Paying For Good Neighbours? Neighbourhood Deprivation and the Community Benefits of Education

Transcription

1 ISSN Payng For Good Neghbours? Neghbourhood Deprvaton and the Communty Benefts of Educaton Steve Gbbons October 2001

2 Publshed by Centre for the Economcs of Educaton London School of Economcs and Poltcal Scence Houghton Street London WC2A 2AE Steve Gbbons, submtted September 2001 ISBN Indvdual copy prce: 5 The Centre for the Economcs of Educaton s an ndependent research centre funded by the Department for Educaton and Sklls. The vew expressed n ths work are those of the authors and do not necessarly reflect the vews of the Department for Educaton and Sklls. All errors and omssons reman the authors.

3 Executve Summary It probably comes as no surprse that home-buyers pay more for a property n hgh-ncome, educatonally-rch neghbourhoods than they do for a smlar property n poorer, loweducaton neghbourhoods. Property crme rates may be lower, streets safer, the physcal envronment may be better mantaned. More mportantly for famles, educaton n the communty may matter because of the nfluence ths has on chldren s acquston of educaton and lfe-sklls. These effects nclude drect effects from adults to chldren through expectatons, role models and skll transfers, alongsde peer group effects that operate through nteractons between chldren n the street and at school. Ths study measures the prce premum attracted by hgher-educaton and hgher ncome communtes, usng property prce data from the Government Land Regstry, qualfcatons data from the 1991 Census, and a commercal data set of local ncomes provded by CACI Ltd.. Ths sample gves us near-unversal coverage of property transactons and neghbourhood characterstcs n England and Wales. Our approach s to estmate how property prces change from one neghbourhood to the next as the educatonal status of resdents changes. We can place a common-sense nterpretaton on the change n household expendture on property that proxmty to more educated neghbourhoods generates: t s the value, n monetary terms, that a household places on mprovements n educatonal levels n the communty. Ths nterpretaton has a sound theoretcal bass, and the technque has been used over many years for valung envronmental goods and the physcal attrbutes of property. If we are prepared to assume that t s really the educaton of resdents that matters and our results suggest ths then we can nfer households valuaton of educatonal mprovements n general. Ths leads us to a rough estmate of the local communty benefts of mprovements n educaton, expressed n monetary terms. The man results show that property prces ncrease by one percent n the South and East of England, and by two percent n Wales, the West and North of England, for each one percentage pont shft n the proporton of hgher-educated resdents. Because mean educaton levels dffer across regons, ths amounts to a 0.24 percent ncrease n prces for a one percent relatve change n the educaton of an average communty n any regon. Ths s equvalent to about 156 on 1995 natonal mean prces. House prces move by 0.52 percent for each one percent change n local mean ncomes. Usng these fgures, and takng nto account the emprcal relatonshp between ndvdual earnngs and educaton, we deduce that educaton s valued as a communty commodty for reasons other than ts mpact on ncomes. We fnd further, that educaton n adult resdents matters over and above other communty characterstcs lke unemployment rates, sck rates, lone-parenthood, age, crme rates and local prmary school qualty. Households pay more for communty educatonal mprovements n areas where there are more owner-occuper chldren. Also, the proporton of home-owners wth chldren s hgher n areas where there are fewer socal tenants. From ths, t seems that famles value communty educatonal status as an nfluence on chldren s development and well-beng. We nfer that households pay about 130 per year to purchase a ten percent mprovement on average communty educaton levels from 19% to 20.9% hgher-educated n Ths reflects the long-run, non-earnngs related, communty benefts of educaton. Ths monetary value of ths beneft s at least as large as the estmates of the average prvate returns the ncrement to earnngs arsng from educatonal mprovement whch domnate the lterature. Gven the sze of these effects, the communty and other wder benefts of educaton deserve further analyss. Focussng only on the prvate returns rsks serously

4 understatng the value of educaton to socety, and any polcy decsons based on these returns alone may result n sub-optmal provson of educatonal servces. Note on methodology: A statstcal assocaton between property prces and local educaton levels or local ncomes s not necessarly evdence of wllngness to pay for neghbourhood educatonal status. Hgher-ncome, more educated households wll be clustered n areas wth better qualty housng and local amentes, smply because they can pay more for property than those on lower ncomes. Property prces and the educaton of resdents wll both be hgher n localtes where there s hgh demand for educated workers. To overcome ths problem, ths paper uses two technques. Frstly, we look only at dfferences between neghbourhoods whch are very closely spatally assocated, so mnmsng the geographcal dfferences. Secondly, we predct neghbourhood educaton levels and ncomes from the proporton and characterstcs of resdents n socal housng. These are fxed pror to the perod of our property prce sample, so do not change n response to property prce changes. And, there seems no reason to beleve that owner-occuped housng qualty, prce, and home-buyer ncome vares n response to the local proporton n socal housng except through home-buyers perceptons of the ds-benefts of lvng near lower-educated, lower ncome people. Comparson wth property-level prce data that allows us to compare house-buyers wth smlar ncomes tells us that these assumptons are correct.

5 Payng For Good Neghbours? Neghbourhood Deprvaton and the Communty Benefts of Educaton Steve Gbbons 1. Introducton 1 2. Context Neghbourhood effects The socal and communty benefts of educaton Precedents n the lterature 5 3. The Hedonc Model 7 4. Descrpton of the Data 8 5. Emprcal Methods Emprcal model Estmaton strategy Estmaton of non-lnear responses to neghbourhood composton Identfcaton strategy Results Summary and assessment of the data Implct prce of neghbourhood educatonal status Senstvty to bandwdth choce Non-lneartes n response Implct prce of neghbourhood mean ncomes Comparng local ncome and educaton efforts Unobserved neghbourhood heterogenety Evdence for human captal externaltes Concludng Remarks 26 Tables 29 Fgures 37 Appendces 40 References 46

6 Acknowledgements Steve Gbbons s a member of the Centre for Economc Performance and Unversty College London.

7 1. Introducton Much of the exstng emprcal work on neghbourhood effects focuses on estmaton of the mpact of a chld s neghbourhood on contemporaneous or subsequent outcomes typcally educatonal outcomes. The usual approach s to fnd mcro data on famly and neghbourhood characterstcs n chldhood, and on outcomes for chldren for these famles, and to apply regresson technques to estmate the effects of neghbourhood condtonal on famly characterstcs. 1 One drawback of ths approach s that the mportant neghbourhood and famly characterstcs are often hghly correlated due to spatal resdental sortng attrbutable to preferences, land prces and housng costs. What s more, the long-run mpact of neghbourhoods may be underestmated f the characterstcs of parents are n part attrbutable to hstorcal neghbourhood-drven processes. Measurement of these effects of neghbourhood on human captal accumulaton s crtcal for addressng ssues of equalty of opportunty and the dstrbuton of educaton, earnngs and work across geographcal space. Nevertheless, by concentratng solely on these effects we rsk gnorng other, potentally substantal, economc costs of neghbourhood deprvaton. The obvous example s the cost assocated wth hgher local crme rates n areas where household permanent ncomes and employment expectatons are low. A dfferent strategy, adopted here, s to sde-step measurement of drect effects on ndvdual outcomes by lookng at the overall value onwer-occuper resdents place on good neghbourhoods. The model s a hedonc property prce model of the type frequently used to value local amentes n the urban, envronmental and housng economcs lterature, to estmate the mplct costs of neghbourhood educatonal and ncome deprvaton. In a hedonc equlbrum, ths mplct prce amounts to a margnal valuaton of the servces provded by educatonally rch or hgh ncome neghbourhoods relatve to educatonally poor or low-ncome neghbourhoods. These servces may nclude neghbourhood-related nputs nto the producton of human captal n resdents and ther chldren, drect and ndrect effects of local crme rates, and any other local consumpton and producton externaltes. A number of theoretcal models propose communty sortng equlbra based on household preferences over some measure of the stock of human captal n the neghbourhood, or mean local ncomes. Benabou (1993) assumes spllovers n the producton of chldren s human captal effect ndvduals wllngness to pay for the proporton of hgh human captal communtes, and those wth hgher margnal benefts bd up land rents n hgher human wealth communtes. In Fernandez and Rogerson (1997) hgher ncome communtes have hgher qualty educaton provded through hgher local taxes, and n Neshem (2001) consumers have preferences over the average schoolng of resdents, because ths determnes local school qualty. The model n de Bartolome (1990) proposes sortng drven by wllngness to pay for peer-group effects, though n hs case the equlbrum property prce premum does not reflect the prce of the better peer group, but stops mgraton between communtes n equlbrum. The approach taken n the current work assumes that evdence of a statstcal relatonshp between neghbourhood status and property prces can be nterpreted as average margnal wllngness to pay for neghbourhood effects from educaton, ncomes, or more general neghbourhood qualty externaltes. A central clam of ths paper s that t s educaton and ts wder benefts that count. The results show that local educaton s much more mportant n the determnaton of property prces than local ncomes, and that t remans 1 For a survey see Gephart (1997). 1

8 a sgnfcant factor when other neghbour attrbutes property crme, ethncty, unemployment, lone parenthood and long-term sck rates are taken nto account. If we accept the evdence for ths, then estmates of the mplct prce of local educatonal composton amount to a measure of the margnal, external benefts of educaton n the communty. On ths bass, we can extend the analyss to explore the scale of the local socal benefts of educaton n relaton to the prvate returns typcally estmated n the labour economcs lterature. A key assumpton underlyng ths work s that dfferences n property prces between neghbourhoods that are closely spatally assocated and are otherwse observatonally smlar can be attrbuted to dfferences n the educatonal composton of the neghbourhood. Clearly, any unobserved dfferences across neghbourhoods n ther utlty-bearng attrbutes physcal sze and qualty of housng, access to amentes for example generates observed dfferences n educatonal composton. The quantty of any local normal good s correlated wth local educatonal composton because educaton s a strong predctor of permanent ncome or lfetme wealth. In a regresson of local property prces on local characterstcs, educatonal composton s endogenous unless all utlty-bearng local attrbutes are ncluded n the regresson. To address ths problem, our emprcal approach explots varaton n the proporton of socal housng across neghbourhoods as an nstrument for neghbourhood educatonal composton, and as t turns out more mportantly, explots varaton between neghbourhoods that are closely spatally assocated. Neghbourhood educatonal deprvaton s measured as the proporton of hghly qualfed resdents n postcode sectors n England and Wales, derved from the 1991 Census of Great Brtan, 10% sample. Ths s the proporton of ndvduals wth hgher educaton qualfcatons, but s almost certanly correlated wth local educatonal attanments n general. Although the proporton of hgher-educated adult resdents s farly crude as a measure of educatonal deprvaton, t does an adequate job of charactersng the man dfferences between areas on the dmensons of deprvaton emboded n the DETR Deprvaton Indces Snce hgher local educatonal attanments mean hgher local average ncomes, any assocaton between local educaton levels and property prces generates a correspondng assocaton between local ncomes and property prces. A natonal local area ncomes data set, collected n the late 1990s for marketng purposes, provdes a unque opportunty for nvestgaton of ths relatonshp at ths neghbourhood level. We wll see that ths relatonshp also holds wth property-level transactons data usng mean ncomes at a broader level of geographc aggregaton. The paper s structured as follows. Secton 2 dscusses the background to ths work and sets t n context wth ts underlyng concepts and exstng lterature. Secton 3 outlnes the standard hedonc property value model n the current context. Secton 4 descrbes the data. Secton 5 explans the emprcal methods used. Secton 6 presents the results. Secton 0 concludes wth an assessment of the sze of communty returns n relaton to mean prvate returns per household. 2. Context 2.1. Neghbourhood effects Resdents value neghbourhood educaton levels and neghbourhood ncomes because of the mpact on a wde range of neghbourhood outcomes. Property crme rates may be lower, streets may be safer, the physcal envronment may be better mantaned, gardens more pleasant, behavour more orderly. More mportantly for famles, educaton levels n the 2

