A Low Cost Methodology for Correcting the Distressed Sales Bias in a Downward Spiraling Housing Market

A Low Cost Methodology for Correcting the Distressed Sales Bias in a Downward Spiraling Housing Market Craig A. Depken, II Department of Economics UNC Charlotte cdepken@uncc.edu Harris Hollans Department of Finance Auburn University hollalh@auburn.edu Steve Swidler* Department of Finance Auburn University swidler@auburn.edu *Corresponding author Steve Swidler 405 W. Magnolia, Lowder 303 Auburn University Auburn, AL 36849 Phone: 334-844-3014 Fax: 334-844-4960 1

A Low Cost Methodology for Correcting the Distressed Sales Bias in a Downward Spiraling Housing Market Abstract This paper examines the impact of distressed sales on single family house prices during a housing market collapse. The innovation here is a methodology to create a proxy for distressed sales when such sales are not identified in the data but are commonplace in the market. We apply our methodology to publicly available data from Las Vegas, Nevada. We find that, during the market collapse in that city, the price impact from REO transactions was greater than other distressed sales, but the difference narrowed over time. Moreover, not identifying distressed sales lowers the measured price impact of REO sales. JEL Classifications: G11, G21, R31 Keywords: real estate, sample selection, short sales, flipping 2

1. Introduction In a typical housing cycle, a metropolitan area experiences growing population, new housing stock replacing old, and steady income growth. If the supply of land is limited, the result is gradually increasing home prices. During a recession, house price gains diminish, sales slow, and eventually there may even be a period during which home prices decline. In contrast, rapidly escalating prices followed by a tapering off of (irrational) exuberance and finally a precipitous fall in home values characterized the most recent cycle in numerous U.S. housing markets. While the causes of this unique occurrence are still open to debate, the last decade witnessed reduced lending standards coupled with a dramatically increasing housing stock that encouraged many new buyers who ordinarily would not have been homeowners. When the market peaked and subsequent price declines were coupled with increased unemployment, numerous recent home buyers were unable to maintain payments on their mortgage or found their loans underwater and strategically defaulted. Ultimately in a number of U.S. cities, housing markets collapsed and foreclosure activity became the majority of residential transactions. In this paper we examine the price impact of foreclosures in a downward spiraling market in which distressed sales become the norm. While the previous literature has mainly studied the contagion effect of a foreclosure within a neighborhood (for example, Towe and Lawley (2011), Campbell, et al. (2011) and Fisher, et al. (2012)), there is growing recognition that it is important to also identify short sales and properties in default when distressed sales dominate the market. 1 This paper provides a low cost methodology for identifying distressed sales from publicly available transaction data and uses the results to correct for an omitted variable bias when estimating the foreclosure effect on price. 3

Clauretie and Daneshvary (2011) were the first to address house price measurement in a distressed-sales dominated market. They examine Multiple Listing Service (MLS) data from Las Vegas, Nevada that differentiates between three types of distressed transactions: a short sale, an in-process-of-foreclosure, and a real estate owned (REO). In a short sale, the lender allows the owner to sell the property even though the proceeds are insufficient to cover the mortgage balance. While the lender might still pursue the seller for any deficiencies, it is common to discharge the entire debt with limited impairment to the homeowner s credit. Clauretie and Daneshvary describe an in-process-of-foreclosure transaction as a property in default that is sold by the owner. Finally, they define an REO sale as one where the lender takes possession of the property and sells the home to a new owner. They posit that all three types of distressed sales are likely to sell at a discount from normal market, no default transactions. Using Las Vegas MLS data from 2008, Clauretie and Daneshvary find that once time on market is taken into consideration, the estimated discount for in-process-of-foreclosure sales is $30,791 (8.95 percent of their sample average house price). The derived discount from a typical normal market sale is $33,153 (9.63 percent) and $40,871 (11.88 percent) for short sales and REO transactions, respectively. In a related paper, Daneshvary, Clauretie, and Kader (2011) argue that many previous studies of the foreclosure effect inadvertently introduce a potential sample selection bias when they do not include either REO or in-process-of-foreclosure sales. Exclusion of these transactions would have the result of overestimating any contagion effect from foreclosures. In their study, they show that dropping REO and foreclosure transactions from their sample increases the estimated spillover effect of foreclosures. However, they note that most studies examine stable housing markets where this sample selection bias may be of little consequence. 4

In contrast, not including other forms of distressed sales in the sample during times of thin markets (e.g., Las Vegas in 2008) will produce estimated spillover effects from foreclosures that are considerably larger than their true influence. In a similar vein, the Federal Housing Finance Agency (FHFA) has recently begun estimating distress-free house price indexes for selected markets. Their methodology removes the effect of short sales and real estate owned transactions from the house price index based on sales price information from Fannie Mae and Freddie Mac mortgages. They state, however, that the government sponsored enterprise (GSE) mortgage data do not identify short sales or REO properties, and thus, it is a significant challenge to detect distressed sales (FHFA (2012), page 12). Ultimately, their methodology relies on private vendor data from DataQuick to distinguish REO sales and CoreLogic to infer short sales or properties in default. 2 In their 2012 report, the FHFA examines twelve metropolitan areas including six from the state of California, but does not include the previously examined Las Vegas market. The selection of the twelve areas, in part, depends upon the availability of default data from their proprietary sources. In the first quarter of 2011, the FHFA finds the average share of distressed sales in the twelve metropolitan areas equal to 41.6 percent. Additionally, from 2007Q2-2012Q2, the FHFA estimates that the average decline for all homes is 5.9 percent lower than for distress-free transactions with Atlanta having the largest differential (12.6 percent). The previous research makes it clear that in the process of valuing homes, it is important to account for distressed sales. Indeed, in downward spiraling markets, identifying distressed sales is imperative. However, this can only be done through the use of proprietary data such as MLS or GSE mortgage information, or with the assistance of expensive private vendor data such 5

