The neighbor is king: Customer discrimination in the housing market

Transcription

1 Te negbor s kng: Customer dscrmnaton n te ousng market Perre-Plppe Combes y Bruno Decreuse z Benoît Scmutz x Alan Trannoy { June 200 Abstract Ts paper provdes a way to detect customer-based dscrmnaton n te ousng market usng survey data. We buld a matcng model wt etnc externaltes were landlords d er n te number of ousng unts tey own wtn te same negborood. In te event of customers prejudce aganst te mnorty group, landlords wo own several apartments dscrmnate more often tan sngle-dwellng landlords because tey nternalze a ger negatve externalty on ter probablty to ll ter oter vacances. Usng te Frenc Natonal Housng Survey, we sow tat tenants wt non-european orgn are less lkely to rent from a multple-dwellng landlord tan oter tenants. We ten sow tat te proporton of multple-dwellng landlords at te local level s postvely correlated wt te probablty of non-europeans to be lvng n publc ousng wle ts s not te case of oter etnc groups. Keywords: Customer Dscrmnaton, Negborood Externaltes, Housng Market JEL Class caton: R2, J7 Ts paper as bene ted from dscussons wt Perre Cauc, Jan Eeckout, Cecla Garca-Peñalosa, and Morgane Laouénan. We also ws to tank semnar partcpants at CREST, Insee, Scences-Po, Pars Scool of Economcs and Unversty of Ax-Marselles, as well as conference partcpants n Djon, Ax-en-Provence and London. Data was made avalable by te Centre Maurce Halbwacs. Ts researc was partly funded by te Drecton de l Anmaton de la Recerce, des Etudes et des Statstques (Dares). Te usual caveat apples. y Greqam-Ax-Marselle Unversté, ppcombes@unvmed.fr, ttp:// z Greqam-Ax-Marselle Unversté, decreuse@unvmed.fr, ttp://greqam.unv-mrs.fr/academc_fr.pp? ID=72 x Greqam-Ax-Marselle Unversté, benot.scmutz@gmal.com, ttp://stes.google.com/ste/benotscmutz { Greqam-Ax-Marselle Unversté, alan.trannoy@unvmed.fr, ttp:// trannoy/ndex.tm

2 Introducton If dscrmnaton n te ousng market appens to be te man drvng force bend urban patterns, ts could dramatcally a ect te desgn of any publc polcy geared towards mprovng te lfe of urban mnortes (Zenou, 2009). For example, far ousng legslaton could ten become a useful tool aganst spatal msmatc on te labor market. Tat sad, provdng conclusve evdence of ousng market dscrmnaton remans a callengng emprcal enterprse (Dymsk, 2006). It eter requres very detaled datasets (to estmate credble edonc prces) or te conduct of randomzed experments (Ynger, 986). Moreover, most researc on ousng market dscrmnaton as not yet bene ted as muc as t could ave from Becker s (957) teoretcal nsgts on te ratonale for customer dscrmnaton, even toug te ousng market s te quntessental customer market (Lang, 2007) and t s now agreed tat "most dscrmnatory beavor n te ousng market s founded upon eter te personal prejudces of agents or ter belef tat t s n ter nancal nterest to cater to te presumed prejudces of ter Anglo customers" (quoted n Farley et al, 99). In partcular, te respectve roles played by landlords and tenants prejudce n te dscrmnaton process are seldom clearly dsentangled, even toug te parallel s easly drawn between landlords and employers and between customers and tenants (or ome buyers). Ts paper addresses bot te teoretcal and emprcal callenges: to our knowledge, t provdes te rst teoretcal model of customer dscrmnaton n te ousng market, wc ten paves te way to a smple test for customer dscrmnaton. Becker s model of customer dscrmnaton focuses on te good market. In Becker (957), rms do not re black applcants for jobs n contact wt customers because wte customers racal prejudce lowers te productvty of black employees. Transposed to te ousng market, Becker s reasonng means tat tenants utlty depends on te etnc composton of te negborood. Prejudced applcants care about te racal makeup of te negborood and tey refuse to move n next to a negbor wose etnc type tey dslke. Landlords know tat and t may a ect ter decson to accept a mnorty tenant regardless of ter own prejudce. However, and unlke te clents of a partcular rm, te ats and ouses n a gven negborood do not necessarly belong to te same owner. Acceptng a mnorty tenant creates a negatve externalty at te negborood level. How landlords react to prejudced wtes depends on ter ablty to nternalze te externalty, wc n turn depends on te number of apartments tey own wtn te same negborood. Namely, te more apartments tey own, te more senstve to applcants prejudce and te more often tey dscrmnate. In ts paper, te negatve externalty created by a mnorty tenant takes place at te buldng level. We focus on te rental market and model landlords decson-makng process n a dynamc framework wt etnc eterogenety and matcng frctons. Rents are xed and some Wtes are prejudced aganst black tenants. Landlords ave to coose weter tey accept an applcant or not. We demonstrate tat tey may gan from refusng some types of applcants, even f ts ncreases te mean delay to ll te dwellng. Our model gves a ratonale for customer dscrmnaton by sowng 2

