Income Segregaton and Suburbanzaton n France : a dscrete choce approach Florence Goffette-Nagot, Yves Schaeffer To cte ths verson: Florence Goffette-Nagot, Yves Schaeffer. Income Segregaton and Suburbanzaton n France : a dscrete choce approach. Workng paper GATE 2011-12. 2011. <halshs-00581139> HAL Id: halshs-00581139 https://halshs.archves-ouvertes.fr/halshs-00581139 Submtted on 30 Mar 2011 HAL s a mult-dscplnary open access archve for the depost and dssemnaton of scentfc research documents, whether they are publshed or not. The documents may come from teachng and research nsttutons n France or abroad, or from publc or prvate research centers. L archve ouverte plurdscplnare HAL, est destnée au dépôt et à la dffuson de documents scentfques de nveau recherche, publés ou non, émanant des établssements d ensegnement et de recherche franças ou étrangers, des laboratores publcs ou prvés.
GROUPED ANALYSEETDETHÉORIEÉCONOMIQUELYON STÉTIENNE WP1112 IncomeSegregatonandSuburbanzatonnFrance: adscretechoceapproach FlorenceGoffette Nagot,YvesSchaeffer Mars2011 Documentsdetraval WorkngPapers
GATEGrouped AnalyseetdeThéoreÉconomqueLyon StÉtenne 93,chemndesMoulles69130Ecully France Tel.+33(0)472866060 Fax+33(0)472866090 6,rueBassedesRves42023Sant Etennecedex02 France Tel.+33(0)477421960 Fax.+33(0)477421950 Messagereélectronque/Emal:gate@gate.cnrs.fr Téléchargement/Download:http://www.gate.cnrs.fr Publcatons/WorkngPapers
Income Segregaton and Suburbanzaton n France: a dscrete choce approach. Florence Goffette-Nagot* +, Yves Schaeffer** Abstract: Ths paper focuses on resdental sortng by socal and ethnc status n large French urban areas. Our obectve s to assess the relatve mportance of two maor determnants of segregaton stressed by the economc lterature (Bartolome and Ross, 2003; Brueckner et al., 1999): () Alonso sortng over space, due to the trade-off between land consumpton and accessblty to the central cty and () Tebout sortng over ursdctons, due to the taste for local publc goods and by extenson for all knds of local publc amentes (e.g. neghborhood externaltes). Our methodology draws on Schmdheny (2006). Frst, a condtonal logt model s estmated for each urban area, n whch movng households are assumed to sort based on ursdcton dstance to the central cty and ursdcton mean of households ncomes (as a proxy for the level of publc amentes). Second, our estmaton results are used to smulate the counterfactual resdental patterns that would preval f, alternatvely, one or the other of these mechansms were nactve (settng the coeffcents of the correspondng varables to zero). The contrbuton of each mechansm to the observed socal and ethnc segregaton s fnally apprecated by comparng the values of dssmlarty ndexes computed on the bass of the counterfactual households dstrbutons and on the observed households dstrbuton. Tebout-sortng emerges as the prmary cause of socal segregaton among wage-earnng households. On the contrary, Alonso-sortng appears to be the man drver of segregaton between economcally actve and nactve households, as well as between Frenchctzen and Foregn-ctzen households. JEL classfcaton: R21, R23, R1. Keywords: Income segregaton, Ethnc segregaton, Suburbanzaton, Local amentes, Mgratons, Condtonal logt, French urban areas. * Unversty of Lyon, CNRS, GATE ** CEMAGREF de Grenoble, Sant-Martn-d'Hères cedex, F-38402. + Correspondng author, GATE-CNRS, UMR 5824, 93 chemn des Moulles, F-69130 Ecully, France; tel: 33+4 72 86 60 78; fax: 33+4 72 86 60 90; mal: goffette-nagot@gate.cnrs.fr We thank Sylve Charlot for her help at varous stages of ths work. 1
1. Introducton Resdental segregaton by ncome and ethncty s a maor feature of contemporary Western ctes (see e.g. Card et al., 2008; Wheeler, LaJeunesse, 2008) and s generally consdered to be undesrable. It may be the source of poverty traps due for nstance to peer effects n educaton and lack of role models. It may weaken socal coheson and redstrbuton mechansms (Borvatn and Capellen, 2003). In short, segregaton may be the source of short-term as well as long-term nequaltes and socal tensons. In the Unted States, at least for ffty years, economsts have been tryng to explan resdental sortng by ncome and race n Amercan metropoltan areas. Two man strands of lterature are competng to ths purpose. The local publc fnance lterature and ts numerous extensons, rooted n Tebout s (1956) model of fscal competton, suggests a sortng across ursdctons that ndrectly contrbutes to the understandng of urban confguratons. The urban economc lterature, orgnally based on the Alonso (1964), Mlls (1967) and Muth (1969) model of a monocentrc cty, suggests a sortng over space. To mprove the understandng of ncome sortng, an ntegrated modellng approach s clearly requred (Nechyba and Walsh, 2004) and has only been ntended recently (Brueckner, Thsse and Zenou, 1999; Bartolome and Ross, 2003, 2004, 2007). The am of ths paper s to provde a contrbuton to ths ntegated approch on the bass of an emprcal analyss of resdental choces and ther mpact on urban segregaton. Our obectve s to assess emprcally the relatve contrbuton of Tebout-sortng and Alonso-sortng mechansms to socal and ethnc segregaton n 37 large French urban areas. Smlar to what s observed n other ndustralzed countres, French urban areas, that host n 1999 77% of the French populaton, have been contnuously sprawlng snce the sxtes and resdental segregaton by ncome and ethncty s commonly acknowledged as a strkng ssue (Pan Ké Shon, 2009; Gobllon and Selod, 2007; Prétecelle, 2006). In ths context, mprovng our understandng of the causes and consequences of households locaton decsons wthn French urban areas s obvously of great poltcal relevance, n partcular because poltcal answers to segregaton are not the same whether t s due to Alonso-lke or Tebout-lke mechansms. Ths nqury requres the emprcal analyss of resdental locaton choces, n terms of dstance to the central cty and local economc condtons, as a functon of households ncome. The lterature nterested n ths queston manly reles on housng prce hedonc estmatons amed at estmatng wllngness to pay for accessblty and local amentes. Nevertheless, some recent works use dscrete choce modelng. For nstance, Ioanndes and Zanella (2008) analyze the demand for neghborhood "qualty" n contrastng locaton choces of households wth and wthout chldren. Schmdheny (2006) focuses on the mpact of local progressve ncome taxes on locaton choces n Swtzerland. We follow the methodology proposed by Schmdheny (2006). The approach conssts n two steps. In a frst step, we estmate on each of the 37 urban areas a condtonal logt model of resdental locaton choce, n whch households are supposed to select a communty dependng on ther ncome nteracted wth dstance to the central cty and mean ncome of potental destnatons. The specfc advantage assocated wth relocatng close to one s former communty s also taken nto account through a specfc former locaton dummy. Ths advantage s due to better nformaton about housng supply, lower movng costs, access to prevously bult local socal networks, etc. Because consderng all the communtes of an 2
