Foreclosure Sales and Recourse

VICTOR WESTRUPP∗

(WORK IN PROGRESS)

September 15, 2016

I document the imbalance of foreclosure sales with respect to recourse laws applied to primarymortgages in the U.S. using geographic state borders as exogenous variation combined withindividual transaction data located close to the border. Under a regression discontinuity design,the relative volume of foreclosure sales (compared to regular sales) in recourse states was 13percentage points lower on average with respect to non-recourse states from 2008 to 2010. Fur-ther evidence from a hedonic pricing model suggests that foreclosures in non-recourse stateswere sold by a larger discount after 2008 (6 percentage points from 2008 to 2010 and 9 per-centage points from 2011 to 2012). Results are robust to differences in borrower characteristicsat the loan origination and suggests differences in the incentives to default under recourse, aswell as a higher volume of foreclosures in non-recourse states.

∗Graduate student, Stanford Graduate School of Business. I would like to thank valuable comments from Tim Mc-quade, Rebecca Diamond, Jules Van Binsbergen, Dirk Jenter, Peter Koudjis, Shai Bernstein, Anat Admati, Piotr Dwor-czak, Zhe Wang, Will Gornall, Marco Giacoletti, Monika Piazessi, Martin Schneider and participants of the FinancialEconomics class of 2014 from Stanford Dept. of Economics. Email for contact: victor westrupp@stanford.edu.

1

1 Introduction

How did the housing market for foreclosed properties behave with respect to different foreclo-sure policies in the recent U.S. housing crisis? It is a well documented fact that the number offoreclosures (properties claimed by lenders as collateral from mortgages in default) increased con-siderably during the recent U.S. housing crisis. Nevertheless, some states reported a higher numberof foreclosures than others (such as California and Arizona). Among potential explanations for thecross-sectional variation of foreclosures in the U.S., we should consider state foreclosure policiesas potential factors, specially given the heterogeneity of such laws across the country. This pa-per then studies differences in the collateral assessment of foreclosed houses with respect to onespecific type of foreclosure policy: the right to recourse over primary residential mortgages.

An important distinction from this paper is the focus on foreclosure sales. Foreclosure policiesare directly related to how foreclosures are set in a specific state, and so focusing on foreclosuresales (as compared to the overall housing market) provides a better understanding about the impactof such laws. Moreover, foreclosure sales can be viewed as collateral assessment from lenderswhose mortgagees were in distress. Understanding what factors may affect how lenders, or residualclaimants, are able to reclaim their collateral is an important question per se.

Another important characteristic is the focus on recourse laws. States with relatively strongerrecourse laws applied to mortgages and deeds of trust allow lenders to pursue previous homeown-ers in deficiency judgments whenever the foreclosure sale was short on the current debt balancein order to recover the difference between the balance and the sales value. Ghent and Kudlyak[2011] provide evidence of recourse affecting the incentives to default conditional on borrowercharacteristics, however, it hasn’t been documented whether such laws are able to explain the av-erage variation of foreclosures across the country. Moreover, given the complexity of the entireforeclosure regulation, it seems wise to focus on a specific foreclosure policy. It is also consistentwith recent developments in the field, such as in Mian et al. [2014], where the mandatory courtprocedure in foreclosures is the main focus.

Not all recourse states are the same and some may vary over the degree in which deficiencyjudgments can claim other unsecured assets, but some states are clearly “borrower friendly” withrespect to mortgage default1. Moreover, recent studies on the historical development of state fore-closure laws showed that many of such laws were passed right after the Civil War and remainedunchanged until the most recent housing crisis, which creates an interesting environment to ex-plore the differences of housing markets with respect to recourse laws. The housing market boomof the early 2000s reduced considerably the expected cost of recourse. If a house was foreclosed,the sales revenue from liquidating the asset as a foreclosure was relatively close to cover the delin-quent’s current balance. With the advent of the current housing crisis however, not only negative

1One of the most important non-recourse states is California, whose civil code prohibits deficiency judgments evenin cases of refinances and HELOCs (home equity line of credit), unless the lender choose to go through a judicialforeclosure process.

2

equity became a regular issue, but income shocks truly affected how recourse played a role inmortgage default2.

In order to conduct the empirical analysis, I use individual transaction level data from DataQuick, a proprietary database that collects housing transaction information from county recordersand clerks3, supplemented with relevant borrower’s characteristics at the time of the mortgage ap-plication, mortgage information at the loan origination and demographic characteristics. I thenselect only transactions close to recourse/non-recourse state borders in order to guarantee a morecontrolled environment, as one should not expect markets to be consistently discontinuous geo-graphically on average ex-ante4.

Although the number of borders covered by Data Quick is limited, it provides enough internalvariation to guarantee robustness of the results. By mapping the latitute-longitude coordinateslinearly until the closest border, I start the analysis with a sharp regression discontinuity designwith discrete changes in recourse law status. The methodology is similar to the one followed byMian et al. [2014], but with different outcome variables. Moreover, the experiment is valid underthe assumption that households did not select their houses with respect to recourse laws ex ante,at the loan origination, which I’m able to test for a specific set of characteristics provided by theHome Mortgage Disclosure Act database (HMDA).

Departing from the literature that analyzed mortgage default decisions, I investigate differencesin log house prices and relative volume of foreclosure sales (with respect to regular sales market)between both recourse and non-recourse sides. Although log prices were mildly upward over therecourse border, the relative amount of foreclosure sales compared to regular sales was around13 percentage points lower on the recourse side from 2008 to 2010. The result is robust under aset of borrower’s initial characteristics and local market characteristics and consistent with recenttheoretical developments regarding recourse and strategic default in which the impact of recourseon default decisions can only be compared using borrowers with similar initial characteristics. Thisresult can be potentially explained by differences in the decisions to default.

I also analyzed how local markets behaved with respect to alternatives to foreclosure, such asa short-sale5 The relative number of short-sales compared to regular sales did not seem to react as

2Anecdotal evidence from the Washington Post shows that recourse was rarely used during the first halfof the previous decade, and its use was intensified both during and post-crisis in order to claim part of themortgage debt, where not only local banks but also Fannie Mae and Freddie Mac filled for delinquencyjudgments against homeowners. Source: http://www.washingtonpost.com/investigations/lenders-seek-court-actions-against-homeowners-years-after-foreclosure/2013/06/15/3c6a04ce-96fc-11e2-b68f-dc5c4b47e519_story.html

3I deeply thank professor Rebecca Diamond for allowing me access to the data.4Evidence will be analyzed later in the paper.5A short-sale is an event in which the lender allows a distressed homeowner to sell his alienated house for a value

that is below the current mortgage value and use the resulting payment to cover its mortgage, basically taking a lossin terms of current mortgage value, but avoiding additional hassles regarding entering a foreclosure process, wheneverconvenient.

3

strongly as the relative volume of foreclosures. However, such tests may suffer from low powergiven the small number of short sales across the state borders in analysis.

In order to provide further clarification with respect to foreclosure sales prices, I analyze pricediscounts of foreclosures with respect to regular sales under a static hedonic pricing model witha few adjustments in order to capture such price interaction. The results show that discounts inrecourse states were smaller than in non-recourse states by 6.5 percentage points between 2008-2010 and 9.0 percentage points between 2011-2012. Model specifications and border selectionissues are addressed as robustness, as well as whether the results could be driven by differences intime-to-sell of foreclosures.

The price impact combined with the differences in relative volume of foreclosures suggest asupply channel story. Since the relative volume of houses entering foreclosure was smaller inrecourse states, REO servicers may have decreased prices even further in non-recourse marketsbecause of increased competition or search frictions (the composition effect as theoretically in-vestigated by Guren and Mcquade [2013]). Naturally, the supply channel inferred here is notnecessarily the only explanation. For instance, one could attribute the price difference to unob-served differences in the quality of foreclosure houses between recourse and non-recourse markets(recourse homeowners may spend more in maintenance in order to minimize the future expectedmortgage debt deficiency). Unfortunately, at this stage the current research is not able to investigateall the pricing channels.

At last, under the results from the hedonic pricing model, there was no impact on short-saleaverage discounts (compared to regular sales). It is inconclusive whether short sale prices shouldbe affected by recourse, since lenders typically provide a deadline for short sales, and if the houseis not sold, it implements the current foreclosure process. Therefore, the option to recourse isembedded in a short sale decision, and it depends on the lender’s time preferences whether hewould push for instantaneous cash or wait until the homeowner offers a better deal.

The rest of the paper is organized as follows: section 2 displays the literature review, sec-tion 3 presents the empirical strategy, the information on recourse, the data and sample selectionsteps. Section 4 shows the empirical results together with robustness checks while section 5 finallyconcludes and discusses further development.

2 Literature Review

Studies involving recourse laws became popular with the advent of the recent housing crisis. Inone of the most cited, Ghent and Kudlyak [2011] analyzed the impact of recourse on loan leveldata. They concluded empirically that, although recourse laws are hardly effective in the UnitedStates, default probabilities may reduce to at most 30% conditional on the borrower’s origination

4

characteristics. The effect is larger whenever homeowners face negative equity. Recent studies thatempirically analyze state foreclosure laws and mortgage defaults when states allow for recourseloans are also Gerardi et al. [2013], Demiroglu et al. [2014]. Broadly speaking, there are otherimportant aspects of state foreclosure laws, such as judicial foreclosure requirements, which wascurrently explored by Mian et al. [2014] in a similar border discontinuity methodology to the oneimplemented in this paper. In Zhu and Pace [2013] data on individual mortgage loans on stateborders are used to analyze state variation laws such as whether states are mainly recourse overdeciding between selling under an REO or performing a short-sale. None of these papers haveconcentrated on the ex-post price implications on foreclosure houses due to recourse, which is themain objective of this paper.

Studies that are closely related to this one in particular are Te Bao [2014] and Nam and Oh[2014], both applied to analyze recourse laws and the housing bubble. In both papers, despitedifferences in methodology, states with non-recourse laws experienced larger bubbles in housingprices. In Nam and Oh [2014], a discontinuity approach is applied, using zipcode level informa-tion, to access different characteristics between recourse and non-recourse states. They rely onthe exogenous crisis shock to identify the effect of prices, but it is not clear that such states weresimilar ex-ante. My analysis differ from them by concentrating on the foreclosure housing market.As recourse laws were directed towards foreclosures, it seems natural that prices and quantities offoreclosures could differ between recourse and non-recourse states. Moreover, the fact that I canuse local border variation with individual transaction level and relative mortgage history allows meto effectively identify the source of the effect based on previous literature on mortgage default andrecourse.

From a theoretical perspective, there has been a recent development on models of mortgagedefault that tries to account for recourse, such as Quintin [2012], John Y. Campbell [2014] andCorbae and Quintin [2014]. Quintin [2012] provides an interesting yet simple framework to sustainthe idea that recourse analysis should take into account the borrower’s characteristics at the loanorigination, since the pool of borrowers may be different. If we are to analyze any real economicimpact caused from recourse laws, it should be in a controlled environment where similar regionswere affected by similar local random shocks.

