Crime, House Prices and Inequality: the Effect of UPPs in Rio *

Transcription

1 Crime, House Prices ad Iequality: the Effect of UPPs i Rio * July 2013 Claudio Frischtak Bejami R. Madel Astract We use a recet policy experimet i Rio de Jaeiro, the istallatio of permaet police statios i low-icome commuities, to quatify the relatioship etwee a reductio i crime ad the chage i the prices of eary residetial real estate. Usig a ovel dataset of detailed property prices, we fid that the ew statios (UPPs) had a sustatial effect o the trajectory of property values ad crime statistics. We also fid that iequality of residetial prices decreased. Both of these oservatios are cosistet with a dyamic valuatio model i which historical crime rates have persistet effects o real estate prices. * C. Frischtak: Iteratioal Growth Cetre ad IterB Cosultoria Iteracioal de Negócios ( claudio.frischtak@iter.com.r). Correspodig author: B.R. Madel: Federal Reserve Bak of New York, Iteratioal Research ( rmadel@gmail.com, tel: ). We thak semiar participats at the World Bak, Iter-America Developmet Bak, Potifical Catholic Uiversity of Rio de Jaeiro, Fudação Getulio Vargas (Rio) ad the Columia Brazil Semiar for helpful commets. This research was coducted with restricted access to real estate listigs data provided y We gratefully ackowledge ZAP ad the support of Eduardo Gama Schaeffer ad Caio Biachi for facilitatig access to these data. Victor Chateauriad, Diego Gilsaz ad Felipe Katz provided super research assistace. The views expressed i this paper are those of the authors ad do ot ecessarily reflect the positio of the Iteratioal Growth Cetre, the Federal Reserve Bak of New York or the Federal Reserve System.

2 Residetial property prices are a importat gauge of ecoomic coditios writ large. They reflect may macroecoomic factors as well as the particular local microecoomy of the property s locatio. Home values also compose a importat part of household wealth, especially i lower icome areas where residetial property is typically a family s primary (or oly) asset. I the Uited States, aout a third of total assets for a give family are accouted for y ower-occupied housig, with that figure closer to two thirds for families elow the media level of et wealth. 1 This statistic might e further skewed i developig coutries, where the capacity of poor families to accumulate fiacial assets is more limited. Take together, these oservatios suggest a powerful mechaism y which ay policy affectig the determiats of house prices ca alter the level ad dispersio of household wealth. I this paper, we ivestigate oe example of this mechaism as it pertais to the coectio etwee crime ad house prices. Our first ojective is to empirically idetify ad documet the relatioship etwee crime ad house prices. As a pulic ad, we fully expect crimes to exert a dowward force o prices; ideed, this is a commo fidig i the related literature o house ameity valuatio ad the ecoomics of coflicts. We quatify the extet to which prices are resposive to crime-related outcomes, as demostrated y a recet policy experimet ad with the use of highly detailed property price data from the olie classified wesite ZAP ( ad fid that these effects ca e quite large ad ecoomically meaigful. Our mai iovatio will the e to documet ad explai the distriutioal cosequeces of removig the pulic ad of crime; that is, the removal of crime may have heterogeeous effects o the prices of differet resideces i a maer which alters the degree of overall iequality amog property values. This would happe, for example, if lower valued properties appreciate or depreciate disproportioately to a give chage i the crime rate. We will discuss the circumstaces uder which that would occur i the cotext of a dyamic model of property valuatio. Our empirical work will show that decreasig crime does, i fact, eefit lower valued properties disproportioately, reducig the iequality amog properties. This relatioship is suggested i Figure 1, which plots idices of homicides, roeries ad a Gii coefficiet of 1 Keickell (2009) presets a road array of statistics o U.S. family icome ad wealth for the years , drawig o the Survey of Cosumer Fiaces (SCF) compiled y the Federal Reserve Board. Family et wealth ad its compoets for various years of the SCF are provided i Tale A2 of that paper. 1

3 house prices for the city of Rio de Jaeiro sice Both homicides ad roeries declied markedly sice mid Though the series for homicides is more volatile, the average decrease for oth types of crime is aout 15 percet y mid The Gii coefficiet measures the level of iequality of house prices across Rio s eighorhoods. It was rather stale at roughly 0.28 through the egiig of 2010 efore fallig to y mid The fact that oth crime ad iequality pivoted ad the fell at aout the same time is suggestive of a relatioship etwee them. Sice there are may differet factors that ca affect house prices simultaeously, we study the housig market aroud the time of a specific policy evet tightly liked to the ojective of crime reductio. This policy is the itroductio of the Uidade Pacificadora da Policia ( Pacifyig Police Uit, or UPP) program i Rio de Jaeiro egiig i late As i may metropolita areas i developig coutries, a sigificat fractio of the populatio of Rio live i very low-icome commuities with a high cocetratio of sustadard, iformal housig; i Rio, home to some of the largest of these commuities i Lati America, they are called favelas. Over the past three decades, the city has ee plagued y coflicts over territory i its favelas with drug gags ad militias, with may favelas effectively eig occupied ad govered y the drug gags. The UPP program, i respose, re-occupies specific favelas y force usig elite police uits, drives out the drug gags ad roots out caches of weapos ad drugs, ad the istalls permaet police statios staffed y highly traied, well-paid ad ewly-recruited officers; eightee such statios have ee istalled sice The asic ojective of reoccupatio is the reewed assertio of the rule of law ad the aatemet of drug gag-related crimes. The program, to the extet it is effective, is resposile for may positive exteralities associated with the accomplishmet of these ojectives. Usig detailed mothly data o residetial property prices i Rio s formal housig market, as well as o homicide ad roery rates i each of Rio s eighorhoods, we formally test the hypotheses that eighorhoods closer to a UPP statio experieced larger tha average decreases i crime ad larger tha average icreases i house prices after the UPP was put ito place. I additio to the variatio across 2 Data sources ad the details of idex costructio for the series i Figure 1 are provided elow i Sectios III ad V. 2

4 eighorhoods ad time, we exploit the staggered timig of the policy across the 18 UPPs y joitly estimatig the idividual effect of each oe o house prices ad crime. We fid that, coditioal o a UPP eig istalled eary, house ad apartmet sales prices icreased y a average of 5-10 percet, homicides decreased y a average of percet, ad roeries decreased y a average of roughly percet. To gai perspective o the ecoomic sigificace of the decrease i crime due to the UPPs, we use our regressio results to costruct couterfactual price ad crime rates ad, with those, city-wide statistics. I the asece of the UPPs, the overall house price idex i Rio would have grow aout 15 percet slower sice 2008, ad homicide ad roery rates would have falle y aout 14 ad 20 percet less tha they did, respectively. We ote that sice we do ot oserve house prices iside the favelas themselves, our estimated price effects are quite likely to e uderestimates of the true city-wide effects. The empirical results, otwithstadig some heterogeeity i the effectiveess of idividual UPP statios, cofirm widely reported aecdotes of aated violece ad of skyrocketig residetial property prices i the formal housig markets surroudig the favelas. Our fidigs complemet ad exted previous work o the effectiveess of the UPPs. Based o household survey data, Neri (2011a) foud that retal prices withi all favelas i Rio rose y aout 7 percet etwee 2007 ad However, those results are ot specific to each commuity protected y a UPP ad do ot cotrol for secular treds i the Rio housig market. The positive exteralities of UPPs are also explored i Cuha ad Mello (2011), which focuses o the formalizatio of services provisio i a favela followig the istallatio of a UPP. I additio to the direct valuatio of disameities due to crime that we emphasize elow, formalizatio ad ura regularizatio are other importat chaels through which crime reductios affect property prices, ad which are captured i our estimates of the effect of the UPPs. Havig estalished that the UPPs iflueced crime ad house prices i opposite directios (that is, that the UPPs seem to e a reasoale istrumet for the effect of crime o house prices), we use our estimates to aalyze the associatio etwee crime ad the dispersio of house prices. We preset a model of property valuatio i which there are dimiishig returs to crime reductio; this implies that properties with either high iitial crime rates or low ameity values have disproportioately large icreases i price for a give declie i crime which, i 3