urban area as potental destnatons n the locaton choce set would gve far too many alternatves n the logt model, we bult a smaller locaton choce set by groupng ursdctons accordng to ther mean household ncome and dstance to the central cty n 1990. As households ncomes are not drectly avalable n our data, occupatonal status s used as a proxy. Ethncty s also accounted for. Our sample encompasses all mgratons that occurred between 1990 and 1999 wthn 37 urban areas of more than 200,000 nhabtants and contans 210,611 households. In a second step, socal and ethnc dssmlarty ndexes of mgrants are calculated for each urban area ether wth the observed household dstrbuton or wth counterfactual household dstrbutons predcted based on the estmated condtonal logt model. The relatve mportance of the two segregaton channels stressed above.e. the choce of dstance to the central cty and of neghborhood economc condtons can be dsentangled by predctng dstrbutons of households between types usng only the coeffcents of our model assocated wth ether the choce of dstance to the central cty or the choce of local economc condtons. As, accordng to the thrd segregaton channel n our model, movng households are ceters parbus more lkely to choose to relocate nearby ther former locaton rather than elsewhere, the segregated pattern prevalng n 1990 s partly translated to 1999. Our man results are the followng. Gong beyond the standard urban model to take nto account ncome-related amentes s strongly ustfed by our results: segregaton among the economcally actve socal groups appears to be manly drven by the ncome channel, especally for the most affluent socal group. Nevertheless, segregaton between economcally actve and nactve households s manly explaned by dstance, as predcted by the standard urban economc model. Regardng ethnc segregaton, the dstance channel domnates the ncome one, but ths s lkely to be due to a non-market spatal structure,.e. the centralzaton of the publc housng supply. More mportant, both of them are domnated by the nerta of the prevously prevalng segregated pattern. The remanng of the paper s organzed as follows. Secton 2 presents the background and the hypotheses of our emprcal analyss. Secton 3 presents the emprcal model. Secton 4 descrbes the data, the sample defnton and a few descrptve statstcs. Results are presented n the last secton. 2. Background, theoretcal model and overvew of the emprcal approach 2.1. Background In the local publc fnance lterature, sortng across ursdctons s drven by households tastes for publc amentes. In the orgnal Tebout s model, publc amentes are the level of publc goods produced by ursdctons. The level of publc good s a normal good, whch means that ts demand ncreases wth ncome. Households wth dfferent ncome demand dfferent publc servce levels and thus choose dfferent ursdctons, each household choosng the ursdcton whch provdes hs preferred publc servce level. Assumng that ursdctons are formed on a featureless plan, ursdctonal boundares may be freely adusted, the publc good s fnanced by a head-tax and a households ncome does not depend on the ursdcton n whch t resdes, ncome stratfcaton across ursdctons should be perfect at the long term equlbrum (Tebout, 1956). Ths ursdctonal sortng s a potental 3
source of suburbanzaton, as people move to suburban ursdctons to get ther desred level of publc servces. Subsequently, the local publc fnance framework has been extended to other knds of publc amentes, produced by local governments or drectly produced by the populaton, ether n ursdctons or n neghborhoods. Some extensons specfcally am at explanng suburbanzaton, ncome and racal sortng,.e. the well-known US urban pattern dsplayng rch households n suburbs and poor households n central ctes. The mpled publc amentes can be dvded nto factors that pull rch whte households nto the suburbs and factors that push them out of cty centers (Nechyba and Walsh, 2004). From the pull-sde, sprawl can be explaned by the possblty of mplementng zonng regulatons n suburban ursdctons to exclude those supposed to brng wth them negatve fscal externaltes (free rdng on tax payments, etc.) or peer externaltes (crme rates, school qualtes, etc.). In turn, f zonng makes hgh densty mpossble n the suburban ursdctons closer to the cty center, ths wll ncrease the development of sprawl on the cty edge. The push-sde corresponds to the flght from blght hypothess (Jackson, 1985; Mlls and Lubuele, 1997): rch households move from the central cty to suburban ursdctons n order to flee the negatve externaltes produced by poor households (hgh crme rates, low school qualty, general fscal dstress, etc). These push-factors has to be related to Schellng s (1971) model of segregaton, whch features externaltes lnked to the neghborhood populaton composton. Ths model shows that even mld preferences for havng smlar neghbors may lead to sharp segregaton patterns. More generally, lterature on neghborhood externaltes s a partcularly vvd feld of research (Ioanndes and Zabel, 2008; Ioanndes and Zanella; 2008). It appears that the demand for the qualty of the neghborhood socal composton n terms of provdng postve externaltes for access to employment and human captal accumulaton for nstance s lnked to the local ncome level and generates ncome segregaton (Benabou, 1993; Durlauf, 2004). In the urban economc lterature, sortng over space s drven by households