9 communty may matter because of spllovers n the producton of human captal n chldren. These spllovers nclude drect effects from adults to kds through expectatons, role models and sklls transfers classfed as collectve socalsaton effects n the socologcal lterature 2. They also nclude peer group effects that operate through nteractons between kds of smlar age n the street and at school. These effects operate to ncrease the expected educatonal attanments of chldren wth hghly educated neghbours, relatve to others. Rather than specfyng all these factors n detal, we may assume that mean educaton levels, or mean ncome provdes a suffcent statstc for the dstrbuton of an unobserved composte neghbourhood good whch s the object of preference (asde from the physcal attrbutes of housng) n the choce of resdental locaton. Let us call the rankng of a neghbourhood on ths scale ts educatonal status. The socologcal lterature and the economcs lterature on educatonal externaltes frequently refers to ths type of composte commodty as socal captal, but as conceved (Coleman, 1988), socal captal descrbes socal nteractons and communty organsatonal structures that are not exclusvely lnked to educatonal attanment of resdents n the communty The socal and communty benefts of educaton Although neghbourhood deprvaton s mult-faceted, the key factors are ncome and educaton. It s well establshed that educatonal attanment s one of the best sngle predctors of long run earnngs and employment. Poor educatonal attanments obvously mean lower expected ncomes for ndvduals and ther famles, but there are also hgh potental external costs. These external costs of neghbourhood deprvaton n educaton mrror the external socal benefts of educaton, whch underpn the prncple of publc subsdy n educatonal provson. The Educaton Reform Act of 1870 whch ntroduced compulsory, publcly funded schoolng to Brtan, was motvated by lberal conceptons of educaton s place n a cvlsed and educated democracy, rather than the need for vocatonal sklls. Nevertheless, most of the emprcal work n labour economcs and the economcs of educaton focuses only on the prvate returns to educaton n the narrowest sense the ncrement to earnngs from addtonal tme n educaton. Others have looked further at socal returns conceved as external effects from human captal on producton, whch ncrease aggregate output. Whlst these are nterestng ssues from the perspectve of polcy drected to mprovng economc performance or addressng nequalty, they say lttle about the value placed by socety on the wder benefts educaton. Prvate returns to educatonal nvestments nclude all the benefts that accrue to the ndvdual who undertakes the educaton and the ndvdual s dependants f we are thnkng of utlty functons at the household level. Some prvate benefts, n partcular ncreased productvty, have well defned markets and are, n prncple, easly measured the ndvdual s wage n the case of productvty. In most emprcal studes, followng early examples by such as Hansen (1963) and Mncer (1974), the prvate returns to educaton are measured as the ncrement to earnngs from addtonal years of schoolng or from dscrete categores of educatonal attanment. Other prvate returns n the labour market nclude effects on employment, job-search, non-wage remuneraton and job-satsfacton, though not all have explct prces. But prvate returns also nclude a wde range of benefts that are not traded n any markets; Haveman and Wolfe (1984) provde a farly exhaustve taxonomy. These non-market effects nclude productvty at home, own-health benefts, the enjoyment 2 Jencks and Mayer (1990). 3

10 value of lesure tme, effects on the educaton and welfare of own offsprng, effects on fertlty, plus the consumpton value of educaton. Socal returns are usually defned as benefts to other members of socety arsng from an ndvdual s school achevements, or partcpaton n hgher educaton. There wll be socal benefts f there are externaltes n producton, whereby the productvty of others s ncreased by assocaton wth more educated workers by more productve work relatons, by drect transfers of knowledge, or where an educated communty stmulates techncal nnovaton. Ths type of model s popular n the growth lterature, followng Lucas (1988). Effects on producton may also operate through externaltes n human captal accumulaton. Indvdual human captal accumulaton may spll over to ncrease the human captal accumulaton of other adults n the communty, and the educatonal attanments of chldren outsde the person s own famly. However, alongsde these benefts whch accrue to socety through ncreased aggregate producton and growth we must consder a catalogue of socal benefts whch are welfare mprovng, but whch may have lttle or no effect on wages, or output. Most of these non-market socal benefts are publc goods that are more or less geographcally localsed: socal coheson, ctzenshp, crme reducton, mproved publc health. These benefts, along wth any productve externaltes n the formaton of human captal, are perhaps better referred to as the communty benefts of educaton. These are the educatonal benefts addressed n ths paper. The clam that educaton or ncome s the key characterstc of nterest to households n the evaluaton of neghbourhood qualty or s a suffcent statstc for neghbourhood qualty provdes a bass for measurng the long-run, socal, communty-based returns to educaton. Ths needs some further justfcaton. A reasonable counter-clam s that educaton merely proxes other behavours of ndvduals whch are unobserved n the data drug abuse, vandalsm, crmnal actvty whch mpose costs on others n the neghbourhood. We must assume that these characterstcs orgnate n lack of educaton and ncome: f these characterstcs are nnate or otherwse fxed pror to educatonal decsons, and an ndvdual s educatonal attanments are determned by these characterstcs, then we cannot nfer the socal returns ths way. One possblty s that parental characterstcs and socal background generate ntal condtons psychologcal or economc whch nhbt an ndvdual s acquston of educaton, or mean that any educaton acqured s valueless n the socal context, even under a supportve polcy regme. In ths settng, educatonal polcy wll have relatvely small effects on educatonal outcomes and wll have few benefts n the short run. Nevertheless, there may be long run effects f even small mprovements n the parents generaton means a better settng for a chld s acquston of educaton. If acquston of educaton s medated solely through genetc or other nnate and unalterable characterstcs, then we cannot nterpret property prce effects that orgnate n preferences for these characterstcs or ther benefts as monetary realsatons of the socal benefts of educaton. Property prces stll reflect the perceved benefts of a neghbourhood cleanup, but the mechansms for achevng ths are not educaton-based. Ths vew mght, for example, fnd support amongst those who consder crmnal behavour as fundamentally nnate, and that lower educatonal attanments amongst partcpants n crme s ndcatve of a preference for crme over legtmate actvty. If the dstrbuton of property prces and educaton levels are related through fear of crme, or the costs of attacks on property, then the mplct prce of educatonal status measures a transformaton of the socal benefts of crme-reducton polcy. Wllngness to pay for hgher-educated or hgh ncome neghbours wll also overstate the communty benefts of educaton f households place value on ther locaton n the dstrbuton of neghbourhood status. If hgh-educaton/hgh-ncome households 4

11 experence no drect costs from lvng amongst low-educaton/low-ncome households, but beneft solely from the status conferred by lvng n relatvely wealthy neghbourhoods, then polces that ncrease educatonal attanments by compressng the dstrbuton may nadvertently generate net socal costs Precedents n the lterature Neghbourhoods and property value Estmaton of neghbourhood ncomes and educatonal status on property prces has a long hstory n the US. Many early studes of the factors affectng property values nclude some neghbourhood characterstcs as covarates, though the response to neghbourhood s not usually the man parameter of nterest. Some examples of early studes n the US lterature that emphasse the role of neghbourhood externaltes on property values follow. Kan and Qugley (1970) estmate that prces of owner-occuped housng ncrease by 7.8%, and rents ncrease by $2.55 for each addtonal year of mean adult educaton n the Census tract, usng a small sample n St. Lous. Berry and Bednarz (1979) found that a $1 ncrease n medan census tract ncomes ncreases the value of sngle famly homes n Chcago by abut $0.70. Both studes condton on a number of neghbourhood and property attrbutes. Freeman (1979) emphasses the mportance of soco-economc and other neghbourhood varables as determnants of property values. A number of studes look specfcally at the effect of socal housng projects and other property development on local property prces. An early example s Schaffer (1972), who looks at the mpact of housng constructon for low ncome famles under the US 1961 Below Market Interest Rate scheme usng treatment and control stes n Los Angeles. He fnds no sgnfcant dfference between the prce trends at the two stes, probably due to the fact that most of the new resdents already lved locally. Dng, Smons et al. (2000) are more concerned wth the mpact of local resdental nvestment on property prces. Usng data on Cleveland, Oho, they fnd a $0.87 ncrease n property prces wth each $1 of medan census tract ncome (n 1990) correspondng to an elastcty of 0.36 at the sample mean. Ther estmates also mply an elastcty of 0.04 wth respect to the census tract proporton of Afrcan-Amercans. They also report a negatve, but nsgnfcant effect from the proporton n poverty. Crme rates attract a strong negatve coeffcent, correspondng to an elastcty of Munneke and Slawson (1999) are nterested n potental negatve externaltes from moble home parks n one parsh n Lousana and estmate a two-step selectvty model to adjust for the endogenety of moble home park locaton. Locaton wthn 0.25 mles of a moble home park n a resdental area leads to a 5% declne n the value of a sngle famly dwellng, relatve to propertes located between 0.25 and 0.5 mles radus. They offer no theory as to the cause of ths externalty, but the perceved type of moble home resdents s presumably a key ssue. Two other recent studes nvestgate the mpact of socal housng programs n the US. Lee, Culhane et al. (1999) consder the effect of publc and asssted housng on property values n Phladelpha. The authors of the frst paper fnd negatve effects from proxmty to publc housng developments and other asssted housng schemes, but these effects largely dsappear, or ther sgn s reversed once neghbourhood composton controls are ncluded. They fnd no statstcally sgnfcant effects from the physcal type of development, whch suggests that t s the characterstcs of resdents and not the physcal structure of socal housng that generates the externalty. Log property prces ncrease by 1.6% for each thousand dollars of neghbourhood medan ncomes an elastcty of Galster, Tatan et al. (1999) look at the mpact on property prces of neghbours n recept of 5

12 Secton 8 certfcates, whch enttle low-ncome households to a housng subsdy. They use a model wth spatal fxed neghbourhood effects to fnd heterogeneous mpacts from asssted housng programs n Baltmore, wth adverse effects n lower prce areas, but postve mpacts from small-scale programs n hgher valued tracts. Interestngly, the authors conducted focus group studes n four communtes wth dstnct soco-economc compostons, to gauge resdents opnons of socal housng developments. Some respondents expressed senstvty to the physcal condton of rental accommodaton, wth a fear that asssted housng brought physcal decay and vandalsm. Many groups expressed clear antpathy to problem tenants, belevng that those n socally asssted housng had dfferent values and standards than what the current resdents desred for ther neghbourhood. Many feared that subsdsed housng brought ncreased crme. Few studes on the value of neghbourhoods exst for Brtan due to the lack of data. One example s Cheshre and Sheppard (1995), who fnd postve amenty values n Readng from local schools, the proporton of whte collar workers and the proporton n non- Afro-Carbbean ethnc groups. They estmate aggregate land values (over the geographcal space of ther sample) of 43,430 attrbutable to schools, and 81,820 attrbutable to socal and ethnc composton, but offer no estmate of the mean benefts per household The socal and non-market returns to educaton No exstng studes propose a lnk between the wllngness to pay for good neghbourhoods and the measurement of communty benefts of educaton. However, there are a number of approaches to measurng other benefts beyond the tradtonal prvate market returns on earnngs. Snce Lucas (1988), who dscussed the potental role of human captal externaltes n economc growth, a strand of emprcal research has emerged whch has tred to measure the mpact of state, regon or country average educaton levels on wages, productvty and growth. A few examples wll gve the flavour of ths research programme. Weale (1992) uses prvate returns and nternatonal comparsons of growth rates and educatonal attanments to suggest that long run socal returns ncorporatng spllover effects on growth rates could be two to three tmes the magntude of the prvate returns. Jaffe (1989) looks at the socal rate of return to unversty research n the form of state-specfc spllovers nto corporate patents, and fnds postve effects wth elastctes as hgh as 0.3 n some ndustres. Acemoglu and Angrst (1999) fnd strong effects on wages from state educaton levels, condtonal on ndvdual educaton usng OLS estmates on US Census data, but these socal returns become weak and nsgnfcant once they nstrument the educatonal varables wth state compulsory school attendance laws and ndvdual date of brth effects. Cccone and Per (2000) fnd negatve effects from cty educaton levels on ndvdual wages usng data from 173 US ctes n 1970, 1980 and Usng data on average wages n ctes, they fnd nsgnfcant, near-zero, effects on wages, but small postve effects on productvty of around 1%. More drectly related to the work n hand s Haveman and Wolfe (1984), who present a meta-analyss of earler work to compute an approxmate fgure for the annual value of an addtonal year of schoolng based on non-marketed effects on the producton of chldren s cogntve development, contraceptve use, effcent budget allocatons, crmnal apprehenson and health. Ther technque s based on obtanng the shadow prce of the nonmarketed nput from the rato of ts margnal product to that of another, marketed nput wth a known prce. Ther fgures suggest a value of socal and prvate non-marketed benefts n the order of $5000 n 1975 a value of a smlar order to the annual value of a year of schoolng n standard prvate rate-of-return estmates. 6