as DataQuick and Corelogic. Even then, for some real estate markets, the supplementary data sources may not have the necessary foreclosure information. In the following analysis, we also focus on Las Vegas, but our work differs from previous studies by using publicly available transaction data from the Clark County assessor s office. Aside from being free, tax assessor data has the advantage of being more comprehensive in its coverage of transactions. If only relying on MLS data, certain homes would be systematically excluded from the sample, e.g., a for-sale-by-owner transaction. Other sales that are sometimes omitted from the local MLS include sales by agents from outside the area and new home sales. 3 Furthermore, proprietary MLS data may also suffer from marketing biases, such as misstating the area, the number of rooms, or the property s time on market. 4 Utilizing the county s property transaction records has the additional benefit of being available over a longer period of time. In our analysis, we are able to follow transactions over four and a half years of rapidly falling prices and examine the discounts associated with foreclosures over the longer period of time. Moreover, the county records form the basis of price indices such as the Case-Shiller index thereby allowing for further comparisons. For many, the only practical data to use for empirical analysis are those that are publicly available from local tax assessors. 5 However, for reasons discussed below, distressed sales are frequently misidentified as arms-length transactions and cannot be excluded along the lines suggested by Daneshvary, et al (2011) or the FHFA (2012). Most likely, the misidentified armslength transactions are included in the sample, and in a market with a large proportion of distressed sales, the omitted variable (an identifier for distressed sale) will bias the discount associated with a foreclosure toward zero. However, in the case that such a distressed sale identifier is not provided, we propose a methodology that identifies those transactions recorded 6

as arms-length but which are very likely distressed sales. We show that including a generated distressed sale indicator variable reduces the bias in the estimated impact of foreclosure on property values during a market collapse and find empirically that the difference between the foreclosure and distressed sale discounts narrows over time. 2. Data Our data describe parcel transactions for single family homes in Clark County, Nevada. Clark County encompasses greater Las Vegas, one of the fastest growing metropolitan areas in the United States last decade. The data come from the county s tax assessor office and is similar to that used in Depken, Hollans and Swidler (2009, 2011). The county records several transaction types, although we focus on three major categories. The first is designated an R transaction or recorded value, and it is typically thought of as an arms-length sale between a willing buyer and seller. The second is a T transaction, and the value refers to the amount bid on the trustee s deed. This represents a home in foreclosure with the highest bidder being the lender in the vast majority of sales. 6 In that case, the subsequent sale would be a real estate owned (REO) transaction. The third is an an F transaction which refers to a resale after foreclosure and primarily follows a T transaction. After peaking in the second half of 2006, Las Vegas house prices began to decline in 2007. 7 Exhibit 1 lists the number of quarterly R, T, and F transactions during this period of falling prices. At the start of 2008, foreclosure-related transactions, those transactions coded as T s and F s, constituted more than half of all single family home transactions in Las Vegas. By the following year, foreclosure activity peaked at 75 percent of all transactions but remained high throughout the end of our sample period in the second quarter of 2011. Initially, most armslength sales occurred at prices above their original purchase price. In the first quarter of 2007, 7

more than 90% of sales recorded as arms-length (R-transaction) exhibited a nominal capital gain. However, the impact of increasing foreclosure activity caused prices to spiral downward, and by 2009, the majority of homeowners recorded losses on their arms-length transactions. While this may be partly the result of a contagion effect, it may also be due to some transactions that are recorded as traditional arms-length really being distressed sales, and is one area of analysis examined below. Before turning to the methodology we employ, recall that the county s coding of transaction type does not allow us to directly determine a short sale or a property in default that is sold by the owner. Since neither of these transactions involves a transfer of the trustee s deed to the lender, both types of distressed sales must be coded as an R transaction by the county. This has the further implication that simply taking the difference between R-transaction and F- transaction prices will understate the true discount of foreclosures from actual arms-length sales. Thus, in the next section we propose a procedure that will identify transactions recorded as armslength that are most likely distressed sales (either short sales or in-process-of-foreclosure sales) with the further result of more correctly measuring the foreclosure discount. 8 For purposes of nomenclature, hereafter the term distressed sales will only refer to short sales or in-process-offoreclosure. Separately we denote a home in foreclosure as a property that is initially auctioned (T transaction) before typically being an REO sale (F transaction). 9 3. Empirical Methodology and Testable Hypotheses In an ideal situation, short-sales and in-process-of-foreclosure sales would be clearly identified in the transaction data making it possible to distinguish them from traditional armslength transactions. However, to our knowledge, this is not the case in a majority of publicly available tax data. 10 Thus, our empirical strategy is to develop a method by which a sub-sample 8

