UNIVERSITY OF MICHIGAN

Transcription

1 UNIVERSITY OF MICHIGAN JOHN M. OLIN CENTER FOR LAW & ECONOMICS BANKRUPTCY AND THE MARKET FOR MORTGAGE AND HOME IMPROVEMENT LOANS EMILY Y. LIN AND MICHELLE J. WHITE PAPER #-13 THIS PAPER CAN BE DOWNLOADED WITHOUT CHARGE AT: THE SOCIAL SCIENCE RESEARCH NETWORK ELECTRONIC PAPER COLLECTION:

2 Bankruptcy and the Market or Mortgage and Home Improvement Loans Emily Y. Lin University o Connecticut and Michelle J. White University o Michigan, and RAND Corporation Corresponding Author: Pro. Michelle White Dept. o Economics University o Michigan Ann Arbor, CA micwhite@umich.edu November 2 Abstract: This paper investigates the relationship between bankruptcy exemptions and the availability o credit or mortgage and home improvement loans. We develop a combined model o debtors' decisions to ile or bankruptcy and to deault on their mortgages and show that the theory predicts positive relationships between both the homestead and personal property exemption levels and the probability o borrowers being denied mortgage (secured) and home improvement loans. We test these predictions empirically and ind strong and statistically signiicant support when evidence rom cross-state variation in bankruptcy exemption levels is used. Applicants or mortgages are 2 percentage points more likely to be turned down or mortgages and 5 percentage points more likely to be turned down or home improvement loans i they live in states with unlimited rather than low homestead exemptions. These relationships also hold when we introduce state ixed eects into the model. Proposed running head: Bankruptcy and Mortgage Loans

3 Bankruptcy and the Markets or Mortgage and Home Improvement Loans 1 1. Introduction Emily Y. Lin and Michelle J. White The US is extremely unusual in having very pro-debtor bankruptcy laws and, alone among the industrialized countries, it has a high and rapidly rising bankruptcy iling rate. The total number o bankruptcy ilings has risen rom under 3, per year in 1984 to 1.1 million in 1996 and 1.4 million in The rapid increase in bankruptcy ilings has led researchers to explore how bankruptcy aects consumers' access to credit. When debtors in the U.S. ile or personal bankruptcy, many types o debts are discharged, causing losses or creditors. Debtors who ile under Chapter 7 o the U.S. Bankruptcy Code are not obliged to use uture income to repay their debts and are only obliged to use wealth to repay debt to the extent that their wealth exceeds predetermined exemption levels. Exemption levels in bankruptcy are set by the state in which the debtor lives and they vary widely. They thereore may aect the terms on which loans are made. In a recent article, Gropp, Scholz and White [12] argued that in states with high rather than low bankruptcy exemptions, demand or credit rises because debtors' wealth is more sheltered rom the consequences o inancial distress and supply o credit alls because debtors are more likely to deault and ile or bankruptcy. They ound empirical support or these predictions. However, Jeremy Berkowitz and Richard Hynes [2] (hereinater B-H) recently argued that these conclusions do not hold up when the analysis distinguishes between dierent types o debts and dierent types o exemptions. Bankruptcy law distinguishes between secured versus unsecured debts. Secured (collateralized) debts such as mortgages and automobile loans--allow the creditor to reclaim the collateral i the debtor deaults on the loan, while unsecured debts--such as credit card and installment loans--have no collateral. Because secured creditors can oreclose on their collateral when debtors deault, regardless o whether debtors ile or bankruptcy, they are in a much stronger position than unsecured creditors to collect. States also provide separate exemptions or equity in owner-occupied homes (homestead exemptions) versus other types o property (personal property exemptions). B-H argue that higher homestead exemptions in bankruptcy help rather than harm secured creditors and thereore increase the supply o credit. Their argument is that, when debtors are in inancial distress, they can ile or bankruptcy, obtain discharge o their non-mortgage debts, and use the unds that would otherwise go to non-mortgage creditors to repay their mortgages and thereby keep their homes. The 1 We are grateul to Jan Brueckner, Masaru Konishi, Kenya Fujiwara and the reeree or very helpul comments. An earlier version o this paper was presented at the 36th Econometrics Conerence held at Lake Biwa, Japan, July Research support was provided by the Law and Social Science and Economics Programs at the National Science Foundation, under grant number SBR Bankruptcy iling data are taken rom the Administrative Oice o the U.S. Courts and rom the Statistical Abstract o the U.S., various editions.