13 3. The Hedonc Model We shall use a standard hedonc property value framework to assess the mplct prce of neghbourhood educatonal and ncome composton. Ths framework has been employed frequently n the envronmental, land and urban economcs lterature to prce local envronmental amentes. Indvduals are assumed to have weakly separable preferences over a set of housng and locaton characterstcs. A dwellng comprses a bundle of these attrbutes. Sellers and buyers wth dfferent ncomes and dfferent preferences over local school performance and other property characterstcs are matched effcently by the property market. Ths leads to an mplct prce surface that traces out the locus of effcent transactons n prce-characterstcs space. See Rosen (1974) for the classc exposton. Followng the standard hedonc, property value model, we specfy household preferences as: c ( c, x, y ( x),q, l) U = U (1) where c s a numerare composte consumpton commodty, x s the measure of neghbourhood status ether the neghbourhood proporton hgher-educated or log mean neghbourhood ncomes y c ( x) s a human captal producton functon for chldren n the household, q s a vector of structural housng characterstcs, l s a vector of locatonal characterstcs. House prces are determned as a functon of the same attrbutes, where the attrbutes are traded at a set of exogenous prces θ fxed by demand and supply equlbrum at a broader geographcal level: ( x, q, l;θ) P = h Ph (2) The household lfetme budget constrant s: ( x, q, l;θ) y = c + P (3) h Assumng the choce space s contnuous so that households can purchase ther optmum bundle the frst order condton for x s: U x + U U y c c y c x Ph = x (4) Ths standard condton justfes the use of an estmated mplct prce functon P h () n the estmate of the margnal wllngness to pay for local educatonal status. If consumers are heterogeneous n ther margnal benefts from x then stratfcaton nto hgh and low x communtes can occur, and P h () can be non-lnear n x. Wthout nformaton on human captal of chldren, t s not possble to dentfy separate contrbutons of educatonal status to human captal formaton and consumpton value n the mplct prce, only the sum of the margnal benefts. But, gven an approprate specfcaton of P h and ndvdual level data on house prces, neghbourhood, housng and locatonal characterstcs, t s possble to estmate the overall wllngness to pay for margnal mprovements n neghbourhood educatonal status or local mean ncomes, for gven neghbourhood qualty x. 7

14 In the case where the communty s valued purely as an nput nto the producton of chldren s human captal and lfetme wealth, margnal wllngness of parents to pay for x wll be the margnal effect on the present value of ther chldren s total future earnngs. Because of ts local publc good nature, famles wth more chldren are wllng to pay more for neghbourhoods wth hgh levels of our commodty. Famles wll bd up the prce of mprovements n the neghbourhood untl the margnal cost equals the sum of the margnal benefts over all ther chldren. The alternatve s to dvde expendtures on prvate goods that mprove the welfare of chldren, or to dstrbute ncome drectly to chldren n the form of transfers. An mplcaton of ths s that the average number of chldren per household, or the proporton of households wth chldren n a neghbourhood, wll be ncreasng as neghbourhood status ncreases exogenously. It follows that the mplct prce of neghbourhood qualty wll ncrease as the mean number of chldren per famly ncreases. 4. Descrpton of the Data Snce the emprcal methods are desgned to sut the data, t wll help to descrbe the data frst. Brtsh data on ndvdual property transactons wth local area dentfers s not readly avalable. Instead, we must use locally aggregated data avalable from the Government Land Regstry. Ths covers most market value property transactons n England and Wales, aggregated to postcode sector level. In the UK, postcodes contan up to seven alphanumerc characters, and contan four herarchcal components. The frst two alphabetc characters defne the postcode area, the broadest postal zone. Examples are N, EX YO representng North London, Exeter, York. Wthn postcode areas, the next level down s the postcode dstrct. A sngle or two-dgt number followng the postcode area defnes ths. Examples are N6, EX24, and YO10. A sngle letter further subdvdes some postcode dstrcts n central London. Below ths, we have postcode sectors. Ths s the unt of observaton n our house prce data set, and the unt adopted here as a neghbourhood dentfer. Ths Land Regstry data s dsaggregated by property type detached, semdetached, terraced, flat/masonette but ths s the only nformaton provded about the characterstcs of propertes ncluded n the prce data. At the tme of wrtng, ths data set s avalable from 1995 to It contans mean house prces and total sales volumes for each dwellng type n each postcode sector, where annual sales numbered 3 or more. Propertes under 10,000 and over 1,000,000 are excluded. Ths amounted to only 0.5% of all property sales n Sales at non-market value transactons are also excluded. Ths s an advantage n our applcaton, because market prces wll not be contamnated by dscounted sales of councl houses to tenants under the Rght to Buy scheme ntroduced n the 1980s. Mcro-spatally aggregated data has an advantage over property level data n the current applcaton because we are only nterested n the varaton n prces attrbutable to mean neghbourhood characterstcs. What we do need though are property prces and neghbourhood attrbutes at the same level of dsaggregaton. Sources of nformaton on educatonal deprvaton n local areas are lmted. The most up to date data s combned nto the Educaton, Sklls and Tranng doman of the DETR Indces of Local Deprvaton Ths ndex s generated at census ward level, not postcode sector level and there s no correspondence between these two geographes. An alternatve source s the 1991 Census 3 Ths ndex combnes a number of dmensons of educatonal deprvaton workng age adults wthout jobs, over-16s not n full tme educaton (DSS), applcants for hgher educaton (UCAS), prmary school performance, chldren wth Englsh as an addtonal language, and prmary school absenteesm. 8

15 Small Area Statstcs. Ths provdes a count of the number of over-18s wth degrees, dplomas and other hgh qualfcatons, based on a 10% sub-sample of the census populaton. 4 The age of the census data and the 10% samplng scheme are a drawback, but n compensaton we have a straghtforward nterpretaton of the relatonshp between property prces and the proporton of hghly qualfed adults, and the fact that re-aggregaton to postcode-sector level rates s farly straghtforward. The 1991 census also provdes an accurate measure of the proporton of households n socal housng requred as an nstrument for educatonal status plus grd references, housng and varous other local characterstcs. For the central results presented n ths paper, I match the Land Regstry property prce data to 1991 census data, re-aggregated from Enumeraton Dstrct level to postcode sector level usng the Postcode-Enumeraton Dstrct lookup tables avalable from the Census Dssemnaton Unt. Local ncome data for Brtan s also scarce. For ths we must turn to a commercal data set produced by CACI Ltd. for marketng purposes. Ther survey s avalable for 1996, 1999 and 2000 (though only the 1996 and 1999 surveys are used here) and each wave s based on over 4 mllon households. Incomes are modelled by CACI down to ndvdual postcode level usng 1991 census data. A postcode sector comprses 2700 households on average. In 1999, the mean number of actual observatons of ncomes used for the postcode sector mean (not mputed) s 436 wth a sample mean ncome of 21,860 and a standard devaton of household ncomes of 15,000. The standard error of the postcode sector mean would be around 720, or 3.3% of the overall mean. Ths gves some confdence that the data, although partally mputed, s relable at postcode sector level. 5 More data on property values s avalable from the Survey of Mortgage Lenders, an annual 5% sample (around 25000) of mortgage transactons. Ths has the advantage of property level prces, dwellng characterstcs and total household ncomes, but the dsadvantage of broader, Local Authorty geographcal dentfers. Stll, we can use estmates from ths data set for comparson wth the baselne results. 5. Emprcal Methods 5.1. Emprcal model As dscussed above, our property prce data set has no household level data. Instead, n the Land Regstry data set, annual housng transactons are aggregated to provde an average of prces n four property-type categores at postcode sector level. A sample of k transactons on a house of type r wll contan a mx of structural characterstcs q. Assumng that the sample value of the prce of houses of type r, n postcode sector, at tme t, wth mean characterstcs q s the market prce of a representatve property, then the household hedonc prce functon s representable by a hedonc prce functon at the neghbourhood-house-type level. There s 4 The relevant queston s number 19 n the Census form, whch asks for detals of post-compulsory-age educatonal qualfcatons of all persons over the age of 18 n the household on Census nght (21-22 Aprl 1991). The Census Small Area Statstcs contan counts of persons wth hgher degrees, degrees, or dplomas, nursng or teachng qualfcatons, based on a random sample of 10% of the responses from each Census Enumeraton Dstrct (a much smaller geographcal unt than the postcode sector). 5 As a safety check, we can compare the CACI data wth New Earnngs Survey data for 1995/1996. Unfortunately the NES records postcode of employment, not resdence, and ncludes earnngs only, not famly ncomes, so we would only expect moderate correlaton. Aggregatng to postcode dstrct level, we fnd a correlaton coeffcent of At postcode area level, ths ncreases to

16 no detaled nformaton on structural characterstcs for our sample of house transactons. One opton s to proxy the characterstcs wth census data on owner occuped housng n The only census varables whch could reasonably be treated as exogenous not, lke housng amentes, subject to change n response to shfts n resdental composton are those that gve the dstrbuton of rooms across households n the postcode sector. In the tradton of the property value lterature, the models presented here use the natural logarthm of property prces as the dependent varable. An unknown functon g( l,t) maps locatonal characterstcs to house prce n each tme perod. Ths specfcaton obvously mposes the constrant of a constant percentage response n house prces to a one percentage pont absolute ncrease n the proporton of qualfed resdents, or a constant elastcty wth respect to local ncomes. Enterng educaton lnearly, and ncome n natural logs, s consstent wth the usual Mncer-type earnngs functon whch specfes log ncomes as lnear n educatonal attanments. The specfcaton of the log-prce of a house of type r n neghbourhood at tme t s then: rt ( l t) + hr urt ln P =, + α + β x xt + g (5) where there are fxed effects for the four housng types h r Estmaton strategy Estmaton of a full structural specfcaton of the mappng of neghbourhood characterstcs l to house prces requres data on local amentes, local housng characterstcs, the proxmty of neghbourhoods to transport servces, local labour demand, envronmental qualty and other unknown local goods. The functon g( l,t) could then be replaced by a specfc functon of avalable covarates. Ths s the tradtonal method used n property value models, wth adhoc ncluson of a broad, though potentally ncomplete, set of explanatory varables. In the absence of ths data and any pror knowledge about exactly what should be ncluded, we can replace the mappng of l to house prces wth some specfcaton that maps neghbourhood to house prces through the locaton of the neghbourhood n geographcal space and tme. The approach adopted here s to estmate g( l,t) as a spatal fxed effect usng a nonparametrc kernel regresson procedure. Ths estmates the average value of property prces and the regressors at a postcode sector, as a dstance-weghted average of the values n the surroundng postcode sectors. Estmaton s then based on lnear regresson usng the devatons of the observed values of property prces and the regressors from the estmated expected value surfaces over geographcal space. The grd reference co-ordnates of a geometrc central pont n postcode sector determne the spatal locaton of an observaton and the weghts that should be appled to other observatons. Observatons n closer proxmty receve the hghest weghts. The advantage of ths method over, say, usng dummy varables for groups of postcode sectors n close proxmty, s that t centres the comparson group on the observaton postcode sector and allows flexblty n the choce of group radus. Ths amounts to decdng on a bandwdth b for the kernel, whch wll determne how rapdly the weghts decrease as we move away n space from a gven neghbourhood observaton. Where there s more than one year of data, we can allow tme effects va a separate non-parametrc surface for each perod, so: ( t) = d g ( l ) g l, (6) t t t 10

17 where d t s a tme dummy. Ths allows for dfferental growth n house prces across geographcal space. Ths smooth spatal effects estmator (SSE) s also used n Gbbons and Machn (2001) for prcng prmary school performance. Expressng the model n devaton from estmated expected values, gven the spatal locaton, c1, c2 and the choce of bandwdth b for the comparson group: [ ln Prt m( ln Prt, b) ] = β[ xt m( xt c, b) ] + [ w rt m( w rt c, b) ] + ϖ rt c γ (7) The locatonal mean of a varable y, m ( y c, b) Watson estmator m ( y c, b) = N y N k 1 k{ ( c c ) Β ( c c )} 1 {( c c ) Β ( c c )} s estmated by the bvarate Nadaraya- 2 where Β s a 2 2 bandwdth matrx, e.g. b I 2, and k{} s a multvarate kernel. For the 1 Gaussan kernel ths s k {} v = 2π exp{ 0. 5v}. Parameters β, γ and ther varance covarance matrx can then be estmated by OLS on the transformed varables. As the bandwdth b tends towards zero, the estmator approaches an estmator wth postcode sector fxed effects. Snce we have no tme-seres varaton n the Census educaton measure, and very lttle n our ncomes data (98.3% of the varaton s cross-sectonal), ths s napproprate. At the other extreme, an nfnte bandwdth s equvalent to the OLS estmator, wth the functon g( l ) estmatng a constant. Snce there s enormous varaton n postcode sector land areas and household densty, a common bandwdth for all observatons wll lead to nconsstent estmates. The estmated prce and regressor surfaces wll be over-smoothed n areas of hgh household densty and low land area postcode sectors, and under-smoothed n rural areas. Consequently, we must weght the neghbourhood bandwdth usng data on household densty matched n from the 1991 Census. Fxng the number of households n n a crcular spatal group of radus b, gves us a bandwdth weghtng rule dependent on housng densty h: (8) b = n πh (9) The baselne results n the paper use bandwdths correspondng to 3400 households, but comparsons are made wth other bandwdth choces 6. Senstvty to bandwdth choce can be tested by the usual Hausman test for equvalence of parameters n alternatve estmators. Too narrow a bandwdth gves a consstent but neffcent estmator; too wde a bandwdth results n nconsstent estmates. 6 Hardle (1990) presents bandwdth adjustment factors for comparng smoothers usng dfferent kernels. The 43% downward adjustment s based on achevng smlar smoothng to a unform kernel wth unt bandwdth,.e. f we want equvalent smoothng to a unform kernel of radus equal, on average, to two postcode sectors (6000 households) we need a Gaussan bandwdth of 0.57*6000 = 3420 households. 11

