggested n Whte (1982) and Davdson and MacKnnon (1993). The major concluson to be drawn from the parameters s a change n the man causes of provncal unemployment rate dfferences. Whle excess of labour costs over productvty, ndustral mx, and human captal dfferences across provnces seem to explan most of the provncal unemployment rates n the md-eghtes, they lose ther explanatory power at the end of the nnetes. Unt labour costs affect postvely, and human captal negatvely, the rate of unemployment n 1985, as expected on a pror grounds. Dfferences n the share of manufacturng employment, and partcularly of agrculture, have sgnfcant coeffcents n They show a negatve effect on unemployment dfferentals whch, despte beng somewhat counterntutve, s n lne wth results obtaned for some other economes (Jones and Mannng, 1992; Taylor and Bradley, 1997; and the dscusson n Elhorst, 2000). However, nether the change n employment nor the demography and partcpaton varables have a sgnfcant mpact on unemployment dfferentals for that year. In sharp contrast, the varables wth sgnfcant coeffcents at the usual levels n 1997 are employment growth, net mgraton, youth populaton and female partcpaton. Provnces that create employment at hgher rates tend to experence less relatve unemployment. The same apples to net n-mgraton, as supply-sde effects seem to surpass the demand-sde effects and, therefore, provnces wth a net ncrease n people had, condtonal to the other factors, lower unemployment rates n The postve effect whch the percentage of youth 17

20 populaton has on unemployment dfferentals s partcularly strong. One extra pont of dfference between a provnce and the natonal average translated nto more than one addtonal pont n the dfference n unemployment rates. Fnally, although the effect of male partcpaton rates s neglgble, female partcpaton reduces unemployment rates. Ths could be due to the fact that female decsons to partcpate n the Spansh labour market are closely related to the current level of unemployment. As a result, female partcpaton would be lower n provnces wth hgh unemployment and hgher where unemployment s low. In any case, there s an mportant dsperson n ths estmated effect as the coeffcent s only sgnfcant at 10%. Thus, none of the factors that appear to be sgnfcant n explanng unemployment rate dfferentals for an average Spansh provnce n the md-eghtes seems to be mportant n the late nnetes. There s another noteworthy result from these estmates, namely, the ncrease n the spatal coeffcent observed over the perod. Its value s estmated to be and sgnfcant at 5% by a t-rato test n However, t s only sgnfcant at 10% when a more approprate lkelhood rato test (LR-LAG) s used. Therefore, we can conclude that most of the spatal dependence detected n the provncal dstrbuton of unemployment rates can be explaned by factors wthn each of the provnces ncluded n our emprcal model. On the contrary, the spatal coeffcent n 1997 s clearly sgnfcant, and s double the one for the ntal year. Furthermore, there s no evdence of remanng spatal autocorrelaton n the resduals (LM-ERR). As a matter of comparson, Table 3 ncludes the estmates of a pure autoregressve spatal model - excludng from our specfcaton the factors wthn each provnce. In ths case, the spatal parameter for both years s qute smlar and as hgh as and 0.751, respectvely, n accordance wth the exploratory results above. The estmated value for the spatal coeffcent s, therefore, much lower when factors wthn each provnce are ncluded, 18

21 ndcatng that spatal dependence n the explanatory varables was mostly responsble for spatal dependence n the dstrbuton of unemployment rates n 1985, although only partally responsble for ths phenomenon n Condtoned dstrbutons. Once an estmate of the parameters n (7) was avalable we could obtan the dstrbuton of relatve unemployment condtonal to the factors defned above for each of the years. In order to separate the effect of wthn-provnce factors from the spatal effect, we computed a condtonal dstrbuton for each. To do ths, we frst had to compute the unemployment dfferentals condtonal to the set of factors. Ths was obtaned by combnng the estmates for the parameters and the correspondng varables plus the vector of resduals, where the values for the varables we wanted to condton for were set to zero. That s to say, we estmated the unemployment dfferentals n case there was no dfference across provnces wth respect to the factors wthn each regon that affect unemployment, leavng unaltered the orgnal values for the spatal lag of unemployment rates. Correspondngly, the dstrbuton condtoned to havng smlar neghbours was obtaned by substtutng the values of the spatal lag for a vector of zeros, whle usng current values for the other varables n the model. The densty functon for the uncondtonal and condtonal dstrbutons could then be computed as descrbed above. Vsual nspecton of both denstes for each year and the calculaton of the OVL coeffcents enabled ther smlarty to be checked and conclusons could thus be drawn about the mpact of the varables on the whole dstrbuton. Fgure 4 depcts the denstes for the current dstrbuton of unemployment dfferentals and the dstrbuton condtoned to no dfferences n wthn-provnce factors, whereas those for the dstrbuton condtoned to the neghbourng effect are shown n Fgure 5. In the frst year (top panel), t can be observed how the geographcal dstrbuton of unemployment 19