4 higher the exemptions, the more that debtors' wealth is protected in bankruptcy and thereore the lower the probability that they will deault on their mortgages. Our goal in this paper is to re-examine the eect o bankruptcy on supply and demand or credit, distinguishing between secured versus unsecured credit and between homestead versus personal property exemptions. In section 2, we analyze debtors' decisions to deault on their mortgages and to ile or bankruptcy and the eect o both types o bankruptcy exemptions on the supply and demand or credit. Our theoretical model does not support B-H's claim that higher bankruptcy exemptions increase the supply o secured credit. Instead the model predicts that higher exemption levels cause lenders to tighten credit rationing, especially when exemption levels are already high. In sections 3 and 4, we test the model using data rom on both mortgage and home improvement loans rom the Home Mortgage Disclosure Act (HMDA) dataset. We ind that debtors are signiicantly more likely to be denied both home purchase and home improvement loans when they live in states with higher bankruptcy exemptions. We ind support or these relationships using both cross section variation in bankruptcy exemption levels and changes in bankruptcy exemptions over time. 2. Theory In this section we develop a model which integrates consumers' decisions to deault on their mortgages with their decisions to ile or bankruptcy. There is a large literature on mortgage deault and our model does not attempt to advance that literature. Instead we ocus on examining the interaction between the two decisions. 3 The model has two periods. In period 1, a representative consumer has an exogenously determined level o non-housing wealth w, which includes her period 1 income, her period 1 inancial wealth, and the value o any non-inancial wealth she owns other than the house that serves as her primary residence. The consumer buys a house o value v, inanced by a mortgage o amount M which is secured by the house. The consumer also borrows an amount P on an unsecured basis. By assumption, the consumer has only one unsecured loan. 4 The interest rates are r m on the mortgage and r p on the unsecured loan. In period 2, both loans come due, so that the consumer owes M' = M(1 + 3 Although the mortgage deault and bankruptcy decision literatures have developed separately, they are strikingly similar. Both decisions are \ruthless" in that the value o deaulting or iling or bankruptcy depends only on the value o particular assets or liabilities. Whether mortgage deault is worthwhile depends only on the value o housing equity and the value o non-housing wealth relative to the personal property exemption; whether bankruptcy is worthwhile depends only on the value o dischargeable debt compared to non-exempt assets. However, both decisions have been shown to be less ruthless in act than in theory, presumably because both have transactions costs and lead to credit impairment and loss o reputation. Quigley and Van Order [17] and Capozza et al [7] show that deault decisions are not ruthless because other household characteristics besides housing equity are signiicant determinants o deault; Fay, Hurst and White [1] show that bankruptcy iling decisions are not ruthless because demographic characteristics and income are signiicant determinants o the bankruptcy decision. Both decisions have also been modelled assuming that borrowers have private inormation concerning their propensities to deault/ile or bankruptcy. See Brueckner [4] and [5] or asymmetric inormation models o the mortgage deault decision and Wang and White [2] and Adler, Polak, and Schwartz [1] or asymmetric inormation models o the bankruptcy decision. 4 See Bizer and De Marzo [3] or discussion o how creditors' incentives are aected by whether debtors have prior loans.

5 r m ) on the mortgage and P' = P(1 + r p ) on the unsecured loan. The consumer's nonhousing wealth in period 2, including her period 2 income, is denoted W. W is uncertain and has the known density unction (W). The value o the debtor's house in period 2 is V. V is also uncertain and has the known distribution unction g(v ). (W) and g(v ) may be either independent or correlated. 5 2a. Borrowers' behavior. Ater learning W and V at the beginning o period 2, the consumer decides whether to deault on the unsecured loan and whether to deault on the mortgage. I the consumer deaults on the unsecured loan, then she is assumed to ile or bankruptcy. When consumers ile or bankruptcy, their unsecured debt is discharged, but their mortgage debt is not. 6 The obligation to repay debt in bankruptcy is determined by the bankruptcy exemption in the consumer's state o residence. Each state has two types o exemptions, one or equity in housing -- denoted E h -- and one or all other types o property -- denoted E p. 7 The consumer must give up non-housing wealth in excess o E p and this amount is used to repay unsecured creditors. Also, the consumer must give up her house in bankruptcy i her home equity exceeds the homestead exemption E h. I home equity is less than the homestead exemption and i the consumer has not deaulted on the mortgage, then the consumer can keep her house when she iles or bankruptcy. Consumers' out-o-pocket costs o iling or bankruptcy -- including iling ees and lawyers' ees -- are denoted C b. We assume that C b must be paid beore iling or bankruptcy -- or obvious reasons, bankruptcy lawyers don't oer credit! I the consumer deaults on her mortgage and does not cure the deault by repaying the amount owed within several months, then the mortgage lender will oreclose on the house. When oreclosure occurs, the consumer must relocate to a new residence at a cost o R. Ater oreclosure, the mortgage lender sells the house or an amount V - C, where C is the transactions cost o oreclosure. 8 The proceeds o selling the house are used, irst, to pay the costs o oreclosure C, and, second, to repay the mortgage and then the second mortgage -- i any -- in ull. (In the empirical work, we examine home improvement loans, which oten take the orm o second mortgages.) Third, an amount up to the homestead exemption E h is returned to the consumer. Fourth, the remainder 5 Whether (W) and g(v ) are independent or not aects the probabilities that the various cases discussed below occur. Using data rom the PSID Wealth Supplements or 1984 and 1989, the correlation between (W) and g(v ) is only This discussion assumes that consumers ile or bankruptcy under Chapter 7 o the U.S. Bankruptcy Code. About 7% o bankruptcy ilings in the U.S. are under Chapter 7. Consumers who ile under Chapter 13 keep all their assets but must propose a plan to repay part o their debts rom uture income. Since consumers have the right to choose between the two Chapters, they cannot be orced to repay more under Chapter 13 than under Chapter 7. We thereore assume that repayment under Chapter 13 would be about the same as under Chapter 7 and we ignore Chapter 13 in the discussion that ollows. See White [21] or discussion. 7 States normally adopt separate exemptions or several types o personal property, such as cash, equity in vehicles, clothing, and household eects. E p equals the sum o these exemptions. 8 C includes the cost o additional wear and tear by owners who anticipate deaulting on their mortgages.