18 As wll be dscussed n Secton 5.4, we need nstruments to dentfy the mplct prces. In IV estmaton, nstruments are taken as devatons from the estmated local means. The fnal partal lnear smooth spatal fxed effects IV estmator s: ˆ β 1 ~ ~ 1 ~ ~ ~ ~ ~ ~ 1 ( Ωˆ Z) Z X X Z( Z ˆ Z) Z ~~ p ~ ~ = X Z Z Ω SSE IV (10) where X s the regressor matrx, Z s the full nstrument matrx and p the house-prce vector. The tlde ndcates devatons from the non-parametrc estmates of the smoothed surface ~ means n the smooth spatal effect models. I estmate the matrx Z ˆ Z ~ Ω usng the Huber- Whte method, wth clusterng on postcode sectors to allow for the fact that we have multple house type (and sometmes tme perods) n each postcode sector. The varance covarance matrx s estmated by the nverse of the frst term n square brackets Estmaton of non-lnear responses to neghbourhood composton The model n (5) mposes a log-lnear relatonshp between property prces and the proporton of hghly qualfed neghbours. Some emprcal verfcaton of ths assumpton s n order. Non-lneartes n the mplct prce functon wll have mplcatons for evaluaton of the aggregate socal benefts of an ncrease n educatonal attanments. Evdence of nonlneartes may enrch the emprcal analyss by revealng threshold effects n the sprt of contagon theores of neghbourhood deprvaton whch have been popular n the socologcal lterature see Jencks and Mayer (1990) or Crane (1991). If contagon or epdemc theores are correct then home-owners should be ndfferent to neghbourhood educatonal composton untl educaton levels fall below some crtcal threshold. In order to check for such non-lneartes we must generalse the sem-parametrc estmaton procedure above, to estmate: r ( x, c, c, b ) + γ w ur ln P = g 1 2 r + (11) Vsual representaton of the relatonshp between ln P and x s nfeasble n the general case. h can be estmated as the average relatonshp between expected prces Instead, a functon ( x) and x, averagng over the dstrbuton of spatal co-ordnates,.e. ( x b) = g( x, c1, c2, b) f ( c1, c2 ) h dc1dc2 C, (12) where C s the support of ( c1, c2 ) n the sample, and f ( c c2 ) computatonal procedure s as follows: 1, s ther jont densty. The 1. Estmate the lnear coeffcents γ n the model (13) below, replacng m () wth a 3- regressor kernel regresson estmates wth observaton-dependent bandwdths b for c,c 1 2, and a fxed bandwdth for x : 7 I compare wth bootstrap standard error estmates n one case to assess the accuracy of the standard errors. 12

19 [ ln Prt m( ln Prt c, x, b) ] = γ [ w rt m( w rt c, x, b) ] + ϖ rt (13) 2. Estmate h ˆ m ( x) by kernel regresson of ( ln γˆw ) r r ponts x&, all at a fxed co-ordnate-par (& c ) g random from the sample. 3. Re-calculate ( x) h m 4. Calculate hˆ ( x) = ( ) P on x, c1, c2 at a number of grd c1 m, & 2m. Ths co-ordnate par s drawn at ˆ at M dfferent co-ordnate pars drawn at random from the sample. m = M 1 h x m, that s the average of the M kernel regressons over the subsample of randomly chosen, wthn-sample, spatal locatons at whch the estmated jont M m= 1 densty f ( x c, c ) 0., 1 2 Snce n ths applcaton, educatonal composton x s treated as endogenous (see Secton 5.4 below), t s replaced by the generated regressor: ( z, c, c, b ) ˆ hˆ 1 + w + r xˆ = xˆ 2 γ (14) where z s a sutable nstrument, x ˆ( z, c1, c2, b ) s estmated by kernel regresson, and the other parameters are estmated usng the partal lnear model descrbed above. If h ˆ ( xˆ ) s to be a good representaton of ĥ ( x), we requre that the nstrument s contnuous, s a strong predctor of x, and that xˆ has a smlar support to x. As dscussed n Secton 5.4, the proporton of socal tenants s the man nstrument, and ths satsfes these requrements. Rlstone (1996) dscusses the use of generated regressors n non-parametrc estmators, and ther asymptotc propertes Identfcaton strategy Incluson of any household personal characterstcs on the rght hand sde of a hedonc model causes problems. The ablty of households to move across space mples that household characterstcs are almost certanly endogenous n a property prce equaton varaton n unobserved determnants of property prces drves the varaton n characterstcs of resdents. Interpretaton of estmated coeffcents on household or ndvdual characterstcs s dffcult, and ther ncluson can lead to based and nconsstent estmates of the parameters of nterest. Ths s partcularly true of characterstcs that are hghly correlated wth household ncome, n that lower land prces wll attract those wth lower ncomes. An analogous problem arses f we nclude communty characterstcs that are correlated wth home-owner household ncomes. The household level relatonshp between ncomes and housng expendture s replcated at an aggregate level f we nclude local mean ncomes, or other local demographcs that are correlated wth household ncomes. A 8 Note that no adjustment s made to the confdence ntervals to take account of ths generated regressor. Snce the asymptotc dstrbuton of ths sem-parametrc estmator s complex, I compare the analytcal standard errors wth bootstrap estmates for one case see Appendx C. The addtonal computatonal effort requred to compute bootstrap standard errors for all the estmates s hardly worthwhle, snce the man purpose of the h x. exercse s to show up any non-lneartes n ( ) 13

20 regresson of property prces on local mean ncome gves us a parameter estmate whch measures the response of property expendture to own ncome, rather than home-owners valuaton of local ncomes as a commodty. Smlarly, neghbourhood educaton levels wll be hghly correlated wth own wealth or permanent components of ncome. The structure of the problem s common to all endogenous regressor models. The relatonshp of nterest s: ( t) + hr + ρν t ε rt ln Prt = α + β x xt + g l, + (15) Where ν t s the component of neghbourhood choce whch s observed to property buyers, but unobserved to the econometrcan, and ε t represents components of property prce formaton whch are unobserved to both optmsaton errors, local estate agent actvtes for example. But neghbourhood status x s partly determned by mgraton of home-owners between neghbourhoods, because of selecton on unobserved components n the determnaton of property prces, underlyng land prces or structural dfferences for example. x t ( t) + ν t ξ t = φ + λ zt + g l, + (16) Hence, E[ ξ x ] 0 ρν t + tr t and regresson estmates are based. Identfcaton of the mplct prce of a local amenty whch s not exogenous to other determnants of resdental land and buldng values requres one of two strateges. Frstly, we can saturate the model wth property descrptors and exogenous local characterstcs, n the hope that unobserved determnants of property prces are purely random, and not drven by unobserved housng characterstcs or local amentes,.e. E[ ρν t + ξ tr xt ] = 0, so ν t = 0. Ths s the method mplctly adopted by most researchers n the housng economcs feld, hence the extended vector of covarates presented n many property-value models. Ths type of estmator tres to acheve condtonal ndependence of the error term and the characterstc of nterest 9. The weakness of ths approach s that we need a lot of property characterstcs and, n the absence of any spatal controls, a fully specfed model of local determnants of land prce dstance to the central busness dstrct, dstance to modes of transport, dstance to other local amentes. Snce property characterstcs exhbt a hgh degree of mutual correlaton, nterpretaton of the parameters can be dffcult. What s more, the determnants of property prces whch are left unobserved to the researcher must also be unobserved to property buyers, or consdered rrelevant, f they are to be truly exogenous to ncomes. Expendture on any attrbutes of the local envronment or physcal structure of local propertes, whch are normal goods and are observed by buyers, wll be postvely correlated wth ncomes. Estmaton on dfferences from local expected values partly overcomes the dentfcaton problem by removng most of the varaton attrbutable to local labour markets, local envronmental goods, and transport servces. Ths assumes that ν t s subsumed n g( l,t). A more robust approach s to combne ths estmaton strategy wth nstrumental varables for the local characterstc of nterest, usng some columns of the vector z t. In the current context, we need local characterstcs whch are correlated wth local educatonal composton or local ncomes, but whch affect the educaton and ncomes of home buyers only through ther nfluence on local educaton or ncome, valued as an amenty. Canddate 9 Ths tradtonal regresson approach s analogous to matchng estmators of treatment effects, but wth restrctons on functonal form. 14

21 nstruments are the proporton and characterstcs of households n socal housng see Appendx A. The dentfyng assumpton s that the ncomes of home-owners are locally uncorrelated wth the proporton, ncomes and educaton of tenants n socal housng except n so far as the presence of low-ncome, low educaton tenants n socal housng generates an externalty that home buyers are wllng to pay to avod. The presence of captal market constrants dctates that home-owner ncomes wll also be lower n areas wth hgh proportons of socal housng, but only because of the nfluence of the neghbourhood on property prces. Although the proporton of socal tenants s a satsfactory nstrument for educatonal attanments and local ncomes, we would prefer some over-dentfcaton, f only to allow a test of the specfcaton. Any characterstcs of socal tenants whch are correlated wth educatonal attanments and ncomes (condtonal on the proporton of socal tenants) wll suffce. As t turns out, a good addtonal nstrument s the proporton of socal tenants from ethnc groups orgnatng n the Indan sub-contnent. These groups have, on average, hgher qualfcatons than socal tenants from other ethnc groups, but the proporton of these groups s uncorrelated wth property prces once we control for local ethnc composton n the property prce equaton. 6. Results 6.1 Summary and assessment of the data Property prce data The results n ths paper are presented separately for three broad geographcal regons of England and Wales. These regons correspond to grouped Standard Statstcal Regons: East and South East: Wales, West and South West: The North: London, South East (rest), East Angla West Mdlands, South West, Wales East Mdlands, Yorkshre and Humbersde, North, North West Ths scheme separates areas wth wdely dfferng property market characterstcs, but retans a mx of rural, urban and metropoltan geographes n each group. Ths enables nvestgaton of dfferences across regons, wthout over complcatng the presentaton of results. Roughly speakng, these areas are grouped accordng to property prce growth n the late 1990s. Table 1 summarses the man varables n our data set. The property prce sample ncludes only those propertes wth recorded postcodes. Ths sub-sample under represents hgher prce propertes n 1996 when compared wth the full sample used by the Land Regstry or the random 5% sample conducted by the Socety of Mortgage Lenders. The postcode sector data under-represents hgher prced detached houses and flats n all regons, probably because t under represents new hgh-end propertes. The Land Regstry confrmed that many new propertes are regstered wthout postcodes, so are mssng from the postcode sector level data. The censorng of these groups n the dependent varable has the potental to downward bas our regresson estmates. Gven that the dfference between the means n the postcode sample and the full sample s only around 5% ths should not be a serous problem. 15