tastes for land and accessblty to obs. In the standard monocentrc cty model, obs are all n the central busness dstrct (CBD), so that a locaton closer to the CBD has the advantage of lower commutng costs but due to land market competton the nconvenent of hgher land prces. Both land and accessblty to obs are normal goods. Rch households have a hgh land consumpton and thus are more strongly attracted than the poor by low land prces. However, they also have a hgh opportunty cost of tme, so that they value more accessblty to obs than the poor. Sortng between rch and poor households depends on the value of the ncome elastcty of commutng costs relatve to the value of the ncome elastcty of the demand for land. If the former s hgher than the latter, then the accessblty effect domnates and the rch households outbd the poor households for locatons closer to the CBD. If, on the contrary, the former s lower than the latter, then the land consumpton effect domnates and the rch households tend to lve further to the CDB than the poor households. Ths case s supposed to explan the commonly observed urban pattern n the Unted States. In both cases, the model dsplays a monotonc relatonshp between households ncome and dstance from the CBD (Wheaton, 1977) and produces ncome segregaton. In addton, a maor strength of the monocentrc model s that t allows to dentfy the prmary causes for urban sprawl n the twenteth century. Frstly, households ncomes rose mportantly durng ths perod (Margo, 1992). If the ncome elastcty of demand for land s greater than the ncome elastcty of commutng costs, as t s tradtonally assumed, rsng ncome s assocated wth rsng land consumpton and thus wth decreased densty n the 4
urban area. Secondly, transportaton costs have dramatcally decreased durng ths perod, especally due to the development of car ownershp (Glaeser and Kahn, 2003). The fall of commutng costs generates an addtonal ncome effect and above all weakens the centrpetal force of the model: the amount of land consumed ncreases and the edge of the cty expands. A model ntegratng these two strands of the lterature must take nto account smultaneously households tastes for publc amentes, land consumpton and accessblty to obs. Ths s the case n the model of Brueckner, Thsse and Zenou (1999) (BTZ n the followng). Income sortng s due to the conventonal forces of the monocentrc cty model, but also to the households tastes for urban amentes. Urban amentes are of three knds: natural, hstorcal and modern. Whle natural and hstorcal amentes are largely exogenous, modern amentes are endogenous, wth ther levels dependng on the current economc condtons n a neghborhood, especally the local ncome levels. Such amentes mght nclude restaurants, theaters, and modern publc facltes such as swmmng pools and tenns courts. (Brueckner, Thsse and Zenou, 1999, p.94). The authors make the tradtonal assumpton that the conventonal locaton forces drve the rch to the suburbs and the poor to the cty center (.e. the ncome elastcty of commutng costs s lower than the ncome elastcty of land demand). They show that f the center s exogenous amenty advantage s suffcently large, the equlbrum outcome can be reversed: the rch households outbd the poor household for locatons n the cty center. Ths addtonal locaton force could be at the orgn of the dfferences between US and European ctes. Europe s longer hstory and dfferences n government nvestment n central cty nfrastructures s lkely to explan the dfferences n the spatal pattern of exogenous amentes. In addton, modern endogenous amentes ntroduce the possblty of multple equlbra: they make the exstng locaton of the rch attractve to them wherever t mght be. Ths possblty may help explan the varety of locaton patterns by ncome observed n realty. The BTZ model consttutes a frst step toward the ntegraton of Tebout-sortng and Alonsosortng models but t unfortunately does not solve ther man caveats. As stressed by Bartolome and Ross (2003), ther predctons are hardly supported by facts: ncome sortng between ursdctons s far from perfect; emprcal estmatons suggest that the ncome elastcty of commutng costs s greater than the ncome elastcty of land demand (Wheaton, 1977; Glaeser, Kahn and Rappaport, 2000); and the relatonshp between households ncome and dstance from the CBD does not appear to be monotonc n general (Glaeser, Kahn and Rappaport, 2000). Another ntegrated model that better matches emprcal facts was developed by Bartolome and Ross (2003, 2004, 2007), hence named the BR model. Ths s a model of a monocentrc cty wth ursdctons provdng publc servces. Households sort over ursdctons based on conventonal locaton forces and publc servce levels. The level of publc servces n a ursdcton s decded by maorty votng and rch households must be the maorty to be able to mplement ther preferred level of local publc goods. When t s not the case, they tend to vote wth ther feet for ursdctons wth hgh publc servce levels. Income elastcty of commutng costs s now assumed to be greater than the ncome elastcty of land demand, but rch households may nevertheless settle n the suburb because of the hgher publc good level there. Therefore, the model s able to predct complete as well as partal ncome sortng between ursdctons and across space, and the relatonshp between households ncome and dstance from the CBD may be, or not, monotonc, dependng on households tastes for local publc goods and of the untary commutng tme. 5