6 is used to repay unsecured debt. In most oreclosures, the proceeds o selling the house are less than the amount owed on the mortgage, or V - C < M'. For simplicity, we assume that mortgage lenders do not have the right to collect deiciency judgments rom debtors, so that they lose whatever portion o the mortgage is not covered by the proceeds o oreclosure. 9 When the consumer both deaults on the mortgage and iles or bankruptcy, the mortgage lender is assumed to incur additional transactions costs o D: We expect that D is positive, since oreclosure in the context o bankruptcy requires approval o the bankruptcy trustee and is thereore likely to be delayed. However D could alternately be zero, i bankruptcy has no eect on the speed or certainty o oreclosure. Our empirical work provides evidence concerning D. Since both V and W are uncertain, there are a range o possible cases, corresponding to varying levels o V and W. We divide the distribution o housing values into our regions and discuss them in order o lowest to highest. Case (A). V <M' - R. Here, housing value is so low that home equity V - M' is more negative than the cost o relocation, - R. First consider the consumer's decision to deault on the mortgage. I she does not ile or bankruptcy and does not deault, then her total net housing plus non-housing wealth will be W - P' + V - M'. I she does not ile or bankruptcy but deaults, then her total net wealth will be W - P' - R. Since V <M' - R, the debtor has an incentive to deault. Now suppose the consumer does not deault but iles or bankruptcy. Then her total net wealth will be either V - M' + W - C b or V - M' + E p, whichever is lower. Finally i the consumer iles or bankruptcy and deaults, then her total net wealth will be either - R + W - C b or - R + E p, whichever is lower. Since V <M' - R, the consumer again has an incentive to deault. Thereore the consumer's decision to deault on the mortgage is independent o her non-housing wealth W. 1 Now consider the consumer's bankruptcy decision, assuming that she deaults on the mortgage. There is a threshold level o non-housing wealth in case (A), denoted where the consumer is indierent between ling versus not iling or bankruptcy. I W is just below W A, then she iles or bankruptcy and her non-housing wealth becomes E p. I W > W A, then she does not ile or bankruptcy and her non-housing wealth becomes W - P. This means that 9 It is straightorward to modiy the analysis o cases (A) and (B) below to allow or deiciency judgments. Note that the oreclosure cost C is borne by whatever party is the last to be paid. 1 This model o the deault decision is obviously simpliied. With more time periods, the borrower might choose not to deault on the mortgage even i V <M' - R, i she expected housing values to rise in the uture.

7 W A = E p + P. The consumer iles or bankruptcy i W - W A and avoids bankruptcy otherwise. Figure 1 shows non-housing wealth W on the horizontal axis and total net housing plus non-housing wealth ater debt repayment, deault, and/or bankruptcy, on the vertical axis. The curve labelled case (A) has three segments. In the let-most segment, consumers ile or bankruptcy and, because W - E p + C b, their total net wealth is W - C b - R. In the middle ( at) segment, E p + C b < W- W A. Consumers ile or bankruptcy and their total net wealth is E p - R. In the right-most segment, consumers avoid bankruptcy and their total net wealth is W - P - R. Table 1 shows repayment to mortgage and unsecured creditors in case (A). Because consumers always deault on their mortgages, mortgage lenders receive either V - C or V - C - D, depending on whether consumers ile or bankruptcy. Unsecured creditors receive ull repayment op i W - W A, partial repayment ow - E p - C b i E p + C b - W < W A, and nothing i W <E p + C b. Case (B). - R - V - M - E h + C and E p - M. Here, housing value is high enough that home equity exceeds the cost o relocation. The consumer would thereore like to repay the mortgage and retain her house. But depending on the realization o W and the personal property exemption level, she may or may not have enough wealth to do so. B-H pointed out that iling or bankruptcy increases the consumer's wealth by reducing the amount thatshemust repay to unsecured creditors and this positive wealth eect increases the consumer's ability to repay the mortgage. To explore this possibility, we assume that consumers use up to 1% o their non-housing wealth to repay their 1 6.mortgages. We also assume that the personal property exemption exceeds the mortgage debt, or E p - M. This assumption is unrealistic in that most states' personal property exemptions are much smaller than a typical mortgage loan. However the amount past due on a mortgage loan may be much smaller than the principle amount and thereore less

