Appraisal inflation: Evidence from the 2009 GSE HVCC intervention

Journal of Housing Economics 27 (2015) 71–90

Contents lists available at ScienceDirect

Journal of Housing Economics journal homepage: www.elsevier.com/locate/jhec

Appraisal inflation: Evidence from the 2009 GSE HVCC intervention Lan Shi a, Yan Zhang b,⇑ a b

Enterprise Risk Analysis Division, Office of the Comptroller of the Currency, United States Compliance Risk Analysis Division, Office of the Comptroller of the Currency, 400 7th St. SW, Washington, DC 20219, United States

a r t i c l e

i n f o

Article history: Received 26 January 2015 Accepted 18 February 2015 Available online 3 April 2015 JEL classification: D82 G21 G28 Keywords: Appraisal Appraisal bias Appraisal inflation Appraisal independence GSE HVCC

a b s t r a c t Appraisal inflation is a prominent aspect of lax underwriting practice. The GSE May 2009 Home Valuation Code of Conduct (HVCC) aims to prohibit lenders from influencing appraisers. Refinance loans, without a transaction price, are potentially more susceptible to appraisal inflation than purchase loans. We use GSE refinance loans as our treatment group and non-GSE refinance loans as the control group, and find that GSE refinance loans originated after May 2009 have lower default rates than non-GSE refinance loans. We further measure the appraisal inflation (bias) as the difference between the appraisal value in a 2009 refinance transaction and the actual transaction price in an earlier purchase transaction for the same property adjusted for local housing value changes. We find that the reduction in appraisal bias was larger for GSE refinance loans than for non-GSE refinance loans. This paper quantifies the ‘‘contribution’’ of appraisal inflation in poor loan underwriting standards and highlights the importance of unbiased and independent appraisal. Published by Elsevier Inc.

1. Introduction Appraisal is an important part of the loan origination process. On the one hand, the loan-to-value ratio (LTV) affects the borrower’s incentive to default (Foote et al., 2008; Elul et al., 2010). On the other hand, in the event of a default, the lender expects to sell the collateral to recover the unpaid loan balance. As a result, the LTV is an important factor in loan underwriting. It is thus important to have an accurate and unbiased appraisal. The appraisal, which is used for underwriting decisions, is usually obtained by evaluating the home’s features and comparing the collateral to recent sales of neighborhood homes with similar features. The quality of the appraisal is influenced by the incentives that appraisers receive. Appraisers get their business ⇑ Corresponding author. Fax: +1 (703) 857 6961. E-mail address: [email protected] (Y. Zhang). http://dx.doi.org/10.1016/j.jhe.2015.02.007 1051-1377/Published by Elsevier Inc.

from loan officers and brokers. A loan officer or a broker, who is often paid partially or wholly on commission based on volume of loan originations, might press the appraiser for a desired property value or a targeted loan amount. Having a higher appraised value than the true property value could potentially lead to a greater loan amount given the LTV, or a lower LTV given the loan amount, which results in a greater likelihood of loan approval or permitting riskier loan terms. Fearing loss of business, appraisers may yield to the pressure and influence from loan officers or brokers to inflate appraisal value. Prior literature documenting appraisal inflation in refinance loans is limited because of data constraints. Unlike purchase loans, for which an actual transaction price can be compared to the appraisal value, refinance loans do not have actual transaction price data (Cho and Megbolugbe, 1996; Nakamura, 2010). A couple of researchers tried to address this challenge. Agarwal et al. (forthcoming) focus on properties that had a subsequent purchase following a

72

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

refinance or purchase. Assuming that the subsequent price, after being adjusted for house price change, reflected the true house value, they are able to assess the potential appraisal inflation in refinance and evaluate its impact on loan performance and pricing. They find that the average valuation bias for residential refinance transactions is above 5% and mortgages with inflated valuations default more often. LaCour-Little and Malpezzi (2003) use a hedonic price model to estimate the ‘‘true’’ value of a property, which they compare to the appraised value. They find that the decreasing appraisal quality, namely over-appraisal, is associated with increased mortgage default.1 Our paper takes an alternative approach to overcome the data challenge. We exploit an event, the 2009 GSE Housing Valuation Code of Conduct (HVCC),2 which imposed appraiser independence. Starting in May 2009 for GSE-purchased loans,3 Fannie Mae and Freddie Mac prohibited lenders from pressing or influencing appraisers to provide a desired valuation, as detailed in the HVCC.4 The GSE HVCC of 2009 is a major change in appraisal practice. The banning of brokerordered appraisal is significant since, in years leading up to the 2007 subprime crisis, broker-sourced loans comprised nearly two-thirds of the market. Also, loan officers can order the appraisal only if they are not influenced by the lender. The pay or selection of appraisers shall not be based on the appraised value and lenders are prohibited from communicating to appraisers a desired value or loan amount. We predict that if the HVCC leads to a more independent valuation, the appraisal will be more accurate and thus the appraisal inflation will be smaller. With more accurate appraisal values, the originated loans after the intervention will be of higher quality. For purchase loans, since the value used to calculate LTV is the smaller of the appraised value and the transaction price, the potential for appraisal inflation is limited. We therefore expect to observe the HVCC effect mainly with refinance loans rather than with purchase loans.5 We test our hypotheses using the difference-in-differences (DID) method. We exploit the fact that the HVCC intervention only applies to loans intended for sale to GSEs, and does not apply to non-GSE loans.6 And we

1 However, LaCour-Little and Malpezzi (2003) did not assess the overappraisal by loan purpose (i.e., purchase vs. refinance). 2 http://www.freddiemac.com/singlefamily/pdf/122308_valuationcodeofconduct.pdf. 3 Government-sponsored enterprises (GSE) include Federal National Mortgage Association (Fannie Mae) and Federal Home Loan Mortgage Corporation (Freddie Mac). 4 http://www.hvccappraisalordering.com/AboutTheHVCC. 5 Cho and Megbolugbe (1996) found that among 600,000 purchase loans purchased by Fannie Mae in 1993, about 30% have zero appraisal bias, 5% have appraisal deflation (an appraisal value lower than the transaction price), 60% have less than 10% appraisal inflation (an appraisal value higher than the transaction price), and only 5% have appraisal inflation above 10%. 6 We compare loans sold to and securitized by GSEs (termed GSE loans) with loans insured or guaranteed under programs sponsored by Federal Housing Administration (FHA), US Department of Veterans Affairs (VA) and the US Department of Agriculture (USDA), and securitized via Ginnie Mae (collectively termed non-GSE loans in this article). Close to 80% of the nonGSE loans are FHA loans. Loans kept on banks’ own sheets (portfolio loans) are very limited in 2009; therefore, they are excluded in the analyses of this paper. We measure loans intended for sale to GSEs using GSE loans since originators usually have master agreements or pool purchase contracts with GSEs and almost all originations are sold to GSEs.

compare the loan performance before and after the HVCC effective date. We apply the DID to refinance loans, and as a comparison, we conduct the same analysis on purchase loans. A critical assumption underlying the validity of DID analysis is that the control and treatment groups do not experience different trends prior to the treatment. For GSE and non-GSE loans, while the levels of default rates differ prior to the treatment, the trends do not, assuring the validity of DID analysis. We also estimate the triple-difference (DDD) estimator, i.e., the difference between the DID estimate for refinance loans and that for purchase loans. In addition, we address the possible different pre-intervention trends for the control and treatment group in two ways. First, we add an interaction of the treatment group and a trend variable to the baseline specification. Second, we use a placebo event and show that there are no pre-intervention different trends for the control and the treatment groups. To further understand how the HVCC affects loan underwriting and to reinforce our conclusion, we complement our DID analysis with a direct measure of appraisal bias. By directly linking GSE refinance loans originated in 2009 with previous purchase transactions for the same property identified, and adjusting for local house price changes over time, we measure appraisal bias as the degree of appraisal value inflation in refinance. Potentially reduced appraisal bias would also show up in LTV for originated loans. Supposing a lender uses a threshold in LTV when underwriting a loan, a more accurate appraisal would lead to more rejections of loans since the more accurate appraisal value will lead to a higher LTV, increasing the proportion of loans that exceed the threshold. For loans whose (accurate) LTV is under the threshold and therefore eligible for underwriting, the reduced appraisal value will lead to a higher LTV (while still being below the threshold).7 We therefore expect that the originated loans under the treatment period have a higher LTV. We conduct the above analyses by merging Home Mortgage Disclosure Act (HMDA) data with an OCC proprietary database, Mortgage Metrics (MM), to conduct loanlevel analyses of loan performance. We find that GSE refinance loans showed a significant decrease in default rate relative to non-GSE refinance loans after the HVCC implementation. The 24-month default rate reduction for GSE refinance loans was about 0.734–2.440 percentage points greater than that of non-GSE refinance loans, ceteris paribus. With the mean value of the default measure being 1.01% for GSE and 6.02% for non-GSE refinance loans, the magnitude of the estimated coefficient is economically large. We match 2009 refinance loans with the previous purchase transactions involving the same house and end up with a sample of 106,077 observations. We find that the appraisal bias of GSE refinance loans decreased by 0.644–1.243 percentage points relative to non-GSE refinance loans after the HVCC, a 6–12% reduction from the mean (10.38%). Reduced appraisal inflation would also show up as higher LTV, which would make denial of loan 7 Given a lower appraisal value, borrowers might also reduce their requested loan amount to keep the LTV below the threshold and help their loan applications get approved. However, this is only feasible if borrowers can come up with a larger down payment.