22 Neghbourhood educaton data The Indces of Deprvaton 2000 publshed by the DETR provde the most up to date ndcators of deprvaton n educaton, ncome, employment, health and housng, chld poverty, and access to servces. The ward level bass of these ndces s not compatble wth our property prce data at postcode sector level, but we can compare the ndces wth our Census educaton measure at ward level. Appendx B shows that the educatonal composton varable from the 1991 Census s moderately correlated wth the current deprvaton measures, and does a good job of predctng educatonal deprvaton when nstrumented wth the proporton n socal housng. Is there any drect evdence that resdents prefer more educated neghbourhoods? Regressng a ward-level ndcator 10 of neghbourhood dssatsfacton on the ward proporton wth hgh qualfcatons suggests a 0.01% (s.e. = %) decrease n the proporton expressng dssatsfacton as the proporton of hghly qualfed resdents ncreases by 1% an elastcty of at the mean. Ths result s unchanged f we nclude other key census varables the proportons professonal, unsklled, unemployed, non-whte, lackng housng amentes, n socal housng, n agrcultural employment, plus household densty and average property sze. The ward proporton wth hgh qualfcatons s the only statstcally sgnfcant coeffcent (at the 5% level) n ths regresson 11. Admttedly, the magntude of the effect s small usng ths data, but educatonal composton seems to be one of the stronger canddates amongst local factors for a contrbutor to resdents self-reported perceptons of satsfacton wth ther neghbourhood Assessng the nstruments As dscussed n Secton 5.4, dentfcaton of the mplct prce of neghbourhood educaton or neghbourhood ncomes requres an nstrumental varables approach. The man proposed nstrument s the postcode sector proporton of households n socal housng. Some overdentfcaton s obtaned by ncludng the proporton of these socal tenants n ethnc groups orgnatng from the Indan subcontnent. Both are hghly sgnfcant n regonal wthn-area 12 regressons of educatonal composton on the exogenous varables and nstruments, wth an F- statstc of 240 for Wales and the South West, 392 for the South and East, and 321 for the North of England. As the proporton n socal housng n 1991 ncreases by 1%, the postcode sector proporton of all resdents wth dplomas, degrees and above decreases: by 0.28% n the South and East, by 0.17% n the North, and by 0.22% n the West. The wthn-area R 2 s are 0.32, 0.41 and 0.38 respectvely, so the proporton n socal housng explans a consderable proporton of the local varaton n qualfcatons. Educatonal attanments ncrease wth the proporton of socal tenants from Indan sub-contnent ethnc groups, except n the Northern regons where the relatonshp s negatve. Appendx B presents evdence on the proportons of these tenancy and ethnc groups wth hgher educaton qualfcatons. Our dentfyng assumpton s that educaton and ncomes of home-buyers and socal tenants are locally uncorrelated, except through the nfluence of the proporton of loweducaton/low-ncome socal tenants on property prces. Obvously ths wll not be true over larger areas, n whch case dfferences n labour market opportuntes and earnngs wll affect 10 Burrows and Rhodes (1998) combne Survey of Englsh Housng and 1991 Census data to model the geographcal dstrbuton of neghbourhood dssatsfacton n terms of the percentage of households n each ward who say they are very dssatsfed wth ther neghbourhood. 11 It s not one of the characterstcs used to model the dssatsfacton varable. 12 Where area s defned as the postcode dstrct. 16

23 home-owners and socal tenants jontly. Estmaton wthn localsed geographcal groups ensures exogenety of the nstruments. Estmates from the 1994 to 1998 Survey of Englsh Housng show that mean ncomes of neghbourng socal tenants and property owners are uncorrelated wthn Local Authorty areas. The coeffcent n a regresson of ward mean socal tenant ncomes on owner occuper ncomes s (s.e ). Incomes of prvate tenants and property owners are moderately but sgnfcantly correlated wth a regresson coeffcent of 0.36 (s.e ). A further check s avalable usng data on new household mortgages n the Survey of Mortgage Lenders data. Regressng the log of ncome (on whch the mortgage s based) on the proporton n socal housng, wth local educaton and county controls, gves us coeffcents of (0.142) n the North, (1.132) n the South East and East, and (0.309) n the South West and West. The statstcal nsgnfcance of the coeffcents s clear ndcaton that the ntrument and home-purchaser ncomes are condtonally uncorrelated. All the IV models nclude a Ch-squared test (Sargan test) of model specfcaton and the valdty of the overdentfyng restrctons Changes between 1991 and 1995 Complete data on postcode-sector property prces s only avalable from the Land Regstry snce The census data on tenancy groups and local qualfcatons dates from Estmates of a property prce model usng 1995 prces on 1991 area characterstcs wll be based f there have been substantal changes n area characterstcs snce Clearly, f the postcode-sector proporton hghly qualfed n 1995 s a multple of the 1991 value, then we need to adjust the coeffcent on the 1991 proporton downwards, though elastctes wll be unchanged. However, changes n the dstrbuton of educaton across postcode sectors wll result n attenuaton of regresson coeffcents and estmated elastctes. Comparng 1991 and 2000 deprvaton ndces at ward level, suggests that the attenuaton due to dstrbutonal changes s probably n the order of 8% see Appendx B for detals of ths calculaton. 6.2 Implct prce of neghbourhood educatonal status Postcode sector data Table 2 to Table 4 presents the central estmates from the smoothed-spatal effects estmators, by three broad geographc regons. In each table, column 1 presents a basc OLS regresson of log mean property prces n postcode-sector-dwellng-type cells n 1995, on the proporton of hghly qualfed adults n the postcode sector n 1991, dwellng type dummes, mean rooms n owner-occuped housng. Column 2 adds controls for ethnc composton and the densty of purpose bult flats. Columns 3 to 6 show estmated parameters and standard errors from the smoothed-spatal effects estmator. In column 4 and column 6 the postcode sector proporton wth hgh qualfcatons s nstrumented. Instruments are the proporton of households n socal housng (councl and local authorty tenants) and the proporton of these socal tenants from Indan, Pakstan and Bangladesh ethnc groups. All the tables show estmated coeffcents and standard errors 13. To llustrate our emprcal method, Fgure 1 and Fgure 2 show the raw and smoothed proporton wth hgher educaton qualfcatons n the London area. Fgure 2 13 Standard errors for the smooth spatal effect models were checked aganst bootstrap standard errors for a London sub-sample. For a coeffcent of the standard error was or when bootstrapped (100 repettons). 17

24 llustrates the functon m( x, b) c n (7). Estmaton s based on devatons of the raw proporton hgher-educated from ths surface. The OLS estmates n columns 1 and 2 n Table 2 llustrate the partal correlatons between property prces and the regressors n the South and East of England. Property prces are around 4% hgher for each 1% absolute shft n the postcode sectors proporton hghly qualfed. But, these numbers should not be nterpreted as structural parameter estmates n a model of property prce determnaton. Dfferences wthn the regon n labour market returns to sklls and employment opportuntes wll smultaneously determne property prces and educatonal composton. When we ntroduce controls for ethnc composton and the densty of purpose bult flats n column 2, these attract postve and sgnfcant coeffcents. Both varables proxy for central London and suburban areas, where property prces are hgh due to the hgh demand for sklled labour n the Captal. As soon as we estmate on devatons from local means usng the sem-parametrc SSE model, the coeffcent on hghly qualfed resdents falls dramatcally, to 1.22 (0.14) n column 3. Takng out the mean dfferences between localtes removes bases n the estmate of the mplct prce ntroduced by local labour market drven property prce and educatonal composton smultanety. The IV estmate n column 4 s only slghtly below ths at (0.218). Ths s a somewhat surprsng result, because any unobserved dfferences between neghbourng postcode sectors n the mean physcal characterstcs of housng should generate varaton n mean property prces, and, we mght expect, varaton n the mean educaton and ncomes of purchasers of these propertes. The smlarty mples that ths source of endogenety s not a serous problem. Ether there s lttle varaton n the prcerelated characterstcs of housng n closely assocated neghbourhoods, or these dfferences n prce are not suffcent to generate dfferences n mean educaton between neghbourng postcode sectors. The varaton n neghbourhood educaton s exogenous, even before we use the nstruments. Measurement error n the educatonal status varable may be another factor that leads to hgher IV estmates than expected relatve to the OLS estmates. The educatonal status varable s taken from the 10% sample of the Census, so the samplng varance s hgh 14. The varables used as nstruments come from the 100% sample. Columns 5 and 6 ntroduce controls for ethnc composton and flat densty. It seems possble that neghbourhood educatonal varaton, or varaton n socal housng captures varaton n the physcal envronment whch are consumpton goods to property buyers. Property owners may prefer to lve away from councl estates because they fnd hgh rse flats and large estates unattractve. Incluson of the densty of purpose bult flats (n 100s per km 2 ) n the postcode sector tests for ths. Although ths varable was sgnfcant n the raw OLS regressons, t s completely nsgnfcant n the SSE or SSEIV models. Another consderaton s the effect of ethnc mnortes on property values, when there s racal prejudce amongst property owners. Ethnc background s assocated wth educatonal attanment, so senstvty of property values to ethnc composton wll affect our estmated coeffcent on educatonal status. When the proporton of black, and Indan ethnc groups s ncluded as a regressor n the SSE model (column 5), we fnd a slght fall (around 10%) n the parameter of nterest. In the SSEIV model, however, we fnd no change. Potentally, we requre the ncluson of ths ethnc group control to justfy our overdentfyng restrctons. The proporton of socal tenants n Indan sub-contnent ethnc groups wll not necessarly be a good nstrument uncondtonal on ethnc overall ethnc composton (ether because of prejudce, or because the proporton from ethnc mnortes s correlated wth mean home-owner ncomes). As t 14 Comparson of 10% and 100% sample unemployment rates suggests that 20% of the varance of the 10% sample s samplng nose. 18

25 turns out, the nstruments and model pass the test, whether or not ethnc composton s 2 ncluded as a man regressor, though the χ statstc moves n the expected drecton. Perhaps the most convncng test of the valdty of the nstruments s the nsenstvty of the IV estmate to the ncluson of ethnc group and flat densty controls. Table 3 shows the estmates for the North of England regonal group. Here we see much less movement n the parameter estmates as we change from OLS on the entre regon (columns 1 and 2), to the SSE models. The coeffcent on educatonal composton falls as expected, but by nothng lke the same amount as n our East and South East sub-sample estmates. In the East and South East group, t s probably the hgh demand for hghereducated workers and hgh property prces n the Captal cty, relatve to outlyng areas, whch generates the hgh coeffcents n Table 2, column 1. In the North, local labour demand factors are less mportant. Nevertheless, the pont estmates fall by around 20% once we abstract from local area effects. Instrumentng the proporton hgher-educated ncreases the coeffcent estmates, an effect whch can only be attrbutable to the samplng nose n the regressor, but the dfference between the SSEIV and SSE estmates of the key parameter s 2 not sgnfcant n a Hausman test ( χ () 1 = ) for column 5 versus column 6). The fnal estmates of our man coeffcent are almost double those obtaned for the South and East. Results for Wales, the West and the South West are much lke those for the North. On ths sample, however, we reject the null hypothess that the overdentfyng restrctons and model specfcaton are correct (the resduals are correlated wth the nstrument vector). The reason for ths msspecfcaton s unclear, but s probably lnked to the hghly rural geography of Wales and the South West pennsula. There s evdence here and n further specfcaton checks (see Secton 0) that the IV results are untrustworthy. Gven the smlarty between the SSEIV and SSE model estmates n the other regons, we can reasonably assume that the non-iv estmates are acceptable here too. Comparng the parameters across regons, an obvous pont s that the response to percentage pont changes n the South East and East s markedly dfferent from the response to a percentage pont change n the other regons. Ths s, of course, largely due to dfferences n mean educaton levels between regons. Convertng the coeffcents nto elastctes at the sample mean we get (0.035) for the South and South East, (0.044) for the North, and (0.041) for the West, SW and Wales A mnmum dstance estmate of the elastctes s (0.039), and we do not reject equalty to the mnmum dstance parameter for all regons (P value =0.356). By these estmates, a 1% relatve mprovement n educatonal status of an average neghbourhood as measured by the proporton wth hgher educaton qualfcatons s valued at around 230 n year-2000 prces. Ths response of prces to local educaton, as predcted by proxmty to socal housng, explans a relatvely small amount of the varaton n property prces wthn local areas. The R 2 s n wthn-property-type-area 17 regressons suggest that around 5% of the varaton n log postcode sector mean prces n the South East and East s assocated wth the proporton n socal housng, around 8% n the North and 4.3% n Wales, West and South West. 15 Treatng the mean as a constant. 16 The elastctes are even closer across regons f we constran the elastctes to be constant wthn regons by estmatng a double-log model: South and SE (0.018); North (0.013); West, South-West and Wales (0.019). The problem wth ths specfcaton s that t mples near zero property prces n areas wth near zero proportons wth hgh qualfcaton, and s nconsstent wth the evdence presented on the log-lnearty of the property-prce/proporton-hghly-qualfed regresson lne. 17 Where area s defned as the spatal group represented by a 6km 6km square. 19