5 tur, lowers iequality amog properties. The mechaism i the model that gives rise to dimiishig returs is the iclusio of historical crime rates as a determiat of curret property values. This treatmet of the dyamic trasmissio of crime rates ito house prices is quite similar i spirit to the way Besley ad Mueller (2011) model the umer of killigs due to coflict as a fuctio of the latet state of the peace process i Norther Irelad. I that model, the persistece of crime i a particular area has a earig o what sigal agets take from a chage i the umer of killigs aout the proaility of eterig a state of peace, ad hece o the trasmissio of the rate of killigs ito house prices. I our model, we have a simpler treatmet of agets' expectatios ut the trasmissio of a chage i the crime rate ito prices depeds similarly o the history of crime, which is a additioal state variale. Thus, curret ad future cosumptio flows from housig deped o oth the level ad duratio of crime rates i the past; lower iitial crime rates with low historical duratio gives rise to the iggest icreases i price whe the crime rate declies. We documet that the disparity i house prices i Rio did i fact declie followig the implemetatio of the UPP policy. A Gii coefficiet costructed with the actual ad couterfactual house prices descried aove shows that the disparity i house prices across eighorhoods has ee fallig faster after istallatio of the UPPs tha for the couterfactual Gii. Moreover, i several eighorhoods with a UPP eary, we fid evidece that the dispersio i property prices withi those eighorhoods arrowed, suggestig that eve withi more homogeeous sets of properties the lowest valued oes are most sesitive to a chage i the crime rate. This paper cotriutes to several areas of active research, ragig from studies of the ecoomics of coflict to the aalysis of the wealth distriutio. Most closely related are the works idetifyig the impact of crime ad violece o property prices, with the paper y Besley ad Mueller (2011) as a closest atecedet; as elow, Besley ad Mueller (2011) exploit oth spatial ad temporal variatio i crime data to idetify the effect o house prices, ad they provide a model i which the respose of property prices depeds o the level ad persistece of historical crime rates. The preset study uses more disaggregate price data, at the level of eighorhoods i Rio, ad has a differet modelig approach that focuses more o the 4

6 implicatios of crime for the dispersio of house prices. To our kowledge, ours is the first study to draw a coectio etwee crime reductio ad wealth iequality. Our empirical measuremet of the crime elasticity of house prices is coected to a sequece of papers estimatig this (largely egative) elasticity. Early examples iclude Thaler (1978), i which a oe stadard deviatio icrease i per capita property crime decreased sigle-family home prices y 3 percet, ad Hellma ad Naroff (1979), i which the elasticity was A kow drawack of these estimates is that they each treated the crime rate as exogeous, which may have iased the elasticity estimates if, for example, crime occurs disproportioately i poorer eighorhoods with low property values or, coversely, if crimials target areas with higher-priced homes. I a recet survey, Ihlafeldt ad Mayock (2009) foud 12 istaces i of a set of 18 empirical studies relatig house prices ad crime that treat crime as exogeous as such, those studies do ot accout for this reverse causality or other sources of edogeeity. Of the recet studies that do istrumet for crime, Gios (2004) ad Tita, Petras ad Greeaum (2006) agai fid a egative sigificat relatioship, a effect that is particularly proouced for violet crimes. We attempt to get aroud issues of cross-sectioal edogeeity y exploitig the time variatio aroud a exogeous policy experimet, the UPPs i Rio. A widely ackowledged ojective of the UPP policy is to icrease the safety aroud key veues for the 2014 soccer World Cup ad 2016 Summer Olympics, the locatios of which are ot systematically related to historical crime rates or the levels of property prices. As such, we will argue that the UPPs are a reasoale istrumet for the effect of crime o house prices. We proceed y estimatig a differece i differeces estimator of property values i eighorhoods with a eary UPP, straddlig the pulic aoucemet that a UPP would e istalled i those eighorhoods. This method for estimatig the (dis)ameities of housig is used aalogously i Lide ad Rockoff (2008) i their study of the effect of the proximity of registered sex-offeders o house prices. 3 Fially, this paper is related to a large literature o the determiats of wealth iequality. Wolff (1992) illustrates that wealth cocetratio ad iequality i the Uited States varied a lot over much of the 20 th cetury (oth icreasig ad decreasig) ad moved fairly closely with 3 There are may other studies that estimate property (dis)ameities more roadly defied to iclude factors such as eviromet. Boyle ad Kiel (2001) provide a thorough, if dated, survey of that literature. 5

7 chages i the icome distriutio. Brazil, i particular, has made great strides recetly to reduce its level of iequality. 4 Our work demostrates a ovel ad potetially importat chael y which policy ca cotriute to chages i the distriutio of wealth. 5 The paper proceeds as follows. The ext sectio provides some ackgroud o the favelas i Rio ad the official madate of the UPP program. Sectio II descries the empirical model, followed y the details of the property price data ad crime data i Sectio III, ad the empirical results i Sectio IV. The valuatio model ad its predictios for the dispersio of house prices, as well as some empirical measures of house price iequality i Rio, ca e foud i Sectio V. Sectio VI cocludes. I. Backgroud o the UPP program Most of Rio de Jaeiro s favelas are situated o the hillsides of the city ad may are located i close proximity to affluet eighorhoods. Both of these factors have made them a favored have for drug gags. By commadig the high groud ad through a mix of cooptatio ad explicit threats, heavily-armed gags have gaied effective cotrol over the residet populatios of certai favelas ad have used these locatios as ases to process, stockpile ad distriute drugs. The profitaility of such traffickig operatios has led to itesifyig territorial dispute, growig levels of violece ad, through ries to protect the operatio from other drug gags ad other illicit activities, the icreasig complicity of the police. Over the past three decades, a complex system has crystallized i which drug-related ad other crimial activities have fed o police corruptio, ad spilled over ito the political area, with drug moey used to fiace politicias i the muicipal ad state legislatures, ad reportedly reachig the highest levels of the state govermet. A ew, reformist state govermet assumed power i 2007, settig improvig security ad reducig the levels of violece as priorities. The selectio of Brazil as the host of the Sice 2001 the Gii coefficiet for Brazil s icome distriutio has decreased mootoically from i 2001 to a estimated 0.53 i 2010 still relatively high compared to 0.36 for Idia or 0.42 for the U.S. For a recet discussio, see Neri (2011). 5 Though our results suggest that crime reductios have eefited the owers of low-priced housig disproportioately, low-wealth households are less likely to ow a home tha wealthier households. I a recet study of U.S. households durig the fiacial crisis, Bricker et al. (2011) show that household wealth declied most severely i the upper percetiles. This reflects oth the dramatic fall i house (ad other asset) prices ad the fact that wealthier families are more likely to ow those assets. I the ottom quartile of wealth, oly 15 percet of families had ay home equity i 2007, compared to 96.8 percet i the top quartile. 6

8 World Cup ad Rio as the seat of the 2016 Olympics added impetus to these ojectives. The ew govermet recogized that achievig these goals etailed dealig simultaeously with the territorial power of drug lords i the favelas as well as police corruptio. There was also a realizatio that a ew security policy for the favelas would have to e more permaet i ature. Previous attempts to comat drug traffic ivolved occasioal icursios ito the favelas for specific operatios, ofte resultig i the deaths of iocet ystaders caught i the crossfire. The core of the ew policy, which took early two years to desig ad deploy, was uilt roud the cocept of territorial occupatio y state forces, ad the istallatio of a large, permaet presece of a ewly traied police force of youg officers utaited y corruptio. 6 presece would maifest itself i a large police statio, a UPP, ad would e preceded y a carefully plaed ad swiftly executed process of expulsio of the drug gags y crack police uits ad special forces. The program, iitiated i Decemer 2008, is cosidered to e a ew paradigm of police actio agaist the ecroachmet of drug gags i the favela commuities. 7 Of the may favelas affected y drug gags, the selectio of which favelas were to receive a UPP was largely a political outcome. This is importat for our empirical idetificatio of the effect of UPPs o crime ad house prices, sice it mitigates the extet of reverse causality etwee our policy variale (the UPP) ad each outcome. I other words, UPPs were ot simply placed i the eighorhoods with the highest crime rates or lowest house prices. Rather, the policy has ee implemeted y prioritizig importat locatios for the World Cup ad Olympic Games, givig geographic factors a domiat role i determiig the locatio of UPPs. This ca e see i the top pael of Figure 2, i which the exact locatios of the 18 existig UPPs are mapped with the gradiet of average apartmet sales prices for each eighorhood i Rio. It is evidet that some UPPs were placed i high-priced eighorhoods while others were placed i low-priced eighorhoods. Similarly for the rate of homicides, show i the ottom pael of Figure 2, UPPs appear i some low homicide eighorhoods i the south zoe, high homicide eighorhoods i the orth zoe, ad eighorhoods with itermediate homicide rates i the 6 Breakig with traditioal repressio techiques, ewly traied officers are taught to e commuity policeme or proximity police y itegratig themselves withi the occupied commuity. Ackowledgig the skepticism ad mistrust that local populatios have historically had with police activity, all UPP staff are ewly admitted ad traied for this specific purpose. This hirig ad traiig practice is cosistet with the idea that UPPs are meat to e the gateway for may other services eyod the suppressio of crimial activity. 7 I a op-ed i the ewspaper Gloo, State Pulic Safety Secretary José Mariao Beltrame compared the UPP program (which he maages) to the Plao Real i 1994, which drastically reduced iflatio ad stailized the ecoomy ( 7 This