2.2. Segregaton mechansms n French urban areas In ths artcle, our obectve s to assess emprcally the relatve contrbuton of Teboutsortng and Alonso-sortng mechansms to socal and ethnc segregaton n 37 large French urban areas. Our emprcal framework reles upon a few mportant assumptons. Put n a nutshell, we assume that wthn urban areas, households sort across ursdctons prmarly on the bass on ursdctons dstances to the central cty and ursdctons means of households ncomes. Due to data constrants, we however do not focus on households sortng by ncome but household sortng by socal status. More precsely, we frst assume that these urban areas are essentally monocentrc n terms of employment. Consderng French urban areas n 1999, we know that central ctes contan 27.2% of the total employed populaton but 41.5% of total obs (Julen, 2001). Secondly, as n the BTZ and BR models, we assume that households sort between ursdctons both on the bass on ther preferred accessblty to the central cty (resultng from ther tastes for land consumpton and accessblty to obs) and ther preferred level of publc amentes. The publc amentes supposed here to be relevant regardng households locaton decsons are the ursdcton s endogenous modern amentes, ncludng notably the publc servce level and the external effects of populaton composton. Thrdly, we make mportant assumptons about proxes. Modern amentes are assumed to be well proxed by the ursdctons mean of households ncomes: hgh ncome ursdctons offer hgh qualty publc and prvate servces, postve peer externaltes, etc. Accessblty to the central cty s assumed to be well proxed by the ursdctons brd s eye dstances to the central cty. 2.3. Methodology overvew Our methodology draws on Schmdheny (2006), who studes the mpact of local progressve ncome taxes on households locaton choces and ncome sortng n the cty of Basel, Swtzerland. Ths methodology s n two steps. The frst step conssts n the estmaton of a condtonal logt model of households locaton decsons, one for each the 37 urban areas consdered n our study. For each urban area, we consder a 1/20 sample of households that moved wthn the urban area between 1990 and 1999 (all urban areas together, there are 210,611 households n our sample). The choce set of each movng household s defned on the bass on all ursdctons ncluded n the urban area n 1999. However, n urban areas of more than 200,000 nhabtants, consderng all ursdctons as potental destnatons would gve far too many alternatves for estmatng a condtonal logt model and hence we group ursdctons of each urban area accordng to ther dstances to the central cty and to ther means of households ncomes n 1990. The choce set fnally contans 17 classes of ursdctons. The man explanatory varables of households choces between these classes are nteractons between households characterstcs and characterstcs of the ursdcton classes. On one hand, we nteract ether the socal or the ethnc status of households wth the classes weghted mean of ursdctons dstances to the central cty (the weght corresponds to the number of houses of each ursdcton). On the other hand, we nteract ether the socal or the ethnc status of households wth the locaton class mean of households ncomes. Importantly, the households ncomes taken nto account are measured n 1990,.e. before mgratons took place. Indeed, as stressed by the BZT model, modern amentes are endogenous: they are both a cause and consequence of the locaton patterns of dfferent ncome groups. There s a 6
smultanety ssue that we try to lmt n our econometrc model by takng the past value of locaton ncome. The second step reles on predcted locaton choces and segregaton measures. For each urban area, we can easly measure the observed level of segregaton by socal and ethnc status across classes of ursdctons of movng households, based on the computaton of dssmlarty ndexes. Then, for each urban area, we can use the coeffcents obtaned n our frst step estmaton to predct the counterfactual locaton pattern of these households and measure the predcted level of segregaton by socal and ethnc status across classes. By comparng the observed and predcted segregaton levels, we are able to assess the predcton power of our model, thus the ont explanatory power of Tebout-sortng and Alonso-sortng mechansms. Furthermore, the relatve mportance of these two segregaton channels can be dsentangled. Indeed, for each urban area, we can select the coeffcents assocated wth explanatory varables correspondng to one segregaton channels only (nteractons wth classes dstances only or wth classes mean ncomes only), settng all the other coeffcents to zero, and predct the correspondng counterfactual locaton patterns of households across classes. We can thus measure and compare the Tebout-sortng predcted level of segregaton and the Alonso-sortng predcted level of segregaton. A thrd segregaton channel s also ncluded n our model, namely, the specfc advantages of relocatng nearby one s former locaton. These advantages may be due to better nformaton about housng supply, lower movng costs, access to prevously bult local socal networks, etc. We expect that movng households are more lkely to choose to relocate nearby ther former locaton rather than elsewhere, and that the segregated pattern prevalng n 1990 s partly translated to 1999 due to ths addtonal effect. 3. Data and sample defnton Before to expose the econometrc model n detal n secton 5, we present n ths secton our data and the varables used n the estmatons. We also gve a few descrptve statstcs concernng the locaton classes that consttute the explaned varable of our model, as well as the sample statstcs. 3.1. Data Our emprcal nvestgaton s essentally based on the 1999 French Populaton Census (produced by INSEE, the French Natonal Insttute for Statstcs and Economc Study). A 1/20 th sample s drawn from the Census, n whch detaled characterstcs of households and ther members are avalable. Resdental locaton of these households at the Commune level s known n 1999 and 1990, allowng to trace households' moves. Unfortunately, households ncomes are not. We wll therefore make use of the occupatonal status of the household's reference person as a proxy for household's ncome. We wll however gve statstcs n the followng n order to show that usng occupatonal status categores s ndeed useful to measure ncome segregaton. As to locaton characterstcs, our analyss requres nformaton on ursdcton average household ncome as well as ursdcton's dstances to the employment center of the urban area. The mean households ncomes of each Commune (.e. French ursdcton) comes from the French Tax Authortes (INSEE/DGI). The French Natonal Geographc Insttute (IGN) 7