26 Comparson wth property level data Some readers may feel uncomfortable wth results based on mcro-spatally aggregated data, wthout any controls for ndvdual housng or owner-characterstcs, despte the dentfcaton strategy employed here. Can we be sure that our neghbourhood educatonal status measure s not smply measurng owner-occuper wealth, whch attracts a postve coeffcent because of unobserved normal-good-type property or area characterstcs? Ideally, to test the robustness of the results, we want to observe the ncomes of home purchasers and estmate the models condtonal on own household ncomes 18. As dscussed n Secton 4, our second property prce data set from the Survey of Mortgage Lenders (SML) has property prces and the household ncomes on whch the mortgage s based, but no neghbourhood dentfers. Nevertheless, we can use t to estmate the relatonshp between prces and Local Authorty educatonal status, for whch we have dentfyng codes. Exstence of effects at ths level of aggregaton cannot be taken on ther own as evdence of neghbourhood effects on property prces. There could be selecton nto local authorty areas by ndvduals at dfferent ponts n the ncome dstrbuton, due dfferng returns to sklls n local labour markets, and local government factors such as councl tax rates. The estmates based on matched census-sml data are tabulated n Table 5, for 1997 property data. Data for Wales s hard to match, so has been excluded. The nstrument for the local authorty proporton hgher-educated s just the proporton n socal housng. The regressons nclude a broad set of property type nteractons and household characterstcs, as lsted n the table notes. Lookng at Table 5, and comparng wth Table 2 to Table 4, we see that the OLS estmates are somewhat hgher for the South East and East and the West and South West regons, but smlar for the North. Instrumentng local educaton brngs the coeffcent down n the South East and East, makes lttle dfference n the North and ncreases t n the South West and West, but none of these coeffcents s sgnfcantly dfferent from the OLS estmates. If we work wth elastctes at the mean and calculate the mnmum dstance estmate from the IV coeffcents across all samples and regons we get an elastcty of 0.250, wth equalty across regons (p-value = 0.254) 19. The estmates usng the SML data renforce the pattern observed by comparng the OLS and SSE estmates n Table 2 to Table 4 that selecton by educaton nto broader geographcal areas s more mportant n the South East and East than n other areas, hence the hgher coeffcents n ths regon when we look at effects at local authorty level. Overall, the coeffcents estmated usng property level data wth own ncome, plus more property and household characterstcs, are not nconsstent wth the estmates of more localsed human captal effects n the man tables. An mportant pont to note from Table 5 s that the IV estmates are smlar whether or not we nclude own-ncomes. Ths s a good ndcaton that the socal housng nstrument s exogenous to home-purchaser ncomes, whch s what we requre. The mnmum dstance IV estmate of the mpact of educatonal status uncondtonal on ncomes s (0.413), or (0.352) condtonal on own ncomes. Wthout IV, the correspondng parameters are (0.274) and (0.051). 18 Although own ncomes may be endogenous. 19 We can also use ths data to compute the local ncome elastcty, usng wthn-sample estmates of local authorty mean ncomes. Ths gves coeffcents whch are strkngly smlar to those n Table 7, wth a mnmum dstance estmate across all samples of (p-value of test for equalty = 0.223). The drawback s that the ncomes are based on those of new owner-occupers, not the populaton. 20

27 6.3 Senstvty to bandwdth choce Snce we have no pror nformaton on the best bandwdth to use to defne the local area groups n the SSE and SSEIV models, we need to check how the parameters vary wth bandwdth choce. We should be worred f the estmates change dramatcally for small changes n bandwdth, as ths would nvaldate the clam that ths uncovers the parameters of a model of property prce determnaton operatng at the household level. In prncple, an optmal bandwdth could be chosen based on a loss functon whch makes a compromse between bas and effcency as the bandwdth ncreases, effcency ncreases but at the rsk of based parameter estmates. A slghtly more ad-hoc approach s to re-estmate the models at ntervals above and below the 3400 household bandwdth used n the man tables. The results of ths exercse are n Table 6. In all regonal groups, the SSE estmates (wthout nstruments) decrease steadly as bandwdth s reduced from 5100 to 850 households. Ths s to be expected, as samplng error n the 10% census sample leads to ncreasng attenuaton n the estmated coeffcents as we remove across-space varaton (just as fxed effect estmaton n panel data exacerbates downward bas due to measurement error). By contrast, for the South East and East, and North regonal groups the SSEIV estmates are remarkably stable. The IV estmate for the South East and East ncreases by only 8.5%, and the estmate for the North ncreases by only 15% as we ncrease the bandwdth by a factor of 6. Hausman tests of the dfference between pars of estmates computed at dfferent bandwdths all fal to reject the null of equalty. For Wales, West and South West of England, the IV estmates are not stable across dfferent bandwdth choces and the Sargan test statstcs suggest a msspecfcaton. The non-iv estmates are, however, relatvely nsenstve to changes n bandwdth around 3400 households, and are consstent wth the elastctes calculated for the other regons, so I take these as the preferred estmates for ths regon. 6.4 Non-lneartes n response The estmates presented n Secton 0 assume a lnear relatonshp between the proporton of resdents hghly qualfed and log mean property prces. Fgure 3 shows the result of estmatng the sem-parametrc model of Secton 5.3 on the regonal groups. It s farly clear that, apart from a few local rregulartes, the relatonshp between the natural log of property prces and the generated neghbourhood proporton wth hgher educaton s lnear, wth no threshold effects 20. Nothng here ndcates that people are wllng to pay proportonally more as educatonal status ncreases, though of course, the absolute amount pad ncreases wth each one percentage pont shft n educatonal status. In the South East and East regon, a one percentage pont mprovement n educatonal status at the 75 th percentle (21% hgher educated) would be worth around 1600, whereas a smlar relatve mprovement at the 25 th percentle (11%) would be worth around 1300 (n 1995 prces). Some specfcaton checks for ths model are presented n Appendx C. 6.5 Implct prce of neghbourhood mean ncomes It follows from ths evdence that home-owners are prepared to pay a premum to lve n a hghly educated neghbourhood, that we should fnd evdence of a premum to hgh ncome 20 Smlar results are obtaned f we estmate a kernel regresson of log-property type on the proporton hghereducated, where the varables are n devatons from the property-type-locaton means, and where locaton s defned by a 6km 6km grd. 21

28 neghbourhoods, uncondtonal on educaton levels. Table 7 confrms ths predcton. The specfcatons n the columns of these tables are dentcal to those n the last two columns of Table 2 to Table 4, but wth log mean postcode sector ncomes replacng the proporton wth hgher educaton qualfcatons as a regressor. Controllng for spatal varaton s mportant. The estmates from the smooth spatal effects models are substantally lower than the OLS estmates wthout area controls (not shown). Except n the North, nstrumentng local ncomes wth the proporton n socal housng and proporton of socal tenants from Indan ethnc groups makes lttle dfference to the estmates. Moreover, ntroducng the ethncty and flat densty varables as regressors reduces the estmated elastcty only slghtly. The fnal estmates are statstcally dentcal across regons. The mnmum dstance estmate s (0.054) and we do not reject equalty of all parameters to the mnmum dstance estmate (pvalue = 0.384). Households n all regons pay around 0.5% for each 1% mprovement n mean neghbourhood ncomes. 6.6 Comparng local ncome and educaton effects Can we determne whether ncome or educaton s more valued n the neghbourhood socal envronment? Educaton mght be mportant f resdents seek out the productve externaltes n human captal formaton; f ncome s more mportant, we mght emphasse consderatons such as lower crme rates and a well mantaned physcal envronment. However, dsentanglng the nfluences of famly ncome and educaton on attanments s dffcult, even usng mcro-data at the ndvdual or household level. The problem s exacerbated n our mcro-spatally aggregated data by the fact that mean ncomes and educaton are hghly correlated 21, they are measured n dfferent perods, the ncome data s partly modelled, our educaton data s based on 10% census samples, and because both are endogenous to property prces. Frstly, what stands out from these tables, s that the elastctes wth respect to ncome are more than double those on educaton at average educaton levels. However, these values are very low relatve to what we would expect f ncomes were the prncple object of preference. We can see ths from conventonal estmates of the returns to educaton n the md to late 1990s, whch are around 0.3 for hgher educaton qualfcatons 22. Snce we have the property prce-local ncome equaton: ln P = β ln + ε (17) y and an earnngs equaton: ln y = δ + ν (18) x then, f hghly educated resdents n the neghbourhood only effect property prces through neghbourhood ncomes, we can substtute (18) nto (17) to get a coeffcent of βδ on the proporton hgher educated n a property prce equaton. Our estmate of β s 0.51 and, assumng most household ncome s earnngs, we know δ s around 0.3, so the coeffcent on the proporton hgher educated n the property prce equaton would be around Our 21 And the raw correlaton between ncomes and educaton s e.g. Harkness and Machn (1999). 22

29 actual estmates are ten-tmes ths fgure! 23. It seems lkely on ths evdence, that educaton s valued as a local commodty for other reasons than just ts mpact on ncomes. Mean neghbourhood ncome on ts own acts as nosy proxy for the underlyng educatonal status of the area. Table 8 shows ths more drectly. The table shows wthn-postcode-dstrct IV estmates from 1996 property value models 1996 neghbourhood ncomes and 1991 educaton on the rght hand sde. The mpled educatonal elastcty s 0.31, the ncome elastcty s 0.56 slghtly hgher than our baselne results, due to the use of less accurate area fxed effects. Instrumentng both local educaton and local ncomes ncreases the coeffcent on educaton slghtly, but the estmated ncome effect becomes near-zero, negatve and nsgnfcant. Ths s despte the fact that the nstruments are more strongly correlated wth ncomes than educaton. Ths s clear evdence that neghbourhood educaton levels domnate neghbourhood ncomes n the preferences of home-buyers. Ths s consstent wth the hypothess that an educated neghbourhood offers real socal and economc benefts to ts resdents. 6.7 Unobserved neghbourhood heterogenety The models presented n the man tables use relatvely few rght hand sde controls, and rely on the spatal effects and nstruments to acheve dentfcaton. As dscussed n 2.2, unobserved neghbourhood heterogenety wll compromse the nterpretaton of the estmates as measures of wllngness to pay for margnal mprovements n the neghbourhood, or as estmates of the margnal socal benefts to educaton. We can, of course, just nclude more neghbourhood characterstcs and observe what happens to the estmated coeffcent on the proporton hgher-educated. 24 Table 9 presents results from ths exercse for Neghbourhood characterstcs are derved from the Census. Addtonal controls are prmary school performance as measured n the sprng of 1996 plus postcode unemployment per household n The table shows coeffcents on neghbourhood educaton or ncome from separate regressons. Row A shows the baselne model, comparable to the results n the man tables, Table 2 to Table 7. The results usng postcode dstrct fxed effects are 15 to 20% hgher than those obtaned usng the sem-parametrc model n the man tables. The results n the table are dscussed below. Secton dscusses further results based on property crme rates from a small sub-sample of postcode dstrcts derved from the Brtsh Crme Survey n The supply of owner-occuped and socal housng Negatve correlaton may exst between property prces and socal housng f socal housng was, hstorcally, bult n areas where land prces were low, and where there s strong seral correlaton n land prces. Row B of Table 9 ncludes the proporton of households n socal housng recorded n the 1981 census as an addtonal regressor, for those postcode sectors where at least fve enumeraton dstrcts match up wth enumeraton dstrcts n The 23 If we work wth years of post compulsory educaton an parameterse δ at 0.07 (from the Famly Expendture Survey) we fnd that our estmates are more than fve-tmes hgher than expected f ncomes alone matter. 24 Some of these addtonal characterstcs may be endogenous. Consstent estmaton of the parameter of nterest reles on the assumpton that condtonng on the vector of neghbourhood attrbutes s suffcent to make the unobserved determnants of property prces ndependent of local educaton levels. Ths s reasonable, snce the OLS estmates based on wthn-local-area varaton were close to the IV estmates n the man tables, even wthout addtonal observable neghbourhood attrbutes. 23