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

applications more likely. Meanwhile, for those that are granted origination given the new accurate LTV, LTV will be higher. We examine LTV and find that LTV of GSE refinance loans increased after the HVCC while that of nonGSE refinance loans did not. We further show that a larger bias is associated with a worse performance, i.e., a higher default rate. Because the LTV calculation of purchases uses the lesser of the appraisal price and the sale price, we do not expect purchase loans to be impacted by the HVCC, which is supported by our findings. There are 12,557 purchase loans in 2009 that involve a previous purchase transaction during 2004–2009 with the same property. The average appraisal bias is less than 3%. Moreover, we find that there is no difference in default rates for GSE and non-GSE purchase loans before and after the HVCC. A concern is that the Home Affordable Refinance Program (HARP), which started in March 2009 and is applied to GSE loans originated before May 2009, might confound our analysis. Specifically, GSE refinance loans originated after May 2009 were subject to the effect of both the HVCC and HARP. We examine our sample and find that 3.7% of our GSE refinance loans are HARP loans.8 Even more interesting is that (i) HARP loans have a higher default rate than non-HARP loans, and (ii) HARP loans originated after May 2009 have a larger increase in default rates than nonHARP loans. We decide to drop all loan originations under HARP in order to accurately assess the effect of the HVCC.9 The last part of our analysis examines whether the denial rate of loan applications is higher after the HVCC, especially for GSE refinance loans, as a result of more accurate appraisal and therefore more informative LTV. Although no significant difference exists in the overall denial rate between GSE refinance loans and that of other loans, we find evidence that the denial rate, due to reasons of collateral or insufficient cash, increased more for GSE refinance loans than for others. Therefore, we interpret our evidence as (i) performance improved for GSE refinance loans more than others, (ii) appraisal bias decreased more for GSE refinance loans than others, (iii) LTV increased for GSE refinance loans more than others, and (iv) denial rate increased for GSE refinance loan applications more than others. Together these factors provide evidence that appraisal bias is reduced with the HVCC implementation. This paper evaluates the HVCC impact on loan performance from the perspective of loan appraisal inflation. It contributes to the limited literature that documents the existence and magnitude of appraisal inflation in refinance loans, illuminating the importance of appraisal regulation.10 It also shows that appraisal inflation is limited for purchase loans. The remainder of the paper is organized as follows. Section 2 provides details of the HVCC and the mortgage 8

Note that this is the first year that HARP was implemented. The take-up rate afterwards was higher than that in 2009 (Urban Institute, Housing Finance Policy Center, Chart-book, March 2014, page 24). 9 The Appendix provides a detailed analysis of the impact of HARP on loan performance. 10 There exists a small but growing literature that documents various forms of information misrepresentation in mortgage securitization process; see Piskorski et al. (2013).

73

industry before and after the HVCC implementation. Section 3 develops the hypotheses. We present the identification strategies in Section 4, and introduce the data and variables in Section 5. Section 6 reports results on loan performance, appraisal bias, and denial decision using the DID and DDD methodologies. We conclude in Section 7. 2. Background 2.1. The HVCC Prior to the 2007 subprime crisis, regulation of real estate appraisers was mainly in the form of licensing and certification. The regulation of real estate appraisers was handled at the state level, with a federal agency, the Appraisal Subcommittee of the Federal Financial Institutions Examinations Council (FFIEC), overseeing the state boards that licensed and certified appraisers. GSEs offer guarantees for their securities and have certain requirements on the appraisal practices.11 The HVCC was born from an agreement between the New York State Attorney General, Office of Federal Housing Enterprise Oversight, Fannie Mae, and Freddie Mac. In 2007 New York Attorney General Andrew Cuomo filed suit against First American Corporation and its appraisal management subsidiary, eAppraiseIT, accusing them of enabling Washington Mutual (WaMu) to pressure appraisers to change values, as well as handpick which appraisers should be used for WaMu’s appraisal reports. Attorney General Cuomo then subpoenaed Fannie Mae and Freddie Mac in order to learn more about loans purchased from banks like WaMu and the valuation processes they used. One outcome of the investigation was the HVCC. Starting May 1, 2009, every loan eventually purchased by GSEs must be in compliance with the HVCC. In the event lenders fail to comply with the HVCC, they will be prohibited from selling mortgages to the GSEs. In addition, if a lender is later found to have violated the provisions of the HVCC, after any such mortgage was sold, it will be required to repurchase any participation interest in any mortgage it may have sold to the GSEs. Based on the HVCC, Fannie Mae and Freddie Mac require the appraisal report be independent from the lenders’ influence. To ensure appraisers’ independence, the HVCC regulates two aspects: requirements on actions of the lender and the process of selecting appraisers. With respect to the actions, the Code prohibits lenders from (i) conditioning the ordering of an appraiser report or the payment on the valuation to be reached, (ii) requesting that an appraiser provides an estimated, predetermined, or desired valuation in an appraiser report prior to the completion of 11 At GSEs, until the mid-1980s, an acceptable appraisal for a residential mortgage involved at least three approaches—replacement (or construction) cost, rental value (or income), and comparable value. Replacement cost and rental value changed slowly as housing prices rose and fell, acting as countercyclical brakes on large increases or decreases in housing prices. In the mid-1980s, the GSEs concluded that the rental value approach was no longer required on the owner-occupied mortgages they bought. In the mid-1990s, they made a similar decision regarding the replacement cost approach.

74

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

the report, or (iii) providing to the appraiser an encouraged or desired value for a subject property or a proposed or target amount to be loaned to the borrower. With respect to selecting the appraisers, it prohibits the practice of appraisals ordered by brokers or real estate agents. Also, in underwriting a loan, the lender shall not utilize any appraiser report prepared by an appraiser employed by the lender, an affiliate of the lender, an entity that is owned by the lender or any entity that owns the lender, or prepared by an appraiser employed by an appraiser company affiliated with the lender, unless (i) the appraiser or the company for which the appraiser works reports to a function of the lender independent of sales or loan production, and (ii) employees in the sales or loan production function of the lender are not allowed to have any substantive communication with the appraiser before the completion of that assignment. Furthermore, lenders are prohibited from using a report prepared by an entity that is engaged by the lender to provide other settlement services. While the HVCC contains many clauses, the goal is to reduce the lender’s influence on appraisers. For external appraisers, the Code tries to eliminate the influence from both the compensation for and the selection of appraisers. For internal appraisers, the Code tries to set up a wall between the lender’s appraisal and loan production functions. 2.2. Mortgage industry before and after the HVCC The issue of appraiser inflation (and more broadly, underwriting standards) became more prominent as the U.S. mortgage finance model evolved from an originateand-hold model to an originate-to-distribute model starting in the 1980s and accelerating in the 2000s. In an originate-to-distribute mortgage financing model, lenders’ incentives to screen the loans are weakened as they focus more on loan origination volume than quality (Ashcraft and Schuermann, 2008; Keys et al., 2010; Demyanyk and Van Hemert, 2011; Purnanandam, 2011; Dell’Ariccia et al., 2012). As a result, lenders’ motivation to evaluate appraisers for their accuracy is reduced as well. In fact, lenders’ pursuits of origination volume give appraisers pressure to inflate property valuation to help approve a loan application. 3. Hypothesis development The value of appraisal is to provide an accurate estimate of the market value of the collateral should the event arise that the collateral be sold to pay back the loan. A typical appraisal involves using recent nearby sales with features comparable to the appraised property to estimate the value of the property. Besides small random measurement bias, there potentially exists systematic appraisal bias driven by appraiser incentives. Appraisals are ordered by lenders. If lenders hold the loans on their own balance sheets, they have incentives to evaluate the risk of the loans including the accuracy of appraisal reports. However, with the originate-to-distribute mortgage finance model, lenders sell loans to the secondary market. Their incentives

are to increase loan volume as long as they can sell. For purchase loans, the policy is that the lower of the transaction price or appraised value be used in calculating LTV. Therefore, the appraiser’s efforts to inflate the appraisal to close a deal are censored by the policy. Only when the appraisal is below the transaction price will the LTV be impacted by the appraisal. However, the lower appraisal could result in a higher LTV under the same transaction price and consequently a potential denial of loan application by the underwriting policy.12 Therefore, it is not in the interest of the lender and appraiser to artificially lower the appraised value of a property for purchase loans. However, for refinance loans, there is no actual transaction and only the appraised value is used as the collateral value in the LTV calculation. Therefore, there is greater potential for lenders to push the appraisers to inflate the price to close a deal. Given a desired LTV, the higher the appraised value, the higher the loan amount permitted, which helps enable the refinance deal if the borrower only had limited equity in the house or would like to cash out home equity accumulated. Or given the loan amount, the higher the appraised value, the lower the LTV, which allows the borrower to meet the desired LTV or receive riskier loan terms. This illustrates the incentives for lenders, specifically loan officers or brokers, to pressure appraisers for a desired appraisal value in refinance transactions. With the 2009 HVCC change, for loans intended for GSE purchase, lenders know that GSE will evaluate whether they use an independent appraiser and thus have to follow the guidance by using independent appraisers. With better protection of appraisal independence, the appraisal is likely to be more accurate and less inflated, thus we predict that for GSE refinance loans relative to non-GSE refinance loans, the appraisal bias will decrease after the HVCC intervention. A more accurate appraised value enables a more accurate calculation of LTV, which is integral in the underwriting policy. Therefore, we predict that compared to non-GSE refinance loans, GSE refinance loans originated after the May 2009 HVCC intervention have better loan performance, due to lower appraisal bias, than loans originated before the intervention. 4. Identification strategy Appraisal bias is the difference between the appraisal and the true value. For purchase loans, there exists the actual transaction, and therefore appraisal bias can be calculated directly using appraised value and sale price. Intuitively appraisal inflation more likely exists in refinance loans, but it is hard to measure the bias directly due to the lack of sale price. LaCour-Little and Malpezzi (2003) use a hedonic price model to estimate the true value. Agarwal et al. (forthcoming) focus on two types of samples with subsequent transactions on the same property: refinance–purchase loans and purchase–purchase loans. They use the appraisal bias calculated from the 12 In rare cases, the borrower can use the appraisal to negotiate and possibly bring down the transaction price (sometimes even get a price match) and thus maintain a relatively low LTV.