of designated arms-length transactions in the tax data can be reclassified as distressed sales. Our approach is similar to propensity score matching (PSM) models in that we estimate a model with which we can "match" transactions recorded as arms-length (coded R) with transactions recorded as trustees deeds (T), recognizing that homeowners often self-select into letting the bank take possession of their property or selling the distressed property themselves. 11 Our intuition is that properties that are chronicled as arms-length but which are very similar to lender takeovers (REO properties) are stronger candidates for distressed sales. To operationalize this intuition, we estimate a probit model where the dependent variable takes a value of one if the transaction is recorded as a T transaction and zero otherwise. We restrict the sample to include only arms-length transactions that follow an arms-length transaction, (no matter how long between transactions) and transactions recorded as lender takeovers that follow an arms-length transaction. To motivate the first stage of our analysis, consider a lender takeover to be the function of a default index which is determined by the homeowner. If an individual homeowner s default index surpasses zero, the homeowner goes into default and the lender takes possession of the property. The individual s default index is a function of personal idiosyncrasies (e.g., unemployment, medical issues, sudden relocation, and risk preferences), characteristics of the overall Las Vegas residential market, and characteristics of the property including terms of the mortgage. While the default index is a latent (unobserved) variable, the probability that a particular property would be subjected to a lender takeover can be modeled as Prob(LT i =1) = Prob(X i α + ε i > 0) = Prob(ε i > -X i α), where LT i is an indicator variable that takes a value of one if a transaction is recorded as a lender takeover and zero otherwise, ε i is a zero-mean error term, X i is a vector of explanatory variables 9

thought to be related to the latent index, and α is a vector of parameters associated with the explanatory variables. We model ε as normally distributed, normalizing the variance to one, and therefore use the probit estimator. The explanatory variables employed in the probit estimation predominantly include overall market characteristics and property attributes. Unfortunately, tax records contain little information about homeowners and terms of the mortgage. 12 The market characteristics used in the probit estimation consist of the percentage change in market median price since the previous sale (PCTCHMEDPRICE), and an interaction with a dummy variable that takes a value of one if the PCTCHMEDPRICE is negative, which allows for an asymmetric effect of overall market price declines, the previous sale price (PREVSALEPRICE), whether the house was new in its previous transaction (NEWHOUSE) and whether the previous sale was the sell-side of a property flip (SELLSIDEFLIP). 13 Property characteristics include the age and age squared of the house (AGE and AGESQ), the total square footage (TOTSQFT), the number of bedrooms (BEDROOMS), the number of full bathrooms (BATHS), and as a proxy for quality, whether the house has a fireplace (FIREPLACE). Recalling from Exhibit 1 that foreclosure activity (transactions recorded as T or F) constitutes the majority of home sales starting in 2008, we estimate the probit model for each quarter from 2008q1 through 2011q2. For recorded arms-length transactions, we interpret high fitted probabilities from the probit equation as indicating a high likelihood that it was, in fact, a distressed sale. On the other hand, a transaction recorded as arms-length that has a low fitted probability from the foreclosure probit is indicative of a true, arms-length transaction. Generally, the quarterly probit models do a good job of discriminating between R and T transactions. 14 As expected, the median fitted probability of R transactions is below the median 10

value of T transactions each quarter. However, the median values draw closer to one another and are nearly indistinguishable at the end of the sample. This suggests that circumstances leading to a recorded arms-length sale become similar to those of a foreclosure and is likely the result of distressed sales constituting a large percentage of R transactions at the end. In order to distinguish between R transactions that are truly arms-length versus those that are distressed sales, we examine the distribution of fitted probabilities for all R transactions in a given quarter. We then classify those transactions with propensity scores greater than the 75 th percentile in a given quarter as distressed sales and those with propensity scores less than the 25th percentile in a given quarter as traditional arms-length sales. This leads to median fitted probit values for classified arms-length transactions that are well below median values of properties taken over by lenders (see Exhibit 2). 15 Having a proxy for the omitted variable that identifies distressed sales, we can now derive the effect of foreclosures on price. Stage two of the analysis entails estimating a standard hedonic model in which the dependent variable is the natural log of sales price, and the sample contains all (classified) arms-length sales, (classified) distressed sales, and (recorded) foreclosures. 16 Following Depken, Hollans and Swidler (2009, 2011) and solving for price (PRICE i ), the hedonic model is: PRICE i = exp(β 0 + β 1 DISTRESS i + β 2 FORECLOSURE i + β 3 BUYSIDEFLIP i + β 4 SELLSIDEFLIP i + β 5 FQTRSBWSALES i + β 6 AGE i + β 7 AGESQ i + β 8 BEDROOMS + β 9 FBATH i + β 10 FIREPLACE i + β 11 TOTSQFT i + δtaxdist + υ i ), where the β's are parameters to be estimated and υ is a zero-mean error term. To test for possible discounts relative to arms-length transactions, the price equation includes a dummy variable for a distressed sale (DISTRESS i ) which takes a value of one for an R 11

transaction in the top quarter of propensity scores obtained in the first-stage analysis and zero otherwise and a second dummy variable that denotes a foreclosure (FORECLOSURE i ) that takes a value of one if the sale is a recorded F transaction and zero otherwise. The price equation also contains several previously defined explanatory variables plus whether the transaction is the front end of a flip (BUYSIDEFLIP i ), the number of quarters between the lender takeover and a foreclosure sale (FQTRSBWSALES i ), and a vector of neighborhood fixed effect for each tax district (TAXDIST i ). The price equation includes FQTRSBWSALES i, the number of quarters between the lender takeover transaction (recorded as a T) and the subsequent foreclosure transaction (recorded as an F), to proxy for the property condition of a foreclosure. Foreclosures in poor condition will likely take longer to sell, ceteris paribus, and we therefore expect β 5 to be negative. As this variable will have a value of zero for all transactions of properties that are not in foreclosure, the full effect of a foreclosure on price is (β 2 + β 5 FQTRSBWSALES i ). Quarter by quarter estimation of the price equation from 2008q1 to 2011q2 enables us to examine the following three empirical hypotheses: H1. In a downward spiraling housing market, the negative impact of foreclosure status may be expected to start out large (and negative) but to decline (get closer to zero) as foreclosures become more commonplace. To test this hypothesis, we plot a trend line of the FORECLOSURE dummy variable coefficient over the various quarters. Standard errors in this auxiliary model are bootstrapped because of the generated regressand problem. 17 We posit a positive relationship between the (negative) price effect of foreclosure and time. H2. We anticipate the impact of a distressed sale to be negative but initially not as great (in absolute value) as the impact of a foreclosure. As a further extension, with foreclosures and 12