3 tat, provded enoug applcants are prejudced aganst anoter group of applcants, any landlord can nd mself n a stuaton were se would gan from refusng access to te members of ts group. Ts result cannot be drectly tested snce bot tenants and landlords may be prejudced aganst te same group of applcants. To come up wt an dent caton strategy between tenants and landlords prejudce, we go one step furter and consder te stuaton were landlords are n fact eterogeneous wt respect to te number of ousng unts tey own wtn te same buldng. Our model predcts tat landlords wo own several contguous apartments dscrmnate more often tan tose wo only own one apartment: ndeed, te latter only care about te mpact of ter selecton decson on ter ablty to rent out te same apartment agan n te future, wle te former also care about te mpact of ter decson on ter current ablty to rent out ter oter vacant lots. Ts leads to te predcton tat, n te event of customer-based dscrmnaton aganst black applcants, black tenants sould less often ave landlords wo own several ousng unts wtn te same negborood. Ts predcton s testable on regular survey data and consttutes a drect test for te presence of customer-based dscrmnatory practces n te ousng market. We ten provde an emprcal applcaton on Frenc data. Te Frenc case s relevant for two reasons. Frst, te d cult ntegraton of etnc mnortes s a major publc polcy concern. Te urban rots of Fall 2005, were more tan nne tousand vecles were arsoned n tree weeks (wt a clmax of almost four tousand n tree ngts), sowed te entre world tat France also ad race-relaton ssues, even toug te mere possblty of a spec cally racal problem ad long been dened by Frenc poltcal plosopy, ntentonally color-blnd. We argue tat some of te d cultes experenced by etnc mnortes n France may be attrbuted to ousng market dscrmnaton and ts consequences, segregaton and unemployment. Te second reason for pckng France s tat tere s lttle legal room for prce dscrmnaton n te Frenc rental market, wc ts well our xed-prce model. Contrary to wat appens on te ome-sale market, te asked rent s generally posted on te adds and landlords are not allowed to ncrease t unlaterally before sgnng te lease. Moreover, a set of laws and regulatory practces prevents tem from xng prces at ter wll on many segments of te prvate rental market. Prce dscrmnaton must be covert: t may nvolve te amount of te securty depost, or temporary dscounts n excange for mprovng te qualty of te dwellng, but ts cannot be te sole force bend te d erental treatment of undesred etnc groups and most dscrmnaton, f any, as to come troug quantty ratonng. Frst, we conduct a drect emprcal test of te teory. Ts test lays on te assumpton tat, condtonal on all oter observable caracterstcs of te dwellng, applcants do not observe or do not care about ter future landlord s type (weter se owns several contguous apartments or not). In oter words, searc s random. Under ts assumpton, te condtonal allocaton of tenants across landlords types can be nterpreted as te result of a natural experment, wc only re ects te supply sde of te market and does not rase any of te usual selecton ssues regardng te coce of resdence. Usng data from te Frenc Natonal Housng Survey, we sow tat rstgeneraton mmgrants of non-european orgn wo lve n prvately-rented apartments are less 3

4 lkely to ave a landlord wo owns te entre buldng. In te absence of any conclusve alternatve explanaton, ts result s nterpreted as te expresson of a supply constrant exerted by multpledwellng landlords upon mnorty applcants and consttutes a strong ndcaton of customer-based dscrmnatory practces n te Frenc prvate rental market. In a second stage, we make advantage of te fact tat, unlke audt data, te Natonal Housng Survey makes t possble to extend te analyss to te macroscopc consequences of customer dscrmnaton on te ousng consumpton patterns of etnc mnortes. In partcular, ter overrepresentaton n publc ousng as generally been accounted for, eter by etnc-spec c preferences, or by te story of mmgraton. We ere provde an alternatve explanaton of suc "socal ousng magnet" (Verdugo, 2009) were customer dscrmnaton plays a key role. In tat purpose, snce ousng consumpton stems from a complex nexus of decsons about many possble outcomes, we restrct ourselves to a very smpl ed dual rental market were some segments are prvately-rented by potentally dscrmnatng landlords and te rest s admnstered by a supposedly color-blnd publc autorty. Ten, usng a metodology derved from te emprcal lterature on spatal msmatc on te labor market, we construct a zone varable wc de nes te relatve wegt of multple-dwellng landlords n eac local ousng market. We nally sow tat te probablty of tenants of non-european orgn to be lvng n publc ousng s postvely correlated wt ts zone varable, wereas te correlaton does not stand for any oter etnc group. Provded ts probablty partly re ects te d culty to access te prvate rental market, ts second result gves valuable nformaton regardng te mpact of customer dscrmnaton on te resdental locaton of etnc mnortes. Te drect test s a new way to provde emprcal evdence on ousng market dscrmnaton. So far, two nterestng metodologes ave been developed n order to detect dscrmnaton n consumer markets but bot are subject to crtcsm (Ynger, 998). Te rst s to make use of avalable survey data and to look for te e ect of te etnc status of consumers on te caracterstcs of te goods wc tey ave access to. Ts s done to answer te followng queston: everytng else equal, do mnortes ave to resort to lower qualty goods and/or pay ger prces tan oter consumers? In te US, prce dscrmnaton n te ousng market as been studed snce te 960s, wen te growng expanson of te Afro-Amercan and Hspanc mddle class was startng to modfy te racal makeup of Suburba (Rapkn, 966, Kng et al, 973). Numerous studes based on edonc metodology and geograpcal dscontnutes sow tat Blacks often ave to pay a premum to enter formerly all-wte negboroods (Ynger, 997). Te use of large-scale representatve surveys s very nterestng from a polcy-orented vewpont snce t gves an dea of te aggregate mpact of dscrmnaton. However, ts metodology requres to dspose of very detaled nformaton n order to mnmze te rsk tat te observed pattern mgt be due to a trd factor, eter from te consumers sde or from te supplers sde; ts condton s even arder to meet wen one tnks tat tastes, wc are largely unobservable, are lkely to be partally etnc-spec c. Moreover, ts metodology does not provde evdence on dscrmnaton n te