provdes the geographc coordnates of each Commune s town hall, allowng to compute the straght-lne dstance between the Commune's townhall and the center of the urban area. Other characterstcs of locatons are taken from the 1999 Populaton Census aggregated at the Commune level. In the descrptve statstcs, we also make use of 1993 local housng tax rates provded by the French Authorty for Local Governments (INSEE/DGCL). We also own data from the French and the Parsan notary socetes (PERVAL, Chambre des notares de Pars) gvng mean housng prces n 2002. 3.2. Choce of urban areas and defnton of the locaton choce set Our study s amed at explanng locaton choces wthn urban areas, that s, n labor-market areas wthn whch households are assumed to choose ther resdental locaton consderng ther workplace as gven. Urban areas are defned n France based on commutng flows as measured n 1999 from Census data. An urban area comprses a cty center and nner suburbs dvded nto several ursdctons and a rng of outer suburbs composed of ursdctons that do not belong to the urban agglomeraton but whch are tghtly ted to t by commutng flows. 1 There were 354 urban areas n France n 1999 wth a total of 45 mllons nhabtants representng 77% of the French populaton. Because sortng s a more strkng ssue the bgger the urban area, we focus on urban areas of more than 200,000 nhabtants n 1999. Wthn each urban area, consderng each ursdcton separately as a potental destnaton would gve far too many alternatves n the locaton choce model. 2 Consequently, wthn each urban area, we form groups of ursdctons that wll be consdered as alternatves n the choce model. These groups are amed at beng as homogenous as possble wth respect to the two man characterstcs of our analyss: dstance to the cty center and household mean ncome. Therefore, n each urban area, we classfy ursdctons: () frst of all, accordng to ther poston n the urban area: cty center, nner suburbs, outer suburbs; () then, n each of these prelmnary groups - -except for the cty center that has only one ursdcton-, accordng to ther poston relatvely to the group medan dstance to the cty center: close and dstant ursdctons; () and eventually, accordng to quartles of the ursdctons average household ncome. Therefore, we obtan 17 groups of ursdctons that are consdered by movng households as potental destnatons. Note that ths classfcaton mposes to have at least eght communes n the nner suburbs and eght n the outer suburbs to be able to defne a choce set of 17 types. Applyng ths crtera to urban areas of more than 200,000 nhabtants yelds 37 urban areas (out of 41 French urban areas of ths sze). 3 We now descrbe the 17 types of locaton thus defned. In partcular, we want to stress the correlaton between the types average household ncome and other characterstcs, namely: the percentage of foregn populaton and local taxes. For the sake of clarty, we present the descrptve statstcs after makng an addtonal poolng of ursdctons: at the last step of the 1 In France, an urban unt (unté urbane n French), s a set of communtes, the terrtory of whch s covered by a bult-up area of more than 2,000 nhabtants, and n whch buldngs are separated by no more than 200 meters. Each urban area s bult around an urban unt havng at least 5 000 obs. 2 For nstance, Lyon urban area has as many as 296 communtes. Toulouse urban area encompasses 342 communtes. 3 The excluded urban areas are those of Brest, Rems, Lmoges and Nîmes. 8
classfcaton, we group the ursdctons whch belong to the second and thrd quartles of average household ncome. Thus, we present statstcs only for 13 types n Table 1. As urban areas can have dfferent average household ncome and dfferent spatal ranges, several statstcs n Table 1 are gven relatvely to the urban area average characterstcs. Table 1: Characterstcs of the 17 locaton classes relatvely to the urban area average Cty center Close nner suburbs Dstant Inner suburbs Close Outer Suburbs Dstant outer suburbs Central ursdcton ncome a dstance b 93.2 0 % foregn 8.0 housng tax rate b 18.20 housng prces 102.3 Low-ncome Medum-ncome Hgh-ncome ursdctons ursdctons ursdctons ncome a 84.4 104.3 139.1 dstance a 42.2 41.7 42.8 % foregn b 7.5 4.1 3.0 housng tax rate b 14.16 12.95 10.84 housng prces a 90.0 107.8 124.5 ncome a 85.0 101.9 129.8 dstance a 97.6 94.9 84.8 % foregn b 6.6 4.4 3.2 housng tax rate b 13.16 12.16 9.98 housng prces a 90.3 105.6 122.6 ncome a 82.0 99.4 127.4 dstance a 124.2 110.2 92.8 % foregn b 3.2 2.4 2.2 housng tax rate b 10.49 9.88 9.70 housng prces a 85.8 101.3 113.8 ncome a 69.1 83.0 101.3 dstance a 204.5 184.3 179.9 % foregn b 2.5 2.4 2.6 housng tax rate b 9.03 9.51 9.26 housng prces a 74.2 84.0 97.7 a These fgures read as follows: on average over the 37 urban areas of our sample, the mean of households ncomes n the close nner suburbs of an urban area equals 84.4% of the average value of the same statstc computed over all types of the same urban area. b These are to be read more drectly: on average over the 37 urban areas of our sample, 7.5% of all households are headed by a foregn person n close nner suburbs and the housng tax rate equals 14.16%. Broadly speakng, except n the close nner suburbs, hgh-ncome ursdctons appear to be closer to the cty center than medum-ncome ursdctons, whch n turn are closer than lowncome ursdctons. Wthn dstant nner-suburbs for nstance, low-ncome ursdctons are at 0.98 from the average dstance, medum-ncome ursdctons at 0.95 and hgh-ncome ursdctons at 0.85. Housng prces are decreasng wth dstance from the cty center and ncreasng wth the ursdcton s average household ncome, rangng from 74% of the urban area mean n low-ncome ursdctons of dstant outer suburbs to 124% n hgh-ncome ursdctons of close nner suburbs. The proporton of foregn ctzen decreases wth dstance and wth the ursdcton s average household ncome. Housng tax rates follow the same pattern, wth hgher tax rates for lower ncome ursdctons. Note that prces used n these 9