distressed sales increasing as a percentage of all transactions, the stigma attached to each might be expected to diminish and lead to similar price effects. Moreover, an influx of investor taking advantage of bargain price REO s may further lead to smaller discounts. To examine this hypothesis, we compare the coefficients of the DISTRESS dummy variable to FORECLOSURE on a quarter by quarter basis, and test for their equality. Comparing β 1 to β 2 indicates the price difference between a distressed sale and foreclosure assuming similar property conditions. However, as noted previously, the full effect of a foreclosure on price is (β 2 +β 5 FQTRSBWSALES i ). Thus we also examine the additional foreclosure effect related to property condition, β 5 FQTRSBWSALES i. H3. In a model in which all arms-length transactions are included without identifying any distressed sales, the negative impact of a foreclosure will be biased upward (toward zero). To test this hypothesis, we estimate the PRICE equation using all arms-length and foreclosure transactions, but do not include the DISTRESS dummy variable in the quarterly hedonic model. We posit that the measured impact of foreclosure is mitigated when distressed sales are not separately identified. 4. Empirical Results: Measuring the Impact of Foreclosures on Housing Prices We estimate the hedonic pricing model for each quarter of the sample period. Generally the results accord with theory and right-hand side variables explain nearly eighty percent of price variation. 18 Property characteristics carry the expected sign and are all statistically significant. The longer between a lender takeover and foreclosure sale (FQTRSBWSALES), the lower the average transaction price in thirteen of the fourteen quarters. 19 This is consistent with the notion that FQTRSBWSALES is a proxy for property condition. 20 With respect to buy side flips, the signs of the coefficients are generally negative and significant for the period of analysis. Similar 13

to findings in previous research, the results suggest that flippers are good at identifying undervalued properties. However, results are mixed for the SELLSIDEFLIP variable as flippers do not consistently receive a premium in a downward spiraling market. From the standpoint of hypotheses 1 and 2, the crucial variables are DISTRESS and FORECLOSURE, and their estimated coefficients for 2008q1-2011q2 appear in Exhibit 3. The results support Hypothesis 1: the FORECLOSURE coefficients are negative and statistically significant. In 2008, the average FORECLOSURE coefficient suggests that a home in foreclosure sold for 21.3 percent less than if it were an arms-length transaction. This compares to a 13.5 percent discount in Clauretie and Daneshvary (2009), and the difference may in part be due to differences in the transactions included in the MLS data used by Clauretie and Daneshvary and those included in the tax assessor's data used here. 21 Further examining hypothesis 1, the trend line in Exhibit 4 shows that the negative impact of a foreclosure starts out large, but the effect diminishes as foreclosures become more commonplace, holding property quality constant. In the final year of the sample, the average FORECLOSURE coefficient is -9.8 percent and implies that the stigma of a foreclosure shrinks as foreclosures and distressed sales become the norm. 22 Turning to the second hypothesis, the coefficients for the DISTRESS variable are negative and significant in all but three quarters. The exceptions to statistical significance are in the first quarter of years 2008, 2009, and 2010, and the pattern suggests a seasonal effect in the data. While distressed sales carry a negative premium, initially the effect is smaller (closer to zero) than the foreclosure discount. In 2008, the average coefficient for the DISTRESS variable is -6.3 percent. As distressed sales increase in number over the sample period, the discount for distressed sales also increases, and becomes similar in magnitude to that associated with 14

foreclosures. Visually, this can be seen in the converging trend lines in Exhibit 5. During the last year, the average DISTRESS coefficient is -10.7 percent and is not significantly different from the foreclosure effect for properties in similar condition. In fact, Exhibit 5 plots the confidence intervals around both the DISTRESS and FORECLOSURE coefficients over the sample period and shows that the two are not statistically different from one another beginning 2009q4. To further examine issues of seasonality as well as check for model robustness, we reestimate the hedonic model as a pooled regression including transactions from 2008q1 through 2011q2. We allow for quarterly and annual price differences for the three types of transactions: arm s length, distressed, and foreclosure. 23 The model includes fixed effects for tax districts and constrains only the coefficients for household characteristics to be the same across quarters. The results appear in Exhibit 6 and are very similar to the quarterly hedonic regressions. Overall, the model explains more than 80 percent of the variation of house prices. With respect to household characteristics, age and age squared are negatively correlated to price, while total square footage, number of fireplaces, and number of full baths all increase the value of the home. With square footage included, the number of bedrooms has a small, but negative coefficient and is comparable to the results in Depken, Harris and Swidler (2009). On average, flips are bought at a discount and sell at a slight premium. With respect to seasonality, distressed sales (foreclosures) sell at a 5.1 percent (3.3 percent) discount in quarter 2, 6.3 percent (3.5 percent) discount in quarter 3 and 3.7 percent (0.1 percent) discount in quarter 4 when compared to prices of distressed sales (foreclosures) in quarter 1. For arm s-length sales, prices in quarter 1 and 2 are nearly identical, but are 3.8 percent lower in quarter 3 and 10.9 percent lower in quarter 4. Thus, the pooled regressions confirm the seasonality result from before. 15