9 west. Media commetary o the UPPs has suggested that while some eighorhoods received a UPP due to their high icidece of crime, UPPs were istalled to garer political support for the UPP program ad protect key World Cup locatios i the high- ad middle-icome South Zoe (or Zoa Sul). 8 Implemetatio of a UPP i a give favela occurs i a four-stage process. A similar protocol has ee oserved for most favelas, though it is ot a official stadard. First, the commuity or set of commuities to e occupied is aouced y the police up to 6 moths i advace, though o specific date is give. Secod, a series of aoucemets idicatig the immiece of the occupatio occur, icludig a aoucemet that it will happe i the ext 1-2 weeks. Betwee 4-7 days prior to the occupatio, the specific date is made pulic ad police egi ecirclig the favela(s). Third, heavily-armed Civil ad Military Police, led y elite forces, ivade a favela i the early twilight hours ad expel the drug traffickers i the eighorhood. Over the ext few days of week, they systematically sweep the area to clear ay remaiig crimials or cotraad ad set up a temporary statio. Fially, the permaet physical statio is istalled ad cotrol is haded over to a ew UPP attalio. I the majority of cases, ad as a iteded cosequece of the pre-aoucemets, this process has led to very little violet cofrotatio as crimials have already left the area. As show i Tale 1, 20 favelas have ee occupied ad 18 UPP uits istalled etwee late ad the ed of 2011, as part of a pla to reach 40 UPPs y The rollout of the program has ee fairly steady over time, with a ew occupatio takig place o average every few moths. 9 The territorial footprit of the curret 18 UPPs ecompasses over 50 commuities, cotaiig more tha half of a millio ihaitats. Over 3 thousad police officers are curretly deployed i the program, which is expected to reach 12 thousad y Tale 1 also shows 8 Several sources idicate that geography rather tha crime rate is the domiat factor i determiig the locatio of UPPs. I a iterview i Jauary 2012, the director of commuicatio for the state police force, Frederico Caldas, liked expadig the umer of police at UPPs with the goal of esurig security ahead of the World Cup ad Olympic Games. A prime example of this goal is the ivasio of Rociha, Vidigal ad Chacara do Céu i the affluet (relatively low-crime) South Zoe of Rio, a operatio which is widely cited for protectig tourist ifrastructure. 9 Whe availale i media reports, Tale 1 icludes the date that the UPP was aouced, though i may istaces these dates, if reported, were oly a few days prior to occupatio. Sice the aoucemet data are ot pulished officially, we rely o media reports to ascertai their timig; ewspaper coverage teds to pick up the story of a UPP oly oce ivasio is immiet (i.e., shortly efore the third stage descried aove). 10 I additio, as of the date of this mauscript, the cluster of favelas kow as Complexo do Alemão, cotaiig approximately 125,000 ihaitats, was occupied y 3,000 highly traied army troops, while aother immese tract 8

10 the ames of the eighorhood (or airro) i which the favelas are located. I some istaces, due to their sprawlig layout over hillsides, they are located i more tha oe eighorhood; for example, the favela Sata Marta is located i Botafogo ad Humaitá. Due to the large umer ad proximity of the eighorhoods (there are 153 official airros i the muicipality of Rio), we also list those which have a order withi 2km of the address of each UPP; this sigificatly wides the umer of eighorhood ad the size of the populatio classified as close to a UPP. Fially, cogizat of the fact that oe eighorhood might share a order with aother where a UPP is located ut oly have a small fractio of its populatio livig close to the order, we use ArcGIS mappig software to compute the distace etwee each UPP ad the cetral poit (called cetroid ) i each airro. The set of eighorhoods with a cetroid withi 2km of a UPP statio is a suset of those with a order withi 2km, ad is used elow as aother measure of UPP proximity. II. Estimatig the effect of UPPs o property prices ad crime rates Our aselie measure of the effect of the UPPs is a differece i differeces estimator for all of the 18 UPPs istalled etwee Novemer 2008 ad Novemer Our elemetal uit of measure is the mothly average listig price of a property with certai characteristics, such as dwellig type (i.e., apartmet or house) ad umer of edrooms, i a give eighorhood. For istace, oe price oservatio would e the average price of a 3-edroom apartmet i Botafogo i Jauary The localized ature of the policy i specific favelas as well as its sequetial rollout leads to variatio i property prices across eighorhoods, time ad property characteristics. The elasticity of house prices to the istallatio of a UPP is estimated usig the followig specificatio: (1) l i i i ( Pt ) = α 0 + α1, UPP + α 2, Dist + α3 UPP * Dist + γ tz + κ + δt + ε, t of commuities i Rociha, Vidigal ad Chácara do Céu, occupied i Novemer 2011, had a ukow umer of crack uits still searchig for drugs ad weapos. 11 Two large clusters of commuities, Complexo do Alemão ad Rociha/Vidigal, are excluded as o UPP was istalled yet y the ed of 2011 (as of Jauary 2012, two ew UPPs were estalished i Vidigal ad Chácara do Céu, to e followed y Rociha). Give sigificat differeces i the size ad scope of those UPPs, it is difficult to apply our fidigs for the first 18 UPPs to them (or, for that matter, to susequet UPPs). However, uder the assumptio that susequet UPPs are o average as effective as previous oes, our estimates of the overall effect of UPPs o house prices ad crime uderstate the effects of the program as a whole. 9

11 where is the average sale price of property type i i airro i moth t, UPP is a dummy takig the value 1 for periods after the occupatio date (or, if availale, the aoucemet date) of the th UPP, Dist is a dummy deotig proximity of airro to UPP, i Z is a vector of cotrol variales for property characteristics, icludig the type of property (i.e., apartmet or house) ad the umer of rooms, ad ε is a mea zero error. A full set of airro ( κ ) ad time i t ( δ t ) fixed effects are icluded to asor commo eighorhood factors ad aggregate moth-tomoth variatio i the housig market, respectively. Sice the average prices of properties i the same airro ut with differet characteristics might have correlated errors, the stadard errors of the estimates are clustered y airro ad period. Our treatmet group cosists of properties that are i eighorhoods withi close proximity of a UPP. Ideally, oe would use the exact distace etwee a property s address ad the UPP as the measure of proximity. However, give that we oserve property prices oly withi the geographic uit of eighorhood, we operate uder a rage of assumptios aout how close adjacet eighorhoods are to a UPP. A set of three specificatios of the variale Dist is used: (i) the eighorhood i which the UPP is located (the peultimate colum of Tale 1), (ii) eighorhoods with a order withi 2km of the UPP address (the fial colum of Tale 1), ad (iii) eighorhoods whose cetroid is withi 2km of the UPP address. The ifluece of each UPP o house prices is give y the coefficiets etwee Dist ad UPP. α 3,, the iteractio α 3, ca e iterpreted as the percet icrease i average property prices i eighorhoods proximate to a UPP after its istallatio relative to the chage i price for properties that are ot located close to a UPP. We idetify the separate cotriutio of 15 out of 18 UPPs to house prices. I three istaces, the overlap of the timig or the eighorhood of multiple UPPs is such that the effect of each UPP caot e separately idetified. For example, the occupatio of the favela i São Carlos (i the eighorhoods Estácio ad Rio Comprido) was aouced i Feruary 2011, the same moth as favelas i Coroa, Fallet ad Fogueteiro (i the eighorhood Rio Comprido). The resultig coliearity led us to merge these two UPPs ito a sigle estimate of α 3,. We did likewise for Borel, Formiga ad Salgueiro, three UPPs i the Tijuca eighorhood itroduced withi moths of oe aother i