statstcs are those of 2002, so that we cannot use them n the estmatons due to smultanety bases. Tax rates are gven n 1993 and the percentage of foregn ctzen n 1990, but both create multcollnearty ssues when ncluded smultaneously wth dstances and mean ncomes of types. 3.3. Sample defnton and descrptve statstcs Followng Schmdheny (2006), we focus on the behavor of households that moved wthn an urban area, gnorng mmoble households and other movng households. Indeed, we suppose that local mgratons (.e. wthn an urban area) are manly drven by motvatons related to housng (broadly understood as ncludng access to employment and amentes), whereas long dstant mgratons are essentally lnked wth the search of study and employment opportuntes, that may be assocated wth dfferent locaton behavors. Furthermore, people newly arrved n an urban area may not know well the characterstcs of locatons avalable n ths urban area. Defned on these crtera, the total estmaton sample (summng the 37 urban areas) contans 210,611 households that moved between 1990 and 1999 wthn urban areas. Some commentares and smple statstcs are gven to ustfy our choces and present the sample. Consderng the 37 urban areas of our study, 51.5% of households moved between 1990 and 1999. Among the movers, 67.7% were lvng n the same urban area n 1990. Descrptve statstcs seem to corroborate the dea that short dstance moves are manly drven by resdental consderatons whereas moves to a new urban area are employment or study related. Indeed, households that moved, but not wthn an urban area, have fve tmes more often a student as reference person of the household (and logcally dsplay a lower mean age and mean sze) (see Table 2). The reference person s also more often a foregn ctzen, what may be partly due to the arrval of new mmgrants lookng for economc opportuntes n French urban areas between 1990 and 1999. Table 2: Characterstcs of the populaton and of moble households Household wth student as RP 1 (%) Househ. wth foregn ctzen as RP (%) Mean age of the RP Mean household sze Household n a new locaton class n 1999 (%) Whole populaton 5.33 7.83 50 2.39 Mgrants wthn urban areas 4.03 8.39 42 2.55 45.79 Others mgrants 20.23 10.21 36 2.15 73.65 2 1 RP: Reference Person 2 Calculated for households that were lvng n another urban area of our selecton n 1990. We defne three ethnc groups and sx socal groups based on the characterstcs of the reference person of the household. 4 The former are foregn ctzens, French ctzens born 4 The reference person of the household s always the man n households where a man and a woman are n couple and ether a man or a woman n all other cases. 10
abroad and French ctzens born n France. The latter are bult on the bass of occupatonal status as follows: () executves and hgh ntellectual professons; () whte-collars n md management postons; () whte-collar subordnates; (v) blue-collar workers; (v) ndependent workers; (v) economcally nactve people,.e. retrees and people who never worked. The szes of these groups are gven n Table 3. As we would lke to be able to nterpret the spatal segregaton of these groups n terms of ncome segregaton, Table 3 also dsplays the results of a smple lnear regresson amed at predctng the average ncome of each category. Ths regresson s performed on data taken from the French Housng Survey 1996 (see Appendx 1). Table 3 shows that the classfcaton nto these four groups translates n a herarchy n terms of mean annual ncome by consumpton unt. Therefore, measurng segregaton on the bass of groups () to (v) provdes us nformaton about ncome segregaton. On the contrary, we know that n groups (v) and (v), ncomes and professonal stuatons are much dspersed (for nstance, ndependent workers can be farmers, craftsmen or bg entrepreneurs). Table 3: Sample dstrbuton by occupatonal status and natonalty of the reference person and estmated mean annual ncome by consumpton unt (n French Francs) French Foregn natonalty All N 1 Mean 2 N Mean N Mean Executves 31,046 152,527 1,041 127,639 32,087 151,720 Intermedate professons 40,779 102,776 1,687 81,454 42,466 101,929 Whte-collar workers 34,673 73,546 2,613 57,042 37,286 72,390 Blue-collar workers 40,616 69,768 7,621 56,116 48,237 67,611 All 147,114 97,273 12,962 65,345 160,076 94,688 1 The number of observatons corresponds to the estmaton sample. 2 The mean ncome s estmated based on French Housng Survey data. 4. Emprcal model of locaton choce and measure of resdental segregaton 4.1. A condtonal logt model of locaton choce We consder a random utlty model, accordng to whch utlty of a household n a locaton s the sum of a determnstc and a random part: V = * V where V s the determnstc part representng the nfluence of observed household and ursdcton characterstcs and s the dosyncratc random term specfc to household and ursdcton. + ε In order to test our hypotheses, utlty s supposed to take the followng form: (1) V * = α + β ln( y ) O 1 + β ln( y ) E 2 + β ln( y ) S 3 + γ d O + γ d O 11 + δ F + ε 12 2 + γ 21 d E + γ 22 d 2 E + γ d 31 S + γ 32 d 2 S (2) 11
where ln(y ) s the average populaton ncome n locaton, d dstance between locaton and the cty center, O and E are two vectors of dummy varables relatng respectvely to the occupatonal status and ethnc category of the household, S represents ts sze and F s a dummy varable ndcatng whether locaton was the former locaton of the household. α, β 1 to β 3, γ 11 to γ 32 and δ are vectors of coeffcents to be estmated. By dong so, we suppose that households dffer n terms of locatonal choces based on ther ncome, ethnc orgn and sze. We also assume that they make ther choce based on average ncome n the ursdcton and sze to the center. All the choce determnants that do not dffer wth household are n the locaton classes fxed effects α. In other words, the mpact of the locaton characterstcs that are constant across households are left n the locaton dummy, together wth unobserved varables. A household chooses among potental locatons by comparng ts utlty level n the dfferent locaton types and select locaton whch maxmzes hs utlty: V * * V k k C = where C s the choce set of K alternatve locatons. ( 1,..., K) We assume that the error terms are dentcally and ndependently dstrbuted followng an extreme value dstrbuton, of whch cumulatve dstrbuton functon s gven by: (3) (4) As a result, the probablty for a household to choose locaton s: Ths condtonal logt model s estmated by the maxmum lkelhood method. (5) Our emprcal model then focuses on nteracton varables between the locaton characterstcs and household's characterstcs. As stated n the theoretcal model, we expect that households sort by ncome relatvely to the dstance to the central cty. We ntroduce nteractons between dummes for household's socal categores (whte-collars n md management postons beng the reference) and locaton dstance to the cty center (and ths varable squared). The same s done wth a dummy for household's ethncty (French born n France beng the reference). To control for household's sze, we also ntroduce an nteracton between household sze and type dstance to the center (and ths varable squared). We also expect that mgrants take account of socal context and related endogenous amentes. The same nteracton varables are bult wth the locaton mean ncome nstead of dstance to the central cty. To sum up, household varables nteracted wth locaton varables are the followng: - dummes for the reference person occupatonal status, - dummy for the reference person not beng a French ctzen, - dummy for the reference person beng a French ctzen born abroad, - number of persons lvng n the household. 12