The pooled regression also demonstrates the narrowing of price discounts over time between distressed sales and foreclosures. In 2008q1, the seasonally adjusted discount to armslength transactions is 0.3 percent for distressed sales and 17.1 percent for foreclosures. By 2011q1, the seasonally adjusted discounts for distressed sales and foreclosures are 7.2 percent and 4.5 percent, respectively, and are not significantly different from one another. This assumes similar property condition between the two sale types. However, the average number of quarters between the T and F transactions in our sample is 2.38, so that adjusting for property condition, the typical foreclosure sells for an additional 2 percent discount. 24 To test hypothesis 3, we re-estimate the quarterly hedonic models including all armslength transactions and all foreclosure transactions. In this specification we only include the FORECLOSURE dummy variable and not the DISTRESS dummy. Exhibit 7 depicts the impact of failing to distinguish between distressed and normal arms-length transactions. In 2007, the impact of a foreclosure is essentially the same regardless of whether the full or restricted sample is used. However, starting in 2008, when foreclosure activity is the rule rather than the exception, the price impact of foreclosures is greater (i.e., more negative) once we adjust for distressed sales. From Exhibit 8, the average difference between the two approaches is 5.2 percentage points from 2008q1 to 2011q2. The main implication from the analysis is that when foreclosures become the norm, not correcting for distressed sales biases the measured impact of foreclosures toward zero. This bias is in the opposite direction of that in Daneshvary, et al (2011), who warn against incorrectly omitting distressed transactions and causing sample-selection bias, and suggests that researchers should be careful to avoid unintentional omitted variable bias. The magnitude of both biases are of less concern in healthy markets in which distressed sales and foreclosures are a minority of 16

all transactions. However, as foreclosures and distressed sales increase in number, both biases become more acute. Whereas Daneshvary, et al, warn against incorrectly excluding foreclosure and REO transactions, our evidence complements and extends this warning to include misidentifying distressed sales as arms-length transactions. 5. Conclusions This paper extends previous work examining the impact of foreclosure on house prices during a localized housing market collapse. During times of relatively stable markets, the exclusion of short sales and in-process-of-foreclosure sales likely has little impact on the estimated effect of foreclosures. However, when distressed sales of all types become the norm, it is essential to include all transactions in the empirical analysis to avoid sample selection bias. Daneshvary, et al (2011) first pointed out that it is important to identify REO transactions, short sales, and in-process-of-foreclosure sales in any study of housing prices, especially when these transactions become the norm rather than the exception. However, it is not common for publicly available data to contain precise identifiers for distressed sales. Therefore, including all transactions without identifying distressed sales runs the risk of omitted variable bias which might lead to an underestimate of the effect of foreclosures. The concern that including unidentified distressed sales might lead to biased localized housing price indexes has led the FHFA to merge GSE mortgage data with costly private sector data to estimate a distress-free house price index. Thus, while it is possible to ascertain distressed sales using proprietary and private vendor data, for most applications these data are either cost prohibitive or not publicly available. Indeed, despite its considerable resources, the FHFA has only produced distress free indexes for twelve metropolitan areas, six of them in California. We propose an innovative and low-cost methodology that uses public records data to identify transactions that are recorded as arms-length transactions but are most likely distressed 17

sales. Even if a researcher has access to MLS data, our methodology provides an alternative estimation of the distressed sales effect to complement the initial findings. The procedure assesses the likelihood that a sale recorded as arms-length might have been a distressed sale by estimating a probit model using only recorded arms-length and lender takeover transactions. Those arms-length transactions that look the most like lender-takeover transactions via propensity scores are coded as being distressed and those that look the least like lender-takeover transactions are coded as arms-length. This methodology solves the omitted variable bias introduced when unable to identify distressed sales, and has the further advantage of using publicly available data which include all transactions in a given time period. Examining parcel transactions recorded in the Clark County, Nevada, tax assessor s office, we find the price impact from real estate owned transactions is greater than other distressed sales from 2008q1 through 2011q2. While the difference narrows over time, once property condition is taken into account, REO sales still have the larger discount. Moreover, we show that including all arms-length transactions but not identifying distressed sales lowers the measured price impact of REO sales by more than 5 percent, on average. Finally, future research should focus on improving the identification methodology for distressed sales. Specifically, minimizing the specification error will entail alternative rules to the 25%/75% selection rule. This might be done with additional explanatory variables in the probit analysis such as mortgage information or through further calibration when proprietary information is available. Even so, this paper provides the first attempt to estimate a separate distressed sales effect using publicly available data and shows the importance of including these sales when analyzing a downward spiraling market. 18