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

purchase–purchase sample to tease out other effects (such as a changing market between the two transactions and selection bias between refinance and purchase) from the total effects of the refinance–purchase sample to obtain the appraisal bias estimate. Our paper takes an alternative approach by exploiting an exogenous change to the policy on appraisal independence—the GSE HVCC of 2009. This policy was adopted by GSEs in May 2009 and applies to all loans sold to GSEs. By documenting the effect of this greater protection of appraiser independence, we highlight the existence of appraisal inflation due to lack of appraiser independence and its impact on loan origination quality. We adopt the DID framework.13 The GSE refinance loans represent the treatment group and the non-GSE refinance loans make up the control group, and we explore loan performance before and after the HVCC event. Under the assumption that unobservable factors impact both GSE and non-GSE refinance loans before and after the HVCC implementation similarly, the DID specification is expected to filter out the effect of unobservable factors to provide an accurate estimate of appraisal bias effect. We did not use GSE purchase loans as our control group as refinance loans and purchase loans are generally believed to have different risk drivers. We are aware that using non-GSE (mainly FHA) refinance loans as the control group for GSE refinance loans introduces the difference between GSE loans and non-GSE loans into the analysis. We use the purchase loans to assess the impact of the difference between GSE and non-GSE. Accordingly, our basic specification for the loan performance analysis is:

Defaultit ¼ b0 þ b1 GSEi þ b2 HVCCt þ b3 GSEi  HVCCt þ bx X it þ eit ;

ð1Þ

where i refers to loan and t refers to loan origination month. The dependent variable default is an indicator variable that takes the value of 1 if the loan originated in month t is in foreclosure (including presale, post-sale, and real-estate owned (REO)) or 60 or more days past due (60 + DPD) within 24 months of origination.14 The HVCC represents two indicator variables, P1 and P2, which take the value of 1 if the loan was originated in the first 13 This is a difference-in-differences framework using repeated crosssectional data. 14 Foreclosure is a specific legal process in which a lender attempts to recover the balance of a loan from a borrower who has stopped making payments to the lender by forcing the sale of the asset used as the collateral for the loan. Formally, a mortgage lender (mortgagee), or other lien holder, obtains a termination of a mortgage borrower’s (mortgagor’s) equitable right of redemption, either by court order or by operation of law (after following a specific statutory procedure). If the borrower defaults and the lender tries to repossess the property, courts of equity can grant the borrower the equitable right of redemption if the borrower repays the debt. Through the process of foreclosure, the lender seeks to foreclose the equitable right of redemption and take both legal and equitable title to the property. When the process is complete, the lender can sell the property and keep the proceeds to pay off its mortgage and any legal costs, and it is typically said that ‘‘the lender has foreclosed its mortgage or lien.’’ REO is a class of property owned by a lender after an unsuccessful sale at a foreclosure auction. A foreclosing beneficiary will typically set the opening bid at a foreclosure auction for at least the outstanding loan amount.

75

period (May–August of 2009) and the second period (September–December of 2009), respectively. The reference time period corresponds to the four-month period (January– April of 2009) before the HVCC implementation. By having two indicators for HVCC, we divide the months of 2009 after the HVCC implementation into two periods to illustrate the possible time-varying effect of the HVCC after it goes into effect. The variable GSE is an indicator variable for GSE refinance loans; the omitted group is the non-GSE refinance loans. The X is a vector of loan and borrower characteristics obtained at origination collected from HMDA and MM. The DID estimator is b3 , recording the effect on default rate change associated with GSE refinance loans after the HVCC implementation, relative to non-GSE refinance loans. We expect b3 to be negative. In the full model, in addition to controlling for a rich set of loan and borrower characteristics, we also include zipcode level fixed effect to capture the local socio-economic factors that potentially affect loan performance. Since it is likely error terms are correlated for loans in the same zip code, we also cluster standard error at the zip-code level. Existing studies highlight the importance of originatorspecific fixed effects in loan originations. We create dummy variables for the top 20 lenders and control for the fixed effects of these lenders. The top 20 lenders originate approximately 80% of the loans. One challenge to our identification strategy is that borrowers as well as lenders could have anticipated the HVCC and rushed in the ‘‘low quality’’ loans prior to the HVCC implementation in May 2009, which may have spuriously generated a result that loans after May 2009 have higher quality than those prior to it. We believe there is a limit in how much borrowers and lenders can ‘‘move’’ loans around. Therefore, we look at the entire year of 2009 in three time periods. These periods include 4 months before and 8 months after the HVCC took effect—a sufficient time period to observe the potential rush-in and the eventual converged level. We thus focus our attention on 2009 originations. A key assumption for the validity of DID is that the control and treatment groups experience the same trend in the dependent variable. While the default measure does not show different trends for the GSE and non-GSE refinance loans as plotted in Fig. 1, to statistically test the validity of this assumption, we include an additional variable to capture the effect of a differing trend in default for GSE versus non-GSE (Acemoglu and Angrist, 2000; Besley and Burgess, 2004),15 i.e., we use the following specification:

Defaultit ¼ b0 þ b1 GSEi þ b2 HVCCt þ b3 GSEi  HVCCt þ b4 GSEi  t þ bx X it þ eit ;

ð2Þ

where t is the time trend variable that takes a value of 1, 2, . . ., 12 for January, February, . . ., December of 2009. An alternative way of evaluating the validity of a DID framework is to assess upfront whether the trend pre-exists for the treatment group relative to the control 15 Acemoglu and Angrist (2000) and Besley and Burgess (2004) used a state-specific trend; we use a group-specific trend.

76

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Fig. 1. Default rate and HVCC. The chart plots the default rates by loan origination month for the GSE refinance, non-GSE refinance, GSE purchase, and nonGSE purchase loans originated in 2009. Default is measured as the percentage of loans being in foreclosure (including presale, post-sale, REO) or 60 + DPD within two years after origination.

group. To shed light on this, a common method is to create a placebo intervention event during the pretreatment period (Heckman and Hotz, 1989). If there is a pre-existing trend that differs between the treatment and control groups, the placebo variable would pick up the effect:

ð3Þ

where t refers to the months before the HVCC, and pseudo_HVCC is a month arbitrarily chosen from the preHVCC period. As a placebo test, we then repeat the analysis (1)–(3) for the purchase loans with GSE purchase loans as the treatment group and non-GSE purchase loans as the control group. We expect to see a neutral or weaker impact of the HVCC on this population as we argue that the HVCC is mainly applicable to GSE refinance loans. The preceding analyses focus on refinance loans with purchase loans as a comparison group. Within refinance loans, GSE loans are the treatment group and non-GSE loans are the control group. One concern with this analysis is that, before and after the May HVCC implementation, GSE loans might have experienced different shocks than non-GSE loans. If the assumption that the shocks affecting GSE loans are the same for refinance and purchase loans, we can use the purchase loans as another layer of control group, as follows:

Defaultit ¼ b0 þ b1 GSEi þ b2 Refii þ b3 HVCCt þ b4 GSEi  HVCCt þ b5 Refii  HVCCt þ b6 GSEi  Refii þ b7 GSEi  Refii  HVCCt þ bx X it þ eit ;

 HVCCt þ b5 Refii  HVCCt þ b6 GSEi  Refii þ b7 GSEi  Refii  HVCCt þ b8 GSEi  t þ bx X it þ eit ;

ð5Þ

and

Defaultit ¼ b0 þ b1 GSEi þ b2 pseudo HVCCt þ b3 GSEi  pseudo HVCCt þ bx X it þ eit ;

Defaultit ¼ b0 þ b1 GSEi þ b2 Refii þ b3 HVCCt þ b4 GSEi

ð4Þ

We are therefore effectively using a DDD estimator and the coefficient of interest is b7 . Our hypothesis is that b7 < 0. We also test the DDD specification with the time trend and pseudo-HVCC, as shown by Eqs. (5) and (6) below:

Defaultit ¼ b0 þ b1 GSEi þ b2 Refii þ b3 pseudo HVCCt þ b4 GSEi  pseudo HVCCt þ b5 Refii  pseudo HVCCt þ b6 GSEi  Refii þ b7 GSEi  Refii  pseudo HVCCt þ bx X it þ eit ;

ð6Þ

To understand how the HVCC impacts loan performance and quantify its effect on reducing appraisal bias, we directly evaluate the appraisal bias of GSE refinance loans and its impact on loan decision and performance. We hypothesize that the HVCC implementation in May 2009 can effectively decrease the appraisal bias for GSE refinance loans. Therefore, we expect to observe a negative a3 in Eq. (7)

Biasit ¼ a0 þ a1 GSEi þ a2 HVCCt þ a3 GSEi  HVCCt þ ax X it þ lit ;

ð7Þ

where the definitions of i, t, HVCC, and X remain the same as those in previous equations. The lender and zip-code fixed effects are also controlled for, and the estimates are obtained with errors clustered at the zip-code level. An alternative way of showing the reduced appraisal bias (inflation) is to see whether the LTV for the originated loans increases. We therefore estimate an identical equation as Eq. (7) except that the dependent variable is origination LTV. As a result of reduced appraisal bias, we predict that the loan performance is improved with stricter appraisal regulation. Therefore, bb in Eq. (8), as shown below

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Defaultit ¼ b0 þ bb Biasit þ b1 GSEi þ b2 HVCCt þ b3 GSEi  HVCCt þ bx X it þ eit ;

ð8Þ

has a positive sign, and b3 a negative sign. Finally, we test whether the reduction in appraisal bias due to the HVCC would lead to a greater denial rate by estimating Eqs. (1) and (4) with the dependent variable being an indicator variable for the loan application being rejected, for which we expect to see a positive b3 and b7 respectively. Due to the lack of zip-code information in HMDA data, the denial analysis controls for MSA level fixed effects with errors clustered at the MSA level. 5. Data and summary statistics 5.1. Data for performance analysis Our main sources of data are HMDA and OCC Mortgage Metrics data. HMDA data are a result of the Home Mortgage Disclosure Act, which requires all U.S. lenders to report loan applications, with few financial institutions exempted (www.ffiec.gov). The HMDA database provides decision outcomes on loan applications. It also has basic loan application information on the loan, borrower, and collateral, but it lacks detailed loan characteristics, such as consumer credit score, LTV, debt-to-income ratio (DTI), and other important loan product features such as adjustable rate mortgage (ARM) or fixed rate and whether interest only (IO) or negative amortization is used. Neither does it have loan performance data. The OCC MM data are collected from 10 large banks supervised by OCC that service about 64% of first lien mortgages in the United States. The MM database records various loan attributes at origination as well as contemporaneous loan performance starting from January 2008. The MM data, however, do not have information on loan purpose; we thus merge MM with HMDA data to enable loan performance analysis by loan purpose (purchase versus refinance).16 Since the HVCC took effect on May 1, 2009, we focus our analyses using calendar year 2009 data. We do not include 2008 since the credit crisis in late 2008 caused disruption to the mortgage market; we stop at December 2009 since, effective January 1, 2010, Federal Housing Administration (FHA) followed suit with GSE and started implementing very similar practices on appraiser independence.17 The Home Affordable Modification Program (HAMP) is applicable to loans originated by January 1, 2009, so it does not impact our analysis period. There are 19.5 million loan applications in 2009 HMDA, among which 8.56 million are loan originations with the purpose of purchasing or refinancing. To enable loan performance analysis, we merge HMDA and MM data using a carefully crafted matching scheme. The two datasets are matched on six variables: the action date, loan amount, 16 Note that HMDA data do not have further detail about refinance, e.g., whether the refinance is for a rate decrease or cash out. For the year 2009 that is under discussion here, likely most of the refinance loans are for rate decrease due to the depressing housing market and low interest rate policy. 17 http://portal.hud.gov/hudportal/documents/huddoc?id=09-28ml.pdf.