12 A importat cocer aout the use of average price chages (of a sample of listigs) to measure treds i the housig market is that the compositio or quality of the sample could e chagig over time. Chagig property compositio, such as the etry of a high-quality, highpriced property ito the listigs, would affect the average price used i the estimatio of (1) eve if the true quality-adjusted price of housig had ot chaged. This is a relevat issue if compositio is chagig systematically i a way that is correlated with the timig of the UPP policy. It is therefore importat to check whether housig compositio might e chagig i respose to the itroductio of UPPs ad to correct for this selectio ias i equatio (1). To do so, we use the two-stage correctio procedure y Heckma (1976, 1979) ad estimate the proaility that a property with certai characteristics (i.e., eighorhood, type, # rooms) is listed. The first stage proit regressio is of the form: (2) z + i t = α + 0 4, α UPP + 1, α UPP * Dist α 2, Dist * StateListigs t + α UPP * Dist 3, + κ + δ + ε t i t + γ Z t i where i i 1 if Pt = > 0 z t. 0 otherwise I (2), the treatmet group is additioally iteracted with the total umer of property listigs o ZAP.com.r i the state of Rio de Jaeiro. We thus allow listigs to e a fuctio of all of the explaatory variales i (1), icludig the UPP policy, as well as ay factors that affect the aggregate listigs o ZAP.com.r. Stated differetly, the first stage allows for the possiility that crime has a outsized ifluece o listigs whe other factors affectig listigs are high. We implemet (2) as a stadard Heckma correctio factor y addig the iverse Mills ratio of this regressio as a additioal explaatory variale i (1). We ote that the sample selectio correctio, while useful as a gauge for the determiats of listigs for groups of properties (e.g., 2- versus 3-edroom apartmets i Botafogo), still does ot accout for compositioal effects withi those groups (e.g., high- versus low-quality 2-edroom apartmets i Botafogo). These compositioal effects might e taitig our results elow, though it is ot ovious whether oe would expect a decrease i crime iduced y the UPPs to cause higher- or lower-quality properties withi each group to e listed. 11

13 Fially, estimates of the UPP effect o house prices from equatios (1) ad (2) are used to costruct a series of couterfactual average property prices; these are the prices that the regressios suggest would have ee oserved i the asece of the UPP policy. The growth rate of the couterfactual price, i P ~ t, i a give period is costructed as follows: ~ i Pt (3) ~ = i exp l( ˆ i P Pt ) t 1 i where ( ) P ~ P i t i t 1 ˆ α UPP * Dist 3, ~ i Pt 1 if if UPP * Dist UPP * Dist l Pˆt is the predicted value of average prices from the regressios (1) or (2), ˆα 3, is the estimated elasticity of property prices due to the UPPs ad i i P0 P0 = 0 > 0 ~ = for every property type ad eighorhood. Couterfactual prices are equal to oserved prices prior to the arrival of a UPP or i eighorhoods that do ot have a UPP eary (i.e., UPP * Dist = 0 ) ut sutract out the estimated effect of the UPP otherwise. Therefore, α 0 implies that couterfactual prices are lower tha oserved prices. ˆ 3, > Our treatmet of crime rates is roadly aalogous to that of house prices i equatio (1), except we sustitute the eighorhood crime rate, cotrols for property characteristics: r t, for property prices ad remove the (4) l ( rt ) = α + α1, UPP + α 2, Dist + α3 UPP * Dist + κ + δt + ε, t 0. Sice crime rates are reported i each eighorhood ad time period, there is o eed to correct for selectio issues. Give estimates of (4), the growth of couterfactual crime rates ca e computed as: 12

14 (5) ~ r ~ r t t 1 = exp l ( rˆ ) t 1 r r~ t t 1 ˆ α UPP * Dist 3, ~ rt 1 if if UPP * Dist UPP * Dist = 0 > 0 which differs from (3) oly i iterpretatio. If UPPs cause a decrease i crime rates, the ˆ 3, < α 0 ad the couterfactual crime rates lie aove the oserved oes. III. House prices ad crime data This sectio descries the sources of detailed property price ad crime frequecy data. a) ZAP.com.r price data Real estate prices are draw from the imóveis sectio of the olie classifieds wesite, ZAP. We were provided with a cofidetial extract from the ZAP dataase cotaiig average mothly offer prices of real estate i Rio for the period March 2007 through August Each oservatio is the simple average of prices across listigs with specific characteristics such as eighorhood, type of property ad umer of rooms. Prices are i uits of Brazilia reais per square meter. We were also provided with the umer of listigs ad the stadard deviatio of the listig prices for each oservatio. Altogether, the extract cotais 54,064 mothly price oservatios of apartmets ad houses for sale spaig all 153 eighorhoods i Rio. Uderlyig these oservatios are 3,302,036 idividual property listigs. The umers of eighorhoods, property types ad trasactio types have ee growig steadily over the course of the sample (that is, the pael is ualaced), which likely reflects the growth of ZAP as a advertisig service provider, ad ot ecessarily the umer of properties ought ad sold i the real estate market over time. Tale 2 summarizes the average listig prices y property type ad umer of rooms, with the average prices at the of the ottom pael weighted y the umer of listigs uderlyig each oservatio. A few patters i the compositio of the data are worth otig. First, aout three quarters of the listigs are apartmets. Secod, for oth apartmets ad houses, the price oservatios are fairly evely distriuted across properties with 1 through 4 edrooms, with roughly percet of 13

15 the oservatios fittig ito each oe of those categories. Third, particularly i the weighted statistics, there are large differeces i price levels across property types; house prices per square meter of R$2,846 are a full third lower tha apartmet prices. Ad fourth, the price per square meter teds to icrease with the umer of edrooms for apartmets ad houses, otwithstadig a relatively high price level for 1-edroom apartmets i the sample. The richess of the dataset is further demostrated y the large umers of listigs i most eighorhoods, with eve the 100 th largest airro y umer of listigs cotaiig over 1,000 listigs over the sample period. There is also a fair amout of heterogeeity amog wealthy ad poorer eighorhoods, with the average property price i the most expesive eighorhood of almost 9,000 reais per square meter dwarfig the average price of 1,300 i the least expesive. Iterestigly, lookig at the raw price data already suggests some effect of the UPPs o property prices. Figure 3 shows oe such example for the average price of houses i the eighorhoods Leme ad Copacaaa. The vertical lies show the aoucemet dates of two UPPs, the first directly adjacet to formal commuities i Leme ad the secod auttig Copacaaa ad Ipaema. Immediately followig the aoucemet of Chapéu-Magueira i Leme, house prices jumped y aout oe third i Leme ad oly edged upwards slightly i Copacaaa. Coversely, after the occupatio of Pavão-Pavãoziho i Copacaaa, house prices jumped y over 25 percet i Copacaaa ad stayed roughly flat i Leme. It is precisely price resposes like these that our differece i differeces estimator will attriute to the UPPs; i each case, the airro cotaiig the UPP (the treatmet group) icreased relative to the other (the cotrol) after the occupatio of the favela. A more sutle, ad more commo, case is illustrated i Figure 4, showig the average apartmet prices i four eighorhoods as well as vertical lies deotig the occupatios of the aforemetioed favelas. Oe differece i Figure 4 is that the prices all have secular upward treds ad so the differece i differeces estimator will attriute to the UPPs icreases i the slope of the treatmet group price idex relative to the cotrol group idices after the istallatio of each UPP. Aother differece i Figure 4 is the iclusio of prices for Botafogo ad Lelo, eighorhoods that share a order with Copacaaa ad Ipaema, respectively. The iclusio of these eighorhoods i the treatmet group (or ot) will clearly have a earig o the measured price effects of the UPPs i this example. We will cosider oth cases elow. 14