Locaton varables are: - mean fscal ncome n 1990 5 dvded by ts average over the 17 types, - dstance to the central cty n km 6 less the mean dstance to the cty center over the 17 types n the urban area, so as to avod colnearty wth ther squared counterparts. - dstance less mean dstance squared. A varable s both a household and locaton varable: - dummy for the locaton beng the former locaton of the household. 4.2. Choce probabltes and measures of socal and ethnc segregaton Our am s not only to test for the dfferent conectured factors of locaton choces, but also to assess the mportance of each of them n segregaton levels observed n French urban areas. We can do so by comparng segregaton levels n dfferent counterfactual ctes that are predcted by the model estmated coeffcents. Segregaton levels are classcally measured by dfferent spatal concentraton ndexes. Among them, the dssmlarty ndex s the most commonly used. Based on soco-occupatonal categores and ethnc orgns that we use n the locaton model, we compute mult-group dssmlarty ndexes (Readon and Frebaugh, 2002) of whch the general expresson s the followng: D = 1 2I M J π π 1 (6) m w m= 1 = 1 m where m ndexes the dfferent groups of populaton and the dfferent locatons. π m s the proporton of group m n the populaton, w s the weght of locaton n the total populaton and π m s the share n group m for the populaton n locaton. I s equal to π m ( 1 π m) and measures the dversty of groups among the populaton. Estmaton of the condtonal logt model n a gven urban area provdes for each household ts probabltes to choose each avalable locaton alternatves. Of course we also know the true dstrbuton of movng households among locatons. Hence, we can compute measures of segregaton n both cases and compare observed and predcted patterns, so as to assess how well our estmated model accounts for the observed resdental segregaton. Then, usng only the coeffcents attached to one dmenson of locaton choce (e.g. dstance to the central cty) and settng all other coeffcents except fxed effects to zero, a new set of choce probabltes can be predcted. Measurng segregaton wth the resultng counterfactual household dstrbuton provdes nformaton as to the contrbuton of ths causal channel to the producton of segregaton. More specfcally, we wll compute three types of predcted locaton choce probablty. The general form of ths probablty s: M m= 1 5 For each type, we sum the total fscal ncome of communtes belongng to the type and dvde by the total number of households of these communtes. 6 For each type, ths dstance s the average over all communtes belongng to the type of the communty level dstances to the cty center, weghted by the number of housng offered by each communty. 13
where V takes dfferent forms as follows: (7) () Probabltes predcted by the full model ~ V = α + β 1 ln( y ) O + β 2 ln( y ) E + β3 ln( y ) + γ d O + γ d O + γ d E + γ d 11 + δ F 12 2 21 22 S 2 E + γ d S Ths s smply the predcton of the full model and allows us to assess the explanatory power of the locaton model. () Probabltes predcted by a gven segregaton channel Our condtonal logt model ncludes only three locatonal characterstcs: household mean ncome, dstance to the central cty and, for each household, f t s her former locaton. The frst two locaton characterstcs are nteracted wth household characterstcs, allowng locaton behavors to dffer along the lne of the latter. The thrd s per se both a locaton and a household characterstc. Takng our estmaton results and settng all the coeffcents correspondng to these varables to zero, except those assocated wth one of these segregaton channels provdes us the choce probabltes that would be relevant f only ths channel produced socal and ethnc segregaton or n other words f households consdered only ths determnant n ther locaton choce. Computng the correspondng dssmlarty ndexes, we can assess the relatve contrbuton to segregaton of each of these channels. For nstance, the predcted probablty based on the dstance segregaton channel would be: 31 + γ 32 d 2 S (8) ~ V = α + γ d O + γ d O + γ d E + γ d E + γ d S + γ d S (9) 11 12 2 21 22 2 31 32 2 We refer n partcular to the nerta model, whch gves the followng predcted probabltes: ~ V = α + δ F (10) () Probabltes predcted by behavors dfferng wth occupatonal or ethnc group In our condtonal logt model, households dffer by ther belongng to a socal group, ther ethnc status, ther sze and ther former urban locaton. Let us focus on socal and ethnc status. Takng our estmaton results and settng the coeffcents of all nteractons mplyng ethnc status (resp. socal status) to zero provde the choce probabltes that would be relevant f behavors dd not dffer between ethnc groups (resp. socal groups) and was the same as the behavor of the reference category n the econometrc model. Thereby, computng the correspondng dssmlarty ndexes, we can check f socal segregaton s manly explaned by the dfferng behavors of households belongng to dfferng socal groups or by the combnaton of the dfferng behavors of households belongng to dfferent ethnc groups and the socal composton of ethnc groups. In the same manner, we can check f ethnc segregaton s manly explaned by the dfferng behavors of households belongng to 14