5 makng: ow and wy agents dscrmnate remans unknown and te penomenon of dscrmnaton, f attested, remans a black box. For all tese reasons, most emprcal evdence ave come from parbased audts, wc glgt te role played by realtors. Many suc audts ave been conducted n te US snce te late 970s. For nstance, usng te results from an audt conducted n 98 n Boston, Ynger (986) sows tat black applcants are o ered up to 30% fewer opportuntes to vst ousng unts: two decades later, ts gap as narrowed but as not closed and by far (Ondrc et al, 2006). Anoter example could be found n te Housng Dscrmnaton Study of 989, were a seres of audts on 25 US metropoltan areas leads to anoter wave of evaluatons. In one of tem, Ondrc et al. (999) explctly dstngus between tree possble causes of dscrmnaton: agents prejudce, customers prejudce and agents mspercepton of Blacks preferences. As for te customer prejudce ypotess, t s tested troug tree d erent cannels. Frst, te autors look at te mpact of te ndvdual caracterstcs of black applcants tat prejudced negbors are most lkely to care about. Second, tey dentfy te caracterstcs of te negborood wc sould n teory mpact customer dscrmnaton as well. Fnally, tey argue tat a larger agency, wc bene ts from a broader clent base, may dscrmnate less. Altoug te researc strategy s plausble, tose tree d erent cannels can be crtczed on te bass tat one may come up wt many alternatve nterpretatons for eac varable. In order to capture te sole e ect of customer dscrmnaton, we argue tat one needs a varable wc spec cally takes nto account te mpact of te negborood externalty on realtors decson-makng process. In addton, wle smlar audts sould be conducted n France, tey are costly and ter partal equlbrum framework makes ter results d cult to nterpret. To paraprase Ynger (998), an audt study only ndcates te dscrmnaton tat occurs durng certan pases of a market transacton wen mnorty customers vst a random sample of rms and are qual ed to buy wat te rm s o erng, not te average dscrmnaton faced by an average mnorty customer. In oter words, audts gve causes, but not results. Our metodology borrows from bot knds of studes. We go back to te teory of dscrmnaton and we extract one spec c ratonale for dscrmnaton out of te black box. We ten derve an dent caton strategy wc reles on farly weak assumptons regardng consumers and supplers tastes. Tanks to te use of an orgnal varable on te geograpy of landlords real estate portfolos, we are able to solate more precsely te mpact of customer dscrmnaton. Ten, we make use of large-scale, easly avalable survey data, wc allows us to pursue analyses at te aggregate level as well. Bot aspects consttute a breaktroug n te lterature on ousng market dscrmnaton. Te paper s organzed as follows: secton 2 presents a matcng model of landlord beavor wt prejudced tenants and ts predctons n terms of dscrmnaton. Secton 3 tests tese predctons on Frenc data. Secton makes use of te prevous results to propose a novel explanaton of wy Recently, eld experments usng newly avalable matcng tecnques, suc as te Internet, ave been conducted. In partcular, Amed et al (2008) ave provded strong evdence for gender and etnc dscrmnaton n te Sweds rental market, by lookng at te reacton of landlords wo ad posted an ad on te web and were contacted by cttous applcants wt dstnctvely etnc and gender names. 5

6 etnc mnortes are over-represented n publc ousng. Secton 5 concludes. 2 Customer dscrmnaton n te ousng market: teory We consder a matcng model of te rental market wt etnc eterogenety and multple-dwellng landlords. Landlords want to ll ter vacances and decde weter tey screen some types of applcants or not. Te systematc refusal of one type of applcant s called dscrmnaton. We sow tat landlords wo own several apartments n te same negborood are more lkely to dscrmnate aganst black applcants. In oter words, tey are better o dscrmnatng under a wder range of crcumstances. Tese crcumstances are de ned by te caracterstcs of te local ousng markets: te matcng and separaton rates, and also te etnc composton and te tastes of te populaton of applcants. 2. Set-up Te cty s made up of dentcal buldngs, eac composed of two dentcal apartments. buldngs de ne te scale of te negborood externalty 2. Tese All apartments are rented out by landlords wo lve elsewere. Landlords are eterogenous wt respect to te number of apartments tey own, wc de nes ter type k (k = ; 2): type-2 landlords own an entre buldng, type- only own one apartment. Importantly, landlords are not temselves prejudced. Applcants are eter Wtes, n proporton p, or Blacks. Prejudce s one-sded: nobody s prejudced aganst Wtes and Blacks are not prejudced aganst Blacks, wc we denote w = 0, wereas a xed fracton b = of te populaton of Wtes s prejudced aganst Blacks. 3 If tey are matced wt a black negbor, prejudced Wtes refuse ter matc and keep searcng. Tey are not tme-constraned n te sense tat ter decson to refuse a matc s assumed to be costless. However, tey do not move out f a black tenant moves n next to tem. Matcng s random. Applcants do not observe te landlord s type, and cannot drect ter searc towards a partcular type as a result. Typcally, tey cannot orentate ter searc towards buldngs were tey are more lkely to be matced wt one type or anoter. Tme s contnuous. apartment tey rent wt probablty q. At eac date, landlords cannot expulse a tenant, but tenants leave te Conversely, landlords receve an applcaton for ter 2 Even f t s very small, t s not unrealstc, snce many dscrmnaton e ects only take place wtn te same buldng (Goux et al, 2007) 3 US studes sow tat more tan 70% of wtes are not wllng to move nto a negborood wc s more tan 50% Afro-Amercan, wereas more tan 80% of Afro-Amercans are wllng to move nto a negborood wt only a few black negbors (Farley et al, 99). From er observaton of current trends n Amercan negboroods, Ellen (2000) draws a two-sded concluson. On te one and, wte wllngness to move nto a negborood really seems to be a ected by te presence of Afrcan Amercans; on te oter and, wte wllngness to reman n ter current negborood s not clearly a ected by Afrcan Amercans movng n. Accordng to Ellen, ts means tat Wtes use racal composton to stereotype negboroods, wle tey do not necessarly prefer to resde n all wte negboroods. A more smple nterpretaton would be to assume tat movng costs, wc are already pad by prospectve tenants, overrde te negborood externaly. 6