16 ) ISP crime data The crime data are compiled y the Istitute of Pulic Safety of Rio de Jaeiro (the Istituto de Seguraça Púlica do Rio de Jaeiro, or ISP). ISP is resposile for cosolidatig ad pulishig a array of official statistical data pertaiig to pulic safety. The crime data that we use are draw origially from icidet reports writte y the state civil police ad the aggregated ito groups of differet types of crime; 12 the assigmet of icidets to categories of crime udergoes quality cotrol checks y the civil police s iteral affairs departmet. From these couts of differet crime types i each moth from Jauary 2007 to Jue 2011, we costruct two aggregates: homicides ad roeries. Homicides is a aggregate of icidet couts for three crime classificatios related to violet acts with the itet to kill: murder, attempted murder ad vehicular homicide. Roeries is a aggregate of icidet couts for the followig crime classificatios: roery of a commercial estalishmet, roery of a residecy, roery of a vehicle, roery of a passer-y, roery of a ak, roery of a ATM, roery where the victim is drive to a fiacial istitutio, extortio, kidappig, extortio through kidappig, ad emezzlemet. The geographic uit of measure for the crime statistics is the coverage area of specific police statios, called delegacias de polícia, or DP. There are forty DPs i the metropolita area of Rio, with most of them coverig more tha oe eighorhood (the level of aggregatio for the ZAP property prices), ad a small umer that oly cover a portio of oe eighorhood. Each DP, i tur, is classified withi a roader geographic uit called a Itegrated Area of Pulic Safety, or AISP. 13 To compute crime rates, we divide the crime couts for each DP y the correspodig mothly AISP populatio data The categories of crime use are stadardized groups defied y the Natioal Secretary of Pulic Safety (SENASP). 13 The official cocordace etwee eighorhoods, DP s ad AISP s ca e foud at the followig wesite: 14 We ote that the computatio of eighorhood-level crime statistics usig DP- ad AISP-level statistics itroduces some imprecisio ito our measure of the icidece of crime. Specifically, each eighorhood is assumed to have the crime data reported y its DP. This iases the crime statistics upwards as each eighorhood withi a DP is erroeously assiged some crimes from other eighorhoods that share their DP. O the other had, the populatio data are oly availale at the AISP level, which typically cosist of multiple DP s. By assumig that each eighorhood has the populatio level of its etire AISP, the crime rate statistic is iased dowward. O alace, the average crime rates that we compile are somewhat elow their more aggregate pulished couterparts, though oly slightly. 15

17 Figure 5 shows the relatioship etwee our two measures of the crime rate ad average sale prices for apartmets. Each poit is the average crime rate ad property price for a particular eighorhood, averagig across time periods. As see i the top pael, homicide rates rage from 4 to 188 homicides per 100,000 people per year, while average roery rates vary from 48 to 4,333 per 100,000 people per year. For oth crimes, the distriutio is positively skewed, with most eighorhoods havig low rates ad oly a few havig very high rates. Turig to the correlatio etwee crime ad apartmet prices, the figures demostrate that havig a low crime rate is ecessary though ot sufficiet for a eighorhood to have a high average property price; while there are o eighorhoods with a high crime rate ad a high property price, there are several eighorhoods with low crime ad low price. Oe might surmise from this oservatio that crime is ut oe ameity affectig property prices, which oly ecomes a domiat pricedetermiig factor at very high rates. The triagles i Figure 5 show the eighorhoods where oe or more of the UPPs are located. It is worth otig that the UPP statios are located i eighorhoods that vary alog oth dimesios. The statios have ee uilt i low ad high crime eighorhoods, as well as i eighorhoods with low ad high average property prices. The square i each pael shows the average price ad crime rate for the etire muicipality of Rio, weightig each eighorhood y its average AISP populatio over the course of the sample. The average price for a apartmets is R$2,152 per square meter, lower tha the weighted average price y listigs computed i Tale 2, while the average crime rates are 27 ad 538 icidets per 100,000 people per year for homicides ad roeries, respectively. Despite the more aggregate data that we use to costruct eighorhood crime rates, the average homicide rate falls withi the rage (aleit o the low ed) of the 42 ad 27 homicides per 100,000 rate reported y Waiselfisz (2011) for the city as a whole i 2007 ad 2010, respectively. 15 IV. Results Estimates of α3, are show i Tale 3 for five differet specificatios of the property price regressios. Each colum cotais 15 coefficiets for the iteractio term of the UPP time 15 A closer aalogue to pulic homicide statistics is the aggregate of DP s (istead of eighorhoods) usig a arrower defiitio of homicide (excludig attempted murder ad vehicular maslaughter). Doig so with the ISP data yields a estimate of 31 homicides per 100,000 per year over the sample period. 16

18 dummy ad proximity measure, oe for each UPP statio with certai UPPs pooled as descried aove. The aselie specificatio, (I), is a OLS regressio of equatio (1) where the proximity measure is the eighorhood(s) i which the favela receivig the UPP is located. We fid that 11 of 15 coefficiets o the differece i differeces term are positive ad statistically sigificat compared to two coefficiets which were egative ad sigificat. A F-test of the hypothesis that the sum of α 3, is zero is strogly rejected ad the sum of price effects divided y the 18 UPP statios is 5.8 percet. I other words, coditioal o eig i the eighorhood of a UPP, property prices icreased y 5.8 percet more tha i the rest of the city after the UPP was aouced. There is also a fair amout of heterogeeity i the price resposes across UPPs, from 6 percet followig the Batam UPP to the 21 percet jump followig the Chapéu- Magueira UPP. The estimates of price effects ad their distriutio across UPPs i (I) are quite stale across estimatio methodologies. Colum (II) shows estimates of equatio (1) usig weighted least squares ad the umer of listigs per oservatio as weights. The ituitio for this choice of weights is that oservatios with very few uderlyig listigs may e more proe to measuremet error tha those where may properties are listed. To further cotrol for the fact that the umer of listigs per oservatio could e edogeous to the UPP policy, we use the total umer of listigs i 2008 (o the eve of the UPP policy) for each airro-type-# rooms oservatio, holdig this costat throughout. The average effect of a UPP icreases to almost 10 percet, with 12 positive ad sigificat estimates ad oly 1 egative ad sigificat oe. A eve more careful treatmet of possile selectio effects ito the ZAP listig sample models the proaility of a give type of oservatio havig o-zero listigs, as i equatio (2). Colum (III) shows the results of the two stage procedure usig the Heckma correctio factor. I the first-stage selectio regressio, a F-test with ull-hypothesis α 0 is rejected ad ˆ 4, < α 17 4, = 0. I other words, the types of properties with positive umers of listigs i the treatmet group vary systematically with the aggregate ZAP listigs i the dataset. Moreover, the estimate of rho, which is the correlatio etwee the errors of the selectio ad outcome (secod-stage) equatios is statistically greater tha zero, which suggests that a sample selectio model efits the price data. Applyig the correctio factor, the resultig margial effects of the

19 UPP are show i colum (III), ad are ot sustatially differet from their OLS aalogues i colum (I). The estimates of price effects are also roust to alterative specificatios of the treatmet group, though these chages i specificatio have a earig o the iterpretatio of the magitudes of the resultig estimates. Colums (IV) reports the UPP price effects for eighorhoods with a order withi 2 kilometers of each UPP statio, respectively, estimated usig weighted least squares. 16 Strikigly, the iclusio of may more eighorhoods i the treatmet group (relative to (II)) yields estimates that, although lower tha efore, are still positive ad very sigificat. I colum (IV), 13 of 15 UPPs have a positive ad sigificat coefficiet. Additioally, the variatio i the size of price effects across UPPs has goe dow, with the majority of estimates i the rage of 5 to 10 percet. It is ituitive that the level of the effect would decrease whe spread across a larger geographic area. This pheomeo is illustrated rather dramatically y Chapéu-Magueira; the order estimate, which icludes Copacaaa, Urca ad Botafogo i the treatmet group (i additio to Leme), is less tha half the size of the airro-ased estimate. O the other had, precisely ecause the effect is applied to a much larger geographic ase of listigs, it may well have a larger overall ifluece o Rio house prices. Our couterfactual price series elow will e iformative of which specificatio implies a larger effect o the overall price idex. Fially, colum (V) shows the results for airros with a cetroid withi 2km of a UPP statio. While the average effect of 5.7 percet is i lie with the other specificatios, o a UPPy-UPP asis results teded to vary without much discerale patter. I some istaces, the effects were i etwee those of colums (II) ad (IV) as oe might expect from the additio of a itermediate umer of airros ito the treatmet group. I others, the results were otaly stroger or weaker. I sum, while the average effect held fairly steady at upwards of 5 percet, the disparities across UPPs varied from specificatio to specificatio. To put the overall effect of the UPP policy ito perspective, we traslate the results of each specificatio ito couterfactual measures of house price growth rates as descried i equatio (3), the aggregate across property types for the etire city of Rio as follows: 16 I the cases of pooled UPPs, the treatmet group icludes all airros with orders withi 2km of ay of the pooled statios. 18