8 than E p. This assumption also allows us to examine the B-H claim in the most avorable circumstances. 11 First consider the consumer's bankruptcy decision, assuming that she does not deault on the mortgage. Suppose W B is now the threshold level o non-housing wealth in case (B) such that the consumer is indierent between iling versus not iling or bankruptcy. She keeps total net wealth o E p +V - M i she iles or bankruptcy and W - P +V - M i she does not. Thereore W B = E p+p, so that W B = W A. I consumers' non-housing wealth is below W B, then they ile or bankruptcy and, i possible, repay their mortgages. However, i W turns out to be suiciently low, then consumers cannot repay their mortgages even ater iling or bankruptcy and devoting all o their non-housing wealth to mortgage repayment. The minimum level o wealth at which consumers can just aord to repay their mortgages ater iling or bankruptcy is W = C b+m. Consumers thereore deault on their mortgages i W <C b + M. Table 1 shows that mortgage lenders receive M i W > W B, M - D i M

9 + C b - W - W B, and V - C - D i W <M + C b. Unsecured lenders receive P i W - W B, W - E p - C b i E p +C b - W < W B, and i W B < E p +C b. The line labelled case (B) in igure 1 shows that consumers have higher non-housing wealth in case (B) than case (A), but are equally likely to ile or bankruptcy. Case (B) thus illustrates the B/H hypothesis that consumers may ile or bankruptcy in order to have their unsecured debt discharged and then use the wealth gain to repay their mortgages. This case may be unlikely in practice, however, since personal property exemptions in most states are low. 12 Case (C). E h +C < V- M - E h +P +C. Here home equity exceeds the homestead 11 Data on personal property exemptions is given below. We assume again that deiciency judgments are prohibited. 12 I Ep <M, then it is straightorward to show that consumers would be less likely to ile or bankruptcy, but more likely to deault on their mortgages. 7.exemption plus the cost o oreclosure, but is less than the homestead exemption plus the unsecured debt. The consumer has relatively high housing equity, but may have large unsecured debts perhaps because she owns an unincorporated business. The unsecured creditor may thereore ind it worthwhile to oreclose on the house i the consumer deaults on the unsecured loan and iles or bankruptcy. 13 I the consumer iles or bankruptcy

10 and the unsecured creditor orecloses, the house is sold or V - C, the mortgage lender receives M and the consumer receives E h. The remaining amount, V - C - M - E h, goes to the unsecured creditor. Because the unsecured creditor orecloses on the house when bankruptcy occurs, the decisions to deault on the mortgage and ile or bankruptcy are tied: the consumer either deaults on both loans or repays both. I she deaults and iles or bankruptcy, her non-housing wealth is E h + E p - R and i she repays both loans, her non-housing wealth is W - P + V - M. The level o non-housing wealth at which she is indierent between these alternatives, denoted W C, is thereore W C = W A -V +M +E h - R. Since under reasonable parameter values, W C < W A, the debtor is less likely to ile or bankruptcy in case (C) than in cases (A) or (B). Figure 1 shows the debtor's net housing plus non-housing wealth in case (C) and table 1 shows repayment to lenders. Case (D). V - C - M > E h + P. Because V is high, the consumer never deaults on the mortgage and never iles or bankruptcy. I her non-housing wealth turns out to be too low to repay the unsecured loan, then she sells her house and uses the proceeds to repay both the mortgage and the unsecured loans in ull. 14 2b. Lenders' behavior. Both mortgage and unsecured lenders are assumed to receive loan applications rom many potential borrowers, all o whom are identical as o period 1. Both types o lenders are risk neutral, so they are willing to lend i expected repayment covers their ixed opportunity cost o unds. 13

11 See Steingold [18], pp. 21/ Note that i the homestead exemption is unlimited, as it is in seven states, then neither case (C) nor (D) can occur. 8.Now suppose one o the bankruptcy exemption levels increases and the probability o deault on one or both types o loans thereore rises. Do lenders respond by raising the interest rate or rationing credit or a combination o both? Most economists tend to assume that, when all inormation is common knowledge, lenders always respond to a change which causes the probabilityodeault to riseby raising the interest rate. Conversely they tend to assume that, when there is asymmetric inormation, lenders respond to a change which causes the probability o deault to rise by rationing credit. However, asymmetric inormation is not a necessary condition or credit rationing. In our context, suppose an increase E p or E h causes the deault rate on mortgage loans to rise. I lenders respond by raising mortgage interest rates, then borrowers repay more when they do not deault. But the rise in the interest rate causes an additional increase in the probability o deault above and beyond the increase caused by the rise in the exemption level. This means that raising the interest rate has both a positive and a negative eect on expected repayment. I the probability o deault is low, then the positive eect outweighs the negative eect and raising the interest rate thereore causes expected repayment to rise. But i the probability o deault is suiciently high, then the reverse holds and raising the interest rate causes expected repayment to all. In the latter situation, lenders would respond to the original change by tightening credit rationing rather than raising the interest rate. Thus any change that reduces expected repayment on loans may cause lenders either to raise the interest rate or to ration credit. More speciically, the model also implies that the relationship between the probability o credit rationing and the exemption level is non- linear the marginal eect o an increase in the exemption level increases as the exemption level rises As an example o credit rationing when all inormation is common knowledge, suppose there is no legal distinction between housing versus non-housing wealth or between mortgage loans versus unsecured loans. Suppose all borrowers borrow $1, in period 1 and their period 2 wealth is distributed normally with a mean o $2, and a standard deviation o $25. Also suppose the cost o bankruptcy is zero and lenders' opportunity cost o unds is.1 per period. I the bankruptcy exemption E is less than $9, lenders respond to an increase in E by raising the interest rate. When E - $9, lenders respond to an increase in E by rationing credit completely (reusing to lend). Thus in this example, credit rationing takes