77

lien status, loan type, owner occupancy, and location of the loan. The matches for the first five variables are exact. For location, HMDA and OCC MM do not have the same location detail: HMDA has census tract information and OCC MM has zip-code information. We use a tract-to-zip lookup table to deduce the corresponding zip code for each census tract. We then match the HMDA zip code to the OCC MM zip code. The merged HMDA-MM sample contains 1.34 million loan originations.18 It is representative of the HMDA origination sample. Panels A through D of Fig. 2 compare the distribution of key variables of the HMDA and HMDA-MM merged samples. Panel A shows that the proportion of originations by loan type (conventional versus nonconventional) and loan purpose (purchase versus refinance) are very similar for the two samples. Panel B shows the patterns of total originations over time for the HMDA and HMDA-MM merged samples. It is reassuring to see that their patterns of originations by month are very similar. Panel C shows the proportion of loans that involve borrowers of different race, ethnicity, and gender. Panel D compares loan amount in a continuous setting of kernel density plots for loans in the HMDA and HMDA-MM samples. Again, the curves for the HMDA and merged HMDA-MM samples exhibit a very high degree of similarity, further attesting to the representativeness of the merged HMDA-MM sample. Through merging with MM, the HMDA-MM sample has information on the investor of the loans each month after origination: GSEs, Ginnie Mae, private label securitization, or staying in the portfolio of a lender. We use the investor information 6 months after origination to further define the sample of loans. Of the 1.34 million HMDA and MM merged loans, about 74% are conventional loans, and the remaining 26% are nonconventional loans. The majority (more than 95%) of conventional loans were sold to GSEs 6 months after origination, suggesting that lenders originate with the intention to sell to GSEs; we term them GSE loans. More than 98% of the nonconventional loans, i.e., FHA/VA/USDA guaranteed or insured loans were sold to investors in Ginnie Mae guaranteed securities. As shown by the HMDA-MM data, in 2009 less than 3% of the loans were held in portfolio and the private securitization diminished to a negligible size to warrant any meaningful analysis. We therefore focus on loans sold to GSEs and loans in Ginnie Mae guaranteed securities. To highlight the contrast between the two, we call the latter non-GSE loans. We also apply the following restrictions to the data to obtain relatively homogenous data to facilitate comparison of loans with different loan purpose and investor type: within conforming loan limit; loan terms no more than 30 years; for single-unit property; non interest-only; with valid property value at origination; and not originated under HARP. The refined sample contains 1.11 million HMDA-MM loans. Panel A of Table 1 provides summary statistics for our sample of performance analysis, as divided into four 18 It would be ideal to use combined loan-to-value ratio (CLTV) instead of LTV. However, MM does not report CLTV before December 2009. Extrapolation for loans originated before December 2009 shows that there were limited (around 5%) loans originated in 2009 with a second lien.

78

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Fig. 2. Representativeness of the merged sample. The four figures plot the proportion of 2009 loan originations by loan type and purpose, origination month, borrower demographic, and loan amount using HMDA versus merged HMDA-MM sample, respectively. The HMDA sample contains 8.56 million loans, and HMDA-MM sample contains 1.34 million loans.

Table 1 Summary statistics. Refinance

Panel A: Sample for performance analysis Default (including foreclosure and 60 + DPD) rate within 24 months after origination (in%) Borrower is Hispanic Borrower is White Borrower is Asian Borrower is Black Borrower is single female FICO < 600, missing or invalid 600 6 FICO < 660 660 6 FICO < 720 720 6 FICO < 780 780 6 FICO 6 950 0% < DTI 6 20% 20% < DTI 6 30% 30% < DTI 6 40% 40% < DTI 6 50% DTI > 50% DTI is missing or invalid 0% < LTV 6 60% 60% < LTV < 80% LTV = 80% 80% < LTV 6 95% LTV > 95% LTV is missing or invalid Subprime loan (vs. prime and Alt-A) Alternative or stated documentation on income Adjustable rate mortgage Owner occupancy indicator

Purchase

GSE

Non-GSE

GSE

Non-GSE

1.01 3% 82% 4% 2% 15% 0% 2% 13% 41% 43% 8% 14% 13% 9% 2% 55% 38% 42% 7% 9% 2% 2% 0% 74% 1% 93%

6.02 6% 78% 1% 9% 17% 23% 21% 29% 20% 7% 6% 12% 14% 14% 7% 48% 2% 9% 0% 38% 51% 0% 4% 74% 2% 99%

0.59 4% 79% 8% 2% 19% 0% 2% 14% 44% 40% 9% 19% 23% 13% 3% 33% 16% 42% 23% 20% 0% 0% 0% 61% 2% 84%

4.50 9% 79% 2% 8% 23% 3% 26% 33% 28% 9% 2% 10% 23% 24% 9% 32% 1% 3% 0% 22% 75% 0% 3% 57% 2% 100%

79

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90 Table 1 (continued) Refinance

Purchase

GSE

Non-GSE

GSE

Non-GSE

Loan amount 6 $100k $100k < loan amount 6 $200k $200k < loan amount 6 $300k Loan amount > $300k Loan term 6 180 months 180 months < loan term 6 360 months Origination channel being retail Origination channel being wholesale Origination channel being correspondent Number of observations

12% 38% 27% 23% 16% 84% 55% 5% 10% 612,684

13% 52% 25% 9% 12% 88% 56% 8% 16% 105,072

16% 36% 24% 23% 7% 93% 56% 11% 12% 198,423

20% 50% 22% 9% 1% 99% 59% 13% 12% 188,907

Panel B: Sample for appraisal bias analysis Default (including foreclosure and 60 + DPD) rate within 24 months after origination (in%) Property value at origination Appraisal bias (in%) Borrower is Hispanic Borrower is White Borrower is Asian Borrower is Black Borrower is single female FICO < 600, missing or invalid 600 6 FICO < 660 660 6 FICO < 720 720 6 FICO < 780 780 6 FICO 6 950 0% < DTI 6 20% 20% < DTI 6 30% 30% < DTI 6 40% 40% < DTI 6 50% DTI > 50% DTI is missing or invalid 0% < LTV 6 60% 60% < LTV < 80% LTV = 80% 80% < LTV 6 95% LTV > 95% LTV is missing or invalid Subprime loan (vs. prime and Alt-A) Alternative or stated documentation on income Adjustable rate mortgage Owner occupancy indicator Loan amount 6 $100K $100K < loan amount 6 $200K $200K < loan amount 6 $300K Loan amount > $300K Loan term 6 180 months 180 months < loan term 6 360 months Origination channel being retail Origination channel being wholesale Origination channel being correspondent Number of observations

1.10 361,849 10.38 2% 83% 5% 1% 14% 0% 2% 13% 42% 43% 5% 10% 9% 6% 1% 69% 22% 48% 7% 18% 5% 0% 0% 90% 1% 92% 6% 34% 31% 28% 9% 91% 64% 1% 3% 87,287

5.37 211,638 13.52 5% 80% 2% 8% 15% 26% 17% 27% 23% 7% 6% 11% 12% 11% 6% 55% 0% 3% 0% 24% 72% 0% 1% 91% 4% 100% 7% 49% 32% 13% 4% 96% 60% 2% 4% 18,790

0.35 309,486 2.79 4% 80% 7% 2% 19% 1% 1% 14% 44% 41% 9% 20% 23% 12% 2% 33% 15% 42% 22% 21% 0% 0% 0% 67% 3% 84% 15% 36% 26% 23% 7% 93% 57% 11% 11% 6,490

3.91 187,191 2.33 9% 79% 3% 8% 23% 2% 25% 33% 30% 9% 2% 10% 21% 24% 9% 34% 0% 3% 0% 23% 73% 0% 2% 63% 2% 100% 17% 52% 23% 9% 1% 99% 61% 12% 10% 6,067

Panel C: Sample for denial analysis Percent of all denials (HMDA action type = 3 or 7) Percent of denials due to collateral Percent of denials due to insufficient cash Borrower is Hispanic Borrower is White Borrower is Asian Borrower is Black Borrower is single female Loan amount 6 $100K $100K < loan amount 6 $200K $200K < loan amount 6 $300K $300K < loan amount 6 $417K Loan amount > $417K Income 6 $40K $40K < income 6 $65K

21.60 7.03 0.51 4% 78% 5% 3% 17% 18% 38% 24% 19% 2% 13% 22%

31.00 7.59 0.85 7% 71% 1% 11% 20% 13% 49% 26% 9% 3% 15% 20%

15.05 2.85 0.96 6% 76% 9% 3% 22% 21% 37% 22% 18% 3% 18% 25%

17.91 2.07 0.93 13% 75% 3% 10% 25% 21% 50% 20% 7% 2% 28% 36%

(continued on next page)

80

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Table 1 (continued) Refinance

$65K < income 6 $100K $100K < income 6 $200K Income > $200K Loan-to-income ratio 6 1 1 < loan-to-income ratio 6 2 2 < loan-to-income ratio 6 3 3 < loan-to-income ratio 6 4 Loan-to-income ratio > 4 Loan-to-income ratio is missing Number of observations

Purchase

GSE

Non-GSE

GSE

Non-GSE

27% 29% 10% 9% 32% 29% 15% 12% 3% 5,753,219

18% 9% 38% 1% 14% 20% 14% 14% 37% 1,575,681

25% 25% 7% 7% 27% 32% 19% 13% 1% 1,311,743

23% 11% 2% 2% 17% 36% 28% 16% 1% 1,690,627

This table reports summary statistics for the samples used for default analysis (Panel A), appraisal bias analysis (Panel B), and denial analysis (Panel C). Sample A is obtained by merging HMDA 2009 with the OCC Mortgage Metrics database. Sample B contains originations in 2009 that can be matched with prior purchase loans originated on the same property during 2004–2009. Sample C is from HMDA 2009. Default is an indicator variable for the loan being in foreclosure (including presale, post-sale, REO) or 60 + DPD within two years after origination. Appraisal bias is measured as the percentage difference between the appraisal value of the 2009 origination and a prior purchase price of the same property, each adjusted by the HPI for the zip code where the property is located. GSE loans refer to conventional loans whose investor 6 months after origination is GSE; non-GSE loans refer to nonconventional loans whose investor 6 months after origination is Ginnie Mae.

subsamples by loan purpose and investor type. In 2009 GSE has a much larger market share than non-GSE. Within GSE, refinance loans compose 76% of all loans. GSE loans have a much lower default rate than non-GSE loans: the 24-month default rate is approximately 1% for GSE loans and 4–6% for non-GSE loans. By loan purpose, refinance loans have a higher default rate than purchase loans. Close to 84% of GSE loans have a FICO score that is above 720, for both refinance and purchase loans, while less than 40% of the non-GSE loans have a FICO above 720. The difference is due to the different goals of GSE versus nonGSE institutions—for example, FHA’s mission is to help sponsor home ownership, especially in the form of low down payments. The difference across the four types of loans in terms of DTI is much less prominent. LTV again differs across the four types. The majority of GSE loans have LTV that is not above 80% while LTV for the majority of non-GSE loans are above 80%. Non-GSE loans have a slightly higher proportion of being subprime and ARM. Almost all homes for non-GSE loans are owner occupied while the number for GSE loans is above 80%. For all types, the major loan origination channel is retail. Fig. 1 plots the 24-month default rate by loan origination month for the four populations as a combination of investor type and loan purpose. Prior to May 2009, the effective date of the HVCC, all four curves are flat over time despite the different levels. However, starting May 2009, the default rates of all loans except GSE purchase loans demonstrate an increasing trend with a much larger increase for non-GSE refinance and non-GSE purchase loans than for GSE refinance loans.