7 vacant apartments wt probablty. From landlords perspectve, te parameter measures te e cency of te ousng market. As for te separaton probablty q, t does neter depend on te applcant s race, nor on te landlord s type. Wen a matc s completed, landlords receve a rent tat does not depend on te tenant race, but tey ave to pay mantenance costs tat are spec c to eac race. We assume tat unobservable caracterstcs correlated wt etncty make Blacks more lkely to deterorate te dwellng. Terefore, te net rent s etnc-dependent and lower for Blacks: R w R b. Ts opens room for statstcal dscrmnaton, wc we want to dentfy separately from taste dscrmnaton. Typcally, landlords may dscrmnate aganst Blacks for ts reason, even toug wte tenants would not be prejudced. Conversely, landlords beng unprejudced, n te absence of prejudce from consumers and for dentcal rents R w = R b tey would never dscrmnate. Fnally, r s te tme dscount rate. 2.2 Value functons and possble strateges Every apartment s of type dependng on ts occupaton status. It s eter vacant ( = v), occuped by a wte tenant ( = w) or occuped by a black tenant ( = b). We ave to caracterze te landlords optmal strategy wen t s vacant. De nton Landlords Strategy A landlord s strategy conssts n a vector of decsons accept/refuse to rent an apartment, condtonal on te applcant and te negborng apartment types. Te strategy s made of sx decsons snce te applcant can be eter wte or black and te negborng apartment can be vacant, rented to a Wte or rented to a Black. For a gven apartment, a landlord can face nne possble stuatons, dependng on te type ( = v; w; b) of er apartment, and te type j (j = v; w; b) of te oter apartment located n te same buldng. Let j k be te expected pro t of a type-k landlord wo owns a type- apartment, located n a buldng were te oter apartment s type-j. Landlord k accepts any applcant a (a = b; w) gven a negbor j (j = v; w; b) f and only f aj k acceptance condton and de ne te followng dummy functons (aj) k vj k. We smplfy te notaton of ts as: 8 (a; j) 2 fb; wg fb; w; vg ; aj k = ( f aj k vj k 0 oterwse () Te nne j k correspondng to an optmal strategy of landlord k are de ned by a system of steady-state equatons, one for eac possble stuaton. In any stuaton, te overall expected pro t s equal to te dscounted sum of te expected contnuaton pro t for eac possble event: Te tenant mgt stay or leave, and wen se leaves, te applcant mgt be wte or black and te landlord may accept er or not dependng on te oter apartment type. For type- landlords 7

8 stragtforward computatons lead to: 5 for all ; j = b; w; n r j = R + q vj r vj = q vv vj r v = R + q vv r vv = ( j 8 < + : v wv p [(wv vv o + v j wj ( j) p wj vj + bj ( p) bj vj + ( w ( ) p w ) + (vw + bv ( p) bv vv + b ( p) b )] + vb vv vv 9 = ; v v ) ) (2) (3) () (5) were aj k are de ned n (). As for te type-2 landlords, te system for ter gans s reduced to sx equatons. Indeed, avng te rst apartment occuped by a Black (or beng vacant) and te second by a Wte (or beng vacant), or te reverse cases, lead to te same outcome, e j 2 j 2 for 6= j. One gets: for all ; j = b; w; n r j 2 = R + R j + q vj 2 j 2 o + v 2 j 2 8 < r vj 2 = R j + q vv 2 vj 2 + : n r vv 2 = 2 wv 2 p [ vw 2 vv 2 ] + bv 2 ( p) wj 2 p ( j) + bj 2 ( p) wj vb 2 vv 2 2 vj 2 bj 2 vj 2 To solve tese systems, we rst assume values for te d erent dummes (j) k and ten we ceck f te solutons we derve are compatble wt tese assumed values. If ts s te case, t means tat strategy s optmal, as proposton summarzes. Proposton Caracterzaton of possble optmal strateges An optmal strategy s a strategy for a type- (type-2 respectvely) landlord for wc tere exst a set of value-functons tat are soluton of te system (2)-(5) (resp. (6)-(8)). A dscrmnatng strategy s de ned by te negbor s and te applcant s types. For a gven applcant, a Black say, egt outcomes may arse. Se can be accepted watever te type of te oter apartment, wc corresponds to a no dscrmnaton case. Ten, we can ave wat we call weak dscrmnaton, were te applcant s refused for one type of te oter tenant, but not te two oter types. For nstance, se s refused f te oter tenant s black, but accepted f se s wte or te apartment vacant. Medum dscrmnaton arses wen te applcant s refused for two of te types of te oter apartment, but not for te trd one. For nstance, se s refused f te oter tenant s wte or te apartment vacant, but accepted f se s black. Fnally, strong 5 See Appendx A for an example o 9 = ; (6) (7) (8) 8

9 dscrmnaton arses wen te applcant s refused watever te oter apartment type. detaled n de nton 2. Ts s De nton 2 Degrees of dscrmnaton () No dscrmnaton: te applcant s always accepted, watever te type of te oter apartment. () Weak dscrmnaton: te applcant s refused for one type of te oter apartment, but accepted for te two oter types. () Medum dscrmnaton: te applcant s refused for two types of te oter apartment, but accepted for te trd type. (v) Strong dscrmnaton: te applcant s refused watever te type of te oter apartment. Snce de nton 2 apples for bot wte and black applcants, all possble combnatons leads to 6 possble cases (for nstance no dscrmnaton of Wtes and strong dscrmnaton of Blacks s one of tem). However, many do not correspond to optmal strateges, wc s studed n te next secton. 2.3 Optmal strateges As long as te net rent from a Wte tenant s postve, R w > 0, a landlord s always better o wt a wte tenant tan wt a vacant unt. Snce furtermore Wtes do not eter exert any negatve externalty, as long as te net rent tey provde to landlord s larger tan for Blacks, tey are never dscrmnated aganst. Ts leaves us wt egt cases of possble dscrmnaton correspondng to de nton 2 appled to Blacks. Proposton 2 states tat four of tese strateges only can be optmal. Proposton 2 Optmal strateges mplemented () Wen R w R b > 0, dscrmnatng a wte applcant s never an optmal strategy. () If R w = R b > 0 dscrmnatng (Wtes or Blacks) s never an optmal strategy wen = 0 or wen tends towards 0: lm!0 bj k > lm!0 vj k for all j = b; w; v and k = ; 2 (9) () Tere are only four mutually exclusve possble optmal strateges regardng black applcants: no dscrmnaton, weak, medum or strong dscrmnaton, wc results from: bb k < vb k ) bw k < vw k ) bv k < vv k for all k = ; 2 (0) Proposton 2 restrcts te number of possble optmal strateges. Dong so, t seds lgt on te varous externaltes nvolved by customer dscrmnaton wt multple dwellngs. Refusng a black tenant features an opton value. In turn, ts opton value depends on te arrval rate of potental tenants, on te proporton of prejudced customers, on weter te oter dwellng s lled or not, and on weter t s rented to a black tenant or not. Te larger te opton value, te ger te rsk of dscrmnaton. 9