Marginally, this paper also fits into the literature that investigates the impact of forced salesand foreclosures on housing markets. Recent work such as Campbell et al. [2011] and Guren andMcquade [2013] focus on the foreclosure mechanism as a way to propagate shocks into regularhousing markets. Instead of pushing towards the impact of foreclosures however, I will remaininterested in analyzing possible alternatives to foreclosure and whether or which lenders are usingthem. Moreover, I also find relative weak evidence of market segmentation between areas that aremore likely to have short sales compared to REO sales, a topic already covered in housing marketssuch as in Landvoigt et al. [2013].

5

Lastly, the theoretical framework could be related to studies analyzing asymmetric informationin the housing market, as lenders may not fully access the amount of assets that lenders haveavailable at the moment of default, and applying for a deficiency may be costly for banks. Somepapers such as Garmaise and Moskowitz [2004], Levitt and Syverson [2008], de Wit and van derKlaauw [2013] and Kurlat and Stroebel [2014]. Garmaise and Moskowitz [2004] analized anexogenous measure of information based on the quality of tax assessments in different regions,applying to the U.S. real estate market, concluding that information considerations are importantand market participants resolve such uncertainties by purchasing nearby properties or avoidingexpert real estate professional brokers. Levitt and Syverson [2008] explored better the agencyproblems regarding real estate brokers, when some might have incentives to force sell their clients’houses quickly at lower prices. As one of the predictions observed, houses marketed by real estateagents stayed in the market longer. Focusing on list prices as potential signals, de Wit and van derKlaauw [2013] analyzed the impact of such list-price reductions on the Time on the market, findinga positive impact of reducing list prices on the average selling speed. Kurlat and Stroebel [2014]analyzed an equilibrium outcome in which both buyers and sellers might have private information.In one of their predictions, an increase in the number of asymmetrically informed sellers mightbe related to lower prices being offered. The research proposal plans to add to this literature byanalyzing the potential asymmetric information that lenders must deal with during the foreclosureprocess.

3 Recourse Laws and Data Description

3.1 Recourse and State Foreclosure Laws

States may differ in some aspects of their foreclosure laws. For instance, some states may allowforeclosures to be pursued only through court (judicial states), while others allow homeowners toRedeem their mortgages even after a sale, for a specific deadline, which also may be different. Suchdifferences undoubtedly are an interesting field of analysis, as they could explain at least part of theheterogeneity of housing price shocks in the U.S.. Table 1 offers information on some of the mostcommon state foreclosure laws. I used as sources the NCLC’s survey on state foreclosure lawsentitled “Foreclosing a Dream” by John Rao and Geoff Walsh, which describes a series of differentstate foreclosure laws, and was issued February 20096. Updated information until 2011 can beobtained in the Office of Legislative Research (OLR) research report issued by James Orlando inDecember 20117.Ghent and Kudlyak [2011] also surveyed state foreclosure laws, which is takenas a final comparison.

6source: http://www.nclc.org/images/pdf/foreclosure_mortgage/state_laws/foreclosing-dream-report.pdf

7source: http://www.cga.ct.gov/2010/rpt/2010-R-0327.htm

6

The dimension that I’ll focus in this paper is the option to recourse on foreclosures. States thatare considered recourse allow lenders (or REOs) to claim in court the remaining of the currentmortgage balance, should a foreclosed house be sold by a price that falls short of its debt. Mostof the recourse states only allow lenders to claim the remaining until the fair market value of thehouse, even if the debt value is higher. Nevertheless, recourse should be thought as an additionalinsurance for lenders in case of default, as an option that lenders may use if the relative payoffcompensates any additional costs of reclaiming their payment, such as court costs for example.

There is a certain degree of subjectivity in defining whether states are mainly recourse, as re-course laws can be different as well. Only 9 continental states were considered as Non Recourse:Alaska Arizona, California, Minessota, Montana, New Mexico, North Carolina, North Dakota, Ok-lahoma, Oregon and Washington. However, some states have its peculiarities. In North Carolina,only mortgage purchases are considered non recourse, where in Nevada the option to recourse wasremoved for mortgages issued after October 2009. These issues will be important when analyzingthe robustness of the results obtained.

Even though there may have been some changes of state laws regarding recourse, these changesare recent and do not apply to most of the mortgages originated prior to the crisis. As reviewed byGhent [2012]. State foreclosure laws were relatively constant after the American Civil War. Thesame argument is used in Mian et al. [2014] to give validity of a similar analysis, but involvingdifferences between judicial versus non judicial states.

3.2 Data

In order to compare house price transactions across states. I’ll use two databases that provideimportant information about such transactions: Data Quick and HMDA.

House transaction prices are provided by Data Quick, along with a series of covariates andtransaction history data. Information is provided at the transaction micro data level, which allowsone to geographically measure its distance towards the nearest state border (under a GIS software).Data availability goes from 1991 to the first quarter of 2012, however, most of the coverage ismade available after 2005. Figure 1 shows the coverage area of historical transactions from 2005to 2012.

Only single houses or condos are considered in analysis. Qualified transactions exclude re-finances or equity extractions, subdivisions and constructions. Distress sales are flagged underData Quick, whether a sale was made by a Real Estate Owner (REO), whether it was a ShortSale or whether it was an Auction sale in a foreclosure auction. Although foreclosure auctions areconsidered in the analysis, most of the foreclosed houses are sold by REO servicers. Also, onlytransactions at arm’s length are considered (houses being sold from a father to a son are not con-sidered for example, since its not possible to distinguish the potential biases in such transactions).

7

The list of covariates is available in Table 2. From the transaction history side, Data Quickcollects most of the transaction and mortgage information. Not only transaction values, but alsoloan values and LTVs at the origination are available. It also informs the type of buyers and sellers(whether it is a corporate/government/trustee) and lenders (if its a commercial bank, credit unionand so long). It also informs whether a loan in specific is insured by the FHA.

From the house-specific side, Data Quick provides the report of houses accessed around 2010to 2011, with several dimensions analyzed: area (sq feet) of several locations including additions,buildings and garages, a condition code which varies from 0 (unsound) to 6 (excellent), the lotdepth and width, age (measure from date of transaction from reported date of effective construc-tion), and number of several installations including bedrooms, bathrooms etc.

Such characteristics are key to isolate housing markets into observable house characteristics.Moreover, the (x,y) coordinate system is provided along with the Census Tract level of location,which allows for location controls at a narrow structure.

I also supplement data on house transactions with demographic information provided by theHome Mortgage Disclosure Act (HMDA), which collects demographic information from everymortgage application in the United States. The applications are separated by year and CensusTract level of location and records mortgage loans, lender information (name and agency type) anddemographic information such as race, gender, ethnic and income informed during the application.Such information may help understand how similar different regions may be with respect to home-owners and mortgage/housing markets from the demand side. The information is incorporated inData Quick by merging loan values under each tract area and year8 .

Individual information on mortgage balances and credit scores are not available, unfortunately(one of the main reasons why are not able to properly test a model of mortgage default, forcingus to rely on identification by state border variation). Although Data Quick provides mortgageinformation at the origination, it does not keep track of any other events regarding the mortgage.I’ll address such limitations whenever it may seem necessary.

3.3 Sample Selection

Since we must rely on state border variation for purposes of identification, ideally we would liketo reach the state borders as close as possible in order to ensure that both markets are relatively“similar”. Here I detail how the sample was selected, while later I address whether such marketsare similar ex-ante. Figure 2 displays which states are defined as recourse compared to no-recourseones. The shaded red areas are border counties covered by Data Quick between 2005 and 2012,when coverage is best. I start by selecting such counties whenever recourse laws vary across states.

8Appendix C provides additional details regarding the merge

8

Current data allows us to investigate a total of 5 state borders: NC - SC, NC - TN, NC - VA, CA- NV, AZ - NV. Data from 2 other borders (NM - TX and AZ - UT) are also available, however, thenumber of transactions is not enough to ensure proper weight in estimated coefficients, in whichcase I remove these borders from the analysis. Non-recourse states are North Carolina, Arizonaand California, while recourse states are South Carolina, Tennessee, Nevada and Virginia.

Table 3 shows the total number of transactions observed. The first two columns display the totalnumber of available transactions under each county border, while the last two columns display thedata used for the discontinuity. Some issues are worth discussing: i) regarding a few counties inthe South Carolina side of the border. In three of the counties, the lack of appraisal informationand the lack of information on mortgage loans responds for almost 90% of the data available insuch locations (compared to roughly 3% on the rest of the data), which seems to suggest lack ofreliable information on mortgages more than lack of mortgage information due to houses boughtunder total down payment. These borders are later removed from the discontinuity analysis andthe missing observations are also removed from the hedonic pricing regression analysis; ii) for thediscontinuity analysis, we do not limit ourselves only to use information from border counties,but we do select data that are close to 15 miles from the border. Moreover, some county-borderpairs may be highly imbalanced (some sides may have close to 0 transaction data available). Sincewe cannot use location-specific fixed effects, we select only county-border pairs whose imbalanceis not worse than 5%. The selection criterion does not affect the main results, but it providesadditional robustness. Table 5, shows the number of observations available by each county-borderpair.

For the static hedonic pricing model regression, however, we must consider transactions whosecovariates are available. Table 4 shows the total volume of transactions whose all covariates areavailable, located at the border counties, used in the linear static hedonic pricing regression. Dataremoved consists of missing data on house characteristics and borrower’s characteristics such asincome, loan information and other demographics such as race, gender and ethnicity.

An important aspect of recourse laws is that such laws mainly target primary mortgages (homepurchase mortgages). Secondary mortgages, refinances and equity extraction loans are junior (withrespect to primary mortgages) and lenders may file for deficiency judgments if the foreclosedproperty value was not enough to cover their balance, although not all non-recourse states allowfor such9. Data Quick provides a history of transactions which allows us to identify at least if (andwhen) a secondary loan was issued (either refinance or equity extraction), and the loan value. Inorder to ensure we are targeting the distress sales which are vulnerable to recourse laws, I removeany distress sale transaction whose history had any secondary loans issued and the house entered

9Such characteristic is prevalent in North Carolina (also documented by Ghent and Kudlyak [2011], in other statessuch as California, the only case in which junior lenders may seek deficiency judgment is when the foreclosure fol-lowed proper judicial procedures (California Civil Procedure Code, paragraphs 580b and 726), which is not commonin the state since its considered mainly nonjudicial.

9

foreclosure10. Unfortunately, the history is not available for all transactions, which may explainwhy only around one fifth of foreclosure sales had any past refinances. Nevertheless, the resultsare not significantly affected by including such refinanced transactions, and so I choose to removethem.

Using the x-y coordinates for each transaction location, I use ArcGIS to map the planar mini-mum distance until the nearest state border, while also identifying which part of the border segmentis the transaction close to. Border segments are separated by the contiguous area in which two bor-der counties (on each side) share. Each transaction is matched with only one segment11. Thisallows us to better regulate the proximity towards the border, as there is clearly a trade-off betweenhaving enough transaction data for robust results or allowing for additional heterogeneity as weexpand the distance range. While 5 borders may seem limiting, transaction coverage provides uswith enough variation to limit our analysis to at least 15 miles from the border.

4 Empirical Analysis

4.1 Average Ex-ante characteristics

Before analyzing local markets with respect to ex-post effects of recourse in foreclosure prices,we must check whether both sides of the border were ex-ante similar. While recourse laws wereconstant until the recent years, the boom period provided the American Housing market withrising house prices and with a very limited number of foreclosures. Deficiencies were extremelyrare, as house values were increasing and foreclosure sales most likely could cover the currentmortgage balance. Under such case, we should not expect major differences in both (local) marketsduring the boom period. Table 6 compares the average of several local market characteristics,between 2004 and 2006, from a distance no further than 15 miles from the border12. Given thehigh number of transactions, the difference in means test is expected to be significant for most ofthe characteristics. We must compare the differences of magnitude and conclude whether thesetwo local markets are comparable ex-ante.