Endnotes 1 Additional papers on the effect of foreclosure on price include Pennington-Cross (2006), Carroll, et al. (1997) and Forgey, et al. (1994). 2 Even with this third party data, the FHFA must make an assumption to identify distressed properties and invokes a twelve month subsequent-to-notice-of-default rule (see FHFA (2012), page 14). The FHFA goes on to say that precise decision rules are subject to modification, and in the future, FHFA may refine these rules. 3 Wood (2008) found that local MLS data excluded 13% of sales in the public records for Bear Lake, CA. He attributed many of these missing sales to REO transactions with agents from outside the area and concluded when it comes to recorded sales, public records are as accurate as you will find. 4 For the game playing strategies of agents and potential for information biases in MLS data, see Frew (1987). An extensive review of literature on the use of MLS data appears in Benjamin, Jud, and Sirmans (2000). 5 We do not mean to imply that proprietary data should not be used to supplement the tax assessors records. For those with access to MLS data, the impact of distressed sales can be estimated directly. Even so, the methodology in this paper provides an alternative set of estimates whereby robustness tests can be performed for the effects of both distressed sales and foreclosures. Nevertheless, the more practical issue is that the proprietary data may not be available for certain markets or it may be sold only at a very high price. 6 Typically the lender will bid the mortgage balance. In the case of default, this will exceed the value of the property; otherwise the homeowner would have sold the home and paid off the mortgage. Since the mortgage balance is greater than the value of the property, no one else will likely outbid the lender. 19

7 See the S&P/Case Shiller Home Price index at : http://us.spindices.com/index-family/real-estate/spcase-shiller 8 The important point here is that these are categories where homeowners are highly motivated to sell their property and likely do so at a price lower than if sold in a normal market. From Clauretie and Daneshvary (2011), this group is more likely to include short sales (2,033 properties in their sample) instead of in-process-of-foreclosure sales (1,202 properties). 9 To summarize the transactions examined here, we consider either the sequence R-T-F or R-R. The first case is a foreclosure sequence where, after an arms-length transaction (R), the lender takes possession (T), which is then followed by an REO sale (F). All three transactions are recorded in the assessor's data although only the arm's length and REO sales have prices that are meaningful. The second case is an initial arms-length transaction (R) followed by another recorded sale (R). However, in this sequence, the second R transaction can be either an arms-length transaction or some type of distressed sale. 10 Florida, for example, has only recently expanded their list to thirty one possible real property transfer codes to determine if the property is qualified and included in sales ratio analysis. (http://dor.myflorida.com/dor/property/rp/dataformats/pdf/salequalcodes12.pdf) However, Florida does not reserve any code specifically for short sales. Moreover, an earlier memorandum states that even though a short sale likely involves duress and may lead to disqualification, merely identifying a transaction as a short sale is not evidence enough for qualification decisions. (http://dor.myflorida.com/dor/property/taxpayers/shortsale/pdf/memo101008.pdf) The memo goes on to say that you must properly qualify or disqualify all sales on the merits of each and document the reasons. This illustrates the subjective nature of determining short sales and more generally distressed sales. Moreover, individual interpretation may very well afflict other data sources such as MLS or private vendors. 20

11 Propensity score matching estimates an average treatment effect by first estimating the likelihood of receiving the treatment for treated and non-treated observations and then "matches" non-treated with treated observations using a specific matching algorithm, such as nearest neighbor. Alternatively, the two-stage process suggested by Heckman (1979) estimates the likelihood of being selected into the treated sample and derives an inverse Mills ratio added to the treatment estimation to consistently estimate the average treatment effect. However, to implement either process it is necessary to know which observations are treated and which are not. In the case of assessors data, it is not possible to identify the type of distressed sale and therefore propensity score matching or Heckman selection models are not estimable. However, our methodology is similar to propensity matching in that we classify arms-length transactions that have similar characteristics to properties in foreclosure as distressed. 12 Clark County does record basic terms of the mortgage, however, this is not part of the Assessor s electronic database. 13 A sell-side flip is a home that has been owned for less than two years and then re-sold. Depken, Hollans and Swidler (2009) show that a flipped home tended to sell for a higher price than an otherwise similar house during the market run-up in prices. Thus, the current owner of a sell-side flip overpaid for the home, and this variable indirectly suggests that these homeowners were less knowledgeable about local market conditions at the time of their purchase. Furthermore, those who purchase from flippers might find themselves underwater faster than other purchasers as the market deteriorates and prices fall. 14 One notable result is that if the previous sale was a sell-side flip, it increased the odds that the transaction is a lender takeover in fifteen of eighteen quarters. In other words, owners who purchased flipped houses at or immediately after the peak of the market in 2006 often found themselves underwater or otherwise unable to service their house related debt, leading to an increased probability 21

of default and a lender takeover. This finding may also speak to the buyers limited knowledge of the real estate market. A complete set of probit results is available from the authors. 15 In selecting the percentile cutoffs, we note that increasing the size of the tails increases the sample size and therefore parameter precision but also runs the risk of misclassifying arms-length transactions as distressed sales. Ultimately, selection of the 25/75 rule depends upon estimation of the hedonic model below. 16 This describes, respectively, recorded arms-length transactions in the top 25% of the fitted values from the foreclosure probit estimation, recorded arms-length transactions in the bottom 25% of the fitted values from the probit estimation, and all foreclosure transactions in a given quarter. We use sensitivity analysis to estimate the hedonic model identifying arms-length and distressed sales using tail widths starting at 10/90 and changing the tails by five percent increments through 30/70. Using the methodology proposed by Browne (1979), in twelve of the fourteen quarters we fail to reject the null hypothesis that the estimated parameters on FORECLOSURE are equal when using the 25/75 and 10/90 tail widths. Thus, while different cutoffs produce similar results, we choose the 25/75 tails because they provide the smallest confidence intervals for the DISTRESS and FORECLOSURE dummy variables. Additionally we note that the 25/75 rule is also consistent with the FHFA s finding that 41% of the transactions they examined in 2011q1 were distressed sales. Finally, the 25%/75% threshold has historical precedence. In the statistical work of Bessel, Galton and Gauss, probable error represents the positive and negative deviation from a central measure that occurs in one half of the cases and may be expected to fall by chance. This concept of identifying the 25% tails is still used in the physical sciences (see Cowles and Davis, 1982). 17 Moreover, we recognize that measurement error introduces slight attenuation bias (toward zero). 18 A full set of the estimation results are available from the authors upon request. 22