5.2. Data for appraisal bias analysis We also attempt to directly measure the appraisal bias for 2009 refinance loans. The MM reports a property value at origination which is the appraisal value for refinance loans and the minimum of purchase or appraisal value for purchase loans. The MM also provides property IDs for active loans starting in December 2009. Using these

property IDs, we search the MM database back to January 2004 and collect the prior loans on the same property that has a GSE refinance origination in 2009.19 By matching these prior loans to HMDA data, we obtain their loan purpose in order to keep prior purchase loans for our analysis. To measure the appraisal bias in refinance loans, we adjust the appraisal value for both the previous purchase and the 2009 refinance loans, using the zip-code level (the finest possible level from the available datasets) Home Price Index (HPI) at the origination time of each loan.20 Specifically, the appraisal bias is calculated as

Appraisal Bias ¼

  PV 2 =HPI2  1  100; PV 1 =HPI1

where PV 1 is the origination property value for the first purchase loan and PV 2 is the origination property value for the 2009 refinance loan. The HPI1 is the zip-code level HPI at the first purchase loan origination and HPI2 is the zip-code level HPI at the refinance loan origination. To provide a comparison, we also apply the same method on purchase loans in 2009 to calculate their appraisal bias. Out of the 1.11 million loans that were originated in 2009, we find that 11% have a prior purchase loan in 2004–2009. The final dataset contains 118,634 loans for the four subsamples by loan purpose and investor type: 87,287 GSE refinance loans, 18,790 non-GSE refinance loans, 6,490 GSE purchase loans, and 6,067 nonGSE purchase loans. Table 1, Panel B provides the summary statistics of the sample for appraisal bias measurement by loan purpose and final investor. Compared to the performance analysis sample, the bias sample illustrates similar borrower and 19 There exists a concern that refinance loans in 2009 that have a prior purchase transaction back to January 2004 for the same property may not be representative as 2009 is a year of depressed house price and therefore, those loans likely started with relatively low LTV or possess certain other features. As we discuss later, the bias sample has similar borrower and loan characteristics as the performance analysis sample. 20 The HPI is the single family combined home price index provided by CoreLogic with calendar year 2000 as the base.

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Panel A: Appraisal Bias by HVCC

Panel B: Origination LTV by HVCC

Panel C: Appraisal Bias by LTV Groups

Panel D: Origination Distribution by LTV Groups

81

Fig. 3. Appraisal bias, LTV, and HVCC. The chart reports the appraisal bias and origination LTV changes before and after HVCC, and the relationship between appraisal bias and origination LTV using the bias analysis sample. Appraisal bias is measured as the percentage difference between the appraisal value of the 2009 origination and a prior purchase price of the same property, each adjusted by the HPI at the time of transaction for the zip code where the property is located.

loan characteristics, suggesting that the bias that was introduced as a result of the matching with prior purchases is limited. The mean bias for GSE refinance loans is 10.38%, i.e., the refinance appraisal value is on average 10.38% higher than the transaction price for the same property, controlling for the HPI change. The mean bias for the non-GSE refinance loans is 13.52%. As expected, the appraisal bias value for both GSE and non-GSE purchase loans is much lower (around 2–3%) than that for refinance loans. We observe a consistent pattern from Panel A of Fig. 3, which plots the appraisal bias curve for each origination month by loan purpose and final investor. We find that while the bias for non-GSE refinance loans only slightly decreased over 2009, that for GSE refinance loans demonstrated a significant decrease in bias from May 2009. The appraisal bias curves for GSE and non-GSE purchase loans do not demonstrate such divergence, with both remaining at a much lower 2–3% level throughout 2009. An alternative way to reflect the less inflated appraisal value is the LTV. A more accurate appraisal value will result in loans with high LTV being rejected and those that were granted originations, compared with previous originated loans, will have lower appraisal value and thus higher LTV. Panel B of Fig. 3 plots the LTV before and after the HVCC by loan purpose and investor type. We see that while the LTV for non-GSE refinance and purchase loans remains high at around 96% and the mean LTV for GSE purchase loans decreases over 2009, the mean LTV for GSE refinance loans increases (especially during the first 3 months) after May 2009.

Panel C of Fig. 3 plots the relation between the appraisal bias and the LTV groups. We have several observations. First, refinance loans have higher appraisal bias than purchase loans. Second, appraisal bias decreases with LTV for refinance loans, while purchase loans do not have this association. Taken together, this is consistent with the hypothesis that, to satisfy the eligibility criteria set by GSE, lenders had greater incentives to inflate appraisal value to reduce the LTV level for refinance loans than for purchase loans. Panel D of Fig. 3 provides the loan distribution corresponding to Panel C of Fig. 3, showing that close to 80% of GSE refinance loans have LTV at or below 80%, while more than 70% of non-GSE refinance loans have LTV greater than 95%. GSE refinance loans show a high concentration of loans with LTV at 80%, as shown in Panel A of Fig. 4. We find that while the number of loans is in the order of 1000–2000 at other LTV levels, more than 10,000 loans are at the 80% LTV level. We also notice that the number of loans with LTV greater than 80 is of materially smaller magnitude, especially right above the 80% LTV.21 This discontinuity in the number of loans at LTV 80% for GSE eligibility is similar in nature to another well-documented discontinuity—FICO at 620 for private-label securitization eligibility (Keys et al., 2010; Piskorski et al., 2010). Different from their case where 21 GSEs have programs for purchase loans with LTV greater than 80%. A prerequisite is private mortgage insurance (PMI) with possibly more stringent requirements of the borrower on other aspects, such as credit score, etc.

82

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Fig. 4. Local discontinuity at the LTV thresholds. The chart reports the appraisal bias and number of originations by origination LTV before and after HVCC for GSE refinance and non-GSE refinance loans using the bias analysis sample. Appraisal bias is measured as the percentage difference between the appraisal value of the 2009 origination and a prior purchase price of the same property, each adjusted by the HPI at the time of transaction for the zip code where the property is located.

FICO is relatively hard to manipulate, appraisal value, and thus LTV, has greater potential to be manipulated. The discontinuity in the number of loans at LTV 80% thus begs the question of whether there exists pressure to reach an LTV of 80%. This pressure is presumed to be particularly great for refinance loans with true LTV at right above 80% (so as to reach an observed 80% LTV). Panel A of Fig. 4 shows that the bias for loans with LTV at 80% is 10% while that for loans with observed LTV of slightly higher than 80% is significantly lower at 4.3%. A similar pattern is observed for LTV kinks at 75, 85, 90, and 95%, albeit at a much smaller magnitude. These pieces of evidence together are consistent with an interpretation that the incentive (pressure) to inflate appraisals is greater when there is a need to meet underwriting eligibility or to gain more preferable pricing. Panel B of Fig. 4 plots the appraisal bias and distribution over the LTV for non-GSE refinance loans. We find that there are two kinks in the number of FHA/VA loans, at the 87% LTV level, which is around the required maximum level for cash-out refinance loans, and at the 97% LTV level, which is the required maximum level for regular-rate refinance loans. There appears to be a decline in bias at these two kinks, but the magnitude is much smaller than observed in GSE refinance loans. While our focus is whether the HVCC reduces the systematic appraisal bias as a result of enforcing appraiser independence, we are also interested in the impact of the HVCC on accuracy in property appraisal. As the HVCC promotes the use of appraisal management companies (AMC),22 there have been concerns that AMC might hire inexperienced appraisers that are not familiar with the area. As a result, incidence of inaccurate appraisals could increase and so could the variance of valuations (Zhu and Kelley Pace, 2012). To address this concern, we compare the standard deviation of appraisal bias for GSE refinance loans before

and after the HVCC implementation and contrast the change against that for non-GSE refinance loans. We find that while the standard deviation of bias for non-GSE refinance loans did not change much before and after the HVCC implementation from 16.25 to 16.96, the standard deviation of the GSE refinance loans decreased by 15%, from 20.76 to 17.64, suggesting that the appraisal was actually getting more accurate after the HVCC. However, the reduction in standard deviation of bias for GSE refinance loans is much smaller than the reduction in bias level, from 14.39 to 7.92% on average, a 45% reduction. We plot the kernel density of the appraisal bias for GSE and non-GSE refinance loans before and after the HVCC implementation in Fig. 5. In Panel A, we see that for GSE refinance loans the shape of the distribution did not change much, yet the position of the distribution moved leftward, i.e., bias decreased. In Panel B, we do not observe a significant shift of bias for non-GSE refinance loans. From these pieces of evidence, we conclude that it is the reduction in bias for GSE refinance loans that is the most prominent change with the HVCC implementation. Moreover, as we conduct regression analysis to analyze bias in Section 4, both the mean and variance of bias will be considered since the regression is at the loan level.

22 To comply with the HVCC and prevent undue pressure from commissioned lending officers or brokers, lenders contract with AMC, which then contract with appraisers for their appraisal service. By using this intermediary, the lender cannot directly pressure an appraiser for a certain valuation.

23 Note that HMDA reporting of denial reason is optional except for OCC respondents who must report. About 75% of our denial sample has reported denial reasons, and the proportion is stable over the year of 2009, suggesting that there is no systematic change of HMDA denial reason reporting.