20 ~ (6) Pt = w where the weight i 19 i 2008 ~ P i t i w 2008 is the total umer of listigs for property type i i airro durig the year The average oserved price level for house ad apartmet sales i the city of Rio is illustrated y the solid red lie i Figure 6. The price level rose approximately 100 percet etwee Jauary 2008 ad August 2011 ad, otwithstadig methodological ad data iput differeces, our computed idex is i the allpark (though elow) pulished figures of a 130 percet icrease i sales price for the city of Rio over the same period. 17 We will focus o the chage i overall average couterfactual prices (computed i the same maer) relative to the oserved aggregate. Each of the couterfactual series i Figure 6 relates to a specificatio i Tale 3. What is strikig aout the couterfactuals is that, with the exceptio of the order measure of proximity, the series are quite close together. This implies that the heterogeeity i: (i) ecoometric techique, (ii) oservatio weights, ad (iii) the geographic scope of the cetroid-ased treatmet group across specificatios does ot amout to much whe applied to the populatio of prices. The fact that the order-ased couterfactual lies elow the rest idicates that the estimated overall ifluece of the UPPs is ideed sesitive to geographic scope whe applied as roadly as i that specificatio. I other words, lower average price effects i the order specificatio are more tha fully offset y the larger ase of eighorhoods that they are eig applied to. All i all, the rage of couterfactuals implies that without the UPP program the cumulative growth rate of property prices i Rio would have ee etwee 12 ad 22 percet lower. Takig a average across specificatios, the UPPs accouted for 15 percet of the growth i prices sice Jauary The aalogous results for the effect of the UPPs o crime rates, estimated with equatio (3), are show i Tale 4. As metioed aove, sice the crime data are iclusive of all airros ad dates, either the selectio model or weighted least squares specificatios are employed. This 17 FIPEZAP, a orgaizatio that pulishes aggregate price idexes ased o the same uderlyig data that we use here (Rio is oe of several regios for which FIPEZAP pulishes a idex), has a differet way of aggregatig across types of property ad airros. While the idex formula we use is the same (i.e., a weighted average of the levels of price oservatios) FIPEZAP uses household icome data from the 2000 cesus to weight across geographic areas ad umer of rooms. I the asece of those weights, we use the umer of listigs withi airro/rooms comiatios i Further, FIPEZAP reduces the amout of oise i the data y oly icludig oservatios with greater tha 5 listigs ad the takig the media value of oservatios over the past 3 moths, steps which we do ot take here. (For more iformatio o costructio of the FIPEZAP idex, see Notas Metodologicas i Fudacao Istituto de Pesquisas Ecoomicas (2011)).

21 leaves us with three specificatios of the scope of the treatmet group for oth roeries ad homicides. The overall effects imply that the UPPs had large egative effects o the levels of each crime, o the order of percet for homicides ad percet for roeries, with the airro-ased estimates at the larger ed of the rage ad the order-ased estimates at the smaller ed. The effect of the UPPs o crime is very heterogeeous across UPPs. I some istaces the effects are very strog, o the order of sixty or eve sevety percet declies i the homicide rate. I other istaces the effect is ot distiguishale from zero. Perhaps due to the lumpiess of the crime data over time, as homicides ad roeries i each airro ad moth ouce aroud ear the zero level, it is difficult to otai precise estimates for each UPP. At the very least, there are o statistically sigificat estimates of crime rates goig up due to a UPP. The ecoomic ad social sigificace of these estimates are illustrated y the couterfactual crime rate estimates i Figure 7. Crime couterfactuals are computed similarly to the price couterfactuals aove, except that they are aggregated across eighorhoods usig 2008 airro populatio measures as weights. Sice the estimates of the UPP elasticity are egative, the actual series (deoted y the red solid lie) lies elow the couterfactual series; the differece idicates the effect of the policy o the city-wide crime rate. The top pael shows that the UPPs were ot the domiat factor i determiig the rate of homicides at the city level, as the couterfactuals track the actual series quite closely. That said, y the ed of the sample the various couterfactual simulatios uaimously show a cotriutio y the UPPs of 14 percet of the declie i homicides sice their peak i May This traslates ito aout 1 homicide per 100,000 people aually, or roughly 60 people i the muicipality of Rio i By ay measure, this is a sustatial umer lives that potetially owe to improved security. Similar oservatios ca e made aout the roery rate i the ottom pael of Figure 7, i which aout 20 percet of the fall i roeries sice mid-2009 was attriutale to the UPPs. Fially, we ca use the property price effects ad crime effects estimated separately i equatios (1) ad (3) to compute the crime elasticity of house prices. A simple way to do so is to compare the estimates of α 3, from each equatio to see whether large crime effects correspoded to large house price effects for a give UPP. Loosely speakig, this is a two-stage istrumetal variale regressio where the first stage regresses crime o the UPP policy variales. Figure 8 illustrates the homicide elasticity of property prices, where each dot matches 20

22 the coefficiets from (1) ad (3) i Tales 3 ad 4, respectively. Due to the imprecisio of the some of the homicide estimates, oly the matches where oth policy coefficiets are statistically sigificat at the 5 percet level are show. There is a clear egative relatioship etwee crime ad property prices, with a te percet declie i the homicide rate correspodig to a 1.8 percet icrease i eary house prices. V. Crime ad house price iequality Havig estalished a empirical relatioship etwee crime ad residetial property prices, we ow derive formal expressios for the chages i the distriutio of house prices due to a discrete chage i the crime rate. We the evaluate the predictios of the model y costructig measures of iequality withi- ad across-airros i the ZAP data. a) A model of house price dispersio The oly source of variatio i the model is a chage i the proaility of a crime occurrig ( r (UPP)) i a give airro ( ) coditioal o the asece or presece of a UPP deoted y UPP {0,1}. Let us assume that the istallatio of a UPP will cause the crime rate to drop to r 1) < r (0). Assumig some disutility from crime, this will have a positive effect o the flow ( of cosumptio services emodied i the property, ad hece o the et preset value of the future stream of those services. The ojective of this exercise is to show that the prices of properties with either low iitial prices or those i eighorhoods with higher iitial crime rates react more strogly to a give chage i the crime rate, compressig the distriutio of property prices. First, we assume that the cosumptio value of housig, deoted U, for a specific property ( i ) depeds o a time-ivariat ameity value, h i, reflectig the property's characteristics, as well as a term summarizig the complete history of crime i the airro i which the property is located: (7) where it i l=0 U = h r ( UPP) t l ( deotes the l -th lag of the crime rate relative to time t. This latter term r t UPP) l 21

23 eters egatively ito the expressio for cosumptio services as the value of the pulic ad due to crime. I that summatio, we also make a importat assumptio aout the dyamics of cosumptio services due to the proaility of a crime; the cotriutio of lagged values of crime to cosumptio services decays geometrically over time, which itroduces some oliearity i the respose of utility ad, as we will show elow, i the respose of house prices to a chage i r. The price of a property is the stream of future cosumptio flows, discouted y a factor β, which for a costat crime rate is: P it U i 0 h i r which rigs us to our first propositio relatig the levels of prices ad crime. Propositio 1: Coditioal o h i, a decrease i the crime rate icreases the price of ay give property. This follows immediately from < 0 P it. r Propositio 1 is a close aalogue to equatio (1) i the empirical exercise aove, which relates house prices to the istallatio of a UPP coditioal o the average price level i each airro ad the average growth rate of prices for properties i the cotrol group. Now let us cosider a chage i the value of UPP from zero to oe. We ca rewrite equatio (7) i terms of oth r (0) ad r (1). Thus, periods after the istallatio of a UPP, the cosumptio services of a property i that period are: 18 (8) U it l ( h, r) = hi rt (1) rt (0) l= 0 l= r(1) = hi + r(1) 1 r(1) 1 r(1) l r(0) r(0) 1 r(0). Notice that the first two terms i this expressio represet the cosumptio services of a property where the etire history of crime i its eighorhood at the level r (1). The comied third ad 18 The airro suscript is dropped i (8) for ease of otatio. 22

24 fourth terms are a fuctio of the relative weight give to more recet history at the lower crime rate ad more distat history at the higher crime rate. The size of the cotriutio of this comied term to total cosumptio services depeds crucially o the size of the differece etwee the old ad ew crime rate; whe there is o differece etwee r (0) ad r (1) the third ad fourth terms cacel out, ad whe r (0) > r (1), the third ad fourth terms are a et egative cotriutio to total cosumptio services. This particular specificatio of housig cosumptio services is a departure from related literature, such as Besley ad Mueller (2011) i which house prices are a liear fuctio of cotemporaeous ameity value ad the crime rate, ut its treatmet of dyamics is quite similar i spirit. I our model, we have a much simpler treatmet of agets' expectatios (i.e., agets assume a costat crime rate goig forward) ut the trasmissio of a oe-time chage i the crime rate ito prices depeds similarly o the history of crime. Specifically, immediately after 1 the chage ( = 1), the value of cosumptio services does ot move all the way up to 1 r (1), ut rather asymptotically approaches that level as the history of r (0) fades. Thus, curret ad future cosumptio flows deped o oth the level ad duratio of crime rates i the past. Propositio 2: Coditioal o a property's ameity value ( h i ), progressively larger decreases i crime lead to smaller margial improvemets i housig cosumptio services. I other words, there are dimiishig margial returs to crime reductio. This follows directly from the cocavity of cosumptio services i the crime rate: U it ( r) < 0,. We ca illustrate this secod propositio graphically usig equatio (8). Figure 9(a) illustrates the crime-related portio of cosumptio services (i.e., the comied secod, third ad fourth terms) for a hypothetical reductio i r from a iitial level of r ( 0) = It is clear from the cocavity of the curves for ay that larger decreases i crime lead to smaller icreases i housig cosumptio services. Further, the degree of curvature i the cosumptio services schedule icreases with, the umer of periods ago that the UPP was istalled. This is a ituitive result. The loger ago the UPP was istalled (i.e., for larger ) agets care less aout icremetal uits of the decrease i crime. 23

25 Dimiishig returs to crime reductio could play a key role i the chagig distriutio of property prices after the estalishmet of the UPPs. For example, two properties with idetical ameity values i airros with differet crime rates could potetially have very differet resposes to the istallatio of a UPP ad a commo decrease i crime. The property i the airro with a high crime rate will oth have lower iitial prices ad a larger chage i its cosumptio services, which will have the effect of compressig the distriutio of property prices across airros. I priciple, this same effect would apply for two properties withi the same airro were they to face either differet iitial levels of crime or differet chages i their crime rate. However, i light of the fact that we oly oserve airro-level data, we have assumed i equatios (7) ad (8) that the crime rate affects the values of all properties withi a airro uiformly. This assumptio eed ot imply that the distriutio of house prices withi airros is uaffected y a uiform reductio i crime. As the ext propositio shows, variatio i the ameity value of properties withi a airro is sufficiet to create cocavity i house prices relative to the crime rate. This is ecause whe ameity values are higher, the percet chage i crime-related cosumptio services after a decrease i crime is applied to a higher (timeivariat) ase, implyig a lower percet chage i the sum of the ameity value ad crimerelated cosumptio services. Propositio 3: Coditioal o the iitial level ad chage i the crime rate (i.e., r (0) ad r (1) ), properties with lower ameity value, h i, will have a higher percet chage i price. Proof: We will first eed to derive a expressio for the price of a property immediately after the estalishmet of a UPP. The et preset value of future cosumptio flows is: P it = = 0 β 1 = hi 1 β ( U ( ) ) it 1 r(1) r(0) + 1 r(1) (1 r(1))(1 βr(1)) (1 r(0))(1 βr(0)) Therefore, the percet chage i a property whe a UPP is estalished is: 24

26 dpit P it = = Pit ( UPP = 1) Pit ( UPP = 0) P ( UPP = 0) 1 1 β it 1 1 r (1) r(0) ( 1 r (1) 1 r (0) ) + ((1 r (1))(1 βr(1)) ) ((1 r(0))(1 βr (0)) ) 1 1 ( h ) 1 β i 1 r (0) which is clearly depedet o the level of dpit hi h, with < 0 i dhi Pit. QED. Figure 9() illustrates these "withi-airro" effects of heterogeeous ameity values, agai for a hypothetical drop i crime from r ( 0) = ad for a rage of h i etwee 6 ad 26. As the ameity value of a property goes up, the percetage chage goes dow at a decreasig rate. The lowest quality properties, those with the lowest ameity value, therefore stad to gai the most from decreases i crime. ) Empirical results Propositios 2 ad 3 suggest that properties of lower value (whether due to high crime rates or low ameity values), will appreciate disproportioately followig a decrease i the crime rate. This would maifest itself oth as a decrease i dispersio of property prices city-wide as well as the dispersio of prices withi-eighorhoods. To evaluate this claim, we egi y computig a measure of house price iequality across all of the eighorhoods i Rio. We choose a commo measure of wealth dispersio, the Gii coefficiet, computed usig a formula similar to the oe i Deato (1997): (9) Gii t = N N N( N 1) N = 1 N w = w ~ X t Pt ~ P t 2008 where N is the umer of eighorhoods, ~ P t = i w ~ i i 2008 Pt is the average price for either apartmets or houses i each eighorhood, ad X t is the price rakig of each eighorhood relative to all others i that period. Prices for a property with a certai umer of rooms are aggregated usig the total umer of listigs of that property type ad eighorhood i 2008 (deoted 25

27 i w 2008 ). The, to aggregate across eighorhoods, prices are weighted y their 2008 populatio (deoted w 2008 ). 19 Higher values idicate greater iequality of prices across eighorhoods. Figure 10 illustrates the Gii coefficiet for actual ad couterfactual city-wide property prices. The solid red lie shows that dispersio has falle y aout two ad a half poits sice the egiig of 2008, from 0.29 to To put these umers ito cotext, recall that the atioal measure of icome iequality i Brazil fell from aout 0.60 i 2001 to 0.53 i The level of icome iequality is higher sice we oly oserve listigs for formal property markets; were we to oserve the lower average real estate prices i the favelas our measure would likely e much more skewed. Istead, oe should focus o the magitude of the chage i iequality over time. Over the course of the 2000 s, which was a very prosperous decade for Brazil i which millios of people where lifted out of poverty, the Gii coefficiet fell y oly 5 poits. We therefore might expect that eve small chages i the Gii coefficiet for wealth are ecoomically meaigful. I terms of the cotriutio of the UPPs to fallig iequality, each couterfactual Gii coefficiet i Figure 10 lies aove the actual Gii y the ed of the sample. This idicates that the UPPs accouted for some portio of the decreasig iequality. That said, there is a lot of heterogeeity across couterfactual specificatios; this is a fuctio of the umer of eighorhoods that the estimates are applied to, with the airro-ased estimates havig the smallest geographic scope ad the order-ased estimates the largest. For the airro-ased estimates the cotriutio equals aout oe quarter of the declie i iequality sice 2008 ad for the cetroid-ased estimates, the cotriutio is aout 45 percet. I the most extreme case, applyig the UPP price effects to all eighorhoods with a order withi 2km of a UPP implies that couterfactual iequality would e ack at its iitial level y the ed of the sample period, a 100 percet cotriutio y the UPPs Sice the pael of price oservatios is highly ualaced (i.e., there are may eighorhoods that were ot iitially i the sample ut etered over the course of the past three years) N icludes oly those that appeared i the ZAP data i Jauary It is importat to have a alaced pael to cotrol for the chagig compositio of listigs o ZAP over time; for istace, the systematic etry of low-priced eighorhoods ito the ZAP listigs could icrease measured iequality eve if the uderlyig level of iequality amog icumet eighorhoods had ot chaged. 20 A decrease i the house price Gii coefficiet of half a poit to 2.5 poits due to the UPPs seems to compare favoraly with other policies aimed at reducig iequality. For example, Soares et al. (2006) decompose the 26

28 Propositio 3, which predicts chages i withi-airro dispersio arisig from heterogeeous ameity values, is ot straightforward to measure with airro-level house prices data. However, usig the stadard deviatio of prices for each oservatio provided y ZAP comied with the estimates of the price effects of the UPPs i Tale 3, oe ca ifer whether low-priced properties respoded disproportioately to the UPPs. Replicatig our aalysis of the price effects o UPP levels, we estimate the followig differece i differeces regressio: (10) l i i i ( stdevt ( Pt ) = α 0 + α1, UPP + α 2, Dist + α3 UPP * Dist + γ tz + κ + δt + ε, t where the oly distictio etwee equatios (10) ad (1) is that the depedet variale is the log of the airro-level stadard deviatio of property prices as opposed to the log of the average price. A egative estimate of α 3, meas that the stadard deviatio of prices declied due to the UPP at the same time as the average price icreased (Tale 3). These two oservatios together ecessarily imply that the lower-priced properties grew faster tha the higher-priced properties i eighorhoods with a UPP eary. Tale 5 shows the estimates of α 3, i equatio (10). For several UPPs the airro-level estimates do suggest that prices codesed after the UPP. For istace, after the first UPP i Sata Marta the stadard deviatio of prices declied y 2-4 percet i Botafogo ad Humaita while average prices icreased y 10 percet i those eighorhoods. Though the same ca e said for aout 5 other UPPs, several airros experieced icreases i price dispersio ad the order- ad cetroid-ased estimates are eve less supportive of decliig price dispersio. The weighted least squares specificatio, alog with the order ad cetroid specificatios have positive average effects across UPPs. I sum, while city-wide price iequality appears to have declied due to the UPPs, the sig of the effect o withi-airro price dispersio is amiguous across specificatios. Oe likely difficulty with testig Propositio 3 is that the iitial crime rates ad chages i crime are ulikely to e uiform withi a eighorhood, much less for differet distaces from a UPP. It is also possile that ameity values are correlated with UPP distace withi a eighorhood i a atioal icome Gii coefficiet ito compoets owig to the well-regarded Bolsa Família program of coditioal cash trasfers. They fid that Bolsa Família was resposile for a reductio of.571 Gii poits, or 21% of the total fall i the atioal Gii coefficiet from 1995 to

29 way that makes the effect of crime o dispersio difficult to idetify. Both of these measuremet issues suggest that iequality could e etter idetified usig more disaggregate data o idividual listigs, which is a worthwhile approach for future work o this suject. VI. Cocludig remarks The asis of the UPP program is reestalishig the rule of law where order was dictated y groups egaged i crimial activity. Destailizig such groups y removig their physical domai over the favelas was meat to: (i) reduce crimiality ad violece; (ii) ease availaility ad access to pulic services, icludig health, educatio, saitatio; (iii) create a etter eviromet for usiess ad commerce (ot limited to popular social tourism); ad (iv) geerally improve quality of life. The UPP policy has ee widely regarded as eig a successful strategy to achieve these goals, ad we have demostrated that the associated positive exteralities maifested themselves i property prices. Betwee 2008 ad mid-2011, we estimate that the UPPs accouted for aout 15 percet of price growth i Rio s formal property markets, a oservatio which we lik to the cotriutio of the UPPs to fallig crime rates. We the descrie ad documet the relatioship etwee crime ad property price iequality. A key theoretical isight is that icorporatig historical crime rates ito a property s valuatio causes the rate of price adjustmet followig a decrease i crime to deped o the property s iitial value. Thus price iequality ca potetially decrease as low-priced properties react more to crime reductios. Ideed, the UPPs accout for a o-trivial portio of the tred declie i iequality across eighorhoods i Rio. The mea of our various empirical specificatios suggests that almost half of the declie i price iequality over the sample period is attriutale to the UPPs. These fidigs illustrate a potetially sigificat ew dimesio of policies aimed at reducig either crime or iequality. Existig attempts to ameliorate ecoomic disparities ted to focus o trasfers of icome ad policies capale of reshapig the wealth distriutio i ay meaigful way are ucommo. The UPPs i Rio demostrate that crime reductio ca play a role i reducig ecoomic iequity, operatig through the distriutio of wealth. I priciple, this should e a importat cosideratio for crime- ad coflict-lade regios elsewhere. 28

30 Refereces 1. Besley, Timothy J. ad Haes Mueller (2011): Estimatig the Peace Divided: The Impact of Violece o House Prices i Norther Irelad. America Ecoomic Review (forthcomig). 2. Boyle, Melissa A. ad Katherie A. Kiel (2001): A Survey of House Price Hedoic Studies of the Impact of Evirometal Exteralities. Joural of Real Estate Literature. 9:2, Bricker, Jesse, Bria Bucks, Arthur Keickell, Traci Mach, ad Kevi Moore (2011): Surveyig the Aftermath of the Storm: Chages i Family Fiaces from 2007 to Fiace ad Ecoomics Discussio Series o , Federal Reserve Board, Washigto, D.C. 4. Cuha, Neiva Vieira da ad Marco Atoio da Silva Mello (2011): Novos coflitos a cidade: A UPP e o processo de uraização a favela. DILEMAS: Revista de Estudos de Coflito e Cotrole Social. 4:3, Deato, Agus (1997): Aalysis of Household Surveys. Johs Hopkis Uiversity Press. 6. Fudação Istituto de Pesquisas Ecoomicas (2011): Ídice FIPEZAP de Preços de Imóveis Auciados: Notas Metodologicas. 7. Gios, Steve (2004): The Cost of Ura Property Crime. Ecoomic Joural, 114: Heckma, James (1976): The commo structure of statistical models of trucatio, sample selectio, ad limited depedet variales ad a simple estimator for such models. Aals of Ecoomic ad Social Measuremet, 5: Heckma, James (1979): Sample selectio ias as a specificatio error. Ecoometrica, 47:1, Hellma, D. A. ad Naroff, J. L. (1979): The impact of crime o ura residetial property values. Ura Studies, 16: Ihlafeldt, Keith ad Tom Mayock (2009): "Property Values ad Crime," i Hadook o the Ecoomics of Crime, edited y B. Beso ad P. Zimmerma, Edward Elgar, forthcomig. 29

31 12. Keickell, Arthur B. (2009): Pods ad Streams: Wealth ad Icome i the U.S., 1989 to Fiace ad Ecoomics Discussio Series o , Federal Reserve Board, Washigto, D.C. 13. Lide, Leigh L. ad Joah E. Rockoff (2008): Estimates of the Impact of Crime Risk o Property Values from Mega's Laws. America Ecoomic Review, 2008, 98(3), Lych, A. K. ad Rasmusse, D. W. (2001): Measurig the impact of crime o house prices. Applied Ecoomics, 33: Neri, Marcelo Cortes (2011a): UPP 2 e a Ecoomia da Rociha e do Alemão: Do Choque de Ordem ao de Progresso. Marcelo Cortes Neri (Coord.), FGV CPS, Rio de Jaeiro. 16. Neri, Marcelo Cortes (2011): Desigualdade de Reda a Década. Marcelo Cortes Neri (Coord.), FGV CPS, Rio de Jaeiro. 17. Rizzo, Mario J. (1979): "The Effect of Crime o Residetial Rets ad Property Values." America Ecoomist, 23: Soares, Faio Veras, Sergei Soares, Marcelo Medeiros ad Rafael Guerreiro Osório (2006): Cash Trasfer Programmes i Brazil: Impacts o Iequality ad Poverty. UNDP Iteratioal Poverty Ceter Workig Paper o Thaler, R. (1978): A ote o the value of crime cotrol: evidece from the property market. Joural of Ura Ecoomics, 5: Tita, George E., Tricia L. Petras ad Roert T. Greeaum (2006): Crime ad Residetial Choice: A Neighorhood Level Aalysis of the Impact of Crime o Housig Prices. Joural of Quatitative Crimiology, 22:4, Waiselfisz, Julio Jacoo (2011): Mapa da Violecia 2012: Os Novos Padrões da Violêcia Homicida o Brasil. Istituto Sagari Wolff, Edward N. (1992): Chagig Iequality of Wealth. America Ecoomic Review Papers ad Proceedigs, 82(2),

32 Homicide ad Roery Rate (Idex, =100) Gii Coefficiet Homicide rate Roery rate Gii coefficiet for Bairro property prices Figure 1: Homicides, roeries ad house price iequality i Rio Notes: The costructio of homicide ad roery rates is descried i Sectio III. The costructio of the Gii coefficiet of property prices is descried i Sectio V. 31

33 Figure 2: Homicides ad property prices y eighorhood Notes: The locatio of the UPP statios are show as gree dots. Average apartmet prices i each eighorhood are the average of oservatios across umer of rooms ad time periods, weighted y the umer of listigs for each oservatio. The homicide rate is the average across time periods weighted y populatio. 32