19 One potential issue is multicollinearity between FORECLOSURE and FQTRSBWSALES. While multicollinearity yields unbiased estimates, interpretation of individual effects might be difficult. As it turns out, the correlation coefficient equals.4420 and suggests that multicollinearity is not a problem. Furthermore, if multicollinearity were an issue, resulting large standard errors would make it hard to reject the null hypothesis. As it turns out, both FORECLOSURE and FQTRSBWSALES are typically significant in the 14 quarterly regressions and both effects are significant in the pooled regression estimated for robustness. 20 Clauretie and Daneshvary (2009) and Daneshvary, et al (2011), are able to identify properties as being in poor, fair, or good condition. 21 Remember that the tax assessor s data include all transactions whereas Clauretie and Daneshvary (2009) only use MLS-recorded transactions from 2008. 22 If we preclude 2011q2 because the data are incomplete for the quarter, the final year s foreclosure effect falls to an average of -7.4 percent. 23 An alternative specification is to have dummy coefficients for each effect for each quarter. The results are virtually the same as those reported in the paper and may be requested from the corresponding author. 24 The discount equals the average number of quarters, 2.38, times the FQTRSBWSALES i coefficient, -.0084. By 2011, the full price effect of a foreclosure is -6.5 percent. 23

References Benjamin, J. D., G. D. Jud, and G. S. Sirmans. Real Estate Brokerage and the Housing Market: An Annotated Bibliography. Journal of Real Estate Research, 2000, 20:1/2, 217-278. Browne, R. H. On Visual Assessment of the Significance of a Mean Difference. Biometrics, 1979, 35:3, 657-665. Campbell, J. Y., S. Giglio, and P. Pathak. Forced Sales and House Prices. American Economic Review, 2011, 101:5, 2108-2131. Carroll, T.M., T.M. Clauretie, and H.R. Neill. Effect of Foreclosure Status on Residential Selling Price: Comment. Journal of Real Estate Research, 1997, 13:1, 95 102 Clauretie, T. and N. Daneshvary. Estimating the House Foreclosure Discount Corrected for Spatial Price Dependence and Endogeneity of Marketing Time. Real Estate Economics, 2009, 37:1, 43-67. Clauretie, T. and N. Daneshvary. The Optimal Choice for Lenders Facing Defaults: Short Sale, Foreclose, or REO. Journal of Real Estate Finance and Economics, 2011, 42:4, 504-521. Cowles, M. and C. Davis. On the Origins of the.05 Level of Statistical Significance. American Psychologist, 1982, 37:5, 553-558. Daneshvary, N., T. Clauretie, and A. Kader. Short-term Own-price and Spillover Effects of Distressed Residential Properties: The Case of a Housing Crash. Journal of Real Estate Research, 2011, 33:2, 179-207. Depken, C., H. Hollans, and S. Swidler. An Empirical Analysis of Residential Property Flipping. Journal of Real Estate Finance and Economics, 2009, 39:3, 248-263. Depken, C., H. Hollans, and S. Swidler. Flips, Flops and Foreclosures: Anatomy of a Real Estate Bubble. Journal of Financial Economic Policy, 2011, 3:1, 49-65. Federal Housing Finance Agency. U.S. House Prices Rose 1.8 Percent From First Quarter to Second Quarter 2012. news release, August 23, 2012. Fisher, L., L. Lambie-Hanson, and P. Willen. Structure Type and Foreclosure Externalities. mimeo, University of North Carolina Chapel Hill, 2012. Forgey, F.A., R.C. Rutherford, and M.L. Van Buskirk. Effect of Foreclosure Status on Residential Selling Price. Journal of Real Estate Research, 1994, 9:3, 313 18. Frew, J. R. Multiple Listing Service Participation in the Real Estate Brokerage Industry: Cooperation or Competition? Journal of Urban Economics, 1987, 21:3, 272-286. 24

Heckman, J. Sample Selection Bias as a Specification Error. Econometrica, 1979, 47:1, 153-161. Pennington-Cross, A. The Value of Foreclosed Property. Journal of Real Estate Research, 2006, 28:2, 193 214. Towe, C. and C. Lawley. The Contagion Effect of Neighboring Foreclosures. Social Science Research Network paper 1834805, 2011. Wood, T. MLS Data vs. Public Records Which Can You Trust? http://www.thetimwoodgroup.com/mls-data-vs-public-records-which-can-you-trust/, 2008. Acknowledgements The authors would like to thank the three anonymous reviewers for their thoughtful comments and the encouragement of the editor, Ko Wang. We would also like to thank seminar participants at Auburn University, Victoria University in Wellington, New Zealand, Stockholm University and the 2012 Western Economic Association annual meetings for their remarks. We further appreciate the insights of Andy Leventis and Will Doerner of the Federal Housing Finance Agency and their discussion of the FHFA s distress-free house price index. 25

Exhibit 1: Home Sales in Clark County, NV by Sales Type. No. of F Transactions No. of T Transactions Pct. F and T Transactions Pct R Transactions with Nominal Gain Year Qtr No. of R Transactions 2007q1 7799 30 647 8.0% 91.0% 2007q2 7456 76 981 12.4% 88.6% 2007q3 6104 155 1432 20.6% 89.4% 2007q4 5528 213 2233 30.7% 88.4% 2008q1 3517 1610 3060 57.0% 82.5% 2008q2 3830 3102 4590 66.8% 78.9% 2008q3 4053 4249 5109 69.8% 73.9% 2008q4 3760 3869 4477 68.9% 66.7% 2009q1 2791 4132 4320 75.2% 53.1% 2009q2 3474 5662 3597 72.7% 48.9% 2009q3 4083 5209 4451 70.3% 47.4% 2009q4 5104 4400 3318 60.2% 46.4% 2010q1 4352 3092 2686 57.0% 39.7% 2010q2 5937 2955 4264 54.9% 45.1% 2010q3 4548 2950 3729 59.5% 38.7% 2010q4 4415 3008 3403 59.2% 37.6% 2011q1 3974 3316 3851 64.3% 31.6% 2011q2 1700 1568 2446 70.2% 34.4% Notes: R indicates arms-length transactions, F denotes a foreclosure sale, and T is a trustee s deed transfer. Data obtained from the Clark County (NV) tax assessors. Totals and percentages calculated by the authors. 26

Exhibit 2 27

Exhibit 3: Effect of Distressed Sale and Foreclosure on Price Quarter DISTRESS SE FORECLOSURE SE Observations R 2 2008q1-0.017 0.018-0.193*** 0.012 2097 0.726 2008q2-0.121*** 0.014-0.259*** 0.010 3646 0.770 2008q3-0.091*** 0.014-0.208*** 0.010 4788 0.769 2008q4-0.036** 0.017-0.192*** 0.013 4514 0.777 2009q1-0.008 0.019-0.179*** 0.016 4610 0.806 2009q2-0.042*** 0.016-0.165*** 0.013 6148 0.812 2009q3-0.108*** 0.014-0.205*** 0.011 5842 0.815 2009q4-0.070*** 0.013-0.092*** 0.011 5545 0.798 2010q1-0.014 0.014-0.037*** 0.012 4324 0.807 2010q2-0.065*** 0.013-0.050*** 0.010 4545 0.812 2010q3-0.122*** 0.013-0.105*** 0.011 4167 0.822 2010q4-0.063*** 0.015-0.071*** 0.013 4248 0.805 2011q1-0.064*** 0.017-0.069*** 0.014 4463 0.743 2011q2-0.178*** 0.026-0.146*** 0.022 1999 0.737 Ln(Price) is dependent variable in estimated regression. Reported coefficients (x 100) have been transformed to represent percentages. *, **, *** significant at the.1,.05 and.01 levels, respectively 28

Exhibit 4 29

Exhibit 5 30

Exhibit 6: Pooled regression Results Coefficient Standard Error DISTRESSED -0.003 0.011 DISTRESSED x 2009-0.054*** 0.010 DISTRESSED x 2010-0.038*** 0.010 DISTRESSED x 2011-0.069*** 0.014 DISTRESSED xqtr2-0.051*** 0.009 DISTRESSED xqtr 3-0.063*** 0.011 DISTRESSED xqtr 4-0.037*** 0.011 FORECLOSURE -0.171*** 0.007 FORECLOSURE x 2009 0.006 0.008 FORECLOSURE x 2010 0.135*** 0.008 FORECLOSURE x 2011 0.126*** 0.012 FORECLOSURE x QTR2-0.033*** 0.007 FORECLOSURE x QTR 3-0.035*** 0.008 FORECLOSURE x QTR 4-0.001 0.009 YR2009-0.306*** 0.005 YR2010-0.385*** 0.004 YR2011-0.439*** 0.005 QTR2 0.000 0.007 QTR3-0.038*** 0.008 QTR4-0.109*** 0.007 BUYSIDEFLIP -0.048*** 0.009 SELLSIDEFLIP 0.018*** 0.006 FQTRSBWSALES -0.008*** 0.001 AGE -0.089*** 0.003 AGESQ -0.011*** 0.001 BEDROOMS -0.032*** 0.002 FULLBATH 0.010*** 0.003 FIREPLACE 0.093*** 0.002 TOTSQFT 0.442*** 0.003 Constant 11.729*** 0.009 Observations 60,936 R-squared 0.812 Dependent variable is the natural log of price. Marginal effects of dichotomous variables reported as exp(β)-1 so that coefficients (x 100) represent percentages. *** p<0.01, ** p<0.05, * p<0.1 31

Exhibit 7 32

Exhibit 8: Foreclosure Effect: Omitted Variable Bias Quarter Restricted Unrestricted Difference 2008q1-0.175-0.162-0.013 2008q2-0.228-0.177-0.051 2008q3-0.188-0.137-0.051 2008q4-0.175-0.129-0.046 2009q1-0.164-0.124-0.040 2009q2-0.152-0.104-0.048 2009q3-0.185-0.108-0.077 2009q4-0.088-0.034-0.054 2010q1-0.036-0.017-0.019 2010q2-0.049 0.000-0.049 2010q3-0.100-0.027-0.073 2010q4-0.069-0.035-0.034 2011q1-0.066-0.021-0.045 2011q2-0.136-0.004-0.132 33