The pool of different borrowers ex ante may affect how markets behaved ex post, so it iscrucial to understand exactly how different these markets were locally. Suppose that banks wereindeed pricing the recourse premium on their mortgages, in which case recourse states wouldhave lower interest rates, while non-recourse banks would charge an additional premium over their

10With respect to short-sales, state laws are not very rigorous, with the exception of California, who forbids recourseover short-sales as well. Mostly, banks are asked to waive the right to pursue a deficiency in case of a short-sale, butthere are cases in which lenders went after homeowners in case of short-sales (Anecdotal evidence from CNN newsin 2010, source: http://money.cnn.com/2010/02/03/real_estate/foreclosure_deficiency_judgement/)

11in the event of numerically identical distances from two segments, randomly select one.12Comparing on an annual basis would yield similar results.

10

loans. This would imply that LTV ratios would be lower in non-recourse states Ceteris paribus.The pool of borrowers, however, could be potentially different. Suppose that borrowers higherpropensity to default would target recourse markets instead, given their possibly more attractivelending conditions. If banks anticipate such behavior, then interest rates and LTV ratios could beactually lower in recourse states.

From the borrower’s side, although we do not observe their credit scores at the mortgage orig-ination, we observe the annual income used in the mortgage application, which differs by approx-imate 3,37% between the recourse side and the non-recourse side. From the lender’s perspective,LTV ratios are slightly higher in recourse states, which is favorable evidence towards a similar poolof borrowers. Nevertheless, the small difference (only 2 percentage points) is additional evidencetowards the fact that recourse premium was relatively small, as negative equity was not an issuewhen prices were booming. Among other demographic characteristics, both markets seem to behomogeneous with respect to gender and race, as such differences should not raise any concerns.

With respect to the housing market, house prices appear to be higher in the recourse side. Ana-lyzing additional characteristics, we also notice that total built square feet is lower in the recourseside, but total lot size is higher. Also, the number of bathrooms and bedrooms are slightly higherover the recourse side. These characteristics combined could potentially explain the difference inhouse prices.

4.2 Border Discontinuity Analysis.

In this section, I describe the main steps taken to analyze the discontinuity across state borders.Estimation and methodology of analyzing sharp discontinuities are based on Imbens and Lemieux[2008] and Lee and Lemieux [2010]. Let y be the variable analyzed (it could be either the outcome,or the set of covariates). y is defined as transaction i in location c from county-border pair b, atyear t. Following the recommendations in Lee and Lemieux [2010], we first “residualize” thevariable in analysis so that later we can analyze if there exists a discontinuity with respect to thetreatment13.

yicbt = yicbt − α− δb (1)

Let Wi = 1(Dib≥0) be the “recourse” assignment, where Dib is the distance towards border b.Negative distances towards the state border refer to non-recourse states.

We start the analysis by visual inspection. Fixing a bandwidth h, we first plot the histogramof observations by each bin in order to show that there is not a significant change of observables

13Lee and Lemieux [2010] argue, however, that either the existence of fixed effects or the residualization processshould not affect the consistency of the estimated average treatment effect (although it could bias standard errors).The advantage of using residuals rather than estimating simultaneously is visual (it is feasible to plot the residuals andshow whether there is a discontinuity). For more of this discussion, please refer to Lee and Lemieux [2010]

11

at the moment of the discontinuity, from a maximum distance of 10 miles from the border. Figure3 displays the histogram of transactions. Each row separates transactions by foreclosures (auctionsales and REO sales), regular sales and short-sales. Histograms are grouped between a set ofyears, we define three groups here: 2005-2007 (the pre foreclosure crisis period), 2008-2010 (theforeclosure crisis) and 2011-2012 (post foreclosure crisis). Data seems to be relatively balancedaround the border.

We then analyze how outcomes/covariates changed over the border. Each set of graphscontain two series as function of D: (i) the average outcome/covariate per bin yk =

1Nk

∑Ni=1 yi1(bk<Di≤bk+1) where bk = 0 − (K − k + 1)h and Nk =

∑Ni=1 1(bk<Di≤bk+1); (ii) a

non-parametric estimation as a function of D (Nadaraya-Watson with Normal Kernel and optimalbandwidth based on Silverman’s rule). 95% confidence intervals are also plotted (standard errorscalculated by bootstrap). Each bin has a equal bandwidth of 0.5 miles.

We focus on two important outcomes: house log prices (divided by total built square feet) andthe relative number of sales. Figure 4 shows the changes in log prices. For all the housing marketsanalyzed, there is not enough power to reject the hypothesis that house prices are not statisticallydiscontinuous at the border. For regular sales, there seems to be a decrease of house prices veryclose to the border (1 mile) on the non-recourse side, however, given the overall description of thegraph, that hardly constitutes a discontinuity, and can be caused by different house characteristics.After analyzing the number of bathrooms and bedrooms, the same pattern repeats, which could berelated to prices14.

Secondly, we analyze the relative volume of foreclosure sales compared to both the regularmarket sales and the short-sales in Figure 5. The first row compares foreclosure sales with regularsales (outcome variable is a dummy when a sale is foreclosure) while the second row comparesforeclosures with short sales. In this case, we observe a discontinuity between recourse and non-recourse markets with respect to foreclosure sales compared to regular sales. Between 2005 to2007, relative foreclosure sales have been similar between both sides of the border. As houseprices were increasing, homeowners did not suffer from negative equity and most mortgages gotfully paid from foreclosure sales. Between 2008-2010 however, the relative number of foreclosuresales is smaller on the recourse side by a difference of approximate 11 percentage points, laterreturning to a similar pattern after 2011. This result may be due to different default incentivesbetween recourse and non-recourse states, but we must be cautious in analyzing this result.

14In addition to mortgage conditions, I repeated the same analysis for each available covariate (since the number ofcovariates is too large, an online appendix is available upon request with the results). Under the list of covariates, onlya few showed a discontinuity around the threshold, but in all cases, the gap appears in all year groups. Such variablesare: house age (+5 years on average); whether a building has a garage (+40 percentage points on average); number ofbathrooms (+1 on average); number of bedrooms (+1 only for the case of REO sales after 2008); whether the buildinghas a pool (+40 percentage points on average); whether it is a condo (+30 percentage points on average). Unfortunately,such gaps might explain price variation, but since it affects both the regular sales market and the foreclosure market,there is little to believe that such gaps would explain the relative volume of foreclosures to regular sales. The regressionpresented in Table 9 helps to analyze differences in prices by controlling for house characteristics.

12

One of the challenges of analyzing the impacts of recourse is that the borrower’s characteristicsat the loan origination matter. Based on Quintin [2012], one of their main results is that defaultrates would fall if recourse laws were strengthened only conditional on the initial borrower’s type.Mortgage interest rates prior to default would be lower in recourse states, which could attract adifferent pool of borrowers.

Intuitively, consider a 2-period economy and suppose recourse laws allow both your assets andcurrent endowment as part of your limit liability L:

L(a0, h′, y1; η) ≡ h′ +max{η(a0R + y1), 0} = (2)

= h′ + η(a0R + y1)

where a0 is the amount of assets invested in period 0, R is the rate of return, y1 is the endowmentavailable at t = 1, h′ is the current house value (in terms of consumption goods) and η ∈ [0, 1]

is the degree in which recourse is effective. Suppose, moreover, that homeowners must either paytheir mortgages (with interest rate Rh) and sell back their houses or default at t = 1. If their housewas worth h at t = 0, the agent is strictly better off by defaulting if and only if

hRh > L(a0, h′, y1; η) (3)

now, assume that y1 is stochastic such that y1 = y0eξ1 , ξ1 ∼ N(0, σ2). The probability of condition

3 to hold is given by:

P (hRh > L(a0, h′, y1; η)) = Φ

[log(hRh − h′ − ηa0R)− log(ηy0)

σ

](4)

Just for simplicity, assume that future house prices are perfectly predictable. In times where h′ ishigh enough so that it can pay off the mortgage debt, recourse laws may not strongly affect thedefault decision, as it would require a relatively high income shock to drive agents not only todefault, but to also have their assets taken. Consequently, if house prices may become underwater,recourse may affect the default decision because of negative equity. This is also related to Ghentand Kudlyak [2011]. Moreover, agents may respond differently with respect to changes in η ex-

ante, as the pool of borrowers may change.

In order to account for this issue, I’ll use past information at the original transaction withrespect to mortgage information such as the LTV ratio and whether the mortgage was insured by theFHA, together with borrower’s income at the moment of application as covariates. Unfortunately,Data Quick does not keep track of all past information, and so information at the originationis limited. Nevertheless, while comparing averages across the border, I’ll use the entire set ofobservations15 Figures 6, 7 and 8 display the results for income at the origination, whether loanswere initially insured by FHA and LTV ratios at the origination. In all cases, there does not seem to

15Results are not affected if we use only transactions whose past information are available.

13

be a significant discontinuity. The result is specially stronger for past income of homeowners whowere foreclosed between 2008-2010, which shows that there is convergence right at the border.In terms of LTV ratios at the origination, it shows that non-recourse borders may have had higherLTV ratios at the origination for foreclosed properties. This could potentially effect the results,since homeowners with higher past LTV ratios may be more willing to default.

Although we do not observe the moment of default, we can at least analyze the entire periodbetween purchasing the house and selling it (either through foreclosure or through regular/short-sales markets). Figure 9 displays the discontinuity analysis. There seems to be no relevant changesacross the border, which makes the variables at origination indeed consistent (looking over thesame time period, basically).

How about the current market conditions? Figures 10 and 11 present the current borrower’sinformation for both income and LTV ratio. Although there are other important characteristics tolook, it seems that local market characteristics at the moment of sale were relatively similar. Thiscould potentially rule out major differences in demand schedules between both sides, and makesit more convincing that the impact on foreclosure sales is coming indeed from a supply channelthrough foreclosures.

4.2.1 Estimation of the Average Treatment Effect.

Although the graphs provide most of the intuition, it still remains the question on how to properlyestimate the average treatment effect over such residuals. Under the assumption of continuity of theconditional expectation around the threshold, the average treatment effect under the sharp designis given by:

τy = limd↓0

E[yicbt − α− δb|D = d]− limd↑0

E[yicbt − α− δb|D = d]

= E[y(Wi = 1)ibct − y(Wi = 0)ibct|Di = 0]

Notice that, in order to identify the effect through the residuals, one must be sure that the fixedeffects considered in the analysis (or their conditional expectations with respect to the forcingcovariate) are continuous at D = 0. Border fixed effects and the demeaning constant both arecontinuous and so one can properly identify the effect by looking over the residuals (even if biased,as long as the bias goes in the same direction across both sides of the border). Such fixed effectsare there only to increase the precision of the estimation.

Among the estimators of τy, the literature provides both a parametric framework and a nonpara-metric. Here I’ll concentrate on the nonparametric estimation only on log prices and the relativevolume of foreclosures sold16. Covariate analysis shall remain visually only, for the sake of sim-plicity.