22 would have been much more concentrated had the provnces not dffered n the rate of employment growth, mgraton flows, unt labour costs, ndustral mx, educatonal attanment, youth populaton and partcpaton rates. In fact, the condtonal dstrbuton almost collapses around the range of no dfferences. As a result, the OVL coeffcent rejects smlarty between the uncondtonal and condtonal dstrbutons, for the whole range and for the three ntervals defned above. Moreover, factors wthn each regon almost completely explan the mass of probablty at the postve dfferentals detected n the real dstrbuton. In contrast, there are no sgnfcant dfferences between the real dstrbuton and the one that results from removng the spatal lag effects. The only noteworthy effect s observed n the nterval of postve dfferentals. When the neghbourng effect s removed the mass of probablty n that nterval shfts to the left. In fact, the OVL HIGH leads to rejecton of smlarty for that partcular nterval, and s strong enough to cause the global OVL to reject smlarty for the whole range, even when smlarty seems to be acceptable for low and medum unemployment dfferentals. The pcture for the end of the nnetes (bottom panel of Fgures 4 and 5) shows, once agan, how factors wthn each provnce account for an mportant amount of the dstrbuton s characterstcs. Once condtoned, most of the probablty s concentrated close to the pont of no regonal dfferences. The cluster of low relatve unemployment dsappears and the one of postve dfferentals shfts to the left. However, t s also clear that these factors cannot fully explan the cluster. Interestngly, t s mostly explaned by the spatal nteracton effect, as shown by the dstrbuton once condtoned to no dfferences across provnces n the spatal lag of unemployment. Summng up, the wthn-provnce factors consdered n our study account for most of the dstrbuton s characterstcs n 1985, the neghbourng effect havng only a moderate nfluence. Ths latter effect only helps to explan some aspects of the cluster of postve 20

23 dfferentals under formaton that year. Of greater mportance, however, s ts explanatory value wth respect to such phenomena n 1997, as here the cluster s almost unexplaned by the explanatory varables. Nevertheless, they stll seem to be responsble for most of the dsperson n the dstrbuton. 5. CONCLUSION Ths paper has analysed the dstrbuton of unemployment n the Spansh provnces from a new perspectve, and has pad specal attenton to the spatal dmenson of the phenomenon. A set of statstcal tools for studyng both changes over tme n the dstrbuton of unemployment rates and the nfluence whch the determnants of regonal unemployment have on the whole dstrbuton has been proposed. Furthermore, spatal effects n that dstrbuton have been specfcally analysed by applyng spatal exploratory technques and spatal econometrc models. Our results for the Spansh provnces show how ths type of study complements the tradtonal regresson analyss and provdes new nsghts nto the geographcal dstrbuton of unemployment. Applyng the above to the Spansh provnces for the last two decades reveals that the ongong processes of economc ntegraton and labour market deregulaton have caused a knd of regonal cluster formaton, as the dstrbuton of unemployment rates n the late nnetes shows a mass of probablty at the nterval of large relatve unemployment, whle another group of regons, where rates are far behnd the natonwde average, may also be under formaton. Ths would confrm dfferences n the regonal reacton to the new economc framework. Interestngly, our results reveal a shft n those factors whch may explan unemployment dfferentals from the md-eghtes to the late nnetes. Whle, n 1985, the sgnfcant varables were dfferences n unt labour costs, the ndustral mx and, to a lesser degree, the educatonal attanment of the labour force, these do not explan the man 21

24 characterstcs of the dstrbuton n In contrast, the dsperson n the dstrbuton for that year seems to be related to a regon s ablty to create employment, to the net attracton of populaton, and characterstcs of the regonal populaton, such as the percentage of youth populaton or female partcpaton n the labour market. It should be stressed, however, that contrary to what happened n 1985, spatal effects play a role at the end of the nnetes. Asde from the fact that the spatal lag of unemployment rates could well be proxyng for other factors wthn each regon that were not ncluded n our analyss, spatal effects prove to be hghly sgnfcant n the regresson analyss and almost completely account for the cluster of provnces wth unemployment rates above the average. Therefore, t would be worthwhle ncludng explctly spatal varables n future emprcal analyses of regonal unemployment n order to elucdate whch knds of mechansm are responsble for the sgnfcant spatal effects detected n ths paper. Fnally, we would lke to stress that other economes n Europe may share at least some of the characterstcs observed n the case of Span. Large dsperson n regonal unemployment rates not only characterses the European Unon as a whole, but s also common to some member states. Polces amed at allevatng ths problem can only be developed f the reasons for such spatal dspartes n these economes are clearly understood. We therefore am to carry out further studes n the future. 22

25 APPENDIX A.1. Data descrpton and sources Varable Defnton Source Unemployment rate u Labour Force Survey (EPA) U = *100 A from the Spansh Statstcal Offce (INE) u : Total unemployed force n regon Employment growth Mgraton Labour unt costs Labour n agrculture Labour n manufacturng Human captal Female partcpaton Male partcpaton Youth populaton A : Total labour force n regon Lt L EMP = L t 5 t 5 L t : Employment n regon n perod t NM OM M = POP NM : n-mgraton OM : Out-mgraton POP : Populaton LC UCL = GDPPW LC : Labour costs per worker GDPPW : Gross domestc product per worker agr % AGR = *100 L agr : Employment n agrculture n regon L : Total employment n regon manu % MANU = L *100 manu : Employment n manufacturng n regon L : Total employment n regon h H = A *100 h : Labour force that has at least started secondary schoolng n regon A : Total labour force n regon FEML FEMALE = *100 FEM FEML : Female labour force n regon FEM16 65 : Females of workng age n regon MAL MALE = *100 MAL16 65 MAL : Male labour force MAL16-65 : Males of workng age you YOU = *100 N16 65 you : Populaton aged 16 to 25 years old n regon N16 65 : Populaton of workng age n regon EPA INE Fundacón BBVA EPA EPA From Pérez and Serrano (1998), takng as prmary source EPA EPA EPA EPA 23