5.3. Data for denial analysis To complete our analysis, we also look at whether a loan is more likely to be denied given the more accurate appraisal with the HVCC implementation. Data for denial analysis is from HMDA. The denial variable is an indicator variable for the loan application being denied (HMDA action type = 3 or 7). HMDA provides specific reasons that a loan is denied, including DTI, employment history, credit history, collateral, insufficient cash (down payment, closing costs), unverified information, etc.23 We are interested

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

83

Fig. 5. Distribution of appraisal bias before and after HVCC. These two figures plot the continuous kernel density of appraisal bias for GSE refinance and non-GSE refinance loans before and after HVCC using the bias analysis sample. Appraisal bias is measured as the percentage difference between the appraisal value of the 2009 origination and a prior purchase price of the same property, each adjusted by the HPI at the time of transaction for the zip code where the property is located.

in whether a loan is denied due to high LTV or too little equity. We thus create two alternative dependent variables: a loan is denied because of collateral, and a loan is denied because of insufficient cash. Table 1, Panel C provides the summary statistics of the sample for denial analysis by loan purpose and final investor. The GSE refinance application number is around four times that of non-GSE refinance applications with the percent of denial lower for GSE refinance (21.60%) than for non-GSE refinance (31.00%). Purchase loans have lower denial rates, 15.05% for GSE loans and 17.91% for non-GSE loans. The percent of denials due to insufficient collateral is 7.03% for GSE refinance and 7.59% for non-GSE refinance loans, and are lower for both types of purchase loans. Less than 1% of all loan types were denied due to insufficient down payment or closing cost. Panel A of Fig. 6 plots the percent of denial by loan purpose and investor during 2009. We do not observe any obvious pattern. Panels B and C plot the denial rates due to collateral and insufficient cash, respectively. We find in Panel C that the denial rate for GSE purchase loans decreased relative to non-GSE purchase loans, and GSE refinance loans did not exhibit obvious change relative to non-GSE refinance loans after the HVCC implementation. This suggests that GSE refinance, relative to purchase loans, showed a greater increase in denial rates due to insufficient down payment than non-GSE refinance loans after the May 2009 HVCC intervention. Panel B has a similar pattern, although with a smaller magnitude. In Section 6.3, our regression results will confirm these observations. 6. Results We first examine whether the performance of GSE refinance loans improved, and then investigate whether

the channel is via a reduced appraisal bias, and finally whether the underwriting standards for GSE refinance loans improved more than others after the HVCC implementation. 6.1. Results on loan performance Using the merged HMDA and OCC MM data, we examine the performance of loans originated before and after the May 2009 HVCC implementation. The loan performance is the default rate 24 months after origination.24 Table 2 reports ordinary least square (OLS) regression results for loan default analyses. Using the default analysis sample, Panel A reports the baseline model results and Panels B and C provide robustness checks by conducting the trend and pseudo-HVCC models, respectively. Each panel reports results on the refinance subsample and the purchase subsample as a comparison under the DID specification, and the entire default analysis sample under the DDD specification. As shown by Column 1 of Panel A, GSE refinance loans on average have a default rate of 1.68 percentage points higher than non-GSE refinance loans 24 months after origination. The default rate of GSE refinance loans decreased 0.734 percentage point more than that of non-GSE refinance loans the first period after the HVCC implementation, and 2.440 percentage points the second period after the HVCC implementation. For comparison, Column 2 shows that GSE purchase loans experienced a larger reduction in default rate than non-GSE purchase loans; however, the difference is smaller than that for refinance loans. Column 3 reports results using the triple-difference specification. We see that the coefficients on GSE  Refi  HVCC periods are negative, indicating that 24 We conducted the performance analysis using various measures of defaults; the conclusion remains unchanged.

84

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Fig. 6. Denial rates and HVCC. These three figures plot the denial rate for 2009 loan applications by loan purpose and investor type using 2009 HMDA. Denial in Panel A refers to the loan application being denied (HMDA action taken = 3 or 7). Collateral denial in Panel B refers to the loan application being denied due to collateral-related reasons. Cash denial in Panel C refers to the loan application being denied due to cash-related reasons (e.g., insufficient down payment or closing cost). GSE loans refer to conventional loans whose investor 6 months after origination is GSE; non-GSE loans refer to nonconventional loans whose investor 6 months after origination is Ginnie Mae.

reduction in default rate due to the HVCC is particularly larger for GSE refinance loans. We conduct a series of robustness checks of our default regression analyses. In Panel B of Table 2, we report regression results including a variable that captures the trend that might differ between the control and the treatment groups. For refinance loans shown in Column 1, we find that with the inclusion of the GSE  t variable, the coefficients on GSE  HVCC remain significantly negative and in fact have slightly larger magnitudes. The coefficient on GSE  t is positive for refinance loans. For purchase loans, the coefficients of GSE  HVCC remain significant but the magnitude becomes smaller. For the combined population, the coefficients on GSE  Refi  HVCC remain significantly negative, and the magnitude is only slightly smaller. We interpret that these results support the validity of using the DID or DDD framework. We then report the regression results from using a pseudo-HVCC for the pre-event period to show that there is no pre-HVCC trend in default rate. Panel C of Table 2 shows the regression results where we look at the preevent sample (January–April 2009) and create a pseudoHVCC—the value equals 0 for January and February of 2009 and 1 for March and April of 2009. We find that all the coefficients related to the pseudo-HVCC variables are statistically insignificant. These findings suggest that there

does not exist pre-intervention different trends for the control and treatment groups, providing support of the validity of using the DID or DDD framework.

6.2. Results on appraisal bias 6.2.1. Appraisal bias, LTV, and HVCC This section provides analysis results on the direct measure of refinance loan appraisal bias. It focuses on GSE refinance loan performance before and after the HVCC implementation. Results are in Column 1 of Table 3. The variable of interest is GSE  HVCC, i.e., the effect of the HVCC on GSE refinance relative to non-GSE refinance. The coefficient on the interaction term is significantly negative, providing formal evidence that GSE refinance loans have a greater reduction in bias after the HVCC implementation than non-GSE refinance loans. The magnitude of the coefficient is not trivial. The coefficient on GSE  P1 and GSE  P2 is 0.644 and 1.243, respectively. Since the mean of the bias measure is in the order of 10 percentage points, the coefficient suggests that GSE refinance loans experienced a 6–12% reduction in appraisal bias. Interestingly, the coefficient on the interaction term for purchase loans, shown in Column 2 of Table 3, is insignificant. Taken together, these pieces of

85

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90 Table 2 Loan default as a function of HVCC. Dependent variable

Default

Sample

Refinance (1)

Purchase (2)

Refinance + purchase (3)

1.677*** [10.665]

0.712*** [5.009]

0.903*** [5.357] 3.002*** [14.770] 0.734*** [4.329] 2.440*** [12.012]

0.323** [2.498] 1.253*** [8.457] 0.424*** [3.217] 1.362*** [9.396]

717,756 0.089

387,330 0.115

1.281*** [10.318] 0.056 [0.341] 0.217* [1.807] 1.181*** [8.819] 0.250** [1.971] 1.152*** [8.292] 0.663*** [3.244] 1.784*** [7.577] 0.317* [1.890] 0.416** [1.982] 1.077*** [4.436] 1,105,086 0.078

1.507*** [9.295]

0.732*** [4.839]

0.913*** [5.417] 3.023*** [14.862] 0.932*** [5.345] 2.914*** [12.587]

0.321** [2.481] 1.250*** [8.362] 0.395*** [2.633] 1.302*** [6.213]

0.058*** [4.305] 717,756 0.089

0.008 [0.402] 387,330 0.115

(1) 0.200 [0.711]

(2) 0.265 [0.958]

0.069 [0.260] 0.036 [0.135]

0.154 [0.690] 0.346 [1.487]

Panel A: Baseline model GSE Refi P1 P2 GSE  P1 GSE  P2 Refi  P1 Refi  P2 GSE  Refi GSE  Refi  P1 GSE  Refi  P2 Observations R-squared Panel B: Trend model GSE Refi P1 P2 GSE  P1 GSE  P2 Refi  P1 Refi  P2 GSE  Refi GSE  Refi  P1 GSE  Refi  P2 GSE  t Observations R-squared

1.118*** [8.736] 0.056 [0.344] 0.229* [1.903] 1.203*** [8.964] 0.475*** [3.559] 1.611*** [9.851] 0.661*** [3.233] 1.782*** [7.568] 0.313* [1.866] 0.392* [1.866] 1.098*** [4.521] 0.059*** [5.367] 1,105,086 0.078

Panel C: Pseudo-HVCC model GSE Refi PP1 GSE  PP1 Refi  PP1

(3) 0.007 [0.030] 0.068 [0.255] 0.107 [0.545] 0.308 [1.449] 0.011 [0.033] (continued on next page)

86

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Table 2 (continued) Dependent variable

Default

Sample

Refinance (1)

Purchase (2)

Refinance + purchase (3)

102,944 0.205

0.108 [0.392] 0.137 [0.406] 381,767 0.115

GSE  Refi GSE  Refi  PP1 Observations R-squared

278,823 0.135

This table reports OLS regression results of loan default rates. Using the default analysis sample, Panel A reports the baseline model results, and panels B and C provide robustness checks by estimating the trend and pseudo-HVCC models. Each Panel reports results on the refinance subsample and the purchase subsample under the DID specification, and the entire default analysis sample under the DDD specification. Default is an indicator variable for the loan being in foreclosure (including presale, post-sale, REO) or 60 + DPD within two years after origination. Variable GSE is an indicator variable for a loan being a GSE loan versus a non-GSE loan. For baseline and trend models, HVCC indicators P1 and P2 refer to the periods of May–August 2009 and September–December 2009, respectively. We select March 2009 as the pseudo-HVCC effective month. Therefore, for the ‘‘pseudoHVCC’’ model, PP1 takes the value of 0 for January and February of 2009 and the value of 1 for March and April of 2009. All the regressions are controlled for borrower and loan characteristics as listed in Panel A of Table 1, as well as lender and zip-code level fixed effect. Standard errors are clustered at the zip-code level. T-values are in brackets. *** Indicates statistical significance level at 0.01 level. ** At 0.05 level. * At 0.10 level.