16In the online appendix, I report the estimation of the average treatment effect for all the outcomes considered,under both the parametric and the nonparametric specification.

14

In the nonparametric local linear regression, let h be a bandwidth such that we only considervalues of D between −h and h. The nonparametric estimation consists on solving the followingminimization problem:

θ = arg minθ∈Θ

N∑i=1

1(−h<Di≤h)[yi − α− τWi − βlDi − (βr − βl)Di ×Wi] (5)

where θ corresponds to the set of parameters estimated. The only parameter that matters is τ ,which is taken as the treatment effect and has the same asymptotic result as τy.

The main problem with this formulation is to define h, in which there is an implicit trade-offbetween precision and bias. The larger the bandwidth gets, the larger is the bias of the treatmenteffect becomes because of the linearity imposed. Imbens and Kalyanaraman [2012] provide analgorithm to calculate the asymptotic optimal bandwidth based on minimizing the squared errorloss, which is used here in the estimation. Moreover, one of the recommendations in Imbens andLemieux [2008] is to test whether the results are too sensitive to the bandwidth choice. Table 7shows the estimated results for the average treatment effect for the optimal bandwidth, togetherwith results for half the band and twice the band size. Among all coefficients, the only oneswho are significantly different from 0 in all cases is the relative sales of foreclosures compared toregular sales, from 2008 to 2010. The magnitude, however, seems to increase as we increase thebandwidth length, in which case I take the value of 13 percentage points decrease in foreclosuresales to be the benchmark result.

4.2.2 Limitations of the discontinuity analysis.

In general, although we showed that results seem to be in a controlled environment with respectto both current and borrower origination characteristics, the discontinuity analysis here is not ableto exactly estimate the rate of defaults between recourse and non-recourse states, and it is not theobjective per se. Instead, it depends on such channel to provide an explanation for the results foundhere. The differences in the relative number of houses offered as an REO sale would be affectedmainly by a supply channel in which fewer homeowners are defaulting in such regions. In order totruly test the differences in delinquency rates, one would need to observe then. Given our dataset,we only observe a fraction of delinquents: those who exited trough foreclosures and those whochose among a couple of alternatives such as ownership transfer to banks or short-sales. We missthe delinquents who were able to liquidate their mortgages after default but before any legal actionwas taken.

Also, another limitation is that sales may not correspond to listings. It is hard to say whethersome banks formed larger inventories in order to push prices up, which could potentially explain

15

the results obtained. Sadly Data Quick does not record the time on sale (the difference of daysbetween the listing and the actual sale), however, for REO sales in specific, Data Quick seems torecord the moment in which the house was transferred to the REO. Counting the days betweenthe moment of sale and the moment of transference to REO could give us a proxy for how manydays the house stayed on the market (assuming REOs list the houses at moment they receivethem). Table 8 shows the average interval between the year groups. Between 2008 and 2010, theaverage difference of days on the market for REO sales between no-recourse and recourse sidesare approximatedly equal to 10 days. Based on estimations from Clauretie and Daneshvary [2011]who estimated the impact of TOM of log prices, each day on the market reduces the house priceby 0.1%. This implies that the difference in days on sale is relatively small in order to explain theimpact observed in the data.

Should we try to estimate the propensity of observing an REO sale, as compared to observinga regular sale? Unfortunately, we may face issues regarding several relevant unobservables suchas credit scores and the current mortgage balance. One of the assumptions implicit in the sharpdiscontinuity analysis is that the unobservable characteristics are continuous at the border. We tryour best to show that some of the origination characteristics are indeed similar, however, we do notobserve the entire mortgage history. Nevertheless, the results are valid if we assume that incomeshocks follow a similar distribution close to the border.

Are the results driven by one specific border? In unreported analyses, I checked whether re-moving one of the borders would change the results systematically. In terms of data, the largestborders are NV-AZ and NC-SC. By removing one of them at a time, none of the results changed,although confidence intervals increased significantly (but up to the point where results are signifi-cant for at least 90%).

4.3 Foreclosure House Prices and Recourse.

Even though prices are not “discontinuous at the border”, it does not mean that there were no pricedifferences between these two markets. For instance, if one believes that the increase in the numberof local foreclosed properties could affect prices through competition, then the previous resultsindicate that prices of foreclosed houses in non-recourse markets could have been lower (comparedto regular sales prices and conditional on local market characteristics and house characteristics).In a simple demand and supply equilibrium framework, a large positive supply shock would causeprices to weakly decrease, given the demand elasticity.

We must, however, compare relative prices of foreclosures based on the alternative local regularmarket. The idea that foreclosure prices sell at a discount can be due to several effects which werealready documented in the literature. Houses may lose value simply because they went throughforeclosure, as foreclosed homeowners would not have incentives to cover up for maintenance

16

costs. Guren and Mcquade [2013] explains that REO servicers are compensated by the volume ofhouses sold, and so they would be willing to take a higher discount in order to get rid of the asset.Other explanations could involve the bank’s capital requirement constraints, as houses are definedas risky asset in their balance sheets.

Also in Guren and Mcquade [2013], the composition of foreclosed and regular houses mayaffect house values in both markets. Foreclosed properties may sell at a larger discount if themarket is flooded by other REO houses, either because of higher competition (REO’s would preferto sell faster and take larger discounts in order to not lose the sale) or because it increases the valueof being a buyer (buyers become more “choosey”, as the probability of meeting an REO houseincrease).

Although it is object of further research to derive a full quantitative model that would explainforeclosure price discounts behaved with respect to both recourse and non-recourse markets, theliterature suggests the following form for price discounts on foreclosure sales:

p(h, d)t =

[1− ξ

(c,

vbtvrt + vdt

,vdt

vrt + vdt

)]p(h, r)t (6)

where d stands for “distress” sale, r stands for regular sale and h stands for the house type. ξ(.)is the functional form of the equilibrium discount of foreclosures over regular sales. Based onthe literature, ξ(.) < 1 for any combination of parameters. We let c ∈ R+ be any costs relatedto transaction fees and house maintenance costs, where c > 0 and ξ1 > 0. v stands for the totalvolume of a specific group, and b stands for buyers. The second element of ξ is the total numberof buyers relative to the total number of sellers, or market tightness, while the third element isthe relative composition of foreclosure houses in the market. As more buyers enter the marketcompare to sellers, it becomes easier to find a match, and so ξ2 < 0, while as more houses enterthe market through distress, the higher competition would drive foreclosure prices even lower, sothat ξ3 > 0. It is clear that such conditions can potentially affect regular sales, which is why thisequation would be the result of an equilibrium between both markets.

Unfortunately, we are not able to fully specify α without a proper model, however, we can atleast try to compare average conditional local foreclosure price discounts by estimating a simplestatic hedonic pricing model. Following Campbell et al. [2011], we use a standard static hedonicpricing equation with a few adaptations to understand how price discounts may differ betweenrecourse and non-recourse states. The analysis will be performed only around the state bordersin which recourse laws differ (in an attempt to provide a better control for unobservables). Thefollowing model is proposed to analyze the differences in prices from foreclosure sales and short-sales:

ln(Pi,c,b,s,t) = α0 + κ× Recourses + δ′Fi,t + γ′Fi,t × Recourses (7)

+αc,t + β′Xi,t + φb + Lender FE + εi,c,b,s,t

17

where i stands for transaction, c stands for Census tract block area, b stands for county-borderpair location, s stands for state and t stands for time, defined as year17. With respect to the inde-pendent variables, F stands for the type of transaction (whether it was an auction sale, REO sale,or Short-sale), f() stands for a linear functional, αc,t stands for location-year fixed effects, Xi,t

gathers several transaction-specific information, including house covariates, buyer’s financing anddemographic information, the seller’s type and what type of lending institution has financed themortgage, if should there be one18, and finally φb represents the cross-border-county-specific fixedeffects.

We also allow for lender fixed effects in our regression. Most of the REOs are also the lenderswho financed the new purchase, and so controlling for lenders might help to control for differentlender behavior over the distressed sales market, as lenders might act differently with respect toinventory and additional liquidity constraints. The standard errors are clustered at the state-yearlevel, since our main variation of importance is between states.

The recourse dummy and the interaction with F allows us to capture the difference in lossesof distress sales over regular sales between both recourse and no-recourse markets, conditional onthe list of covariates. For instance, the conditional difference of losses for REO sales is given by:

γREO = {E[log(Price)|REO = 1,Recourse = 1;X]− E[log(Price)|REO = 0,Recourse = 1;X]}︸ ︷︷ ︸Conditional discount in recourse state

(8)

−{E[log(Price)|REO = 1,Recourse = 0;X]− E[log(Price)|REO = 0,Recourse = 0;X]}︸ ︷︷ ︸Conditional discount in non-recourse state

which matches the coefficient from equation 8. Also, notice that the coefficient γ would providean equilibrium comparison between the price discounts of foreclosures ξ(.) between both recourseand non-recourse markets.

Table 10 shows the results for 3 different groups of selection areas: i) entire border county; ii)limited towards 20 miles from the border; iii) limited towards 15 miles from the border. We takeas benchmark the third one since it is the closest towards the border. First, notice that foreclosure(and short-sale) discounts are present throughout all the year groups used for the analysis. Thisis consistent with the previous literature that claims that foreclosures are sold under a discount,conditional on several characteristics. Another important aspect is that, between 2005-2007, therewere no differences in discounts between both recourse and non-recourse states. Since the volumeof foreclosure sales was considerably low (and assuming that the relative number of buyers wassimilar among both sides of the border), the estimation provides an interesting controlled environ-ment in which we can base our future analysis.

17Time could be consistently defined as quarters, but since we are grouping the observations by specific time periods,it seems more intuitive to look over possible yearly trends. Also, one should not expect local market conditions to varyconsiderably during a year interval.

18For a complete list of covariates, please refer to Table 2

18

From 2008 to 2010, discounts on REO sales are smaller in recourse states by approximate6.6 percentage points. Based on our previous session on border discontinuities, the proportionof REO sales compared to regular sales seem to be smaller in the recourse side of the border,conditional on several local market characteristics. This result goes in favor with the explanationthat foreclosure sellers may sell under larger discounts if the market is flooded by foreclosures.Moreover, since foreclosures lock defaulted homeowners out of the housing market, the relativeproportion of buyers would decrease relatively faster in the non-recourse market, which wouldalso favor the results. From 2011 to 2012, recourse state discounts are smaller by approximate 9.0percentage points. There could be several channels in which the difference increases from whichwe are not able to disentangle given our current analysis (but it will be object of future research).

Also surprisingly, short sale discounts were not statistically different between recourse andnon-recourse states, which seem to suggest that these price impacts observed were specific toforeclosure sales only. However, the imbalance in terms of observables seems to be strongeramong foreclosure sales. In Figure 5, the second row compares the volume of short-sales withregular sales. Although there is some impact of recourse laws (homeowners would prefer to exitthrough short-sales instead). Moreover, the recourse dummy would capture the price differencesbetween recourse and non-recourse states for the group of control. The regression does not allow usto understand whether there were any potential implications on regular house prices. The literatureon price spillover effects of foreclosures over the local regular housing market would suggest thatprices would decline stronger in states with relatively more foreclosure sales, but the recoursedummy is not significant from 2008 to 2010.