10 We already comment out te reasons wy not dscrmnatng wte applcants can be a domnant strategy, as stated n pont () of proposton 2. Ts s te case wen tere s room for statstcal dscrmnaton aganst Blacks, e R w R b. Blacks also are never dscrmnated for some values of te parameters as stated n pont (). If tere s no room for statstcal dscrmnaton (R w = R b ) or f tere are no prejudced applcants, and gven tat landlords are not temselves prejudced, dscrmnaton does not occurs. Te lmt property on, te probablty to receve an applcaton to a vacant apartment, sows tat dscrmnaton requres market power. Landlords cannot dscrmnate wen tey ave no cance of ndng anoter tenant. Rejectng an applcaton s a gly costly strategy as te correspondng opton value s nl n ts case. Pont () n proposton 2 ranks opton values and sows tat tere are only four possble optmal strateges for a landlord of any type. If bv k > vv k, Blacks are accepted wen te oter apartment s vacant, and terefore also wen t s rented by a Wte and even more by a Black. Ts may occur even wen R w R b > 0 for nstance wen s low. It s possble tat bv k < vv k, but bw k > vw k, a case of weak dscrmnaton were Blacks are dscrmnated wen te oter apartment s vacant but not wen t t s rented to eter a Wte or Black. Owng to customer dscrmnaton, te probablty tat a prejudced wte applcant come s g wen te oter apartment s vacant, wc leads to refuse Blacks n ts stuaton. Wen te oter apartment s occuped, ts probablty s su cently lower (because one rst needs te current tenant to move out) suc tat Blacks are accepted. Medum dscrmnaton were Blacks are not dscrmnated only wen te oter apartment s rented by a Black emerges wen bb k > vb k but bw k < vw k and bv k < vv k. Te value of a vacant dwellng remans relatvely low wen te oter dwellng s occuped by a black tenant but te value of rentng to a Black s lower wen te oter apartment s eter rented to a Wte or vacant. Fnally, strong dscrmnaton arses wen bb k < vb k. Te possblty tat prejudced Wtes wll refuse to move n, tereby ncreasng te probablty tat te dwellng wll stay un lled, s so g tat landlords always dscrmnate. 2. Landlords strateges comparsons Prevous results are vald for any type of landlord. We study now to wat extent optmal strateges of landlords d er dependng on te number of apartments tey own. Proposton 3 D erental dscrmnatng beavour of type- and type-2 landlords () Type-2 landlords dscrmnate more tat type- landlords: bj < vj ) bj 2 < vj 2 for all j = b; w; v () () If = 0, type- landlords dscrmnate f and only f type-2 landlords dscrmnate: bj < vj, bj 2 < vj 2 for all j = b; w; v (2) 0

11 Pont () n Proposton 3 reveals a fundamental property. Gven te oter tenant type, te fact tat a type- landlord dscrmnates mples tat a type-2 landlord dscrmnates. Customer dscrmnaton mples tat acceptng a black tenant entals two externaltes. Te rst externalty s statc. Havng a black tenant today reduces te cances tat te oter dwellng wll be rented out by a wte tenant. Terefore te value of te oter dwellng goes down. Type- landlords do not take ts e ect nto account. Te second externalty s dynamc. Acceptng a black tenant today a ects te future composton of te buldng. In turn, ts composton may alter te cances to nd anoter tenant n case of separaton. Bot types of landlords face te dynamc externalty, wereas only type-2 landlords nternalze te statc externalty. Tey are more lkely to dscrmnate as a result. Pont () n Proposton 3 sows tat type- and type-2 landlords beave smlarly wen tere s no customer dscrmnaton. In tat case, te only reason wy Blacks may be dscrmnated aganst s tat te net rent tey nduce s lower (Proposton 2). Statstcal dscrmnaton becomes equally lkely for bot types of landlords. Ts result provdes a smple strategy to test for te presence of customer dscrmnaton on te rental market. If Blacks are dscrmnated aganst by multpledwellng landlords and not by sngle-dwellng landlords, ten t means tat tere are prejudced tenants (Wtes) and tat landlords take tem nto account pror to rentng te apartment to a black tenant, weter tere s room for statstcal dscrmnaton or not. Predcton If black tenants are less lkely to ave a type-2 landlord tan wte tenants, ten tere are prejudced Wtes n te rental market. Te model predcts te dstrbuton of landlord s type by etnc group of tenants. In te absence of prejudced Wtes, black and wte tenants are equally lkely to ave a multple-dwellng landlord. Te probablty s equal to te proporton of dwellngs owned by multple-dwellng landlords. Wen tere are prejudced Wtes, multple-dwellng landlords may dscrmnate more tan sngle-dwellng landlords. In tat case, Black tenants are less lkely to ave a multple-dwellng landlord. Tere may be up to tree d erent stuatons tat enable us to dentfy te role of te landlord s type on te probablty of Blacks to access te rental market: Wen type-2 landlords weakly dscrmnate and type- landlords do not dscrmnate at all, wen type-2 landlords strongly dscrmnate and type- landlords stll do not dscrmnate, and wen type-2 landlords strongly dscrmnate and type- landlords also dscrmnate but weakly. As long as one of tese stuatons at least as an emprcal counterpart, Blacks sould be less lkely to ave a type-2 landlord. 2.5 Comparatve statcs Wereas one could go drectly to our emprcal test, we can move before deeper nto ntuton regardng te comparatve statcs of our model by parameterzng and smulatng t, keepng n mnd te study of racal segmentaton n d erent local ousng markets. We abstract from statstcal dscrmnaton and assume R w = R b = snce ts mpact s farly clear. We completed smulatons