dfferng ethnc groups or by the combnaton of the dfferng behavors of households belongng to dfferent socal groups and the ethnc composton of socal groups. The correspondng predcted probabltes are for nstance, consderng that households only dffer n terms of socal categores: ~ V = α + β y O + γ d O + γ d O (8) 1 ln( ) 11 12 2 5. Results 5.1 Estmaton results. The estmated coeffcents for the fve largest urban areas, namely Pars, Lyon, Marselle, Llle and Toulouse, are presented n Table 4 as examples. Jont sgnfcatvty tests amed at assessng the sgnfcance of each of the nteractons n the model are presented for the 37 urban areas n table 5. The sgns of estmated sgnfcant coeffcents are presented n Table 6. Frst of all, fxed effects are always ontly sgnfcant (Table 5). Estmated fxed effects for the 17 locaton classes are all negatve: the central cty s always more attractve, what obvously reflects dfferences n housng stocks, whch are also dfferences n housng opportuntes. More densely populated locatons mechancally receve more mgrants. As expected, t s also observed that outer suburbs whch exhbt a very low densty compared to others generally dsplay the lowest coeffcents. These fxed effects thus control, among other effects, for the sze dfferences between locatons. We comment only brefly the coeffcents correspondng to the effect of household sze and prevous locaton. Household sze sgnfcantly nfluence dstance choce n the vast maorty of the urban areas. The effect s the one predcted by urban economc models: large households locate further away from the cty center, although the effect s not lnear, as shown by the coeffcent of quadratc dstance. The nteracton between household sze and average locaton ncome s sgnfcant and postve n 13 urban areas, whch can be nterpreted n partcular as the fact that the presence of chldren ncreases the preference for wealthy locatons. Fnally, there s a very sgnfcant effect of prevous locaton on locaton choces: locatons where the household was ntally are very more lkely to be chosen as the new destnaton when movng. The nteractons of socal category dummes wth locaton average ncome are ontly sgnfcant n 29 out of 37 urban areas. Ths result ponts out the power of the search for ncome-related amentes n the producton of resdental socal segregaton. When sgnfcant, the coeffcent of the "executve" category s postve, whereas t s always negatve for whtecollar subordnates and blue-collars (Table 6). Ths s true n partcular for the fve urban areas n Table 4. As expected, the most affluent socal groups are more attracted by the most affluent locatons than the less favored socal groups. However, n most urban areas, the coeffcent of the nteracton wth locaton ncome s hgher for blue-collar than for whtecollar subordnates: although the ncome dfferental between these two categores s very small (see Table 3 n Secton 3.3), they behave somewhat dfferently regardng ncomerelated amentes. The effect of socal status on the choce of dstance s slghtly less often sgnfcant. Nevertheless, these nteractons are ontly sgnfcant n 27 out of 37 urban areas (Table 5). 15
On average, whte-collar subordnates have the most negatve coeffcents: they are the least attracted by dstant locatons, compared to ntermedate categores (Table 6). Apart from Pars, t s the case n the large urban areas consdered n Table 4. In Pars urban area, the executves are the category that s the most reluctant to settle n dstant locatons, ceters parbus. Ths can be explaned by the level of traffc congeston, that makes wealthy households locate close to the center. Blue-collar workers behave dfferently from whtecollar subordnates: they behave ether lke the ntermedate category (reference) or are more attracted by dstant locatons (n only three urban areas, among whch Pars). Ths result s n lne wth what we already know from urban confguratons n France: t has been observed that blue-collar workers are more prone to locate n outer suburbs than whte-collars (Goffette-Nagot, 2000). What we show here s that t remans true after controllng for preferences regardng ncome-related amentes. The effect of ethnc orgn, after controllng for socal status, s slghtly weaker than that of occupatonal status. Stll, the coeffcents of the nteractons between ethnc orgn and dstance from the central cty are sgnfcant and negatve n the maorty of the urban areas (21 urban areas). Ths centralzaton of foregners could be the consequence of a strong concentraton of publc housng n the nner suburbs and the fact that foregn households are often housed n publc housng. Further, only n half of the urban areas (18) do foregn ctzens behave dfferently than French ctzen born n France regardng locaton ncome, by locatng n less affluent muncpaltes. French ctzens born abroad behave more often as ther fellow-ctzens, except n the four bggest urban areas (Pars, Lyon, Marselle and Llle) and n three others of medum sze. Ths result also can be the consequence of publc housng accommodaton, as the average ncome n the muncpalty s lkely to be correlated wth the percentage of publc housng n the housng stock. The sgnfcantly dfferent behavor of foregn households can also be the consequence of network effects, as mmgrants often choose to settle near ndvdual belongng to the same ethnc group. 5.2. Analyss of socal and ethnc segregaton. In the three followng sectons, we present and analyze smple statstcs for the dssmlarty ndexes computed for each of the 37 urban areas. In the frst secton, we look at how well our full model can predct the observed socally and ethncally segregated patterns. Based on predctons obtaned wth partal models, we then try to assess the relatve contrbutons to segregaton of the choces of neghborhood ncome and dstance to the central cty. Lastly, we try to dsentangle the socal and ethnc determnants of the observed segregated pattern. 5.2.1. Observed segregaton patterns and predctons of the full model Table 7 and 8 present the observed and predcted dssmlarty ndexes for the fve largest urban areas taken as examples, for ncome and ethnc segregaton respectvely. The followng tables dsplay ndexes averaged over the 37 urban areas: observed ndexes n Table 9 and the rato of predcted to observed ndexes n Table 10. Broadly speakng, we can frst note that the mean values of the dssmlarty ndexes computed here at a supra-muncpalty geographc level are low compared to what s generally obtaned n studes workng wth muncpalty or nfra-muncpalty levels. Ths s expected as socal and ethnc segregaton s lkely to be stronger the smaller the spatal scale consdered. Note also that ndvduals consdered here are mgrants and t s not clear a pror whether 16