4.3.1 Robustness analysis and additional explanations

The reduced number of state borders available could potentially undermine the results if the resultswe are capturing are caused by an outlier border. Moreover, other state foreclosure laws, such asjudicial foreclosures, could also affect the validity of the estimation analysis. Among all the statesin the sample, the ones who are also judicial are South Carolina and Texas. Since the number ofobservations in the Texas border is relatively small, we should concentrate mainly over the SC-NCborder. Table 11 shows the results for the main coefficients. There are no differences with respectto the main coefficients. Estimations are done using all transactions from border counties in orderto guarantee robustness of the results. Secondly, I remove the AZ-NV border. Results, displayedby Table 7 are still resilient, giving additional robustness to the estimation.

4.3.2 Limitations

Linear regression model tends to observe equilibrium differences over price discounts across re-course and non-recourse states only, based on a controlled experiment that such differences were

19

basically null during the boom period. It does not allow us to pin down exactly what are the forcesbehind such results. The previous border analysis may suggest a supply shock story, but we wouldrequire a model structural version of the representation to exactly claim that the decrease in dis-counts was due to fewer foreclosures available. This is the main limitation of this approach, and itis the objective of further research to build over a quantitative model in which prices are flexibleand may react to search costs and matching frictions, which should be able to capture this channel.

Another potential explanation can be related to the unobservable quality of the houses. Ifhomeowners expect to be foreclosed, their incentives to still invest in house maintenance costs isrelatively higher in recourse states, simply because houses may increase their value. The channelcan be easily seen in our previous analysis on the probability of default by including maintenancecosts and the potential impact on future house prices. So long as the cost is not high enoughand the relative benefit on prices compensate the cost, this could potentially explain the results.Once again, this feature can only be captured in a structural framework, since we do not observemaintenance expenses.

However, the fact that foreclosure prices may be different between recourse and non-recoursestates is essentially an empirical one, which was pursued in this paper. Understanding all thepossible channels that may affect prices is a challenge for further research in the area

20

5 Remarks and Future work

The current paper explored foreclosure house markets between recourse and non-recourse states.Among the main findings obtained, the discontinuity design showed that there were fewer housessold as foreclosure compared to the regular sales market in recourse states. Moreover, averageforeclosure discounts on recourse states (compared to local regular sales) seem to be lower inrecourse states, possibly because of fewer competition among fewer foreclosures.

In terms of implementation, the geographic discontinuity design under recourse states usinglocal individual transaction variation can be considered a recent feature in the literature and hasnot been fully explored yet. Many discontinuity analysis concentrated only on zip code leveland aggregate price variables, not separating foreclosure sales from the regular housing market.By analyzing the foreclosure sales market in specific, one can draw conclusions regarding howcollateralized house values were assessed in the recent financial crisis depending on whether stateshave stronger recourse laws or not.

As mentioned previously, the study suffers from a series of limitations. The supply channelon relative prices could be explained by differences in unobserved quality or from REO servicerswaiting for better offers to appear in recourse states. Although the latter does not have a strongempirical support given the analysis on the total average time to sale from the REO record entry,the former is not testable given the data observed. In order to disentangle such hypothesis fromthe main channel, one could use a structural modeling approach of the housing market applied torecourse law variations and price composition effects as caused by the relative volume of foreclo-sures observed, together with maintenance expenditure as an important price determinant. It is theobjective of future research to propose such a fully fledged model.

Moreover, we were able to observe the relative volume of foreclosure sales as being affectedby recourse laws, in which I proposed the strategic default channel as a plausible mechanism toexplain such differences. A further step would be to ascertain whether the default mechanism isindeed playing a role by comparing total delinquencies in such border areas under the same design.

21

References

John Y. Campbell, Stefano Giglio, and Parag Pathak. Forced sales and house prices. American

Economic Review, 101(5):2108–31, 2011. doi: 10.1257/aer.101.5.2108. URL http://www.

aeaweb.org/articles.php?doi=10.1257/aer.101.5.2108.

TerrenceM. Clauretie and Nasser Daneshvary. The optimal choice for lenders facing defaults:Short sale, foreclose, or reo. The Journal of Real Estate Finance and Economics, 42(4):504–521, 2011. ISSN 0895-5638. doi: 10.1007/s11146-009-9201-3. URL http://dx.doi.

org/10.1007/s11146-009-9201-3.

Dean Corbae and Erwan Quintin. Leverage and the foreclosure crisis. Journal of Political Economy

(forthcoming), pages 1–86, 2014.

Erik R. de Wit and Bas van der Klaauw. Asymmetric information and list-price reductions inthe housing market. Regional Science and Urban Economics, 43(3):507 – 520, 2013. ISSN0166-0462. doi: http://dx.doi.org/10.1016/j.regsciurbeco.2013.03.001. URL http://www.

sciencedirect.com/science/article/pii/S0166046213000240.

Cem Demiroglu, Evan Dudley, and Christopher M. James. State foreclosure laws and the incidenceof mortgage default. Journal of Law and Economics, 57(1):pp. 225–280, 2014. URL http:

//www.jstor.org/stable/10.1086/674868.

Mark J. Garmaise and Tobias J. Moskowitz. Confronting information asymmetries: Evidence fromreal estate markets. Review of Financial Studies, 17(2):405–437, 2004. doi: 10.1093/rfs/hhg037.URL http://rfs.oxfordjournals.org/content/17/2/405.abstract.

Kristopher Gerardi, Kyle F. Herkenhoff, Lee E. Ohanian, and Paul S. Willen. Unemployment,negative equity, and strategic default. Working paper, pages 1–50, 2013.

Andra Ghent. America’s mortgage laws in historical perspective. Working paper, pages 1–40,2012.

Andra C. Ghent and Marianna Kudlyak. Recourse and residential mortgage default: Evidencefrom us states. Review of Financial Studies, 24(9):3139–3186, 2011. doi: 10.1093/rfs/hhr055.URL http://rfs.oxfordjournals.org/content/24/9/3139.abstract.

Adam Guren and Tim Mcquade. How do foreclosures exacerbate housing downturns. Working

Paper, pages 1–53, 2013.

Guido Imbens and Karthik Kalyanaraman. Optimal bandwidth choice for the regression dis-continuity estimator. The Review of Economic Studies, 79(3):933–959, 2012. doi: 10.1093/restud/rdr043. URL http://restud.oxfordjournals.org/content/79/3/933.

abstract.

22

Guido W. Imbens and Thomas Lemieux. Regression discontinuity designs: A guide to practice.Journal of Econometrics, 142(2):615 – 635, 2008. ISSN 0304-4076. doi: http://dx.doi.org/10.1016/j.jeconom.2007.05.001. URL http://www.sciencedirect.com/science/

article/pii/S0304407607001091. The regression discontinuity design: Theory andapplications.

Joao F. Cocco John Y. Campbell. A model of mortgage default. Working paper, pages 1–78, 2014.

Pablo Kurlat and Johannes Stroebel. Testing for information asymmetries in real estate markets.Working Paper 19875, National Bureau of Economic Research, January 2014. URL http:

//www.nber.org/papers/w19875.

Tim Landvoigt, Monika Piazzesi, and Martin Schneider. The housing market(s) of san diego.Working Paper, pages 1–51, 2013.

David S. Lee and Thomas Lemieux. Regression discontinuity designs in economics. Journal of

Economic Literature, 48(2):281–355, 2010. doi: 10.1257/jel.48.2.281. URL http://www.

aeaweb.org/articles.php?doi=10.1257/jel.48.2.281.

Steven D. Levitt and Chad Syverson. Market distortions when agents are better informed: Thevalue of information in real estate transactions. The Review of Economics and Statistics, 90(4):599–611, 2008.

Atif Mian, Amir Sufi, and Francesco Trebbi. Foreclosures, house prices and the real economy.Working paper, pages 1–59, 2014.

Tong Yob Nam and Seungjoon Oh. Recourse mortgage laws and the housing bubble. Working

Paper, pages 1–51, 2014.

Erwan Quintin. More punishment, less default? Annals of Finance, 8(4):427–454, 2012. ISSN1614-2446. doi: 10.1007/s10436-012-0203-4. URL http://dx.doi.org/10.1007/

s10436-012-0203-4.

Li Ding Te Bao. Non-recourse mortgage and housing price bubble, burst and recovery. Working

Paper, pages 1–24, 2014.

Shuang Zhu and R. Kelley Pace. Factors underlying short sales. Working paper, pages 1–33, 2013.

23

Appendix A Figures

Figure 1: Coverage Area provided by Data Quick dataset, from 2005 to 2012

24

Figure 2: Recourse and Non-Recourse States, together with selected neighbor couties.

25

Figure 3: Discontinuity analysis: Histogram of current transactions available.

26

Figure 4: Discontinuity analysis: Transaction log prices For the residualized variable based on 1, eachgraph contains two series as function of D (distance): (i) the average outcome/covariate per 0.5 mile bin; (ii) a non-parametric estimation as a function of D (Nadaraya-Watson with Normal Kernel and optimal bandwidth based onSilverman’s rule). 95% confidence intervals also plotted (standard errors calculated by bootstrap). negative distancefor non-recourse states.

27

Figure 5: Discontinuity analysis: Total volume of foreclosure sales based on regular sales andshort-sales For the residualized variable based on 1, each graph contains two series as function of D (distance):(i) the average outcome/covariate per 0.5 mile bin; (ii) a non-parametric estimation as a function of D (Nadaraya-Watson with Normal Kernel and optimal bandwidth based on Silverman’s rule). 95% confidence intervals also plotted(standard errors calculated by bootstrap). negative distance for non-recourse states.

28

Figure 6: Discontinuity analysis: Income at the origination ($ 1000)

Figure 7: Discontinuity analysis: Loans insured by FHA at the origination.

29

Figure 8: Discontinuity analysis: LTV ratio at origination

Figure 9: Discontinuity analysis: Total number of days from original purchase to currentsale.

30

Figure 10: Discontinuity analysis: Current Borrower’s Income ($ 1000).

Figure 11: Discontinuity analysis: Current Borrower’s LTV ratio.