Table 3 Appraisal bias, origination LTV, and HVCC. Dependent variable Sample

Appraisal bias

Origination LTV

Refinance (1)

Purchase (2)

Refinance (3)

Purchase (4)

GSE

9.376*** [28.531] 4.051*** [14.761] 5.013*** [14.955] 0.644** [2.156] 1.243*** [3.484] 106,077 0.426

4.605 [1.464] 0.457 [0.276] 1.077 [0.754] 0.091 [0.045] 0.639 [0.361] 12,557 0.559

0.237*** [124.159] 0.000 [0.083] 0.010*** [4.510] 0.036*** [16.496] 0.085*** [32.865] 105,855 0.534

0.174*** [26.052] 0.011** [2.243] 0.011* [1.815] 0.009 [1.140] 0.019** [2.183] 12,541 0.780

P1 P2 GSE  P1 GSE  P2 Observations R-squared

This table reports the OLS regression results of appraisal bias and origination LTV over the HVCC time periods using the bias analysis sample. The sample size for origination LTV is slightly smaller than appraisal bias due to missing origination LTV for some loans. Appraisal bias is measured as the percentage that the appraisal value is above the transaction price for the same property, each adjusted by the HPI at the time of transaction for the zip code where the property is located. Variable GSE is an indicator variable for a loan being a GSE loan versus a non-GSE loan. HVCC indicators P1 and P2 refer to the periods of May–August 2009 and September–December 2009, respectively. All the regressions are controlled for borrower and loan characteristics as listed in Panel B of Table 1, as well as lender and zip-code level fixed effect. Standard errors are clustered at the zip-code level. T-values are in brackets. *** Indicates statistical significance level at 0.01 level. ** At 0.05 level. * At 0.10 level.

evidence suggest that GSE refinance loans experienced a large reduction in appraisal bias after the HVCC implementation. If bias is reduced due to the HVCC, the appraised value would be more conservative, rendering the reported LTV higher. This would have two effects. First, it would reduce

the approval rate of an originally inflated loan application, especially one with true LTV above the decision threshold, which we discuss in Section 6.3 on denial analysis. Second, for those loan applications where the ‘‘correct’’ LTV is still below 80% and thus were approved, the ‘‘correct’’ LTV would be higher than its LTV before since its original appraisal price was inflated. Therefore, a prediction is that LTV would increase for originated loans. Results on LTV are listed in Columns 3 and 4 of Table 3. We find that for refinance loans, GSE loans have an increase of LTV in post-HVCC periods more than non-GSE loans, consistent with our hypothesis. The coefficient on GSE  P1 is 0.036, suggesting that the mean LTV of GSE refinance loans increased an extra 0.036 than that of non-GSE refinance loans after the HVCC implementation. Considering that the mean of LTV is close to 0.80, the magnitude is noticeable. Meanwhile, for purchase loans, GSE loans have a decrease in LTV in post-HVCC periods more than non-GSE loans. Taken together, this is further evidence that appraisal bias decreased, thus correcting the LTV to a higher level while still qualifying the loans for origination. 6.2.2. Subsample analysis of appraisal bias We conduct a subsample analysis where we examine whether states that had higher appraisal bias for GSE refinance loans before the HVCC experienced a larger reduction in bias after its implementation. The idea is that the close-to-exogenous shock from the HVCC would have a greater effect in states where lender-appraiser independence was the weakest. Results are in Panel A of Table 4. We find that states with higher bias prior to the HVCC experienced a larger reduction in bias after the HVCC implementation. The other subsample analysis we conduct exploits the idea that the potential for appraisal bias is larger in areas where it is harder to appraise. One such case is when the heterogeneity of houses in an area is greater. That is, we

87

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90 Table 4 Subsample analysis of the HVCC impact on appraisal bias.

Table 5 Appraisal bias and loan default.

Dependent variable

Appraisal bias

Dependent variable

Default

Sample

Refinance (1)

Purchase (2)

Sample

Refinance (1)

Purchase (2)

Appraisal bias 1.567** [2.342] 2.023*** [2.887] 2.987*** [3.827] 1.304 [1.556] 3.116*** [3.475] 61,373 0.444

6.233* [1.846] 5.060 [1.099] 2.238 [0.637] 8.925* [1.739] 6.875 [1.561] 7,697 0.536

0.008*** [2.912] 2.322*** [6.444] 1.140*** [3.067] 3.937*** [8.043] 0.841** [2.258] 3.256*** [6.573] 106,077 0.212

0.012 [1.415] 0.352 [0.268] 1.020 [0.867] 0.234 [0.179] 1.089 [0.884] 0.144 [0.110] 12,557 0.506

1.383*** [2.870] 0.232 [0.436] 0.640 [1.072] 0.162 [0.271] 1.410** [2.100] 105,887 0.424

3.130 [1.093] 0.134 [0.036] 1.648 [0.603] 3.232 [0.771] 0.127 [0.036] 12,538 0.559

Panel A: By state’s appraisal bias level GSE  T15 P1  T15 P2  T15 GSE  P1  T15 GSE  P2  T15 Observations R-squared Panel B: By heterogeneity of local housing market GSE  H P1  H P2  H GSE  P1  H GSE  P2  H Observations R-squared

This table reports the OLS regression results of appraisal bias and origination LTV over the HVCC time periods using a subset of the bias analysis sample. Panel A uses a sample of states with high versus low appraisal bias. The indicator variable T15 takes the value of 1/0 for the 15 states with the highest/lowest appraisal bias for GSE refinance loans before the HVCC. The indicator H in Panel B equals 1/0 if the local housing market is considered to be of high/low heterogeneity. The heterogeneity is measured by the variation coefficient of origination property appraised value for the zip code where the property is located, and is considered as high/ low if the zip-code level variation coefficient is in the top/bottom 50th percentile. Observations from the zip codes where transactions are too limited to calculate coefficient of variation are omitted. Appraisal bias is measured as the percentage that the appraisal value is above the transaction price for the same property, each adjusted by the HPI at the time of transaction for the zip code where the property is located. Variable GSE is an indicator variable for a loan being a GSE loan versus a non-GSE loan. HVCC indicators P1 and P2 refer to the periods of May–August 2009 and September–December 2009, respectively. All the regressions are controlled for borrower and loan characteristics as listed in Panel B of Table 1, as well as lender and zip-code level fixed effect. Standard errors are clustered at the zip level. T-values are in brackets. *** Indicates statistical significance level at 0.01 level. ** At 0.05 level. * At 0.10 level.

expect that the appraisal bias is larger in areas with more heterogeneous houses and consequently, the reduction in appraisal bias is larger for these areas after the HVCC implementation. To test this hypothesis, we use the coefficient of variation of house appraisal value in a zip code to measure the heterogeneity of houses within that area.25 We create an indicator variable ‘‘H’’ that takes the value of 1 if the coefficient of variation of a zip code is at or above 25 It is admittedly a crude measure and one that we are able to measure using our data.

GSE P1 P2 GSE  P1 GSE  P2 Observations R-squared

This table reports OLS regression of loan default as a function of appraisal bias using the bias analysis sample. Default is an indicator variable for the loan being in foreclosure (including presale, post-sale, REO) or 60 + DPD within two years after origination. Appraisal bias is measured as the percentage that appraisal value is above the transaction price for the same property, each adjusted by the HPI at the time of transaction for the zip code where the property is located. Variable GSE is an indicator variable for a loan being a GSE loan versus a non-GSE loan. HVCC indicators P1 and P2 refer to the periods of May–August 2009 and September–December 2009, respectively. All the regressions are controlled for borrower and loan characteristics as listed in Panel B of Table 1, as well as lender and zip-code level fixed effect. Standard errors are clustered at the zip-code level. T-values are in brackets. *** Indicates statistical significance level at 0.01 level. ** At 0.05 level. * At 0.10 level.

the sample median. We create the interaction variables of this variable with GSE, P1, P2, GSE  P1, and GSE  P2. Regression results are in Panel B of Table 4. We find that consistent with our hypothesis, in periods after the HVCC, especially in P2, the reduction in appraisal bias for GSE refinance loans is larger for houses in neighborhoods with a higher heterogeneity. These findings for refinance loans are absent for purchase loans, confirming that the findings are unique to GSE refinance loans. We interpret these findings as providing additional support to our hypothesis. 6.2.3. Relation between loan performance and appraisal bias Extant literature shows that default increases with the existence of appraisal bias (Kelly, 2006; LaCour-Little and Malpezzi, 2003). We replicate their findings in Table 5 where the dependent variable is loan default, and the explanatory variables include appraisal bias in addition to those used in previous performance analysis in Table 2 (investor, HVCC periods, investor and HVCC interactions, and loan and borrower characteristics). We find that a higher bias is positively associated with a higher default rate. The coefficient of 0.008 suggests that an increase in bias of 10 percentage points is associated with an increase in the default rate of 0.08 percentage points (0.008  10), which is about a 7.2% increase from the mean value of the default rate for refinance loans (1.10 percentage points). Results in Column 2 are for purchase loans in 2009 that have a matching purchase transaction in

88

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Table 6 Loan denial decision and the HVCC. Dependent variable

All denial

Denial due to collateral reason

Denial due to insufficient cash

Sample

Refinance (1)

Purchase (2)

Refi + Pur (3)

Refinance (4)

Purchase (5)

Refi + Pur (6)

Refinance (7)

Purchase (8)

Refi + Pur (9)

GSE

7.966*** [18.792]

0.195 [0.086]

2.851*** [18.514]

0.832*** [13.737]

0.336*** [2.637]

2.204*** [3.654] 3.040*** [2.802] 0.358 [0.670] 0.332 [0.390]

0.027 [0.167] 0.081 [0.333] 0.150 [0.976] 0.228 [1.551]

0.137*** [4.190] 0.237*** [6.433] 0.438*** [7.513] 0.395*** [5.919]

0.064*** [3.621] 0.185*** [9.224] 0.112*** [5.291] 0.039 [1.572]

0.036 [0.601] 0.109*** [7.034] 0.202** [2.538] 0.196*** [5.547]

7,328,900 0.147

3,002,370 0.078

7,328,900 0.052

3,002,370 0.014

1.417*** [14.136] 8.466*** [23.803] 0.034 [0.762] 0.073 [1.228] 0.216*** [3.372] 0.041 [0.665] 0.261 [1.268] 0.162 [0.559] 4.182*** [22.109] 0.463*** [2.843] 0.126 [0.766] 10,331,270 0.046

0.481*** [21.621]

0.618** [2.125] 1.599*** [5.024] 0.201 [0.755] 0.490** [2.264]

1.619 [0.657] 16.439*** [5.320] 2.042*** [3.500] 2.284*** [2.683] 0.366 [0.657] 0.687 [1.085] 1.249* [1.837] 3.732*** [3.612] 10.431*** [4.388] 0.360 [0.559] 0.478 [0.691] 10,331,270 0.112

7,328,900 0.008

3,002,370 0.007

0.290* [1.937] 0.071 [0.326] 0.069 [1.072] 0.132*** [7.547] 0.184** [2.319] 0.179*** [4.715] 0.008 [0.110] 0.305*** [11.778] 0.735*** [4.901] 0.305*** [3.161] 0.148*** [3.476] 10,331,270 0.007

Refi P1 P2 GSE  P1 GSE  P2 Refi  P1 Refi  P2 GSE  Refi GSE  Refi  P1 GSE  Refi  P2 Observations R-squared

This table reports the OLS regression results of the denial analysis. Denial refers to the loan application being denied (HMDA action taken = 3 or 7). Collateral denial refers to the loan application being denied due to collateral-related reasons. Cash denial refers to the loan application being denied due to cashrelated reasons. Variable GSE is an indicator variable for a loan being a GSE vs. non-GSE loan. HVCC indicators P1 and P2 refer to the period of May–August 2009, and September–December 2009, respectively. All the regressions are controlled for borrower and loan characteristics as listed in Panel C of Table 1, as well as lender and MSA level fixed effects. Standard errors are clustered at the MSA level. T-values are in brackets. *** Indicates statistical significance level at 0.01 level. ** At 0.05 level. * At 0.10 level.