12 for a large range of values 6 for r and q. Bot parameters act as deterrents aganst dscrmnaton. If r s g, landlords care a lot about current vacancy and do not dscrmnate; f q s g, acceptng a black applcant as a lower mpact on pro t snce te turnover s faster. Ts s farly stragtforward and tese two parameters sould not vary so muc across local ousng markets, te reason wy we dsplay results only for an annual nterest rate of 5% and for an average stay of 6 years. 7 We focus on te oter tree parameters p, and, wc are lkely to drve most of te eterogenety across local ousng markets. Fgure draws n te space (p; ; ) te solds for wc dscrmnaton s a domnant strategy for eac landlord type. From now on, ntutons beng smlar and tese stuatons leadng bot n te long run to te excluson of Blacks of te rental market, we call strong dscrmnaton te stuatons correspondng to bot cases () and () n de nton 2 (strong and medum dscrmnaton). Fgure : Trplets fp; ; g for wc dscrmnaton s a domnant strategy Fgure llustrates several features of te model. Wat strkes rst s tat dscrmnaton s often a domnated strategy. If Wtes were locally outnumbered (p. =2) or f enoug of tem were unprejudced (. 0:), no unprejudced landlord would ever dscrmnate. Te same appens n te event of a completely frozen rental market (. 0:05). However, tese tresold values would decrease wt te addton of a pecunary motve to dscrmnate (ger net rents for Wtes) or prejudced landlords. landlords dscrmnate. tan type-2. It s con rmed tat type-2 landlords always dscrmnate wen type- Wen bot types dscrmnate, type- never dscrmnate more strongly Te e ect of and s unambguous: f wte applcants are more prejudced, customer dscrmnaton ncreases and, smlarly, a more ud market makes t less rsky to dscrmnate by ensurng 6 Between 0.% and % for r, wc s a consdered to be a montly nterest rate, and between 0.0 and 0. for q, wc corresponds to an average stay between 00 and 0 monts. 7 Bot values were cosen to be close to our data. Furter smulatons are avalable upon request. 2

13 tat oter applcants wll be met sortly. In almost all cases, an ncrease n p also leads to an ncrease n dscrmnaton: f more applcants are wte, acceptng a black applcant s more costly snce t wll be d cult to meet anoter black applcant n te future. All of ts corresponds to te ntuton. However, wen p gets close to, ts e ect becomes ambguous, especally for type- landlords. Ts re ects te fact tat te dynamc externalty s no longer a concern. If almost all applcants are wte (and provded tere stll are some wte applcants wo are not prejudced), acceptng a black applcant wll not ncrease te probablty tat te oter apartment wll be occuped by a black tenant wen te landlord as to nd anoter tenant. 2.6 Dscusson In ts subsecton, we dscuss te teoretcal robustness of our test strategy to detect customer dscrmnaton n te ousng market by consderng varous possble extensons about landlords prejudce, pecunary externaltes, endogenzaton of some parameters, stronger prejudce from consumers and te ome sale market Taste-based dscrmnaton and statstcal dscrmnaton Introducng taste-based dscrmnaton does not alter te workng of te model. Suppose tat a proporton of landlords are temselves prejudced and systematcally refuse to rent ter apartment to a black tenant. As far as ts proporton s broadly te same among type- and type-2 landlords, te fact tat landlords ave a taste for dscrmnaton does not a ect te teoretcal predcton accordng to wc black tenants sould be less lkely to ave a multple-dwellng landlord. Actually, te same remark olds for statstcal dscrmnaton. Our model assumes tat te potentally negatve e ects tat Blacks may ave on landlord s pro ts do not cange wt landlord s type. To put t anoter way, our strategy s robust to varous omtted factors provded tat suc factors are not correlated wt landlord s type. However, te test strategy s not robust to omtted externaltes at te buldng level. Consder te case of pecunary externaltes. Suppose tat te net rent decreases for bot apartments as soon as a black tenant as been accepted n te buldng. Ts mgt arse f unobservable caracterstcs correlated wt etncty make Blacks more lkely to deterorate te common property of te buldng for nstance. Acceptng a black tenant n one of te two apartments lowers te value of te entre buldng. Type- landlords do not nternalze ts externalty, wle type-2 do. Terefore, Blacks could be less lkely to ave a type-2 landlord, even toug tere s no customer dscrmnaton and even toug a drect e ect of statstcal dscrmnaton s at work for bot types of landlords. However, suc externaltes are very assocated wt te formaton of wte prejudce. If black tenants deterorate te common parts of te buldng, wte tenants are lkely to avod black negbors. From ts perspectve, te consderaton of omtted externaltes at te buldng level as more to do wt te orgns of racal prejudce tan wt a competng teory for te under-representaton of Blacks n rentals owned by multple-dwellng landlords. 3

14 2.6.2 Endogenzaton of te parameters Many parameters of te model could be made endogenous. Te applcants arrval rate mgt depend on te proporton of black applcants and on etnc-spec c rents net of mantenance costs. Ts would requre specfyng a matcng functon and te supply of buldngs. Prejudce may also depend on te etnc make-up of te populaton. Moreover, te proporton of black applcants may respond to dscrmnatory beavor. In partcular, f p were made endogenous, dscrmnaton mgt lead to segregaton troug te consttuton of a dual ousng market were landlords would specalze n one type of tenant or anoter. Tose varous extensons would enrc te teoretcal model and elp understand dscrmnaton ssues better. However, tey would not a ect te test strategy tat reles on ndvdual dscrmnatory beavor. Optmal tenant acceptance or rejecton, as descrbed n Proposton 2, do not depend on te partcular way of closng te model. Rents also could be made endogenous. For nstance, tey could be barganed between te tenant and te landlord. Barganng requres to set an outsde opton for te potental tenant. In case of agreement, te barganed rent would mply tat te tenant s utlty s between te reservaton utlty and te gest level compatble wt landlord s acceptance. As te latter utlty level must be lower for a black tenant tan for a wte tenant, black tenants would pay ger rents at gven reservaton utlty. Wen te matc surplus becomes negatve, tere s no rent compatble wt landlord s acceptance and black applcatons get rejected. Assumng tat matc surplus s larger wt wte tan wt black tenants (tat condton does not seem too demandng, but t depends on te reservaton utlty tat may vary across etnc groups), te test strategy would be una ected. Indeed, multple-dwellng landlords would stll account for te negatve externalty tat a black tenant orgnates. Matc surplus would be lower for tose landlords tan for sngle-dwellng landlords. In oter words, multple-dwellng landlords would reject black applcants more often tan sngle-dwellng landlords. We do not elaborate more on ts extenson, snce statstcal regressons presented n te next secton do not conclude tat mnorty tenants pay ger rents, regardless of landlord s type Stronger de nton of prejudce We could add te possblty of a wte gt, wt prejudced Wtes movng out as soon as tey ave a black negbor. How ts stronger prejudce would a ect landlords beavor would depend on weter tenants prejudce s observable or not. However, n bot cases, type-2 landlords wll keep on dscrmnatng more. For nstance, n case prejudce s observable on tenants, type-2 landlords wt a prejudced wte tenant n ter second apartment wll always reject black applcants. Type- landlords, on te oter and, mgt care about knowng tat tey are about to make te negbor move out, but ts wll not always prevent tem from acceptng a black applcant. In case prejudce s unobservable, bot types wll dscrmnate more often tan wt te prevous de nton of prejudce. However, type- landlords stll do not care as muc as type-2 landlords about te mpact of ter acceptance decson on te probablty tat te oter tenant mgt leave as te result.