31

Appendix B Tables

State ANSI Name Judicial Non Judicial Recourse Redemption Red. Period Comments

AL 1 Alabama 1 1 1 1 365AK 2 Alaska 1 1 0 1 365AZ 4 Arizona 1 1 0 1 180AR 5 Arkansas 1 1 1 1 365CA 6 California 1 1 0 1 730CO 8 Colorado 1 1 1 0 0CT 9 Connecticut 1 0 1 0 0DE 10 Delaware 1 0 1 0 0DC 11 District of Columbia 0 1 1 0 0FL 12 Florida 1 0 1 0 0

GA 13 Georgia 1 1 1 0 0HI 15 Hawaii 1 1 0 0 0ID 16 Idaho 1 1 1 0 0IL 17 Illinois 1 0 1 1 90IN 18 Indiana 1 0 1 0 0IA 19 Iowa 1 1 1 1 365KS 20 Kansas 1 0 1 1 365KY 21 Kentucky 1 0 1 1 365LA 22 Louisiana 1 0 1 0 0ME 23 Maine 1 0 1 1 90MD 24 Maryland 1 0 1 0 0MA 25 Massachusetts 1 0 1 0 0MI 26 Michigan 0 1 1 1 365

MN 27 Minnesota 1 1 0 1 180MS 28 Mississippi 1 1 1 0 0MO 29 Missouri 1 1 1 1 365MT 30 Montana 1 1 0 1 365NE 31 Nebraska 1 0 1 0 0 Recourse only for Non Judicial mortgagesNV 32 Nevada 1 1 1 0 0NH 33 New Hampshire 0 1 1 0 0NJ 34 New Jersey 1 0 1 1 180

NM 35 New Mexico 1 0 0 1 270NY 36 New York 1 0 1 0 0 Recourse only for NYC MSANC 37 North Carolina 1 1 0 0 0 No Recoure only for purchase mortgagesND 38 North Dakota 1 0 0 1 60OH 39 Ohio 1 0 1 0 0OK 40 Oklahoma 1 1 0 0 0OR 41 Oregon 1 1 0 0 0PA 42 Pennsylvania 1 0 1 0 0RI 44 Rhode Island 1 1 1 0 0SC 45 South Carolina 1 0 1 0 0SD 46 South Dakota 1 1 1 1 365TN 47 Tennessee 0 1 1 1 730TX 48 Texas 1 1 1 0 0 Recourse only for Non judicial mortgagesUT 49 Utah 0 1 1 0 0VT 50 Vermont 1 0 1 0 0VA 51 Virginia 1 1 1 0 0WA 53 Washington 1 1 0 0 0WV 54 West Virginia 0 1 1 0 0WI 55 Wisconsin 1 1 1 0 0

WY 56 Wyoming 1 1 1 1 90

Table 1: State Foreclosure Laws

32

Variable Description

Buyer/transaction covariatesOrigination buyer’s LTV Buyer’s LTV ratio at the moment of transactionCorporate Buyer Dummy whether buyer is a corporate institutionGovernment Buyer Dummy whether buyer is a government institutionTrustee Buyer Dummy whether buyer is a trust institutionAsian Dummy whether buyer is defined as Asian (HMDA)Black Dummy whether buyer is defined as Black (HMDA)Hispanic Dummy whether buyer is defined as Hispanic (HMDA)Male Dummy whether buyer is defined as Male (HMDA)White Dummy whether buyer is defined as White (HMDA)Buyer’s Income Income from buyer at the moment of transaction ($ 1000, HMDA)High LTV Dummy whether LTV ¿ 1.Loan is FHA Dummy whether loans are insured by the FHA

Lender/transaction covariatesBank Dummy whether lender is a commercial bankCredit Union Dummy whether lender is a credit unionDeveloper Dummy whether lender is a developerFinancial Institution Dummy whether lender is a financial institution (other than the ones listed)Holding Dummy whether lender is a bank holding companyGovernment Dummy whether lender is a government institutionInsurance Company Dummy whether lender is an Insurance companyMortgage company Dummy whether lender is a mortgage companyMortgage banker Dummy whether lender is a mortgage bankerMERS Dummy whether lender belongs to the MERS systemFSB Dummy whether lender is a Federal Savings BankCommercial Corp Dummy whether lender is a Commercial Corporation

Seller covariatesCorporate Owner Dummy whether owner is a corporate/trust/estate/government institutionGovernment Owner Dummy whether owner is a government institutionTrustee Owner Dummy whether owner is a trust institution

House CovariatesSingle/Condo 1 if house is a Single residence, 2 if house is a Condo/Coop residenceSq ft Addition Square feet of identified additionsSq ft Building Building square feetSq ft Unfinished basement Unfinished basement square feetSq ft Garage Garage Square feetSq ft total Total Square feetCondition code Property condition code (ranging from 0 = Unsound to 6 = Excellent)Lot Depth Lot DepthLot Width Lot WidthLot Size Lot Size (in square feet)Garage 1 if it has a garagePorch 1 if it has a porchPool 1 if it has a pool# bathrooms Number of bathrooms for all structures# bedrooms Number of bedrooms for all structures# rooms Number of rooms for all structures# stories Number of stories for all structures# units Number of units for all structures# structures Number of structures on propertyAge Age of specific house, based on effective year built. (in years)

Table 2: List of covariates.

33

Border Counties Discontinuity Analysis (15 miles from border)State-Border pair Recourse Non-recourse Recourse Non-recourseAZ-NV 241553 12902 19910 5729CA-NV 86242 20534 10888 10664NC-SC 71586 109980 30361 40443NC-TN 4118 1474 1093 2006NC-VA 11540 1945 3684 1398Total 415039 146835 65936 60240

Table 3: Total Volume of Transactions located at the border counties, separated by border-state pairs, from 2005 to 2012. First two columns show the total number of transactions available by state-border pair using only border counties. Last two columns show the total number of transactions available for thediscontinuity analysis. For the sharp discontinuity, data is selected to be at least 15 miles close to the border (norestrictions on counties being at the border). Moreover, in order to avoid large transaction imbalances between bothsides, county-border pairs with a difference of observations higher than 5% are removed from sample.

Border CountiesState-Border pair Recourse Non-recourseAZ-NV 145070 6023CA-NV 60464 11377NC-SC 32377 68928NC-TN 1442 141NC-VA 8712 1159Total 248065 87628

Table 4: Total Volume of Effective Transactions located at the border counties, from 2005 to2012. Total volume of effective transactions located at the border counties used in the linear static hedonic pricingregression, presented later in the paper. Data removed consists of missing data on house characteristics and borrower’scharacteristics such as income, loan information and other demographics such as race, gender and ethnicity.

34

Border Counties Discontinuity Analysis (15 miles from border)State-Border pair Recourse Non-recourse Recourse Non-recourseAZ/15-NV/3 241553 12902 19910 5729CA/17-NV/5 1129 11121 1132 4002CA/27-NV/3 18658 2 - -CA/3-NV/5 3010 256 - -CA/35-NV/31 344 1892 242 216CA/49-NV/31 3 267 - -CA/51-NV/19 586 8 120 8CA/51-NV/5 104 72 105 72CA/57-NV/31 4095 2683 3444 2287CA/61-NV/31 2159 3118 2165 3562CA/61-NV/5 75 183 75 180CA/61-NV/510 3497 600 3605 337CA/71-NV/3 20963 15 - -CA/91-NV/31 31619 317 - -NC/11-TN/19 - - 294 324NC/115-TN/171 101 270 179 526NC/115-TN/29 - - 387 42NC/119-SC/57 419 10862 - -NC/119-SC/91 2395 54824 - -NC/161-SC/21 123 426 1149 353NC/161-SC/83 4127 455 2994 408NC/175-SC/45 3231 590 505 1352NC/175-SC/73 45 102 275 103NC/175-SC/77 3000 85 1421 85NC/179-SC/57 5154 14390 6787 27328NC/189-TN/91 16 976 155 933NC/19-SC/51 30981 13856 13891 5258NC/199-TN/171 - - 78 181NC/45-SC/21 1203 1195 1303 1334NC/45-SC/91 212 142 153 292NC/53-VA/550 - - 1168 391NC/53-VA/810 11540 1945 2516 1007NC/71-SC/91 311 8148 - -NC/87-TN/155 2307 26 - -NC/87-TN/29 1694 202 - -NC/89-SC/45 20381 4869 1879 3894NC/99-SC/73 4 36 4 36Total 415039 146835 65936 60240

Table 5: Total Volume of Transactions located at the border counties, separated by border-county pairs, from 2005 to 2012. First two columns show the total number of transactions available bystate-border pair using only border counties. Last two columns show the total number of transactions available forthe discontinuity analysis. For the sharp discontinuity, data is selected to be at least 15 miles close to the border (norestrictions on counties being at the border). Moreover, in order to avoid large transaction imbalances between bothsides, county-border pairs with a difference of observations higher than 5% are removed from sample.

35

Recourse No-RecourseN mean N Mean Difference

Log house price 221,649.00 12.30 75,672.00 12.04 0.256***REO sale 221,754.00 0.02 75,839.00 0.04 -0.022***Short-sale 221,754.00 0.01 75,839.00 0.01 0.002***Auction-sale 221,754.00 0.00 75,839.00 0.00 0.001***Ownership transfer 221,754.00 0.09 75,839.00 0.04 0.054***FHA-insured 175,762.00 0.018 57,821.00 0.031 -0.013***LTV 221,543.00 0.71 75,670.00 0.69 0.018***LTV (mortgage) 175,551.00 0.898 57,652.00 0.910 -0.012***LTV > 1 175,762.00 0.038 57,821.00 0.052 -0.014***Buyer: Income 172,978.00 108.34 57,436.00 104.83 3.515***Buyer: Asian 172,978.00 0.07 57,436.00 0.03 0.047***Buyer: Black 172,978.00 0.06 57,436.00 0.12 -0.064***Buyer: Hispanic 172,978.00 0.16 57,436.00 0.07 0.090***Buyer: Male 166,420.00 0.66 55,187.00 0.67 -0.017***Buyer: Corp 221,754.00 0.05 75,839.00 0.07 -0.018***Buyer: Gov 221,754.00 0.00 75,839.00 0.00 -0.000***Buyer: Trustee 221,754.00 0.02 75,839.00 0.03 -0.00Seller: Corp 221,754.00 0.07 75,839.00 0.11 -0.037***Seller: Gov 221,754.00 0.00 75,839.00 0.00 0.000**Seller: Trustee 221,754.00 0.08 75,839.00 0.07 0.011***Lender: Commercial Bank 217,329.00 0.13 73,683.00 0.21 -0.076***Lender: C.U. 217,329.00 0.01 73,683.00 0.02 -0.006***Lender: Development 217,329.00 0.00 73,683.00 0.00 -0.000***Lender: Fin. Inst. 217,329.00 0.11 73,683.00 0.07 0.045***Lender: FSB 217,329.00 0.05 73,683.00 0.04 0.011***Lender: Holding comp. 217,329.00 0.04 73,683.00 0.01 0.026***Lender: Gov 217,329.00 0.00 73,683.00 0.00 0.000***Lender: Insurance comp. 217,329.00 0.00 73,683.00 0.00 -0.000***Lender: MERS 217,329.00 0.03 73,683.00 0.03 -0.002**Lender: Mort. Banker 217,329.00 0.03 73,683.00 0.04 -0.009***Lender: Mort. Company 217,329.00 0.38 73,683.00 0.34 0.043***House: Age 213,213.00 16.13 69,767.00 21.68 -5.549***House: Condition 221,744.00 2.02 75,835.00 2.41 -0.396***House: Garage 221,754.00 0.66 75,839.00 0.49 0.168***House: Lotsize 221,749.00 32,000.00 75,827.00 38,000.00 -5.9e+03***House # Bathrooms 221,748.00 2.31 75,836.00 2.08 0.230***House # Bedrooms 221,750.00 2.87 75,835.00 2.54 0.326***House # Units 221,727.00 0.90 75,836.00 0.73 0.170***House: Pool 221,754.00 0.13 75,839.00 0.08 0.057***House: Condo 221,754.00 0.24 75,839.00 0.14 0.104***House: Sq.ft. 221,749.00 1,700.16 75,832.00 1,858.52 -158.363***Previous: Construction 202,179.00 0.01 73,233.00 0.01 -0.010***Previous: Resale 202,179.00 0.36 73,233.00 0.37 -0.00Previous: Subdivision 202,179.00 0.09 73,233.00 0.03 0.061***Future transaction: Is Refi/Equity ext. 160,917.00 0.38 61,083.00 0.45 -0.073***Future transaction: % Refinanced 51,416.00 -0.20 23,605.00 -0.23 0.024***Future transaction: Time until Refi. 61,068.00 623.22 27,656.00 747.23 -124.003***REO: Time until sale 2,116.00 191.57 2,081.00 162.17 29.405***

Table 6: Ex-ante Average of Observable Characteristics, Border Counties only, from 2004to 2006. Average ex-ante summary statistics for relevant observables, using information provided by Data Quicktogether with HMDA from 2004 to 2006. Limited demographic observations due to either lack of a proper matchbetween Data Quick and HMDA, or absence of mortgage information (missing or non-existent). Other differencesin observations are due to missing information. REO time until sale is calculated whenever Data Quick flags thedifference between the date of sale and the date in which the specific house is transferred to the Real Estate Owner.