30 1991 proporton n socal housng nstruments neghbourhood educaton, or ncome. Snce the vast majorty of socal housng was bult pror to 1981, f the hstorcal supply of socal housng s nfluencng our results, we would expect the coeffcent on the proporton of socal housng n 1981 to be sgnfcant, and for the coeffcent on educatonal or ncome status to fall. In fact, the coeffcent on the 1981 proporton n socal housng s nsgnfcant, and the IV estmates are not sgnfcantly dfferent from those n row A usng the standard Hausman test. Ths procedure also checks effects of socal housng on the pre-1981 supply of dwellngs for owner occupaton. If property developers sted lower qualty developments near areas of socal housng, or were more lkely to convert houses nto flats n neghbourhoods close to socal housng, or less lkely to upgrade dwellngs, then the coeffcent on neghbourhood educaton may be based upwards by unobserved dfferences n housng sze or qualty. Controllng for the 1981 proporton n socal housng shows that ths s not a serous consderaton, for propertes bult before 1981 at least t s the subsequent measure of educatonal status that s mportant. A further test for effects from the supply of flats from converted propertes s to re-estmate the models excludng property transactons on flats. The new coeffcents n the area reported at the foot of Table 2 to Table 4. Excludng flats ncreases the coeffcents slghtly, though not sgnfcantly so. Agan we would conclude that the relatonshp between local educaton levels and property prces s not attrbutable to unobserved varaton across neghbourhoods n the qualty of housng suppled. An mportant pont to note here s that deteroraton n the qualty of housng occurrng as a result of the negatve externalty from low human captal neghbourhoods for example as poorer home-owners move nto the area does not result n upward-based estmates of the mplct prce. Instead, f the supply of housng measured n qualty unts falls back n response to fallng demand, then the estmated mplct prce, condtonal on housng characterstcs, wll under-estmate the mpact of the externalty. What we want to measure s the full dervatve of prces wth respect to local human captal, ncludng the effect on prce resultng from property deteroraton Other envronmental and adult characterstcs We can gauge the extent of the mportance of unobserved neghbour heterogenety by ntroducng more neghbour attrbutes n the regressons. Row B ncludes some more controls proportons unemployed, the long-term sck, and lone parents to proxy the type of ndvdual typcally housed n socal accommodaton. These are frequently used as ndcators of area deprvaton. These three characterstcs explan around 75% of the varance n the proporton of socal tenants, and nearly 40% of the varaton n the proporton hghereducated. Addtonal regressors are the proporton of resdents not at ther current address one year earler, the proporton of agrcultural workers, the proporton over 65, and household densty. Controllng for these characterstcs brngs down the estmated mpact of local educaton and ncomes, but the elastctes are now almost dentcal to those obtaned usng the sem-parametrc spatal fxed effect models n the man tables. Although unreported n the table, t s worth notng that the coeffcents on other neghbourhood characterstcs have t- statstcs whch are less than half those on educaton and ncomes. Elastctes on the proportons of lone parents, long-term sck and unemployed are all well below 0.1 (n absolute value) at the mean. 24

31 Local schools In Gbbons and Machn (2001) we report strong local property prce effects from prmary school performance, as measured by Natonal Currculum Key Stage 2 test results at age 11, whlst secondary school has no measurable effect at postcode sector level. Row C n Table 9 ncludes the postcode sector mean prmary school test results (the proporton achevng level 4 n the tests), plus the proporton of chldren determned as havng specal educatonal needs, or wth local educaton authorty statements of specal needs. Although these are sgnfcant n the regressons, the coeffcent on the proporton of hgher-educated adults or on local ncomes hardly moves. Ths suggests that f more hghly educated neghbourhoods matter to home-owners because of ther concern for the human captal accumulaton of chldren, then the antcpated nput nto human captal s operatng outsde the prmary schoolng envronment Local crme rates Good local crme rate data s not easly avalable n the UK. A crude measure of neghbourhood crme can be constructed from the 1992 Brtsh Crme Survey, whch ncludes 575 postcode sector dentfers. The sample sze wthn postcode sectors s small the mean s 21 respondents and there are only 67 postcode dstrcts wth more than one postcode sector. A crme rate proxy constructed as the mean number of property crmes n the last year recorded per respondent n each postcode dstrct attracts a negatve and sgnfcant coeffcent ( 0.014, s.e =.007) when entered on ts own n a postcode dstrct level property value model. However, the coeffcent becomes near-zero and nsgnfcant (-0.004, s.e ) once we control for postcode dstrct educaton levels. Agan, we must conclude that local educatonal status s the more mportant factor Socal tenant-specfc effects We can check f t s somethng partcular about socal tenants, or propertes near socal housng whch generates the observed educaton-prce relatonshp usng an alternatve nstrument the locaton of hgher educaton nsttutons, whch generate hgh educaton enclaves. Ths s a weak nstrument (t-statstc of 1.67 n the predcton equatons), and only 1.5% of the sample sectors have hgher educaton nsttutons located wthn them. Nevertheless the pont estmate of the educaton effect s almost dentcal to the natonal average effect mpled by the man results. The IV coeffcent s not, however, very sgnfcant (t statstc =1.29). Stll, ths ndcates that we are pckng up educaton related effects n our man results, rather than pure prejudce aganst socal tenants. 6.8 Evdence for human captal externaltes As ponted out n Secton 3, one predcton from a model where neghbourhood educatonal status generates an externalty n the producton of chldren s human captal, s that the mplct prce of mprovements n neghbourhood educatonal composton must be ncreasng n famly sze (treatng famly sze as exogenous). We should fnd that the mplct prce of educatonal status s hgher n neghbourhoods wth more chldren per household, or wth a hgher proporton of households wth chldren. By nteractng neghbourhood educatonal composton wth an above/below medan famly sze ndcator n a wthn-postcode dstrct property prce model, we fnd that that the mplct prce of neghbourhood educatonal status s (s.e ) n below-medan famly sze sectors, but (0.063) n above-medan 25

32 famly sze sector. The dfference s sgnfcant (t = 2.66). Ths s consstent wth the hypothess that households value good neghbourhoods because of the benefts to chldren, though there s no drect evdence here that these benefts accrue n terms of chldren s educatonal attanments, or acquston of other productve sklls. The proporton of home-owners wth chldren s also ncreasng n the proporton of hghly qualfed resdents once ths s nstrumented by the proporton of socal tenants n the neghbourhood. Agan ths suggests that famles wth chldren beneft more from good neghbourhoods, and that they are more wllng than others to bd up property prces n hgheducaton neghbourhoods. Regressng the postcode sector proporton of home-owners wth chldren on the proporton of resdents wth dplomas and degrees (where these varables are n devatons from postcode dstrct means) gves an nsgnfcant coeffcent of (s.e. = 0.025). Instrumentng the proporton wth hgh qualfcatons wth the proporton n socal housng drves the coeffcent up to (0.026). A reducton on the proporton of socal tenants equvalent to a one percentage pont rse n the proporton of resdents wth dplomas and degrees s assocated wth a 0.14% rse n the proporton of home owners wth chldren (from a mean of 30%). One nterpretaton of ths s that home-owners wth chldren are wllng to bd more for margnal mprovements n the educatonal status of neghbourhoods because of the mpact on the educatonal attanments of ther chldren. The relatonshp s not evdent n the OLS relatonshp, because low home-owner household ncomes are assocated wth larger famly sze, whch obscures the relatonshp of nterest. 7. Concludng Remarks These results demonstrate that neghbourhood property prces respond to exogenous varaton n neghbourhood educaton levels and local ncomes generated by varaton n the local proporton n socal housng. Households value resdence n educatonally rch and hgherncome neghbourhoods. The estmated elastctes n the average neghbourhood are stable across regons, at around 0.24 for the proporton hgher-educated, and 0.52 for mean ncomes uncondtonal on educaton. A sem-parametrc, wthn-area, IV approach dentfes these effects. We get smlar elastctes on local authorty educatonal status usng property-level mcro-data and condtonng on home-purchaser ncomes. Our local ncome elastcty s of the same order as the estmates bured n hedonc regressons n the US lterature (around ). Educatonal dfferences between neghbourhoods seem far more mportant than ncomes the coeffcent on local human captal s much hgher than expected f ncome alone mattered, and the ncome coeffcent vanshes once we nclude educaton and ncome together n the IV regressons. Usng addtonal Census, school, unemployment and crme data, we can see that the senstvty of prces to local educatonal status s undmnshed once we abstract from other observable characterstcs of ndvduals n the neghbourhood. We conclude that households place partcular mportance on the educatonal status of a neghbourhood n choosng a resdental locaton, and that households wth more chldren seem prepared to pay more. Ths hghlghts the potental mportance of communty spllovers n the producton of human captal. These results have drect relevance to the cost-beneft analyss of measures to mprove the educatonal status of deprved neghbourhoods, as well as to educatonal polcy n general. Unfortunately, wthout detaled nformaton on neghbourhood educatonal attanments other than hgher educaton qualfcatons, we can make no assessment of the extent to whch hgher educaton matters over and above educaton n general. 26

33 On the assumpton that t s the educatonal status of communtes that matters to households, we can make a tentatve assessment of the long-run socal, communty-level benefts of educaton and compare ths wth the prvate returns. I focus on households headed by someone under the age of forty 25. Mean household annual earnngs for these households was around 19000, and the mean property prce was n A 10% relatve ncrease n the proporton of adults wth hgher educaton qualfcatons n 1995 would have meant a 1.9% absolute ncrease n the proporton wth hgher-educaton qualfcatons. Assumng the prvate returns to hgher educaton qualfcatons are n the order of 25% to 30%, ths mprovement n educaton mples an ncrease n mean earnngs of around 0.53%, or 100 on average household ncome. From the estmates presented n ths paper, ths change n educatonal attanments would be valued at 1500 at the 1995 mean property prce. Average mortgage nterest rates n 1995 were around 7% and the mean loan perod 22.5 years, so 1500 s equvalent to 130 per annum n mortgage payments. By ths calculaton, the money-metrc value of the non-pecunary benefts to socety from an ndvdual ganng hgher qualfcaton s hgher than the mean return n terms of ncreased earnngs. Note, ths fgure s based on changes to the proporton hgher-educated only. If we assume that the proporton n all post-school attanment groups ncreases n proporton to the proporton wth hgher educaton qualfcatons, then a 10% relatve mprovement represents a 10% ncrease n the mean number of years of post-school educaton. Takng a generous estmate of the returns to years of schoolng as 0.07, ths relatve ncrease n mean non-compulsory years of educaton (1.3 years for 25 to 40-year-olds) n 1995, gves a return per household of 170. Agan, ths s smlar to our value on the pure communty benefts. If we beleve that households value communty educatonal status purely as an nput nto chldren s human captal accumulaton, and that parents can transfer ncome drectly to chldren, t follows that the average household, whch has one chld, expects a 10% relatve mprovement n qualty of the average neghbourhood to ncrease a chld s expected household ncome by around 0.7% 26. We can take ths s an upper bound to the average mpact of neghbourhood qualty on a chld s future household ncome. If all mprovements n ncome are lnked to better ndvdual educatonal attanments, and returns to educaton are expected to reman unchanged, then parents expect ths 10% relatve change n neghbourhood status to mprove ther chld s chances of ganng hgher educaton qualfcatons by a smlar proportonal amount 27. Ths unt elastcty s substantally hgher than the chld outcome-neghbourhood educatonal elastctes estmated n the neghbourhood effects lterature, whch are n the order of (see Kremer, 1997 or Gbbons, 2001). Clearly, not all the expected benefts of a better neghbourhood relate to 25 Sources of the fgures that follow are varously: Famly Resources Survey 1995/6, Survey of Englsh Housng 1995, Survey of Mortgage Lenders Age 40 s the 75 th percentle n the age dstrbuton of those takng out mortgages n the survey of mortage lenders. Returns to educaton control for gender, ablty and famly background calculatons from Natonal Chld Development Survey and 1970 Brtsh Cohort Survey, but see also Blundell, Dearden, et al. (2000), or Harkness and Machn (1999). Mortgage nterest rates n 1995 were 6.15% accordng to the 5% Survey of Mortgage Lenders, though the fgure gven n Buldng and Constructon Statstcs s 7.83% per annum. Returns to years of educaton s the fgure from a smple regresson of log earnngs on age left full tme educaton, age, age squared and gender for ndvduals aged between 1992 and 1996 n the Famly Expendture Survey. 26 The mean number of chldren per owner-occuper household for owner occupers headed by someone under 40 n the Survey of Englsh Housng n 1995 s one. The calculaton assumes that expected chld s household ncome s the same as current mean household ncome, so we just dvde the value of the benefts ( 130) by household ncome. 27 Because the change n the proporton of people wth hgher qualfcatons necessary to ncrease earnngs by 0.7% s roughly 2.3%, assumng the return to hgher qualfcatons s around 0.3. The current proporton wth hgher qualfcatons s tendng towards 23%. 27