15 2.6. Heterogenety n buldng sze and colluson beavor Te dscrmnaton strateges we depct ave consequences n terms of long-run patterns of segregaton. Under weak dscrmnaton, Blacks are under-represented n te rental market; under strong dscrmnaton, tey are completely barred from t. However, f te goal of te model were to predct te allocaton of tenants across types accordng to etncty, many oter factors sould be ncluded, suc as te story of ownersp n te buldng and te eterogenety of te ousng supply wt respect to te number of apartments wtn eac buldng. Ts s not te purpose of te model, wc only ams to dentfy ndvdually optmal strateges and derve a predcton out of tem. In tat sense, te assumpton of omogenety of buldng sze s just ed. 8 Te man teoretcal drawback of a framework wt only two apartments n te buldng s tat t makes t d cult to rule out te possblty of colluson between te two type- landlords of te same buldng: f bot landlords cooperate, tey can no longer be dstngused from a type-2 landlord. Bot features (eterogenety n buldng sze and te possblty of colluson between type- landlords), f ncluded n te model, would decrease te probablty to observe predcton n te data. However, snce te bas can only be downward, ts does not a ect te relevance of te test Etnc matcng between d erent buldngs If landlords own apartments n several buldngs, tey sould be able to drect applcants to a buldng or to anoter, accordng to ter group. Ts knd of spatal targetng may act as a substtute for dscrmnaton. In tat case, te set of landlords strateges becomes qute complex because eac strategy s now a combnaton of a dscrmnaton strategy and an allocaton strategy. Te exstence of an allocaton strategy nges on a seres of assumptons ensurng tat te reallocaton of applcants between d erent buldngs can be made nstantaneously at no cost. Ts s seldom te case. Te apartments may not be substtutable because ter caracterstcs, suc as locaton or sze, d er too muc from one anoter. Also, tere may be vst costs, snce vstng d erent places takes tme. Last but not least, te varous apartments may be attrbuted by a real estate rm tat accounts for te varous landlords dscrmnaton strateges, but, gven te magntude of at portfolo, does not necessarly reallocate potental tenants towards apartments belongng to te same landlord. However, snce our emprcal strategy reles on te predcton of te model, t s wort examnng f suc an extenson s lkely to jeopardze t or not. In Appendx B, we develop te smplest case of "type-" landlords, wo own two buldngs. We descrbe ter most e cent tenant allocaton strategy tat can take applcants prejudce nto account. Gven te complexty of te system, we perform smulatons on, p and to ceck weter ts strategy reduces te ncentves to dscrmnate. Tese smulatons, avalable upon request, lead to te followng tree results: () regardng strong dscrmnaton, type- landlords are very smlar to type-2 landlords: 8 Te next secton wll sow tat usng real data mples to compare type-n landlords (n 2), wo own te entre buldng, to type-k landlords (k 2 f; :::; n g), wo do not. For a gven buldng of sze n, former type- landlords may ten greatly d er from one anoter. 5

16 tey wll nd pro table to strongly dscrmnate under very smlar crcumstances, only margnally more restrctve; (2) on te contrary, as far as weak dscrmnaton s concerned, tey wll dscrmnate less often tan bot type- and type-2 landlords: for weak dscrmnaton to start beng a domnant strategy for type- landlords, te market as to be close to frctonless ( > :5); (3) owever, n most cases, f type- landlords weakly dscrmnate, ten type- landlords strongly dscrmnate. As a consequence, even toug we cannot rule out te possblty tat type- landlords sometmes dscrmnate less tan type- landlords, ts stuaton s rater unlkely. Moreover, snce type- landlords cannot be dstngused from type-2 landlords n te data, te bas can only be downwards and ts extenson does not substantally a ect te relevance of te test Home-sale market Separaton parameter q s wat mostly dstnguses our framework from a model of te omesale market. Te matcng for a sale s one-sot. Once te lot s sold, te seller s no longer nterested n ts future evoluton, ence q = 0 for ever. Te modellng of te ome-sale market does not requre many addtonal canges. One can stll consder a cty of dentcal buldngs wt two dentcal apartments, wt apartments for sale belongng to unprejudced owners. We dstngus between type- sellers, wo only sell one apartment, and type-2 sellers, wo sell bot apartments smultaneously yet n two separate lots. Absent statstcal dscrmnaton and own prejudce, type- sellers wll never dscrmnate out of te prejudce of wte buyers because tere s no more dynamc externaltes. On te contrary, f te sale process s sequental and f applcants observe te type of ter potentally future negbor before acceptng to buy te apartment, type-2 sellers may dscrmnate aganst Blacks wo apply for te rst of ter two apartments (once tey ave sold one apartment, type-2 sellers become type- and tey stop dscrmnatng). We do not elaborate more on ts extenson snce te type of seller from wc omeowners ave bougt ter dwellng s not present n te dataset we use. 3 Detectng customer dscrmnaton Ts secton tests for te presence of customer dscrmnaton n te Frenc ousng market. We descrbe our dataset and notably te fact tat tenants state weter ter landlord owns te entre buldng or not. Te man test conssts n confrontng Predcton to data. Te test sows tat tenants wt non-european orgn are less lkely to rent from a landlord wo owns te entre buldng. We ten ceck tat ts result s robust to several possble ssues. 3. Data Our dataset conssts n te poolng of te last tree 9 waves (996, 2002 and 2006) of te Frenc natonal ousng survey ("Enquête Natonale Logement", encefort ENL). Te ENL s a detaled 9 Prevous waves lack crtcal nformaton about te orgn of te respondent. 6