36

2005-2007 2008-2010 2011-20120.5× h

coeff t-stat h coeff t-stat h coeff t-stat hlog(P )foreclosure 0.08 0.28 1.13 0.37 5.20 1.23 0.19 0.79 1.23log(P )regular 0.01 0.13 0.60 0.11 0.71 1.28 -0.15 -0.61 1.45log(P )short−sale 0.60 1.26 1.93 0.63 1.58 1.19 0.07 0.40 1.161(Foreclosure over regular) 0.00 0.04 0.47 -0.10 -2.08 0.78 0.10 1.39 0.921(Short-sale over regular) -0.02 -1.42 0.57 -0.03 -1.00 0.91 -0.01 -0.11 0.90

1× hcoeff t-stat h coeff t-stat h coeff t-stat h

log(P )foreclosure 0.56 1.43 2.26 0.33 2.54 2.47 0.35 1.77 2.46log(P )regular 0.21 1.82 1.20 0.21 1.22 2.55 0.09 0.42 2.90log(P )short−sale 0.42 0.71 3.86 0.39 1.23 2.38 0.18 0.66 2.311(Foreclosure over regular) 0.01 0.80 0.95 -0.13 -4.47 1.57 0.00 0.09 1.841(Short-sale over regular) 0.00 0.07 1.14 -0.03 -1.49 1.82 0.01 0.27 1.79

2× hcoeff t-stat h coeff t-stat h coeff t-stat h

log(P )foreclosure 0.12 0.51 4.52 0.08 0.61 4.94 0.16 1.64 4.93log(P )regular 0.17 0.60 2.40 0.00 0.02 5.10 -0.01 -0.10 5.80log(P )short−sale 0.35 0.75 7.71 0.15 0.72 4.76 0.13 0.76 4.631(Foreclosure over regular) 0.01 0.83 1.90 -0.18 -3.92 3.14 -0.01 -0.33 3.671(Short-sale over regular) -0.01 -2.49 2.27 -0.05 -3.50 3.65 0.05 1.53 3.59

Table 7: Nonparametric Estimation (Local Linear Regression with Optimal AsymptoticBandwidth) - Average Treatment Effect Sharp regression discontinuity average treatment effect estima-tion, following the local linear regression methodology described by equation 5. Optimal asymptotic bandwidth basedon Imbens and Kalyanaraman [2012] is used, together with a downscale and upscale of the order of 2. Standard errorsare based on Imbens and Lemieux [2008] and are robust to heterocedasticity. MATLAB code provided by prof. GuidoImbens in his webpage and adapted to our specification.

15 Miles from State Border 10 Miles from State BorderPeriod Recourse side No-recourse side Recourse side No-recourse side2005-2007 177.362 155.378 182.1266 159.279

(147.236) (114.703) (120.324) (130.441)2008-2010 150.282 159.919 150.723 163.011

(109.560) (117.866) (109.477) (122.621)2011-2012 169.935 193.137 176.495 186.964

(134.833) (141.991) (139.978) (132.310)

Table 8: Average total number of days between the house being transferred to the REO andsold as an REO sale Standard errors in parenthesis.

37

Dep. Var: log(Price) Border Counties only Distance to border < 20 miles Distance to border < 15 miles2005-2007 2008-2010 2011-2012 2005-2007 2008-2010 2011-2012 2005-2007 2008-2010 2011-2012

Recourse -0.184∗ ∗ ∗ -0.442∗ ∗ ∗ -0.382∗ ∗ ∗ -0.366∗∗ -0.575∗∗ -0.416 -0.090 -0.310 0.526∗ ∗ ∗(-2.78) (-6.40) (-3.16) (-2.54) (-2.69) (-1.55) (-0.57) (-1.53) (10.41)

REO Sale -0.175∗ ∗ ∗ -0.196∗ ∗ ∗ -0.230∗ ∗ ∗ -0.167∗ ∗ ∗ -0.190∗ ∗ ∗ -0.224∗ ∗ ∗ -0.175∗ ∗ ∗ -0.205∗ ∗ ∗ -0.229∗ ∗ ∗(-16.39) (-10.27) (-9.07) (-21.40) (-7.94) (-9.94) (-28.19) (-8.11) (-11.84)

Short-sale -0.091∗ ∗ ∗ -0.119∗ ∗ ∗ -0.134∗ ∗ ∗ -0.089∗ ∗ ∗ -0.108∗ ∗ ∗ -0.108∗ ∗ ∗ -0.087∗ ∗ ∗ -0.106∗ ∗ ∗ -0.112∗ ∗ ∗(-4.50) (-14.97) (-12.62) (-3.86) (-11.19) (-16.25) (-2.93) (-10.15) (-12.21)

REO × Recourse 0.013 0.085∗ ∗ ∗ 0.134∗ ∗ ∗ -0.008 0.066∗∗ 0.107∗ ∗ ∗ -0.011 0.065∗∗ 0.089∗ ∗ ∗(0.63) (2.95) (4.69) (-0.39) (2.15) (4.39) (-0.40) (2.15) (3.63)

Short-sale × Recourse -0.009 -0.006 0.008 -0.028 -0.010 -0.018∗∗ 0.006 0.006 -0.001(-0.30) (-0.51) (0.59) (-0.94) (-0.84) (-2.76) (0.20) (0.40) (-0.06)

Auction Sale 0.000 -0.044 -0.064∗ ∗ ∗ 0.000 -0.099∗ ∗ ∗ -0.107∗ ∗ ∗ 0.000 -0.069 -0.148∗ ∗ ∗(.) (-1.19) (-4.50) (.) (-3.30) (-3.88) (.) (-1.43) (-4.16)

Ownership Transfer -0.018∗∗ 0.047∗ ∗ ∗ -0.196∗ ∗ ∗ -0.013 0.030 -0.583∗ ∗ ∗ -0.015 0.036 -0.290∗ ∗ ∗(-2.12) (4.06) (-27.82) (-0.99) (1.11) (-4.96) (-1.08) (1.24) (-9.72)

LTV > 1 -0.069∗ ∗ ∗ -0.015 -0.024∗ -0.055∗ ∗ ∗ -0.034∗∗ -0.037∗ ∗ ∗ -0.056∗ ∗ ∗ -0.041∗ ∗ ∗ -0.037∗∗(-4.81) (-0.99) (-1.91) (-4.50) (-2.50) (-3.46) (-4.00) (-2.93) (-2.79)

FHA insured -0.037∗ ∗ ∗ -0.014∗∗ -0.015 -0.023∗ ∗ ∗ -0.012∗ -0.014 -0.023∗ ∗ ∗ -0.013 -0.019(-5.31) (-2.46) (-1.31) (-3.82) (-1.85) (-0.94) (-5.34) (-1.58) (-0.99)

LTV -0.144∗ ∗ ∗ -0.095∗∗ -0.039 -0.156∗ ∗ ∗ -0.107∗ -0.055 -0.170∗ ∗ ∗ -0.133∗∗ -0.074(-7.31) (-2.09) (-0.88) (-9.69) (-2.00) (-0.95) (-9.88) (-2.52) (-1.04)

Income 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗(5.63) (5.48) (4.90) (6.92) (5.52) (4.59) (7.36) (5.17) (4.38)

Asian -0.020∗ ∗ ∗ -0.030∗ ∗ ∗ -0.038∗ ∗ ∗ -0.022∗ ∗ ∗ -0.033∗ ∗ ∗ -0.045∗ ∗ ∗ -0.030∗ ∗ ∗ -0.036∗ ∗ ∗ -0.049∗ ∗ ∗(-4.41) (-6.53) (-6.25) (-3.22) (-4.22) (-6.51) (-3.72) (-4.19) (-4.06)

Black -0.044∗ ∗ ∗ -0.027∗ ∗ ∗ -0.017 -0.029∗ ∗ ∗ -0.027∗ ∗ ∗ -0.011 -0.043∗ ∗ ∗ -0.031∗∗ -0.039(-5.86) (-3.73) (-1.63) (-5.51) (-4.58) (-0.59) (-6.04) (-2.69) (-1.64)

Hispanic -0.008 -0.012∗ -0.030∗ ∗ ∗ -0.015 -0.020∗ ∗ ∗ -0.036∗∗ -0.024∗ ∗ ∗ -0.035∗ ∗ ∗ -0.071∗ ∗ ∗(-0.94) (-2.03) (-3.10) (-1.45) (-2.91) (-2.80) (-3.17) (-7.09) (-4.29)

Male 0.011∗ ∗ ∗ 0.007∗∗ 0.013∗ ∗ ∗ 0.011∗∗ 0.011∗∗ 0.002 0.012∗∗ 0.015∗ ∗ ∗ 0.007(3.42) (2.16) (3.45) (2.71) (2.73) (0.29) (2.39) (2.87) (1.21)

Buyer corporate -0.180∗ ∗ ∗ -0.078∗ ∗ ∗ -0.059∗ ∗ ∗ -0.214∗ ∗ ∗ -0.110∗∗ -0.089∗ ∗ ∗ -0.240∗ ∗ ∗ -0.111∗∗ -0.118∗ ∗ ∗(-2.78) (-2.78) (-7.23) (-3.81) (-2.40) (-4.06) (-2.96) (-2.60) (-6.46)

Buyer Gov -1.427∗ ∗ ∗ 0.385∗ ∗ ∗ 0.309∗ ∗ ∗ -1.101 0.362∗ ∗ ∗ 0.430 -0.115 0.534∗∗ -0.155(-5.93) (3.70) (4.96) (-1.55) (3.25) (1.61) (-0.10) (2.51) (-1.05)

Buyer Trustee 0.049∗ ∗ ∗ 0.089∗ ∗ ∗ 0.061∗ ∗ ∗ 0.055∗ ∗ ∗ 0.071∗ ∗ ∗ 0.044∗ ∗ ∗ 0.060∗ ∗ ∗ 0.075∗ ∗ ∗ 0.055∗ ∗ ∗

38

(6.61) (10.15) (12.23) (5.96) (10.49) (4.23) (6.27) (7.83) (3.80)Seller Corporate -0.053∗ ∗ ∗ -0.049∗ ∗ ∗ -0.047∗ ∗ ∗ -0.050∗ ∗ ∗ -0.053∗ ∗ ∗ -0.044∗ ∗ ∗ -0.055∗ ∗ ∗ -0.040∗ ∗ ∗ -0.035∗∗

(-11.66) (-6.18) (-6.19) (-7.87) (-9.27) (-5.33) (-7.69) (-4.74) (-2.14)Seller Gov. -0.357∗ ∗ ∗ -0.499∗ ∗ ∗ -0.319∗ ∗ ∗ -0.198∗ ∗ ∗ -0.374∗∗ -0.281∗ ∗ ∗ -0.084∗∗ -0.370∗ -0.177∗ ∗ ∗

(-7.39) (-4.83) (-7.08) (-3.64) (-2.35) (-10.78) (-2.31) (-1.74) (-3.25)Seller Trustee 0.010 -0.016∗ 0.004 0.004 -0.020∗ 0.003 0.003 -0.011 0.004

(1.35) (-1.78) (0.44) (0.50) (-1.85) (0.26) (0.31) (-0.92) (0.25)House: Age -0.003∗ ∗ ∗ -0.004∗ ∗ ∗ -0.005∗ ∗ ∗ -0.004∗ ∗ ∗ -0.004∗ ∗ ∗ -0.005∗ ∗ ∗ -0.003∗ ∗ ∗ -0.004∗ ∗ ∗ -0.005∗ ∗ ∗

(-6.75) (-7.21) (-8.75) (-5.68) (-5.93) (-7.50) (-5.66) (-5.95) (-7.27)House: Condition 0.008 -0.000 -0.002 0.013 -0.005 -0.006 0.019 -0.011 -0.012∗

(1.42) (-0.22) (-0.92) (1.53) (-1.13) (-1.53) (1.57) (-1.66) (-2.02)House: Garage 0.102∗ ∗ ∗ 0.145∗ ∗ ∗ 0.172∗ ∗ ∗ 0.105∗ ∗ ∗ 0.116∗ ∗ ∗ 0.141∗ ∗ ∗ 0.109∗ ∗ ∗ 0.123∗ ∗ ∗ 0.126∗ ∗ ∗

(14.75) (6.03) (4.84) (13.90) (8.03) (3.95) (12.66) (6.63) (4.18)House: Lot size 0.000 0.000 0.000 -0.000 0.000 0.000 -0.000 -0.000 0.000

(0.12) (1.11) (0.86) (-0.43) (0.55) (0.57) (-1.64) (-0.41) (0.35)House: # Bathrooms 0.044∗ ∗ ∗ 0.017 0.005 0.055∗ ∗ ∗ 0.037∗ ∗ ∗ 0.023∗∗ 0.065∗ ∗ ∗ 0.046∗ ∗ ∗ 0.029∗ ∗ ∗

(4.01) (1.32) (0.26) (6.23) (5.57) (2.50) (7.55) (6.82) (3.59)House: # Bedrooms 0.013 -0.002 0.005 0.022∗ ∗ ∗ 0.007 0.007∗∗ 0.024∗ ∗ ∗ 0.010 0.013∗ ∗ ∗

(1.65) (-0.34) (1.45) (2.93) (0.95) (2.30) (3.79) (1.23) (4.33)House: # Units -0.025∗ 0.003 0.007∗ ∗ ∗ -0.023∗ -0.002 0.006∗ ∗ ∗ -0.016 -0.002 0.006∗ ∗ ∗

(-2.02) (0.63) (3.77) (-1.78) (-0.40) (3.43) (-1.49) (-0.38) (4.78)House: Pool 0.068∗ ∗ ∗ 0.114∗ ∗ ∗ 0.119∗ ∗ ∗ 0.058∗ ∗ ∗ 0.099∗ ∗ ∗ 0.098∗ ∗ ∗ 0.053∗ ∗ ∗ 0.073∗ ∗ ∗ 0.068∗

(6.79) (9.40) (7.03) (5.60) (6.25) (4.59) (3.65) (3.42) (2.11)House: Condo -0.180∗ ∗ ∗ -0.214∗ ∗ ∗ -0.248∗ ∗ ∗ -0.197∗ ∗ ∗ -0.238∗ ∗ ∗ -0.306∗ ∗ ∗ -0.185∗ ∗ ∗ -0.229∗ ∗ ∗ -0.297∗ ∗ ∗

(-20.68) (-12.60) (-24.81) (-20.19) (-8.75) (-11.22) (-15.15) (-9.75) (-12.47)House: Sq ft. 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗ 0.000∗ ∗ ∗

(40.46) (23.82) (20.54) (31.78) (34.60) (42.70) (28.59) (30.16) (35.36)Cons 12.126∗ ∗ ∗ 12.007∗ ∗ ∗ 11.726∗ ∗ ∗ 12.125∗ ∗ ∗ 12.162∗ ∗ ∗ 11.939∗ ∗ ∗ 11.933∗ ∗ ∗ 12.060∗ ∗ ∗ 11.464∗ ∗ ∗

(336.10) (224.39) (113.21) (209.95) (94.52) (63.60) (142.86) (97.78) (127.90)County-border pairs FE Yes Yes Yes Yes Yes Yes Yes Yes YesTract-year FE Yes Yes Yes Yes Yes Yes Yes Yes YesLender FE Yes Yes Yes Yes Yes Yes Yes Yes YesR-squared 0.842 0.822 0.832 0.851 0.841 0.854 0.849 0.841 0.859N 160957 123698 56222 90155 57954 26344 60374 37694 16775

39

Table 9: Linear regression of log transaction prices near the State borders. Estimation of the linear regression model described by Equation 8. Each columnrepresents the results obtained by regressing the log of prices over the set of covariates, together with the variables of interest. The regression is separated between groups ofyears in order to observe potential dynamics of parameters. Groups are divided between 2002-2004, 2005-2007, 2008-2010 and 2011-2012. The first 4 columns contain alltransactions from covered counties at the recourse/no-recourse border of each state in analysis, while the last 4 columns limits the data to 20 miles from each border. Unreportedauction sales coefficient due to lack of variation. Fixed effects control for lenders, tract-year interaction and county-border pairs. T-statistics reported in parenthesis. Standarderrors clustered by state-year level.

40

Dep. Var: log(Price) Border Counties without NC-SC border2005-2007 2008-2010 2011-2012

Recourse 0.165∗∗ -0.049 -0.351∗∗(2.70) (-0.69) (-2.96)

REO Sale -0.150∗ ∗ ∗ -0.164∗ ∗ ∗ -0.173∗ ∗ ∗(-9.81) (-9.58) (-8.06)

Short-sale -0.126∗ ∗ ∗ -0.099∗ ∗ ∗ -0.102∗ ∗ ∗(-4.08) (-7.53) (-5.47)

REO × Recourse 0.007 0.069∗∗ 0.090∗ ∗ ∗(0.29) (2.67) (4.28)

Short-sale × Recourse 0.008 -0.025 -0.013(0.25) (-1.35) (-0.73)

Auction Sale 0.000 -0.037 -0.041∗ ∗ ∗(.) (-1.06) (-4.16)

Ownership Transfer -0.024∗ ∗ ∗ 0.023∗∗ -0.188∗ ∗ ∗(-5.66) (2.08) (-20.41)

County-border pairs FE Yes Yes YesTract-year FE Yes Yes YesLender FE Yes Yes YesR-squared 0.842 0.838 0.847N 101134 92767 44169

Table 10: Linear regression of log transaction prices near the State borders: Removing NC-SCborder. Estimation of the linear regression model described by Equation 8. Each column represents the resultsobtained by regressing the log of prices over the set of covariates, together with the variables of interest. The regressionis separated between groups of years in order to observe potential dynamics of parameters. Groups are divided between2002-2004, 2005-2007, 2008-2010 and 2011-2012. Unreported Fixed effects control for lenders, tract-year interactionand county-border pairs. T-statistics reported in parenthesis. Standard errors clustered by state-year level.

41

Dep. Var: log(Price) Border Counties without AZ-NV border2005-2007 2008-2010 2011-2012

Recourse -0.250∗ ∗ ∗ -0.478∗ ∗ ∗ -0.550∗ ∗ ∗(-3.29) (-6.52) (-7.20)

REO Sale -0.139∗ ∗ ∗ -0.187∗ ∗ ∗ -0.234∗ ∗ ∗(-7.29) (-9.20) (-14.41)

Short-sale -0.147∗ ∗ ∗ -0.118∗ ∗ ∗ -0.117∗ ∗ ∗(-12.01) (-18.09) (-19.53)

REO × Recourse -0.016 0.074∗∗ 0.122∗ ∗ ∗(-1.08) (2.77) (4.57)

Short-sale × Recourse -0.086 0.007 -0.002(-1.14) (0.58) (-0.25)

Auction Sale 0.000 -0.047 -0.084∗ ∗ ∗(.) (-1.26) (-3.35)

Ownership Transfer -0.002 0.038∗ ∗ ∗ -0.183∗ ∗ ∗(-0.14) (2.78) (-13.17)

County-border pairs FE Yes Yes YesTract-year FE Yes Yes YesLender FE Yes Yes YesR-squared 0.759 0.810 0.826N 100131 62931 27192

Table 11: Linear regression of log transaction prices near the State borders: Removing AZ-NVborder. Estimation of the linear regression model described by Equation 8. Each column represents the resultsobtained by regressing the log of prices over the set of covariates, together with the variables of interest. The regressionis separated between groups of years in order to observe potential dynamics of parameters. Groups are divided between2002-2004, 2005-2007, 2008-2010 and 2011-2012. Unreported Fixed effects control for lenders, tract-year interactionand county-border pairs. T-statistics reported in parenthesis. Standard errors clustered by state-year level.

42

Appendix C Merging Data Quick and HMDA

The following steps are taken in order to clean the data:

• Keep only negotiations that are at arms length

• Keep only condos or single residences

• Keep only transactions (resales), regular and distress sales. The distress sales are formedby the following types of transactions: REO sales, Auction sales, short-sales and ownershiptransfers (deed in lieu). REO sales and auction sales are considered foreclosure sales, whileshort-sales and ownership transfers are considered alternatives to foreclosure.

• Loan values are rounded by 1000.

C.1 Merge with HMDA

Given the initial data selection, we match all the loan information from resales into the HomeMortgage Disclosure Act database in order to obtain demographic information from mortgagepurchasers. Since there is no specific unique identifier between both Data Quick and HMDA,the merge process will be fuzzy, and so the steps must be detailed in order to provide additionalrobustness to the procedure.

The merge uses three levels of information to identify the current mortgage owner: census tractarea identifier; the mortgage loan value, and the name of the lender/subsidiary. From the HMDAdatabase, we select only loans that were effectively taken, and perform the merge using the censustract numbers and loan values. After the merge, we verify the quality by comparing both names oflenders (from Data Quick and HMDA). The algorithm is intrinsically related to how bank namesare related, whether we have subsidiaries being matched with their main branch, and how manymatches did we observe for a specific tract number and loan value.

We map lender names in Data Quick to a list of alternative names. The objective is to linkthe bank name in the Data Quick files to an alternative name, as if lenders in Data Quick weresubsidiaries or shared a professional relationship with such financial institution. We then matcheach of the identified alternative names with a list of subsidiaries.

Whenever the quality of the merge is relatively poor, or when there are multiple matches pertransaction, I use probabilistic methods to select the one with higher probability of having a match.From a local sample of successful matches, I compare the relative income level, loan values andcharacteristics to select which of the multiple merges are more likely to be successful. The per-centage of the sample that is duplicated is relatively small and does not affect the overall results. Ialso compare whether the matched sample shows any direct biases with respect to the unmatchedsample, results show no effects of manipulation in terms of merged information.

43