34 better earnngs-related outcomes for chldren, or else the exstng lterature ms-measures ths effect. These back-of-an-envelope calculatons are, of course, approxmate. Stll, the message comes across that the resdents are prepared to pay for neghbourhoods wth hgher stocks of human captal, and that the aggregate non-earnngs related communty benefts per household are of a smlar order to the aggregate prvate returns per household as measured by the ncrement to earnngs from hgher educaton qualfcatons. It should be borne n mnd also that the socal benefts measured here are only those that accrue locally, so wll not nclude spllovers n producton, n workplace relatons, n technologcal nnovaton and n other areas where acton s at a broader geographcal level. Gven the sze of these effects measured here, these communty benefts warrant further analyss. Focussng on the prvate returns to educaton serously understates the value of educaton to socety, and any polcy decsons based on these returns alone may result n sub-optmal provson of educatonal servces. 28

35 Table 1: Summary statstcs for local ncomes, educaton and property prces South East and East North of England Wales, S-West & West Mean/s.d Mn/Max Mean/s.d Mn/Max Mean/s.d. Mn/max 1995 sector mean prce 1996 sector mean prce 1999 sector mean prce (50139) (52570) (74843) (22901) (23786) (33452) (25931) (26630) (40135) Mean annual sector sales volume 131 (74) (59) (65) Mean sector detached prce Mean sector semdetached prce Mean sector terraced prce Mean sector flat/mas. prce (74958) (56517) (59889) (50438) (31363) (13868) (12875) (18379) (36955) (18354) (18417) (24433) Mean postcode sectors Proporton wth dplomas, degrees Proporton n socal housng (0.081) (0.147) (0.077) (0.178) (0.064) (0.128) sector mean ncome 1996 postcode sectors 1999 sector mean ncome 1999 postcode sectors (4486) (3913) (3399) (5224) (4359) (3779) Means of property prces are means of postcode sector means, weghted by sales volumes. Means of qualfcatons and socal housng weghted by households. Means of ncomes are means of postcode sector means, weghted by number of households. 29

36 Table 2: Property prce response to neghbourhood educatonal composton, South East and East England, 1995 OLS OLS OLS- SSE IV-SSE OLS- SSE IV-SSE Mean Proporton hghly qualfed n 1991 Densty of purpose bult flats (100s/km 2 ) Proporton Black, Indan, P stan, B desh Mean rooms n owner-occuped housng (0.089) (0.083) e -3 (0.9 e -3 ) (0.077) (0.011) Detached (0.006) Flat/Masonette (0.006) Terraced (0.005) Constant (0.055) (0.013) (0.005) (0.006) (0.004) (0.071) (0.139) (0.218) (0.143) e -3 (0.8 e -3 ) (0.184) (0.015) (0.005) (0.005) (0.004) (0.015) (0.005) (0.005) (0.004) (0.015) (0.005) (0.005) (0.004) (0.209) 0.2 e -3 (0.8 e -3 ) (0.189) (0.015) (0.005) (0.005) (0.004) Overall R Wthn R P-value test of restrctons Dependent varable s log of postcode sector mean property-type prce. Sample sze (sectors x property type) = Mn, mean, max bandwdth:.24 km, 1.27 km, 8.75 km. Mean house prce = Mean of dependent varable (log-mean-prce) = Mean Eastngs 52854, Northngs Instruments n columns 4 & 6 are postcode-sector proporton n socal housng and proporton of socal housng tenants from Indan, Pakstan, and Bangladesh ethnc groups. Estmate of educatonal composton parameter s (0.222) n column 4 f socal housng s the only nstrument. Excludng transactons on Flats and Masonettes from sample gves estmate of educatonal composton parameter of (0.200) n column 6. 30

37 Table 3: Property prce response to neghbourhood educatonal composton, North of England, 1995 OLS OLS OLS- SSE IV-SSE OLS- SSE IV-SSE Mean Proporton hghly qualfed n 1991 Densty of purpose bult flats (100s/km 2 ) Proporton Black, Indan, P stan, B desh Mean rooms n owner-occuped housng (0.064) (0.064) e -3 (0.7 e -3 ) (0.079) (0.011) Detached (0.005) Flat/Masonette (0.009) Terraced (0.004) Constant (0.055) (0.011) (0.005) (0.009) (0.004) (0.057) (0.175) (0.321) (0.170) e -3 (1.5 e -3 ) (0.216) (0.030) (0.004) (0.009) (0.004) (0.033) (0.004) (0.009) (0.004) (0.030) (0.004) (0.009) (0.004) (0.341) -1.2 e -3 (1.5 e -3 ) (0.219) (0.033) (0.004) (0.009) (0.004) Overall R Wthn R P-value test of restrctons Dependent varable s log of postcode sector mean property-type prce. Sample sze (sectors x property type) = Mn, mean, max bandwdth:.28 km, 1.44 km, km. Mean house prce = Mean of dependent varable (log-mean-prce) = Mean Eastngs 41272, Northngs Instruments n columns 4 & 6 are postcode-sector proporton n socal housng and proporton of socal housng tenants from Indan, Pakstan, and Bangladesh ethnc groups. Estmate of educatonal composton parameter s (0.323) n column 4 f socal housng s the only nstrument. Excludng transactons on Flats and Masonettes from sample gves estmate of educatonal composton parameter of (0.264) n column 6. 31

38 Table 4: Property prce response to neghbourhood educatonal composton, Wales, West and South West of England, 1995 OLS OLS OLS- SSE IV-SSE OLS- SSE IV-SSE Mean Proporton hghly qualfed n 1991 Densty of purpose bult flats (100s/km 2 ) Proporton Black, Indan, P stan, B desh Mean rooms n owner-occuped housng (0.111) (0.107) (0.211) (0.323) (0.200) (0.303) e e e e e e (0.077) (0.017) Detached (0.006) Flat/Masonette (0.009) Terraced (0.005) Constant (0.089) (0.018) (0.006) (0.009) (0.005) (0.090) (0.162) (0.033) (0.005) (0.008) (0.004) (0.034) (0.005) (0.008) (0.034) (0.035) (0.005) (0.008) (0.004) (0.173) (0.033) (0.005) (0.008) (0.004) Overall R Wthn R P-value test of restrctons Dependent varable s log of postcode sector mean property-type prce. Sample sze (sectors x property type) = Mn, mean, max bandwdth: 0.40 km, 1.69 km, km. Mean house prce = Mean of dependent varable (log-mean-prce) = Mean Eastngs 34637, Northngs Instruments n columns 4 & 6 are postcode-sector proporton n socal housng and proporton of socal housng tenants from Indan, Pakstan, and Bangladesh ethnc groups. Estmate of educatonal composton parameter s (0.322) n column 4 f socal housng s the only nstrument. Excludng transactons on Flats and Masonettes from sample gves estmate of educatonal composton parameter of (0.308) n column 6. 32

39 Table 5: Property prce response to local educaton: Survey of Mortgage Lenders data, by regon, 1997 South East and East North West and SW OLS IV OLS IV OLS IV Local authorty proporton wth hgher educaton (0.194) Own log ncome (0.017) ) (0.018) (0.050) (0.013) (0.319) (0.013) (0.393) (0.016) (0.899) (0.017) R Sample sze Mean log-prce Mean prce Dependent varable s log of property prce. All models nclude: man purchaser age, number of males, number of females, bugalow, detached, sem-detached, terraced, flat/masonette (converted), flat/masonette (purpose bult), other dwellng type, bult , , , after 1980, new, number of rooms, dwellng type number of rooms, dwellng type property age, county dummes. Mnmum dstance estmate of IV coeffcent on educatonal status = (0.352). Test of equalty across regons p-value = Wthout own-ncomes as control, IV coeffcents on local educatonal status are: SE&E: (0.877), North: (0.387), West and SW: (1.205). Mnmum dstance estmate = (0.413). Test of equalty across regons p-value =

40 Table 6: Senstvty of educaton parameter estmates to bandwdth choce South and East North West, SW and Wales OLS IV OLS IV OLS IV 850 households (0.277) (0.383) (0.233) (0.452) (0.397) (0.667) Overdentfcaton test Mean bandwdth 0.64 km 0.72 km 0.85 km 1700 households (0.214) (0.280) (0.218) (0.428) (0.277) (0.410) Overdentfcaton test Mean bandwdth 0.90 km 1.02 km 1.19 km 3400 households (0.143) (0.209) (0.170) (0.341) (0.200) (0.303) Overdentfcaton test Mean bandwdth 1.27 km 1.44 km 1.69 km 5100 households (0.119) (0.184) (0.144) (0.280) (0.169) (0.276) Overdentfcaton test Mean bandwdth 1.56 km 1.77 km 2.08 km Hausman test (1700 aganst 850) Hausman test (3400 aganst 1700) Hausman test (5100 aganst 3400) χ 2 (1) = 0.00 χ 2 (1) = χ 2 (1) = 0.74 χ 2 (1) =0.34 χ 2 (1) = 0.36 χ 2 (1) = 5.1 χ 2 (1) =0.26 χ 2 (1) = 0.37 χ 2 (1) = 3.4 Dependent varable s log of postcode sector mean property-type prce. Table shows estmates of coeffcents n model of columns 5 an 6 n Table 2 to Table 4 under dfferent bandwdth choces. Hausman tests are for tests of parameter equalty under dfferent bandwdth assumptons. 34

41 Table 7: Property prce response to neghbourhood ncomes, by regon, 1996 and 1999 South East and East North Wales, West and SW OLS- SSE IV-SSE OLS- SSE IV-SSE OLS- SSE IV-SSE Log mean postcode sector ncomes Densty of purpose bult flats (100s/km 2 ) Proporton Black, Indan, P stan, B desh Mean rooms n owner-occuped housng (0.035) (0.051) (0.034) (0.053) (0.057) (0.066) 0.2 e e e e e e -3 (0.5 e -3 (0.5 e -3 (1.2 e -3 (1.0 e -3 (1.9 e -3 (2.0 e -3 ) (0.166) (0.012) Detached (0.003) Flat/Masonette (0.004) Terraced (0.003) (0.158) (0.013) (0.003) (0.004) (0.003) (0.096) (0.018) (0.003) (0.006) (0.003) (0.162) (0.019) (0.052) (0.006) (0.003) (0.120) (0.029) (0.004) (0.005) (0.003) (0.138) (0.026) (0.004) (0.005) (0.003) Overall R Wthn R P-value test of restrctons Sample sze Mean log-ncomes Mean log-prce Mean prce Mean bandwdth 1.30 km 1.50 km 1.80 km Dependent varable s log of postcode sector mean property-type prce. Instruments n columns 4 & 6 are postcode-sector proporton n socal housng and proporton of socal housng tenants from Indan, Pakstan, and Bangladesh ethnc groups. Reducng bandwdth by factor of 2 gves estmated parameter on educatonal composton of: (0.051) n column 1, (0.082) IV n column 2 and Sargan test statstc p-value of (0.042) n column 5, (0.071) IV n column 6, wth Sargan test statstc p-value of (0.085) n column 5, (0.099) IV n column 6 and Sargan test statstc p-value of

42 Table 8: Comparson of educaton and ncome effects: Wthn-postcode-dstrct IV estmates, all regons, 1996 property prces Postcode sector proporton hgher educated (0.092) Log mean postcode sector ncomes (0.025) (0.592) (0.160) Sectors property types Both models nclude dwellng-type dummes, mean rooms n owner occuped housng, proporton nonwhte, purpose-bult flat densty, postcode dstrct dummes. Instruments are the proporton of socal tenants, the proporton of Indan subcontnent and the proporton of blacks n socal housng. F-statstc on three nstruments n ncome equaton = F-statstc on three nstruments n educaton quaton = Sargan test of overdentfyng restrctons p-value = (column 3 ). Table 9: Property prce response to neghbourhood: estmates wth addtonal controls Educaton Log ncomes Coeffcent Elastcty Elastcty A B C D Dwellng type, year, proporton non-whte, average rooms, purpose-bult-flat densty, postcode dstrct fxed effects (0.063) (0.009) (0.021) Sample sze Model A, plus proporton of socal households n 1981: IV usng 1991 proporton of socal households (0.226) (0.033) (0.055) Sample sze Model A, plus md-year unemployment per household, household densty per km 2, proportons n agrcultural employment, one year mgrants, lone parents, long-term sck, age 65 plus (0.078) (0.011) (0.028) Sample sze Model C, plus prmary school performance, proporton of pupls wth statements and specal educatonal needs (Wales excluded) (0.079) (0.012) (0.026) Sample sze Dependent varable s log of 1996 property prce 2. School data taken from DfEE 1996 Prmary School Performance Tables. 3. Unemployment counts from Noms. 4. All other controls from 1991 census. 36

43 Fgure 1:Raw postcode sector proporton wth hgh qualfcatons, 1991: London and surroundng area Fgure shows postcode proportons wth hgher educaton qualfcatons n postcode sectors n London area, Eastngs to 56000, Northngs to Source 1991 Census