17 cross-sectonal survey on a natonally-representatve sample of around trty tousand ouseolds, trty- ve tousand dwellngs and seventy- ve tousand ndvduals. Te man drawback of te ENL s nerted from a Frenc poltcal tradton, wc makes t stll controversal to collect racal or etnc statstcs. Consequently, we solate a group of "Blacks" wo are n fact rst-generaton mmgrants of non-european orgn: bot non-european ctzens and people born out of Europe and not Frenc at brt. Tree quarters of tem come from Afrca, and most of te quarter left come from Sout and Souteast Asa. Ts measure of etncty msses a large number of people, because of colonal story (people born n te colones were gven Frenc ctzensp at brt), of Frenc West Indes and of te ncreasng number of second, trd and even fourt-generaton mmgrants of non-european orgn n France. Moreover, t does not clearly dsentangle etncty and mmgraton status. We address ts ssue n two ways: rst, we always consder te group of "European orgn" (bot non-frenc European ctzens and people born out of France, n Europe and not Frenc at brt) as a second control group, ntermedate between "te Frenc-born" and "non-europeans". Ts group sould be subject to te smlar d cultes as all mmgrants n terms of language or cultural knowledge of France, but ts members are not expected to be dscrmnated aganst out of race and skn colour. Second, we drop from our sample all te ouseolds wose respondent was not lvng n France or was lvng at someone else s place four years before. By dong so, we focus on mmgrants wo are really settled n France and may ave started to ntegrate n te labor market. Eac one of tese two groups of not-too-recent rst-generaton European and non-european mmgrants represents around.5% of te populaton of ouseolds wose respondent ad a place of er own four years before te survey. As sown n Table 8 n Appendx C, non-europeans are over-represented bot n te prvate rental market and n te market of apartments (broadly de ned ere as dwellngs wc sare a buldng wt anoter dwellng). A a consequence, te sare of non-european mmgrants n te populaton of tenants n prvately-rented apartments s twce as g as te sare of Frenc-born or European tenants. Wtn ts sub-populaton of tenants, Table 9 n Appendx C sows tat non-european tenants d er n terms of caracterstcs, wc s terefore mportant to control for n regressons. Non-european tenants are less often women and are less educated, ter ouseold s less rc per consumpton unt, as more members and cldren. Apart from ndvdual caracterstcs, te ENL provdes a lot of nformaton about te caracterstcs of te dwellng. Notably, t ncludes a dummy varable wc ndcates weter te apartment s located n a buldng owned by a sngle landlord or not. Ts varable s nformed by te respondent mself or, f e does not know, by s negbors or by te caretaker of te buldng. 0 Even f ts varable does not allow us to dentfy te cases were te landlord owns part but not te entre buldng, t gves an dea of te magntude of multple ownersp. It regards 0% of prvately-rented apartments. Ts rate s surprsngly g and vares a lot across regons. From now on, a "multple-dwellng landlord" s a landlord wo owns an entre buldng, 0 In te model, we assume tat te applcant does not observe te landlord type. Altoug, n te dataset, te nformaton s provded by te tenant tself, t s probably revealed after te matc s completed. 7

18 wle a "sngle landlord" does not. Table 0 n Appendx C sows tat bot types of apartments are smlar n terms of sze and comfort. However, rents are somewat lower and buldngs are bot older and smaller n case of multple-dwellng landlords. Fnally, multple-dwellng landlords are not randomly allocated across France: tey are fewer n more populated départements and n te départements were sngle-parent famles are more present. All tese features are accounted for n te spec caton we use to test for customer dscrmnaton n te next subsecton. 3.2 Test of te man predcton Predcton states tat tere s evdence of consumer-based dscrmnaton n te rental market f black tenants less often ave a landlord wo owns several apartments wtn te same buldng. In order to test ts predcton, we use te sub-sample of tenants n prvately-rented apartments to estmate a probt model of te probablty to ave a landlord wo owns te entre buldng. We regress ts probablty on a dummy varable wc ndcates weter te respondent s of non-european orgn or not. If te coe cent on ts varable s negatve, tere s consumerbased dscrmnaton accordng to our model. As already mentoned, ts varable of multple ownersp does not dentfy all te ntermedate cases were landlords own several apartments but not te entre buldng. If te latter dscrmnate as muc as te former, te coe cent s based downward. Smlarly, te non-european varable msses many racal mnorty ouseolds wo are not rst-generaton mmgrants. If tese mgrants are as dscrmnated as muc as rst generaton mmgrants, our coe cent s agan based downward. In any case, we ope tese measurement ssues not to a ect too muc our strategy, te man rsk beng a lack of sgn cance due to te somewat narrow categores of dscrmnatng and dscrmnated categores we consder. Table sows tat te margnal e ect of non-european orgn remans sgn cantly negatve, regardless of te spec caton. In partcular, t does not decrease wen we control for every avalable caracterstcs of te tenant, locaton and apartment (column 5): non-europeans reman less lkely to ave a landlord wo owns te entre buldng by.2% ponts. We can nterpret ts negatve margnal e ect n terms of customer dscrmnaton f we are con dent tat te oter varables ncluded n te regresson control adequately for te man d erences n te ousng supply and te marketng process of bot types of landlords. In ts perspectve, t s nterestng to see ow te margnal e ect vares wt te set of controls. and te e ect goes up. Column 2 controls for tenant caracterstcs As explaned before, Non-European ouseolds are poorer and larger, wle ousng unts owned by multple-dwellng landlords are typcally ceaper and contan more rooms. Terefore, non-europeans sould be over-represented n ousng unts owned by a multpledwellng landlord. Wen we control for ter caracterstcs, te e ect s now as g as 8.3% ponts. On te oter and, controllng for département xed e ects and cty sze 2 n columns and 5 reduces te margnal e ect (n absolute value). It s due to te fact tat multple-dwellng Tere are 9 départements n Contnental France. Te département level s rougly comparable to te dstrct level n te US. 2 Cty sze s avalable n nne d erent classes n te ENL, wc s a pror way enoug to control for ts role ere. 8