2004–2009 that involved the same property. It is reaffirming to see that there is no statistically significant relation between loan defaults and appraisal bias for purchase loans. Consistent with the results on the performance sample, based on this bias sample, the default rate of GSE refinance loans decreases more than that of non-GSE refinance loans, while there is no difference in the defaults between GSE and non-GSE purchase loans. This evidence supports the interpretation that appraisal inflation leads to artificially lower LTV, which results in either qualifying loans that should be denied or reducing the need for compensating factors for ‘‘realistically’’ high LTV loans,26 and therefore generating worse performance.27 26 Examples of compensating factors that could have been required include higher FICO or lower DTI. 27 We also conducted an analysis on the interest rate using the same specifications as in the baseline analyses and did not find that interest rates for GSE refinance loans decreased more than those for other loans after the HVCC implementation (despite their larger reduction in default rates). Agarwal et al. (forthcoming) use different specifications and find some evidence that the appraisal inflation is somewhat priced. We suspect that our finding on the interest rates could be due to the fact that at least Fannie Mae, in December 2008, raised its rates for loans that it purchased after April 1, 2009, possibly due to the increase in default rates for loans amid the height of the credit crisis (https://www.fanniemae.com/content/announcement/0838.pdf).

6.3. Denial analysis We use whether a loan application is denied to measure the stringency of loan origination; the idea is that, given the same characteristics, the fact that a loan is denied suggests that the lender uses more information in making the decision. Our hypothesis is that after the HVCC implementation, appraisal inflation becomes smaller and LTV becomes higher or more equity is required for the refinance; therefore, more GSE refinance loans will be denied than other types of loans following the HVCC. The regression results for denial analysis are in Table 6. We find that there exists little pattern for the overall denial. However, for the denials due to collateral or insufficient down payment, the denial rate for refinance loans after the HVCC implementation is stagnant while that for purchase loans after the HVCC is steadily declining. Table 6 further shows the triple difference, i.e., DID results for refinance loans compared to DID for purchase loans. We find that the coefficients on GSE  Refi  HVCC are positive, confirming that the denial rate for GSE refinance loans increased more than that for other loans after the HVCC. The magnitude of the estimated coefficient is not negligible. For example, in the DDD specification for denial due to a collateral reason, the coefficient on GSE  Refi  P1 is 0.463. Considering the mean of the dependent variable is

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

89

Fig. A1. The performance of HARP versus non-HARP loans. This figure plots the default rate for HARP versus non-HARP loans by the HVCC for GSE refinance loans. Default is measured as the percentage of loans being in foreclosure (including presale, post-sale, REO) or 60 + DPD within two years after origination.

around 7%, this is a 6.6% increase from the mean. This is consistent with an interpretation that more accurate appraisal value, as a result of improved incentives for appraisers, leads to better screening of loans. 7. Conclusion

Office of the Comptroller of the Currency or the U.S. Department of the Treasury. The authors would like to thank Chau Do for her comments, the audience at the OCC Risk Analysis Division seminar and the FSU-OCC conference for their helpful feedback, and R. Kelley Pace for his discussion of our paper at the FSU-OCC conference. The authors take responsibility for any errors.

This paper exploits a close-to-exogenous GSE intervention on enhancing appraiser independence to highlight the existence of appraisal inflation in refinance loans and its impact on loan performance. We find that the requirement of using independent appraisers appears to have tightened underwriting standards and raised loan origination quality of GSE refinance loans. Using the DID and DDD specifications, we show that GSE refinance loans performed better than other loans after the May 2009 intervention. We further show that the improved performance is due to lower appraisal bias and thus more accurate LTV, and that the denial rate increased more for GSE refinance loans than that for others due to more accurate appraisals. This analysis helps us to assess the existence and impact of appraisal inflation before the intervention. The effectiveness of this new code is yet another piece of evidence of the existence of problems caused by information asymmetry in the mortgage securitization process. This paper uses the HVCC intervention as a way to identify the presence of appraisal inflation. As a by-product, our paper also shows the effect of the HVCC in reducing appraisal inflation and raising loan origination quality. The 2010 Dodd-Frank Act kept most of the HVCC tenets in enacting a new set of comprehensive regulations regarding appraisal practices. Our finding that the HVCC improved loan underwriting standards by reducing appraisal bias provides support for this regulatory move.

HARP was set up in March 2009 to help borrowers who had LTV above 80% and as a result had difficulty refinancing, missing out on the benefits of a lower mortgage rate. Only GSE loans originated before May 2009 were eligible and they shall not have defaulted and have current LTV above 80%. One concern is that GSE refinance loans after the HVCC intervention, i.e., after May 2009, were impacted by the implementation of HARP as well as by the HVCC, potentially causing an inconsistent estimate of the effect of the HVCC. To evaluate this potentially valid concern, we search the GSE refinance loans and identify those that were HARP loans. The MM data provide a data field that directly identifies HARP loans. There are 22,538 loans, or 3.7% of all GSE refinance loans, that are HARP loans. Fig. A1 plots the performance, showing that HARP loans have higher default rates than non-HARP loans. This is not surprising, given that they are high-LTV loans. What is more relevant is that HARP loans originated after May 2009 had a larger increase in default rates relative to HARP loans originated before May 2009 than non-HARP loans for the same two time periods. This suggests that the existence of HARP in fact biases us in finding an HVCC effect for GSE refinance loans; we thus drop HARP loans from our analyses.

Acknowledgments

References

The views expressed in this paper are those of the authors alone and do not necessarily reflect those of the

Acemoglu, Daron, Angrist, Joshua, 2000. How large are human-capital externalities? Evidence from compulsory schooling laws. NBER Macroecon. Annu. 15, 9–59.

Appendix A. The impact of HARP on our analysis

90

L. Shi, Y. Zhang / Journal of Housing Economics 27 (2015) 71–90

Agarwal, Sumit, Itzhak Ben-David, Vincent Yao, forthcoming. Collateral valuation and borrower financial constraints: evidence from the residential real-estate market. Manag. Sci., . Ashcraft, Adam B., Schuermann, Til., 2008. Understanding the securitization of subprime mortgage credit. Found. Trends Finance 2 (3), 191–309. Besley, Timothy, Burgess, Robin, 2004. Can labor regulation hinder economic performance? Evidence from India. Q. J. Econ. 119 (1), 91–134. Cho, Man, Megbolugbe, Isaac F., 1996. An empirical analysis of property appraisal and mortgage redlining. J. Real Estate Finance Econ. 13, 45– 55. Dell’Ariccia, Giovanni, Igan, Deniz, Laeven, Luc, 2012. Credit booms and lending standards: evidence from the subprime mortgage market. J. Money Credit Bank. 44 (2–3), 367–384. Demyanyk, Yulia, Van Hemert, Otto, 2011. Understanding the subprime mortgage crisis. Rev. Financ. Stud. 24 (6), 1848–1880. Elul, Ronel, Souleles, Nicholas S., Chomsisengphet, Souphala, Glennon, Dennis, Hunt, Robert, 2010. What ‘triggers’ mortgage default? Am. Econ. Rev. Pap. Proc. 100 (2), 490–494. Foote, Christopher L., Gerardi, Kristopher, Willen, Paul S., 2008. Negative equity and foreclosure: theory and evidence. J. Urban Econ. 64 (2), 234–245. Heckman, James, Hotz, Joseph V., 1989. Choosing among alternative nonexperimental methods for estimating the impact of social

programs: the case of manpower training. J. Am. Stat. Assoc. 84 (408), 862–874. Kelly, Austin, 2006. Appraisals, automated valuation models, and mortgage default. In: Federal Housing Finance Agency Working Paper, April 2006. Keys, Benjamin J., Mukherjee, Tanmoy, Seru, Amit, Vig, Vikrant, 2010. Did securitization lead to lax screening? Evidence from subprime loans. Quart. J. Econ. 125 (1), 307–362. LaCour-Little, Michael, Malpezzi, Stephen, 2003. Appraisal quality and residential mortgage: evidence from Alaska. J. Real Estate Financ. Econ. 27 (2), 211–233. Nakamura, Leonard, 2010. How Much Is That Home Really Worth? Appraisal Bias and House-Price Uncertainty, Philadelphia Federal Reserve Bank, Business Review, First Quarter, 11–22. Piskorski, Tomasz, Seru, Amit, Vig, Vikrant, 2010. Securitization and distressed loan renegotiation: evidence from the subprime mortgage crisis. J. Financ. Econ. 97, 369–397. Piskorski, Tomasz, Amit Seru, James Witkin, 2013. Asset Quality Misrepresentation by Financial Intermediaries: Evidence from RMBS Market, Working Paper. Purnanandam, Amiyatosh K., 2011. Originate-to-distribute model and the sub-prime mortgage crisis. Rev. Financ. Stud. 24, 1881–1915. Zhu, Shuang, Kelley Pace, R., 2012. Distressed properties: valuation bias and accuracy. J. Real Estate Financ. Econ. 44, 153–166.