12 an all-or-nothing orm. However i borrowers dier according to some commonly observed characteristic related to the probability o deault such as income then credit rationing will occur more gradually as E rises. This is because lenders cease lending to low income borrowersat a low level o E and cease lending to high income borrowers at a higher level o E. See Fan and White [9]. Stiglitz-Weiss [19] developed the original model o credit rationing under asymmetric inormation. In their model, lenders respond to a change which raises the deault rate by rationing some borrowers but lending to others. 9.Now consider how changes in the bankruptcy exemption levels E p and E h aect expected repayment on mortgage and unsecured loans. To start with, assume that the E =, i:e:; the cost o oreclosure is unaected by whether borrowers ile or bankruptcy. Under these assumptions, neither the homestead exemption nor the personal property exemption aects the terms on which mortgage lenders are willing to lend. Table 1 shows that mortgage lenders always receive partial repayment ov - C in case (A), and they always receive ull repayment om in cases (C) and (D). In case (B), they may receive either partial or ull repayment, but the condition under which partial versus ull repayment occurs (whether W is greater than or less than M + C b) is independent o the exemption levels. Thus neither a change in E h nor E p is predicted to aect the terms o mortgage loans. When debtors deault on their mortgages, mortgage lenders are repaid only rom the proceeds o oreclosure and their claims rank above the exemption. So repayment is independent othevalues o both E p and E h. Now suppose D >, but all the other assumptions remain the same. Because mortgage lenders pay costs o D when borrowers ile or bankruptcy, expected repayment now depends negatively on the probability o bankruptcy. When E p rises, the probability o bankruptcy rises in cases (A), (B) and (C), because W A, W B, and W C all depend positively on E p. Also, the probability o cases (A), (B) and (C) occurring remains the same. Thus a rise in E p causes lenders to tighten the terms o mortgage loans. Now consider a rise in E h. AriseinE h causes the probability o bankruptcy to increase in case (C), because W C depends positively on E h. In addition, a rise in E h makes case (C) more likely and case (D) less likely to occur, and also makes case (B) more likely and case (C) less likely to occur. Since the probability o bankruptcy is higher in case (C) than case (D) and

13 higher in case (B) than case (C), both changes reduce expected repayment to mortgage lenders because lenders are more likely to incur the additional delay costs D. Thus when costs D are positive, a rise in either E h or E p reduces mortgage lenders' expected return and is predicted to cause the terms o mortgage loans to tighten. This is because lenders bear additional costs when borrowers ile or bankruptcy, and increases in either exemption level make bankruptcy more likely. Even when borrowers use all available unds to avoid losing their homes, an increase in the either o the exemption levels is predicted to cause increased credit rationing on mortgage loans. 1.Now consider how changes in E h or E p aect expected repayment o unsecured loans. When E h rises, table 1 shows that expected repayment o unsecured loans alls in case (C), because repayment is lower when borrowers ile or bankruptcy and the probability o bankruptcy rises. Expected repayment remains unaected in all o the other cases. When E p rises, table 1 shows that expected repayment o unsecured loans alls in all cases except case (D), while the probability o being in case (D) remains the same. Thus the model predicts that increases in both E h and E p unambiguously reduce expected repayment o unsecured loans. These results re ect the act that lenders must collect rom debtors' housing or non-housing wealth, but their claims rank below the homestead and personal property exemptions. These results remain the same regardless o whether D is zero or positive. To sum up, the model predicts that an increase in either the homestead or the personal property exemption causes expected repayment o mortgage and unsecured loans to all, as long as delay costs in bankruptcy D are positive. These predictions hold even i we assume that borrowers ile or bankruptcy to obtain discharge o their unsecured debt and use the entire wealth gain to repay their mortgages. The model also predicts that the marginal eect o an increase in either exemption level on the probability that lenders ration credit (rather than increasing the interest rate) rises as the exemption level increases. We test these predictions in the next section, using evidence o credit rationing on mortgage and home improvement loan applications. 3. Data and Empirical Speciication We use the Home Mortgage Disclosure Act (HMDA) data or 1992 through This dataset contains inormation on a large number o applications or home purchase loans and home improvement loans. 16 Home purchase loans (mortgages) are always secured by the house and always have highest priority or repayment i the house is sold in a oreclosure. Home improvement loans may be used or home improvements or other expenditures and 16 See Canner and Passmore [6] or a general description. The data have mainly been used to analyze

14 discrimination in lending to low-income and minority households. See, or example, the articles in Yezer [22]. 11.may either be unsecured or secured as second mortgages. 17 The dataset does not indicate whether individual home improvement loan observations are secured or unsecured. Because home improvement loans rank below mortgages in priority, they are less completely secured, even i they take the orm o second mortgages. Lenders thereore are orced to rely more heavily on debtors' non-housing wealth or repayment. As a result, home improvement loans are a mixture o secured and unsecured, so they allow us to test the predictions o the model or unsecured loans. 18 Our sample o both types o loans consists o all loan applications rom the largest metropolitan area in each state. 19 We exclude applications that were withdrawn, closed or incompleteness, or purchased by an institution. For home purchase loan applications, we limit our sample to applications or loans used to purchase (rather than re-inance) owner-occupied homes. Because sample sizes are large, we draw a 5% random sample o home purchase loan applications and a 1% random sample o home improvement loan applications. 2 The HMDA data include inormation on whether the loan application was turned down and the applicant's state o residence, but only a ew additional variables. We include dummies or the applicant's sex, whether the applicant is Arican-American, and whether there was a co-applicant or the loan, as well as the applicant's income (in thousands) and income squared. As an indicator o high risk, we also include a dummy variable which 17 The legal deinition o home improvement loans requires either that the proceeds o the loan be used or repairs/improvements to the house or that the loan be secured by the house. (See 12 C.F.R. 23.) 18 B-H also used the HMDA data, but they analyzed only home purchase loans and they used data rom The HMDA home improvement data have not previously been analyzed. Note that some debtors have an incentive to make their home improvement loans secured, since doing so makes the interest payments deductable rom taxable income or borrowers who itemize on their tax returns. But only about hal o all homeowners itemize. 19

15 See below or a test o whether this sample is representative. 2 A problem with our sample is that less credit-worthy applicants who anticipate being denied loans are more likely to apply to make multiple applications to lenders. This means that less credit-worthy borrowers are over-represented in the dataset, which could aect the estimated coecients o the homestead exemption variables in our model o credit denial. Using a 5% or 1% random sample o the original HMDA dataset has the advantage o reducing this potential bias, because the probability o more than one application by the same borrower occurring in the random sample is lower than in the original HMDA dataset. Because there is no individual-speciic identiier in the dataset, we cannot correct this problem by weighting. 12.equals one i the loan amount applied or divided by the applicant's income is greater than three. 21 To augment the sparce demographic inormation in the HMDA data, we merge it with census tract (neighborhood) inormation rom the 199 U.S. Census o Population and Housing. This allows us to include the median house value in the census tract where the house is located, the percent o houses in the census tract which are single amily, and the percent o residents in the census tract who are Arican-American. We also include several state-level variables as controls or local macroeconomic conditions. These are the state's unemployment rate, the change in the average income level in the state since the previous year, and the aggregate bankruptcy iling rate per 1, population in the applicant's state o residence during the previous year. Creditors are assumed to use last year's bankruptcy iling rate in the state to predict uture deault rates by residents o the state. A higher bankruptcy iling rate last year causes creditors to raise their prediction o next year's deault probability andthus to reduce credit availability in the state. (See Fay, Hurst, and White [1] or discussion.) We also add inormation concerning the homestead and personal property exemptions in each consumer's state o residence. The personal property exemption is deined to be the sum o states' exemptions or personal property and equity in vehicles and their \wildcard" exemptions, which can be applied to any type o property. Most states have low personal property exemptions, but homestead exemptions vary widely, rom zero in two states to unlimited in eight states. 22 About one-third o the states also allow their residents to choose between a uniorm Federal bankruptcy exemption and the state's exemption. In these states, we assign the highest o the two exemption values. Many states also allow married

16 couples who ile or bankruptcy to take higher exemptions, usually double. Because the dataset does not include inormation concerning marital status but does indicate whether there was a co-applicant or the loan, we assume that co-applicants are actually married couples and we double (or otherwise raise) the exemptions or co-applicants who live in states that allow doubling. We divide the homestead exemption into our categories and 21 Omitting this variable rom the regressions does not change the results or the exemption variables. 22 There are sometimes other limitations on homestead exemptions, such as maximumacreage limitations. See White [21] and B-H [2] or discussion. 13.use dummy variables or each category above the lowest. This speciication allows us to test the theoretical hypothesis that the marginal eect o an increase in the exemption level rises as the exemption level increases. The median homestead exemption values in the three lowest categories are $7,5, $15,, and $75,. The highest category consists entirely o unlimited homestead exemptions. Because the personal property exemption has little variation across states, we use the dollar value. An additional dummy variable equals one i the relevant state allows home purchase and home improvement lenders to collect deiciency judgments against debtors who deault on their mortgages. This variable is predicted to be negatively related to credit denials. 23 Our basic speciication is a linear probability model explaining whether applicants or loans were turned down or home purchase (home improvement) loans as a unction o the exemption levels and other variables. 24 Two important econometric issues are (1) whether bankruptcy exemption levels might be endogenous and (2) whether is a long enough time period to allow us to test the model's predictions using only changes in exemption levels, rather than variations across states in exemption levels. Consider the endogeneity issue irst. The U.S. Congress adopted a new Bankruptcy Code in 1978 which speciied uniorm Federal bankruptcy exemptions that were applicable all over the U.S. But the Code allowed states to opt out o the uniorm exemptions by adopting their own exemptions and, within a ew years, all 5 states had done so. Since the early 198's, the pattern has been that only a ew states change their exemption levels each year and these changes mainly involve correcting nominal dollar amounts or in ation. From 1992 to 1997, states changed their homestead exemptions eleven times and changed their personal property exemptions ten times. Many o these changes were very small. In addition, the Federal bankruptcy exemption was raised in 1994 and this raised exemption

17 levels in six states that allow their residents to use the Federal exemption. See table 2. The 23 Only ive states, Caliornia, Minnesota, Montana, North Carolina and Washington, prohibit deiciency judgments. The data sources or the legal variables are Elias et al [8] and Leonard [15], page 7/6 and appendix. 24 The results are virtually identical using probit or logit rather than OLS. 14.act that the pattern o exemption levels was set well beore the time period o our study and has remained airly constant since then suggests that individual states' bankruptcy exemptions can be treated as exogenous in the model o credit denials. Now consider the issue o using levels versus changes in bankruptcy exemptions as explanatory variables. Gropp, Scholz, and White [12] used cross-section data rom 1983 to test whether dierences in exemption levels across states aect access to credit and they ound signiicant eects. However cross-section results are vulnerable to criticism because the exemption variables may be acting as proxies or non-bankruptcy variables at the state level which are omitted rom the regression. The usual response to this problem in the program evaluation literature has been to use pooled cross-section or panel data rather than single year cross-section data and to introduce both state and year ixed eects. B- H ollowed this approach in their study. But suppose counter-actually that no changes in bankruptcy exemption levels occurred between 1992 and In that case the bankruptcy exemption variables and the vector o state dummy variables would be perectly correlated and both could not be used simultaneously. 25 In this case, pooled cross-section data would tell us no more than cross-section data or a single year. Now suppose a ew states changed their exemption levels between 1992 and Then i we use pooled cross-section data and introduce state dummy variables into the estimation, the state dummies will capture the eect o variation across states in exemption levels, while the exemption variables themselves will capture only the eects o changes in exemption levels between 1992 and Since only a ew changes in exemption levels actually occurred over the period, this speciication makes it very unlikely that the exemption variables will be statistically signiicant. But, surprisingly, B-H ound a negative and signiicant relationship between the homestead exemption and the probability o credit denial. They also ound a positive but insigniicant relationship between the personal property exemption and the probability

18 o credit denial. We report results using both approaches In our model, the exemption variables would actually dier rom the state dummy variables depending on whether the state allows doubling o exemptions by married couples who ile or bankruptcy. 26 I we regress exemption levels or each state rom 1992 to 1997 on a constant term and state dummy variables using OLS, then the R 2 o the regression is.9889 or homestead exemptions and.9779 or personal property exemptions. These high R 2 values illustrate the diculty o relying exclusively on changes in exemption levels over time to establish whether bankruptcy exemptions aect credit denials Results Table 3 shows summary statistics. 27 The average rejection rate or home improvement loan applications is much higher than that or home purchase loans 3% versus 15%. Also applicants or home improvement loans have lower incomes and live in census tracts with lower housing values than applicants or home purchase loans. The higher rejection rate or home improvement loans is not surprising. Because home purchase loans carry lower interest rates, households have an incentive to use them to inance both their home purchases and their home improvements. Only less credit-worthy households are orced to use home improvement loans at all. (See Jones [13] and [14].) The median homestead exemption is about $3, and the average personal property exemption is about $5, in both samples. About 2% o applicants in both samples live in states with unlimited homestead exemptions. In table 4, we report the results o running a linear probability model explaining whether applicants were turned down or home purchase or home improvement loans. Year ixed eects are included, but not state ixed eects. t-statistics are reported in parentheses. 28 All o the exemption variables have positive signs, or both types o loans. In the home purchase loan model, the personal property exemption and the unlimited homestead exemption variables are both signiicant at the 1% level and, in the home improvement loan model, the third category and unlimited homestead exemption variables are both signiicant at the 1% level. The large and highly signiicant coecients or the unlimited homestead exemption dummies support the theoretical hypothesis that lenders are more likely to choose credit rationing rather than interest rate increases when exemption levels are high. The personal property exemption variable is just short o statistical signiicance at

19 the 1% level in the home improvement loan model. The coecient o the state bankruptcy iling rate in the previous year is also positive as expected and statistically signiicant at the 1% level in both regressions. 27 We useweights to adjust our sample to the number o mortgage applications by state in We use robust standard errors, which allow or dependence within state/year clusters. See Greene [11], pp Not that the results in table 4 are virtually the same i we use logit or probit rather than linear probability models. 16.Other results in table 4 are reasonable. Applicants with higher incomes are less likely to be turned down or loans, although the eect diminishes as income rises. Single applicants and Arican-American applicants are signiicantly more likely to be turned down or both types o loans. 29 Applicants who live in orplan to live in neighborhoods with more single amily homes or in neighborhoods with ewer Arican-American residents are signiicantly less likely to be turned down or loans o both types, but the eects are small. Applicants who live in states with higher unemployment rates are signiicantly more likely to be turned down or both types o loans. The dummy variable or high loan/income ratio is positive in both regressions and signiicant in the home purchase loan regression. Applicants who live in states that allow deiciency judgments are less likely to be turned down or both types o loans, but the relationship is only signiicant or home purchase loans. Table 5 gives predicted probabilities o being turned down or credit, evaluated at the mean values o the variables. I an applicant lives in a state with a homestead exemption o $7,5, the applicant's probability o being turned down or a home purchase loan is.141, but it rises to.151 i the applicant instead lives in a state with a homestead exemption o $75, and to.162 i the homestead exemption is unlimited, respectively. I an applicant with the same characteristics applies or a home improvement loan, then the probabilities o being turned down are.277 in a state with a homestead exemption o $7,5 and.331 in a state with a homestead exemption o $75, or unlimited. The act that the marginal eect o an increase in the homestead exemption on the probability o being turned down is larger or home improvement loans than or home purchase loans is consistent with applicants or home improvement loans being less credit-worthy based on unobserved characteristics (since we hold constant the eect o observed characteristics). I the same applicant lives in a state with a personal property exemption o $1,, then his/her probability o being turned down or a home purchase loan is.14, but it rises to.149 i the exemption increases to $1,. For home improvement loans, the corresponding

20 increase is smaller rom.298 to.32. Finally i the applicant lives in the state with the highest rather than the lowest average bankruptcy iling rate (Georgia versus Hawaii), then 29 Other studies have previously ound evidence that Arican-American households are more likely to be turned down or home purchase loans. See Yezer [22] and Munnell et al [16]. 17.the probability o being turned down or a home purchase loan is predicted to rise by 4.2 percentage points and the probability o being turned down or a home improvement loan is predicted to rise by nearly 8 percentage points. We wished to test or whether our dataset o loan applications rom residents o large metropolitan areas might be a biased sample o all home purchase and home improvement loan applications. To do so, we constructed a second dataset consisting o all loan applications in 1996 rom the smallest metropolitan area in each state or, or states that contain only one metropolitan area, all loan applications rom non-metropolitan areas. 3 We reran the model in table 4 using this dataset and the results are reported in table 6. The exemption variables again have positive signs, except or the second category homestead exemption in the home improvement loan regression. The unlimited homestead exemption variable and the personal property exemption variables are both statistically signiicant at the 1% level in the home purchase loan regression. Thus the results suggest that bankruptcy exemptions are important determinants o credit denial or residents o both large and small metropolitan areas. In table 7, we rerun the model rom table 4, both with and without state ixed eects. To reduce collinearity between the discrete homestead exemption categories and the state dummy variables, we use a continuous rather than discrete speciication or the homestead exemption. For states with unlimited homestead exemptions, the value o the continuous homestead exemption variable is coded at $4,. (The largest \limited" homestead exemption in the sample is $2, or singles and $4, or married couples iling or bankruptcy.) We also drop the deiciency judgment variable because it does not change over the time period. The results in table 7 show that, without state ixed eects, the continuous homestead exemption variable is positive and statistically signiicant at the 5% level in both regressions, but the unlimited homestead exemption dummy is negative and insigniicant. The negative sign or the unlimited homestead exemption dummy implies that the probability o credit denial is highest in states with high but not unlimited 3 The District o Columbia is excluded rom this dataset, since it is entirely in a single metropolitan area. Note that applicants rom small metropolitan areas and rural areas have a higher probability o being turned down or home purchase loans than applicants rom large metropolitan areas the igures are.255 or mortgages and.2651 or home improvement loans.

21 18.homestead exemptions. For example, the probability o credit denial rises by.14(2) = 2.8 percentage points when the exemption increases rom zero to $2,, but by only.14(4) = 1.83 percentage points when the exemption increases rom zero to unlimited. When state ixed eects are added, the continuous homestead exemption variable remains positive and signiicant at the 5% level in the home purchase loan regression and the unlimited homestead exemption dummy shits rom being negative and insigniicant to positive and signiicant at the 1% level in the home improvement loan regression. 31 These indings suggest that the positive relationship between the homestead exemption and the probability o credit denial holds even when we rely on evidence rom changes in exemption levels, rather than variation in exemption levels. For the personal property exemption, adding state ixed eects to the model makes the coecient become negative and insigniicant in the home purchase loan model, while it remains positive but insigniicant in the home improvement loan model. Because all o the changes in personal property exemption levels over the period are relatively small, there is little possibility o identiying a signiicant relationship between the personal property exemption and the probability o credit denial. Finally, in order to reproduce B-H's speciication as closely as possible, we reran the model in table 7, but dropped the unlimited homestead exemption dummy and coded unlimited exemptions as $1 million rather than $4,. Speciying a linear relationship between the homestead exemption and the probability o credit denial orces the marginal eect o increases in the homestead exemption to be constant over the entire range and, since the dollar value assigned to the unlimited homestead exemption is very high, the marginal eect necessarily is small. 32 The results or the exemption variables both with and without state ixed eects are shown in table 8. In this speciication, the homestead exemption is positive and statistically signiicant in both models when state ixed eects 31 Only one state, Minnesota, changed rom a limited to an unlimited homestead exemption and no states changed in the opposite direction. (See table 2.) Thereore identiication o the unlimited homestead exemption variable is based on the dierence between the probability o credit denial in Minnesota beore versus ater Other dierences between our speciication and B-H's include slightly dierent time periods, our use o data rom the largest metropolitan areas rather than a random sample rom the ull dataset, and our