26 A.2. Overlappng Coeffcent Bradley (1985) and Inman and Bradley (1989) proposed an overlappng coeffcent (OVL) as an ntutve measure of the smlarty between two probablty dstrbutons. In our case, we used the OVL to compare frequences throughout the range of a varable for two samples. The dea behnd the OVL can be summarsed n the followng fgure, where the range of values of two varables, x1 and x2, s on the x-axs, and the densty on the y-axs. The OVL s the area where the denstes of the two dstrbutons overlap when they are plotted on the same axes. The expresson for ths coeffcent n the dscrete case s the followng: where f ( x 1 ) and ( 2 ) [ 01, ] [ f ( x ), f ( x )] OVL = mn 1 2 OVL x f x are the emprcal densty functons. In the case of contnuous dstrbutons, summaton s replaced by ntegraton. A value of 1 for the OVL means that the two densty functons are exactly the same, whereas a null value ndcates the absence of overlappng n the densty functon at any pont n the range of the varable. The closer the OVL to 1 the more smlar the two dstrbutons beng compared. If we wsh to assess the contrbuton of the dfferent ndvduals n the sample to dfferences n the dstrbutons, t s possble to compute the OVL for dfferent ntervals of the total range of the varable, usng the followng expresson: OVL α x α = x α mn max [ f ( x ), f ( x )] 1 [ f ( x ),f ( x )] where α denotes a specfc nterval. We have computed the OVLα for three dfferent ntervals of the unemployment rate dfferentals (α=low, medum and hgh). OVL low consders values from the mnmum to the average mnus one standard devaton of the unemployment rate, OVL hgh goes from the average plus one standard devaton to the maxmum of the unemployment rate. OVL med measures the dscrepancy of the dstrbuton n between [ 01, ] The statstcal propertes of the OVL coeffcent depend on those of the data under analyss. Thus, the way to approach the ssue s va smulaton. Furthermore, the OVL s a based statstc, because any samplng varaton n the denstes of two samples obtaned from the same populaton causes the OVL to be strctly less than one. We used the bootstrap method to obtan the mean and varance of the OVL. We dd ths by resamplng both the orgnal data and a smulated sample of the same sze from a (A.1) (A.2) 24

27 ( x, ) N, (=85, 97). The number of replcatons s m= Tabulated values n Tables A.1 s and A.2 were used to construct a knd of confdence nterval n order to test the hypothess of equalty of two dstrbutons. The rule of thumb was to reject the hypothess of smlar dstrbutons f the value estmated for the OVL was lower than the expected value for the OVL n each case mnus twce the standard devaton. The null hypothess was rejected when the overlap was lower than that whch would be expected by allowng for sample devatons gven the sze of our sample. On the contrary, when the OVL was closer to 1 than the crtcal value we could be more confdent about assumng smlarty. 25

28 REFERENCES Anseln L (1988) Spatal Econometrcs: Methods and Models. Kluwer Academc Publshers, Dordrecht. Anseln L (1996) The Moran scatterplot as an ESDA tool to assess local nstablty n spatal assocaton. In: Fsher M, Scholten H, Unwn D (eds) Spatal Analytcal Perspectves on GIS. Taylor and Francs, London. Anseln L, Bera A, Florax, RJGM, Yoon M (1996) Smple dagnostc tests for spatal dependence. Regonal Scence and Urban Economcs 26: Bean C (1994) European unemployment: a survey. Journal of Economc Lterature 32: Bentolla S, Blanchard O (1990) Spansh Unemployment. Economc Polcy 10: Blanchard O, Jmeno JF (1995) Structural unemployment: Span versus Portugal. Amercan Economc Revew 85: Bradley EL Jr (1985) Overlappng coeffcent. In Kotz S, Johnson NL (eds.) Encyclopeda of Statstcal Scences, (6), Bronars SG, Jansen DW (1987) The geographcal dstrbuton of unemployment rates n the US. A spatal-tme seres analyss. Journal of Econometrcs 36: Burda MC, Proft S (1996) Matchng across space: evdence on moblty n the Czech Republc. Labour Economcs 3: Burgess S, Proft S (2001) Externaltes n the matchng of workers and frms n Brtan. Labour Economcs 8: Davdson R, MacKnnon JG (1993) Estmaton and Inference n Econometrcs. Oxford Unversty Press, Oxford. Dolado JJ, Jmeno JF (1997) The Causes of Spansh Unemployment: A Structural VAR Approach. European Economc Revew 41: