Chapter 1 Personal Bankruptcy and Household Debt

MASSACHUSETS INSTRWE OF TECHNOLOGY Essays in Financial Economics MAY 15 2014 by LIBRARIES Felipe Severino

B.Sc., Pontificia Universidad Catolica de Chile, 2005 M.Sc., Pontificia Universidad Catolica de Chile, 2007

Submitted to the Alfred P. Sloan School of Management in partial fulfillment of the requirements for the degree of

Doctor of Philosophy

at the

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

June 2014

Signature redacted Author...... Alfred Sloa chool of Management May 2, 2014 Signature redacted

Certified by...... Antoinette Schoar

Michael Koerner '49 Professor of Entrepreneurial Finance Thesis Supervisor Signature redacted Accepted by...... Ezra Zuckerman Director, Sloan School of Management PhD Program 2 Essays in Financial Economics by Felipe Severino

Submitted to the Alfred P. Sloan School of Management on May 2, 2014, in partial fulfillment of the requirements for the degree of Doctor of Philosophy

Abstract

This thesis consists of three empirical essays in financial economics, examining the consequences of imperfect financial markets for households, small business and house prices. In the first chapter (co-authored with Meta Brown and Brandi Coates) we explore the effect of personal bankruptcy laws on household debt. Personal bankruptcy laws in the US, and many other countries, protect a fraction of an individual's assets from seizure by unsecured creditors in case of default. An increase in the level of bankruptcy protection diminishes the collateral value of assets, and can therefore reduce borrowers' access to credit. However, it might also increase the demand for credit especially from risk averse borrowers by improving risk-sharing. Using changes in the level of protection across US states and across time, we show that bankruptcy protection laws increase borrowers' holdings of unsecured credit, but leave secured debt -mortgage and auto loans- unchanged. At the same time we find an increase in the interest rate for unsecured credit, but not for other types of credit. The effect is predominantly driven by lower-income areas and regions with higher home ownership concentration, for which an increase in the level of protection explains between 10% and 30% of the growth in their credit card debt. Using detailed individual data, we find no measurable increase in delinquency rates of households in the subsequent three years. These results suggest that changes in bankruptcy protections did not reduce the aggregate level of household debt, but they might have affected the composition of borrowing. In the second chapter (co-authored with Manuel Adelino and Antoientte Schoar) we document the role of the collateral lending channel in small business employment and self-employment in the period before the financial crisis of 2008. Small businesses in areas with a bigger run up in prices experienced a stronger increase in employment than large firms in the same industries. This increase in small business employment was more pronounced in industries that need little startup capital and can be financed more easily using housing as collateral. The increase is not limited to the non-tradable sector and is also present in manufacturing industries, in particular in those that ship goods over long distances. This indicates that this channel is separate from the aggregate demand channel by which home equity based borrowing leads to higher demand and employment creation. In aggregate, the collat-

3 eral lending channel explains 15-25 % of employment variation. In the third chapter (co-authored with Manuel Adelino and Antoinette Schoar) we use exogenous changes in the conforming loan limit as an instrument for lower cost of financing, and show that cheaper credit significantly increases house prices. Houses that become eligible for financing with a conforming loan show an increase in value of 1.16 dollars per square foot (for an average price per square foot of 220 dollars). These coefficients are consistent with a local elasticity of house prices to interest rates that is lower than some previous studies proposed (below 10). In addition, loan to value ratios around the conforming loan limit deviate significantly from the common 80 percent norm, which confirms that it is an important factor in the financing choices of home buyers. In line with our interpretation, the results are stronger in the first half of our sample (1998-2001) when the conforming loan limit was more important, given that other forms of financing were less common and substantially more expensive. Results are also stronger in zip codes where personal income growth is low or declining, and in regions with lower elasticity of housing supply.

Thesis Supervisor: Antoinette Schoar Title: Michael Koerner '49 Professor of Entrepreneurial Finance

4 Acknowledgments

I always thought that writing the acknowledgments to my thesis was not going to be easy, because I received encouragement and support from so many people along the way. Even if they are not mentioned here, I am truly grateful to each of them. I am deeply indebted to Antoinette Schoar: she has been an outstanding mentor. Her advice, comments and support were always insightful; our many discussions and conversations largely shaped the way I now think about research and finance. She has always been there. Working with her and learning from her has been a true privilege. I am extremely grateful to Nittai Bergman and Andrey Malenko, who provided invaluable advice. They always pushed me to deepen my understanding and focus on the important things. I also want to thank Xavier Giroud for his constant support and willingness to help. I also benefited from discussion and guidance with Hui Chen, John Cox, Sharon Cayley, Raj Iyer, Leonid Kogan, Gustavo Manso, Jun Pan, Stephen Ross, Hillary Ross, Adrien Verdelhan and Jiang Wang. Thanks you all for your time and dedication to make me a better researcher. My research has benefited from working with many people; my conversations with Manuel Adelino helped me understand the way research works. I will also want to thank Meta Brown and the Federal Reserve Bank of New York for their generous support. I cannot fail to mention my undergrad professors that encouraged me to start this adventure, especially Jaime Casassus, Gonzalo Cortazar and Nicolas Majluf. I am also grateful to Patricio Agusti, for his support during my first undergrad years. I had the great pleasure of sharing my experience with an incredible group of friends. I can still remember the first years, crammed into in the study room trying to make sense of our problem sets. I am very grateful to Marco Di Maggio, Sebastian Di Tella, Juan Passadore, Vicent Pons, Yang Sun, Tyler Williams, Luis Zermeno and especially to Will Mullins thanks a lot for always being there. Their help and friendship are something that I will always remember with affection, and I hope it will continue in the future. I have always felt the love and support of my family. I want to thank my par- ents, Fernando Severino and Fresia Diaz, for always believing in me, and for their encouragement to always give the best of me: you taught me all that I know, and are a true inspiration. To my brother and sister, Fernando and Francisca, for many years of friendship, conversation and joy together. To my daughter, Ema, and my son, Mateo, for bringing that special and unique happiness to my life: when you smile nothing else matters, and I feel truly blessed to have you. Last, but certainly not least, I would like to thank my wife Daniela Agusti. She has been by my side every step of the way. Since the beginning you believed in me, and left everything that was important to you to start this adventure with me. These have been years of hard work, but also of wonderful experiences, but none of this would have been the same without you. You make me want to be a better man. Thank you for everything that you have done. For your unconditional support and love, I will be forever grateful.

5 To Daniela, Ema and Mateo.

... en la calle codo a codo somos mucho mas que dos ... "

(Mario Benedetti)

6 Contents

1 Personal Bankruptcy and Household Debt 13 1.1 Introduction ...... 13 1.2 Bankruptcy Procedure and Related Literature 19 1.2.1 Institutional Framework ...... 19 1.2.2 Related Literature ...... 22 1.3 Data and Summary Statistics ...... 24 1.3.1 Data Description ...... 24 1.3.2 Summary Statistics ...... 26 1.4 Empirical Hypothesis ...... 27 1.5 Empirical Strategy ...... 29 1.6 Results and discussion ...... 32 1.6.1 Bankruptcy Protection and Household Leverage and Interest R ates ...... 32 1.6.2 Robustness Test ...... 34 1.6.3 Magnitude of the effect . 35 1.6.4 Borrowers, Delinquency and Self-Employment . 35 1.7 Conclusion ...... 37 1.8 Bibliography ...... 38 1.9 Appendix A. Model of Effect of Bankruptcy Protection on Household B orrow ing ...... 42

2 House Prices, Collateral and Self-Employment 73 2.1 Introduction ...... 73 2.2 Data and Empirical Methodology ...... 77 2.2.1 Data Description ...... 77 2.2.2 Summary Statistics ...... 80 2.2.3 Empirical Model ...... 81 2.3 Empirical Results ...... 84 2.3.1 House Prices and Employment at Small Establishments .. .. 84 2.3.2 Sole Proprietorships ...... 90 2.3.3 Crisis Period (2007-2009) ...... 90 2.3.4 M igration ...... 91 2.3.5 Credit Conditions and Elasticity of Housing Supply ...... 91 2.4 C onclusion ...... 92 2.5 Bibliography ...... 93

7 2.6 Appendix. Calculating the magnitude of the collateral effect 105

3 Credit Supply and House Prices: Evidence from Mortgage Market Segmentation 115 3.1 Introduction ...... 115 3.2 The User Cost Model ...... 119 3.3 Data and Methodology ...... 120 3.3.1 Summary Statistics ...... 120 3.3.2 Hedonic Regression ...... 121 3.3.3 Empirical Approach ...... 122 3.4 Cost of Credit and House Prices ...... 128 3.4.1 Main Regression Results ...... 128 3.4.2 Credit Supply and Income ...... 129 3.4.3 Robustness and Refinements .... . 130 3.4.4 Economic Magnitude of the Effect .. 133 3.5 Conclusion ...... 135 3.6 Bibliography ...... 137 3.7 Appendix A. Robustness and Refinements - .dditional Tests 153 3.7.1 Restrict LTV Choices ...... 153 3.7.2 Different Bands ...... 153 3.7.3 Timing of the Control Group . ... 154 3.7.4 Pos-October Effect ...... 154 3.7.5 Value per Square Foot by ZIP Code Income 154 3.8 Appendix B. Data Manipulation . .. 155 3.8.1 Data Cleaning ...... 155 3.8.2 Variable Construction .. .. 157

8 List of Figures

1-1 Debt Growth and Bankruptcy Filings ...... 44 1-2 States that Changed their Level of Bankruptcy Protection ...... 45 1-3 Ilustration of Different Demand and Supply Responses ...... 46 1-4 Ilustration of a Solution of the Model ...... 47

3-1 Transaction-Loan Value Surface ...... 139 3-2 Borrower Composition for the Regression Sample ...... 140 3-3 Frequency of Transactions as Percentage of CLL Threshold ...... 141 3-4 Share of Unused Mortgage Applications ...... 142 3-5 Fraction of Transactions with a Second Lien Loan by Year ...... 160 3-6 Value per Square Foot by House Value and by ZIP Code Income . . . 161 3-7 Income as a Percentage of CLL Threshold ...... 162

9 10 List of Tables

1.1 Summary Statistics Data...... 48 1.2 Summary Statistics Protection Level ...... 49 1.3 Effect of Bankruptcy Protection on Debt. Credit Card Debt .. .. . 50 1.4 Effect of Bankruptcy Protection on Debt. Mortgage Debt ...... 51 1.5 Effect of Bankruptcy Protection on Debt. Auto Debt ...... 52 1.6 Determinants of Bankruptcy Protection Levels and Changes ... .. 53 1.7 Dynamics of the Change in Protection Levels on Credit Card Debt . 54 1.8 Local Business Conditions. Neighboring County-pairs across State Borders. Credit Card Debt ...... 55 1.9 Heterogeneous Treatment of Bankruptcy Protection on Credit Card Debt: Income and Home ownership ...... 56 1.10 Effect of Bankruptcy Protection on Interest Rates: Personal Unsecured Loans and Credit Cards ...... 57 1.11 Effect of Bankruptcy Protection on Interest Rates: Mortagage Credit 58 1.12 Effect of Bankruptcy Protection on Debt. Number of Credit Cards and E ntry ...... 59 1.13 Effect of Bankruptcy Protection on Credit Card Delinquency . .. .. 60 1.14 Effect of Bankruptcy Protection on Self-Employment ...... 61 1.15 Effect of Bankruptcy Protection on Credit Card Debt. Alternative Specifications ...... 62 1.16 Other Heterogeneous Treatment of Bankruptcy Protection. Credit C ard D ebt ...... 63 1.17 Determinants of Bankruptcy Protection Levels and Changes. Eventu- ally Treated ...... 64 1.18 Dynamics of the Change in Protection. Mortgage Debt ...... 65 1.19 Dynamics of the Change in Protection. Auto Debt ...... 66 1.20 Local Business Conditions. Neighboring County-pairs across State Borders. Mortgage Debt ...... 67 1.21 Local Business Conditions. Neighboring County-pairs across State Borders. Auto Debt ...... 68 1.22 Heterogeneous Treatment of Bankruptcy Protection: Income and Home- ownership. Mortgage Debt ...... 69 1.23 Heterogeneous Treatment of Bankruptcy Protection: Income and Home- ownership. Auto Debt ...... 70 1.24 Effect of Bankruptcy Protection on County Delinquency Proportions 71

11 1.25 Effect of Bankruptcy Protection on Debt After Bankruptcy Reform 2005 72

2.1 Sum m ary Statistics ...... 96 2.2 Employment Growth, Firm Size, and House Price Appreciation .. . 97 2.3 Employment Growth and House Prices: Excluding Construction, Non- Tradable, and Finance Industries and Considering Manufacturing Only 98 2.4 Breakdown of Manufacturing Industries by Distance Shipped .. ... 99 2.5 Employment and House Price Appreciation across Industry Types . . 100 2.6 Proprietorships and House Price Appreciation ...... 101 2.7 Employment Growth, Firm Size, and House Price Appreciation, Crisis Period (2007-2009) ...... 102 2.8 Total Employment, Unemployment, and Migration ...... 103 2.9 D enial R ates ...... 104 2.10 Employment Growth, Firm Size, and House Price Appreciation: Indi- vidual Industries by Firm Size ...... 107 2.11 Robustness Test: Difference between High and Low Start-up Capital 108 2.12 Effect of One Standard Deviation Change in the Independent Variable 109 2.13 Dollar-weighted Average Distance Shipped in Manufacturing (miles) . 110 2.14 Detail on Average Start-up Amount by 2-digit NAICS Sector . ... . 111 2.15 Distance Shipped and Share of Employees at Large Establishments . 112 2.16 House Price Growth and Creation of Establishments ...... 113 2.17 List of 3-digit NAICS Industries Excluding Non-tradables, Manufac- turing, F.I.R.E., and Construction ...... 114

3.1 Sum m ary Statistics ...... 143 3.2 Summary Statistics by Geography and Year ...... 144 3.3 Verification of the Impact of the CLL on Financing Choices ...... 145 3.4 Impact of CLL on Number of Transactions ...... 146 3.5 Effect of the CLL on House Valuation Measures ...... 147 3.6 Effect of the CLL on House Valuation in Different Income Growth Areas148 3.7 Placebo Test for Coefficient of Interest ...... 149 3.8 Effect of the CLL on the Valuation of Different Groups of Transactions 150 3.9 Effect of the CLL on House Valuation in Low Supply Elasticity Areas ( Elasticity< 1) ...... 151 3.10 Elasticity Estim ates ...... 152 3.11 Data Cleaning Description ...... 155 3.12 Effect of the CLL on House Valuation Measures, Constrained Sample (0.5

12 Chapter 1

Personal Bankruptcy and Household Debt

1.1 Introduction

The last two decades in the US have seen a massive increase in household leverage, from 320 billion dollars in 1994 to 1060 billion dollars in 2010, and at the same time an increase in personal bankruptcies, which peaked in 2005 with 2.04 million filings. 1 These trends have brought renewed attention from academics and policy makers on the role that bankruptcy rules play in helping people manage their debt load, but also the incentives they provide to take on leverage in the first place. Personal bankruptcy laws in the US protect a fraction of a household's assets from seizure by unsecured creditors; under Chapter 7 bankruptcy, households are protected from creditors up to a monetary limit set by each state - the personal bankruptcy exemption. An increase in the level of this exemption (referred to as protection henceforth) may strengthen the demand for credit but can also decrease the supply of credit. In case of default, the lender cannot seize the borrower's assets if their value does not exceed the protection level dictated by law, while if they do the lender can only seize the excess value. Consider any simple model of a credit market with financially constrained, risk-averse borrowers, and a risk-neutral lender. If borrowers have a stochastic income, increased bankruptcy protection makes defaulting attractive to borrowers in more states of the world. As a result it diminishes the collateral value of assets, forcing lenders to charge a higher interest rate ex ante to break even (Hart and Moore, 1994). Therefore, this is akin to reducing the supply of credit, increasing prices, and/or reducing quantities. In addition, such a change in the supply of credit could increase the riskiness of the pool of loan applicants; increases in lending rates might foster borrowers' incentives to undertake riskier projects, or could intensify the entry of riskier borrowers (Stiglitz and Weiss, 1981)2.

'Debt amounts converted to year 2000 constant dollars to reflect change adjusted by inflation, see Figure .1-1 2Furthermore, lenders' willingness to supply credit will vary depending on their ability to screen borrowers.

13 Most of the existing empirical literature has focused on the effects described above that tend to reduce the supply of credit (the seminal paper in the area is Gropp et al., 1997). However, a higher protection level will also improve risk-sharing by increasing the insurance function of bankruptcy: in bad states of the world the borrower declares bankruptcy and, as a result of the higher protection level, is allowed to keep a larger proportion of their assets - the protection amount (Dubey et al. 2005, Zame 1993)3. This increases the demand for credit at a given interest rate. Changes in the level of protection will also affect the composition of borrowers: more risk averse borrowers might choose to use more debt since they weight the loss of their assets more severely. Therefore, an increase in level of asset protection might also lead to a change in the mix of borrowers, but in this case by drawing in new (more risk-averse borrowers), or by encouraging existing borrowers to take on more debt. Interest rates must therefore rise in equilibrium; but depending on which effect dominates (demand or supply), there can be an increase or decrease in the amount of credit extended.4 We use the timing of state changes in the levels of bankruptcy protection in a difference in difference design to identify their effect on the credit market equilibrium. We find that bankruptcy protection laws increase borrowers' unsecured credit holdings, mainly credit cards, leaving their level of secured debt - mortgage and auto loans - unchanged. At the same time we find an increase in the interest rate for unsecured credit, but not for other types of credit. These results are predominantly driven by low-income areas, and suggest that bankruptcy protection levels provide important downside insurance, which has first order effects on the supply and also on the demand for credit. Interestingly, using detailed individual data, we do not find an increase in default rates, which suggests that households are not necessarily over-borrowing or risk shifting as a response to the increase in protection. Empirically identifying the true effect of bankruptcy protection levels on household leverage is challenging, as these levels are correlated with unobservable borrower and lender characteristics that might simultaneously affect credit availability, and the level of protection. For example, states with higher protection levels may be states in which households are less financially savvy, or they might be states with higher house prices, and therefore more willing to take on more debt. This in turn will lead to a positive correlation between debt and protection. Therefore, we exploit changes in the dollar amounts of asset protection under bankruptcy to identify the effect of this protection on household debt5 . Our identification benefits from the fact that states increased bankruptcy protection at different times and by different amounts over our sample period. We show that changes in

3Non-state contingent contracts are a key friction here; in the absence of this friction, the effect of personal bankruptcy protection on household borrowing disappears. One possible explanation for why lenders do not offer more flexible contracts (more protection in "bad" states, or state contingent repayment) is that these lenders could face a collective action problem: if only one lender offered such a contract it would attract predominantly bad type borrowers, which is not an equilibrium. Alternatively, customized state contingent contracts could be hard to enforce. 4 For a more developed model see Appendix A. 5 Asset protection in our empirical implementation is the sum of homestead exemption and personal assets exemption levels for each state and year. Our results are invariant to the use of only homestead exemption.

14 protection levels are uncorrelated with macroeconomic conditions and other determinants of credit equilibrium, most importantly changes in state level house prices and unemployment rates. This allows us to disentangle the effect of bankruptcy protection levels on household leverage from other determinants of household debt that may be changing as well. We then estimate the effect of the changes in the levels of protection on changes in household debt. In doing so, we compare the change in the level of household debt between counties in a state that increases the level of protection between t and t+1, with other counties in a state that did not change their level of protection during the same period. The variation in bankruptcy protection changes over time and across states, which helps us to deal with two crucial assumptions of any difference in difference estimator. First, that the timing of the changes in the levels of protection are uncorrelated with other determinants of household leverage, as discussed above. And second, that after controlling for observed time-varying characteristics, linear county trends, and time-invariant county characteristics, changes in protection at the state level only affect the states which adopted the change, making the exogenous change in the level of protection the only determinant of the difference in household debt across states. Our empirical strategy is therefore similar to Cerqueiro and Penas (2011) and Cerqueiro et al (2013) who examine the effect of bankruptcy protection on start-up performance and innovation respectively. Our results show that the exogenous variation in the levels of protection causally increases the level of credit card debt held by households during our sample period (1999-2005)6, leaving secure debt (mortgage and auto) unchanged. This is consistent with the fact that personal bankruptcy allows households to discharge only unsecured debt 7 . Using novel bank branch-level data on credit rates for different types of credit, we explore the effect of bankruptcy protection changes on interest rates, and we find that an increase in bankruptcy protection leads to an increase in the interest rate on unsecured credit, which is consistent with a credit market equilibrium, where supply decreases and demand increases but the net effect is dominated by the demand response. A possible concern may be that states which did not change the level of protection within our sample period are not a good control group, as they could be systematically different from the group which did opt to change their level of bankruptcy protection, and this would therefore invalidate our empirical inference. However, the staggered nature of our empirical strategy, whereby each state which changed its level of protection is a control for past and future periods for other changes, allows us to

6We focus on the Pre-Bankruptcy Abuse Prevention and Consumer Protection Act of 2005 (BAPCPA), where the cost of filing for bankruptcy was low, and therefore the intensity of the treatment was higher. The bankruptcy reform makes the process of filing for bankruptcy harder, which ex ante diminished the incentives to take on more credit. The nature of the subprime crisis of 2007 and financial shock of 2008 may have affected household willingness to take on credit, and lenders' ability to supply credit, contributing to the lack of the effect during the post-reform period. We empirically investigate this by extending our sample until 2009; we find that changes in the law have no effect on unsecured debt held after the reform, see Appendix B8. 7The fact that the levels of protection only affect unsecured credit holdings helps to rule out that protection levels do not endogenously increase when the credit market becomes looser.

15 replicate our findings focusing only on the states where changes in protection levels were implemented in our sample period (i.e. "eventually" treated). In this case the effects we estimate are unchanged. We also look at the dynamics of the changes. By analyzing the timing, we can rule out that the level of protection may be correlated with pre-existing state specific trends that survive our controls, and thus that our results are a reflection of these differential pre-trends rather than changes in the levels of protection. We show that our estimates are not affected by the inclusion of lag changes in the levels of protection, and that the coefficients on the lags are small in economic terms, and statistically insignificant.' We now explore the heterogeneity of the average treatment effect. Exploiting within-state variation on the levels of debt held by counties, we find a stronger increase in the level of unsecured debt held by lower-income counties'. These results are consistent with the fact that increases in personal bankruptcy protection levels improve risk-sharing; this improvement should be stronger for lower-income regions, as they have fewer resources to diversify their risk exposure than wealthier ones, for which the differential impact of the increase should be smaller. Personal bankruptcy levels of protection are heavily concentrated on home equity; a big fraction of the protected nominal amount is exclusively linked to the home equity of the borrower. In line with a demand driven channel, we find that the effect is almost three times stronger in areas where homeownership is higher, after we condition on the level of income. Also conditioning on the level of income10 , we find that the increase in credit is stronger in areas where the banking industry is more concentrated (fewer banks), which is consistent with the relationship lending model proposed Petersen and Rajan (1995), where creditors are more likely to finance a credit constrained borrower when credit markets are concentrated, because it is easier for these creditors to internalize the benefits of assisting these borrowers; although this is only suggestive evidence. Overall we find that the average credit card balance in a county in our period is 290 million dollars in credit card debt, and the average increase in credit card debt is 7.6%. Our main estimate explains 10% of this balance growth". However, this value more than triples for low-income homeowners and for our micro-level sample, for 8 Considering that our exogenous variation is at the state level, we cannot control for state-time unobserved heterogeneity that is contemporaneous to the effects we observe. 9 Within each state, counties are divided into terciles based on total wages and salary levels in 1999. 0 ' Homeownership and bank concentration are correlated with income at the county level. There- fore, looking at cross-sectional variation without controlling for income is not informative, as it provides confounding information within all the correlated variables. In order to overcome this limi- tation of the data, we replicated the specification of interest for each income subgroup; this strategy proved to be useful. For example, under this setup, unemployment heterogeneity within income groups has no cross-sectional implications. However, homeownership and bank concentration still provide meaningful variation within income groups. 1 " This percentage is estimated using the average change in protection in our sample period, approximately 40k dollars, which represents a 54% change with respect to the average exemption level of 70k dollars. This value is a more conservative measure than using one standard deviation of level (70k dollars).

16 which our estimate explains 34% and 47% respectively of the increase in credit card balance. This heterogeneity seems to suggest that this affects only a subset of people: 2 homeowners who are expecting to be close to distress level on their credit cards . There is also the possibility that our estimates are biased downward (attenuation bias), due to measurement errors of our treatment. Finally, local economic conditions could produce spurious effects due to geographical heterogeneity that is uncorrelated to changes in the levels of protection. To overcome this endogeneity we compare neighboring county-pairs across state borders, within the same income bucket. The results of the estimation of changes in protection within each county-pair are very similar to the main estimates, and stronger when we concentrate on county-pairs in the lower end of the county income distribution. The aggregate results raise important questions about how credit expands in response to bankruptcy protection, and by whom; and whether it affects the overall composition and default probability of borrowers. We use detailed individual data containing debt levels and specific account information to understand and empirically test household behavior. We find that changes in protection levels increase the number of credit cards per household; this increase is stronger among households that had ex ante credit card accounts and those that had a positive balance. Finally, changes in protection are uncorrelated with entry into the credit card market, defined as the time when a member of a household opens their first account, or as the time when a credit card balance goes from zero to positive. All these results provide evidence that in this sample, the effect is driven by existing debtors expanding their current balance, or their number of accounts, rather than new households entering the credit market. Focusing on the same sample, we explore their delinquency behavior up to three years after the increase in credit card usage induced by the change in protection. Within this sample there is no measurable increase in the level of delinquency; if anything, the probability of being delinquent in the future decreases. If the households which are increasing their level of debt are over-borrowing, or taking on more risky projects, we would expect delinquency rates to increase. Although we cannot completely rule out over-borrowing or risk shifting behavior, the results described are more consistent with risk-averse borrowers increasing their debt as a result of the increase in downside protection 13 Furthermore, using county self-employment information, we show that areas that experienced an increase in the level of credit card debt also experienced an increase in the level of self-employment creation, specifically in industries that use more credit cards as start-up capital". It is important to point out that these outcome variables are only suggestive evidence of the real effect of the increase in the level of unsecured

1 2 Appendix B2 shows that within low-income areas the effect is differentially stronger for areas with a higher proportion of credit card delinquency (90+). 13Also, at the county level, delinquency rates do not seem to increase, which implies that also at the aggregate level, increases in the level of protection did not lead to an increase in the level of delinquencies. 4 1 For example, construction, photography, and other low capital-intensity industries that can be financed with credit card debt.

17 debt, as they represent county aggregates. The results are also robust to restricting the sample to states which changed the level of protection only once during the sample period, to considering only states with large changes in protection as treated states, and to the use of an indicator instead of the magnitude of the change. Given the nature of our empirical strategy, as we argue before, time-varying changes at state levels may be omitted variables explaining our results; one candidate is the level of unemployment insurance in each state (Hsu et al., 2012). However, the inclusion of this variable has no impact on the estimated coefficient. 5 Our results suggest that existing borrowers increase their leverage without increasing their ex post delinquency, consistent with risk-averse, constrained borrowers reacting to the increase in insurance. We cannot say anything about the welfare effect of these changes. In a world with complete markets, increases in protection will constrain the contract space and therefore may lead to inefficiencies. Furthermore, in the presence of limited commitment, harsher penalties for defaulting could improve welfare ex ante (Kehoe and Levine 1993, Alvarez and Jermann 2000). However, if state contingent contracts are not available (i.e. incomplete markets), a pro-debtor bankruptcy code could lead to welfare gains (Link 2004). Therefore, theoretically the effect of increased bankruptcy protection on welfare is undetermined, and dependent on modeling choices. A number of earlier papers have looked at the cross sectional relationship between the level of bankruptcy protection and consumer credit. See for example Gropp et al. (1997), the first paper to examine this relationship. Using household data from the 1983 Survey of Consumer Finances, they found that higher levels of protection were associated with both reduced credit availability for low-asset households and increased debt balances among higher-asset households. Similarly Berger et al. (2010) found that higher protection is associated with lower access to credit for unlimited liability firms. Also, Lin and White (2001) found the same relationship for mortgage credit. The recent legislative history of staggered introduction of bankruptcy exemptions in combination with household data allows us to identify the effects of changes in bankruptcy protection on the change in the supply and demand of credit for different types of debt. Most importantly, we find that an increase in personal bankruptcy protection leads to an increase in the amount of unsecured debt held by households, leaving secured debt unchanged. Therefore, using an improved empirical strategy, we see that the demand effect of bankruptcy protection, arguably driven by improved risk-sharing, dominates its supply-deterring effects. Hence increased bankruptcy protection increases equilibrium debt reliance, particularly for low-income homeowners. Increases in personal bankruptcy protection results in a weakening of creditor rights. There is a vast literature in corporate finance that has examined the effect of

15 As a case study during our relevant sample period, 1999-2005, one state went from having some level of protection to unlimited protection. When we include this time-varying dummy in the regression, we find that the main effect is unchanged, but the unlimited protection dummy is negative and significant for mortgage and credit card debt. This suggests that the effect of protection is a non-linear function of the level of exemption, and therefore above a certain threshold lenders increase prices to a magnitude which decreases quantities.

18 changes in creditor protection on debt (La Porta et al. 1998, Levine 1998, Djankov et al. 2007). Most related to this paper is Vig (2013), which looks at increases in the seizability of assets for large firms in India, and how this triggers a drop in the demand for secured debt. Vig (2013) suggests that this demand response is driven by an increase in the threat of early liquidation due to the increase in creditor protection. Our paper focuses on a different channel, i.e. changes in the self-selection of households with different risk aversion levels, or their willingness to default strategically. The rest of the paper proceeds as follows: Section 2 explains the institutional framework of personal bankruptcy laws and related existing literature; Section 3 outlines the empirical hypothesis with a theoretical focus; Section 4 describes the data; Section 5 develops the empirical strategy; and Section 6 shows the results before the conclusion.

1.2 Bankruptcy Procedure and Related Literature

1.2.1 Institutional Framework Personal bankruptcy procedures determine both the total amount that borrowers must repay their creditors and how repayment is shared among individual creditors. An increase in the amount repaid may benefit all individuals who borrow, because higher repayment levels may cause creditors to lend more, and at lower interest rates. However, a larger repayment amount implies that borrowers need to use more of their existing assets and/or post-bankruptcy earnings to repay pre-bankruptcy debt, therefore reducing their willingness to borrow and their incentive to work 6 . US bankruptcy law has two separate personal bankruptcy procedures, which are named as they appear in bankruptcy law, Chapter 7, and Chapter 13. Under both procedures, creditors must immediately terminate all efforts to collect from the borrower (such as letters, wage garnishment, telephone calls, and lawsuits). Most consumer debt is discharged in bankruptcy, however most tax obligations, student loans, allowance and child support obligations, debts acquired by fraud, and some credit card debt used for luxury purchases or cash advances are not. Mortgages, car loans, and other secured debts are not discharged in bankruptcy, but filing for bankruptcy generally allows debtors to delay creditors from retrieving assets or foreclosure. Prior to the Bankruptcy Abuse Prevention and Consumer Pro- tection Act of 2005 (BAPCPA), debtors were allowed to freely choose between the two.

Bankruptcy Law Before 2005 The most commonly used procedure before 2005 was Chapter 7. Under it, bankrupts must list all their assets. Bankruptcy law makes some of these assets exempt, meaning that they cannot be seized by creditors. Asset exemption amounts are determined by

16See Dobbie and Song (2013) for a more detailed description of this issue.

19 the state in which the borrower lives. Most states will have personal asset protection, which exempts debtors' clothing, furniture, "tools of the trade", and sometimes equity in a vehicle. In addition, nearly all states have some level of homestead protection for equity in owner-occupied homes, but the levels vary from a few thousand dollars, to unlimited amounts in six states, including Texas, Florida, and DC". This exemption level is what we refer to here as the protection level. Under Chapter 7, debtors must use their non-protected assets to repay creditors, but they are not obliged to use any of their future income to make repayments. Under the alternative procedure in Chapter 13, bankrupts are not obliged to repay from assets, but they must use part of their post-bankruptcy income to make repayments. Before 2005, there was no predetermined income exemption; on the contrary, borrowers who filed under Chapter 13 proposed their own repayment plans. They often proposed to repay an amount equal to the value of their non-protected assets under Chapter 7. Also, borrowers were not allowed to repay less than the value of their non-protected assets and, since they had always the option to file under Chapter 7, they had no incentive to offer any more. Judges did not need the approval of creditors to approve repayment plans.18 The cost of filing for bankruptcy before 2005 was low: about 600 dollars under Chapter 7, and 1,600 dollars under Chapter 13, as of 2001 (White 2007). The punish- ment for bankruptcy included making bankrupts' names public and the appearance of the bankruptcy filing on their credit records for 10 years subsequently. In addition, bankrupts were not allowed to file again under Chapter 7 for another six years, (but they were allowed to file under Chapter 13 as often as every six months)19 . Overall, these features made US bankruptcy law very pro-debtor. Since debtors could choose between the procedures under Chapters 7 and 13, they would select the procedure which would maximize their gain from filing. Around three quarters of all those filing for bankruptcy used Chapter 7 (Flynn and Bermant, 2002). Most debtors who filed under Chapter 13 did so because their gains were even higher using this

17 See Table 1.2 for summary statistics of the level of protection. 18 Even when households file under Chapter 13, the amount that they are willing to repay is affected by Chapter 7 bankruptcy protection. For example, suppose that a household that is considering filing for bankruptcy has 40,000 dollars in assets and is located in a state in which the protection level is 20,000 dollars. Since the household would have 20,000 dollars of unprotected assets if filing under Chapter 7, it would be willing to repay no more than 20,000 dollars (in present value) from future income if it were to file under Chapter 13. As a result of this close relationship between Chapter 7 and Chapter 13 bankruptcy filings, we assume that changes in Chapter 7 protection levels will affect household willingness to file for bankruptcy (either under Chapter 7 or 13). 19 US bankruptcy law allowed additional debt to be discharged under Chapter 13. Debtors' car loans could be discharged to the extent that the loan principal exceeded the market value of the car (negative equity). Also, debts acquired by fraud and cash advances obtained shortly before filing could be discharged under Chapter 13, but not under Chapter 7. These characteristics were known as the Chapter 13 "super-discharge", and some households took advantage of the situation by filing first under Chapter 7, where most of their debts were discharged, and then converting their filings to Chapter 13, where they proposed a plan to repay part of the additional debt covered under Chapter 13. This two-step procedure, known as "Chapter 20", increased borrowers' financial gains from bankruptcy as opposed to filing under either procedure separately.

20 procedure than under Chapter 7.

The Bankruptcy Abuse Prevention and Consumer Protection Act

The Bankruptcy Abuse Prevention and Consumer Protection Act (BAPCPA) of 2005 made several major changes to bankruptcy law. First, it abolished the right of debtors to choose between Chapters 7 and 13; now debtors must pass a new "means test" to file under Chapter 7. Debtors qualify for Chapter 7 if their monthly family income average over the six months prior to filing is less than the median monthly family income level in the state in which they live, adjusted for family size. In some places households could be allowed to file under Chapter 7, without satisfying the means test, as long as their monthly "disposable income" was lower than 166 dollars per month. Thus, the 2005 law prevents some wealthy debtors from taking advantage of the unlimited income exemption in Chapter 7. The reform also imposed new restrictions on strategies used to protect high value assets in bankruptcy. For example, state of residence home-equity protection is only valid after two years of residency in that state, and within 2.5 years the level is capped at 125,000 dollars. Finally if borrowers convert non-exempt assets into home-equity by making a down payment on their mortgage, they must do so at least 3 and one third years before filing (White, 2007). The second major change under the BAPCA is a uniform procedure that determines repayment obligations under Chapter 13. Debtors must now use 100 percent of their "disposable income" for five years following their bankruptcy filing to make repayments 20 . Third, BAPCPA greatly raised bankruptcy costs, and households are now required to take a financial management, and also a credit counseling course before their debts are discharged. They must file detailed financial documents, including copies of their tax returns for the previous four years, which may force them to prepare unfiled tax returns. Filing fees have also increased. These new requirements have increased debtors' out-of-pocket costs of filing to around 2,500 dollars to file under Chapter 7 and 3,500 dollars under Chapter 13 (Elias, 2005), not forgetting the cost of the two training courses, and the preparation of tax returns.2 1 BAPCPA among other things also increased the minimum time that must pass between bankruptcy filings from six to eight years for Chapter 7, and from six months to two years for Chapter 13 filings 22 . Therefore, fewer debtors than before are eligible for bankruptcy at any given period. Overall, the adoption of BACPA increases the cost of bankruptcy, decreases the possible amount of debt discharged in bankruptcy, while implicitly decreasing income protection. Therefore, setting a maximum income level above which debtors can no longer gain from filing, making the US bankruptcy law more pro-creditor.

2 0 BAPCPA defines disposable income as the difference between debtors' average monthly family income during the six months prior to filing, with a new income exemption. 2 'A large proportion of the cost is attributable to the fact that bankruptcy lawyers can be fined if debtors' information is not accurate. 22 BAPCPA also imposes a four-year minimum period, where no such minimum existed previously, for filing first under Chapter 7 and then under Chapter 13; and it also eliminates the "super- discharge" effect.

21 1.2.2 Related Literature

Gropp et al. (1997) was the first paper to use household level debt data to look at the difference on credit availability for different levels of protection. Using the Sur- vey of Consumer Finance of 1983, they found that higher protection under personal bankruptcy is associated with a lower probability of access to credit, and a lower level of debt for low asset households, in states with more generous bankruptcy exemptions. Using detailed bank information, Berger et al. (2010) found that unlimited liability small businesses have lower access to credit in states with more debtor-friendly bankruptcy laws. In addition, these businesses face harsher loan terms: they are more likely to pledge business collateral, have shorter maturities, pay higher rates, and borrow smaller amounts. Also, Lin and White (2001) looked at how the protection levels affect the availability of mortgage credit application granting, finding that accepted applications are negatively correlated with the level of protection. However, all these studies use cross-sectional variation on protection to look at how these levels correlate with credit availability. Hynes et al. (2004) find that state levels of exemptions are correlated with bankruptcy filing rates and state redistributional policies to help the poor, among other variables that can be correlated with the supply of credit, suggesting that the examination of the impact of bankruptcy laws should not treat protection levels as exogenous variables. This paper contributes to this literature using state time variation in bankruptcy protection levels to overcome these endogeneity concerns when looking at relationship between bankruptcy protection and credit markets. Using this empirical strategy we find that increases in bankruptcy protection did not lead to a reduction in the amount of debt held by households. Our empirical strategy is more closely related to the work of Cerqueiro and Pe- nas (2011), who use state level variation in the level of bankruptcy protection to look at start-up creation, finding that increases in protection decrease start-up performance; and to Cerqueiro et al. (2013), who uses a similar strategy to look at the effect of personal bankruptcy laws on innovation, finding that there is an aggregate decrease in the level of innovative activity among small firms in places in which protection increased. The effect of the use of credit cards in entrepreneurial activity has also been studied by Chatterji and Seamans (2012). Using states' removal of credit card interest rate ceilings in 1978 they show that this deregulation increases the probability of entrepreneurial entry, arguably through an access to finance channel. Finally, Fan and White (2003) find that personal bankruptcy protection motivates entrepreneurial activity using cross-sectional variation in the level of protection. In this paper, we show that increases in bankruptcy protection are correlated with increases in self-employment. Although we cannot rule out a demand channel, it seems that bankruptcy laws could have an expansive impact on self-employment through an increase in the credit channel. Bankruptcy laws directly affect unsecured debt, given that secured debt cannot be discharged. Therefore this paper is related to the literature on credit card borrowing. Agarwal et al. (2013), analyze the effectiveness of consumer financial regulation in the credit card market, using the 2009 credit card reform. They find that regulatory limits on credit card fees reduce the overall borrowing cost to consumers by 2.8% of average

22 daily balances. Gross and Souleles (2002a) use credit card account data to analyze how people respond to increases in the supply of credit; they find that increases in credit limits generate an immediate response to debt, which implies a big sensitivity of households to credit market changes. Gross and Souleles (2002b) use credit card accounts to analyze credit card delinquency to highlight the importance of time- varying household characteristics on their ex post behavior. Our paper contributes to this literature, showing new evidence of how bankruptcy protection affects the demand for credit card debt. This paper also relates to the studies that focus on the effect of personal bankruptcy on filings and delinquency rates. Gross et al. (2013) use tax rebates to find that households have a significant sensitivity of income to probability of filing, which is consistent with the high sensitivity of financially constrained agents to increase leverage as credit availability increases, found by Gross and Souleles (2002b). White (2007) looks at the effect of the interaction between personal bankruptcy filings and credit card growth before the adoption of the new Bankruptcy Abuse Prevention and Consumer Protection Act (BAPCA), arguing that the increase is due to the debtor friendly bankruptcy laws in the pre-2005 period. In a related article, Jagtiani and Li (2013) focus on the ex post effect of filing, and find that after a consumer files for bankruptcy, there are long-lasting effects on their availability of credit. This paper contributes to this literature providing suggestive evidence of how bankruptcy protection affects the mix of borrowing with no impact on delinquency behavior. Furthermore, the protection of assets under bankruptcy affects the amount of household collateral, and thus, their access to credit. Since Bernanke and Gertler (1989), or Kiyotaki and Moore (1997), a number of theories have suggested that improvements in collateral values ease credit constraints for borrowers. The collateral lending channel builds on the idea that information asymmetries between lenders and borrowers can be alleviated when collateral values are high (Hart and Moore, 1994). From an empirical point of view, the collateral channel has been explored in its effect on firms, by Benmelech and Bergman (2011), and Chaney et al. (2012); and credit availability for small businesses, by Hurst and Lusardi (2004), and Adelinot et al. (2013). The effect of housing collateral on household leverage has also been analyzed, by Mian and Sufi (2011). Increases in bankruptcy protection can also be seen as decreases in creditor rights, which connects this paper to a large literature tracing the link between creditor rights and financial development, pioneered by La Porta et al. (1998), and including Levine (1998); Djankov et al. (2007); and Haselmann et al. (2010). Overall, this literature reports a positive correlation between increases in creditor rights and the amount of credit.2 4 Most relevant to the current paper is Vig (2013), which looks at the increase in creditor protection for secured debtors in the context of large firms in India. The main difference between Vig (2013) and this paper (besides the fact that this paper looks at US households, as opposed to firms in India), is how demand responds to

2 3 Rampini and Viswanathan (2010) in the context of a firm's access to credit. 2 4 Most recently, there are other papers which have looked at the same relationship but using cross- country settings: Gianetti (2003); Qian and Strahan (2007); Acharya et al. (2011); and Davydenko and Franks (2008).

23 changes in creditor protection. In Vig (2013), the decrease in the amount of secured debt is driven by an increase in the threat of early liquidation, which firms face due to the increase in creditor protection.25 In the current paper, the demand response (increases in the demand for credit card debt), is based on an insurance channel which relies on household risk aversion, and/or an increase in the number of strategic borrowers. 26 This paper is also related to previous studies that have looked at the effect of bankruptcy laws design in the context of corporate bankruptcy (Baird and Ras- mussen, 2002; Bolton and Scharfstein 1996). In this context there is a large literature that describes the tension between ex ante and ex post efficiency in any bankruptcy design. For instance, Gertner and Scharfstein (1991), and Hart (2000), show the incentives of the debtor and creditors under corporate resolution in a theoretical framework, and demonstrate how debt contracts can lead to inefficient liquidation and underinvestment. This framework is also relevant when thinking about the incentives for households to file for bankruptcy. Empirically, Chang and Schoar (2013) look at the judge-specific fixed effect, showing that pro-debtor judges have worse firm outcomes after Chapter 11, suggesting that this is a result of managers and shareholders' incentives misalignment, highlighting how bankruptcy codes can have a significant impact on ex post outcomes. Furthermore, Iverson (2013) looks at the effect of bankruptcy courts' reduction in court caseloads due to the consumer bankruptcy reform in 2005, finding that firms in more pro-debtor courts allow more firms to reorganize and liquidate fewer firms. Finally, this paper is complementary to studies looking at the effect of personal bankruptcy laws on labor markets. Dobbie and Song (2013) find that filing for bankruptcy under Chapter 13 has a significant effect on increasing earnings and employment, and also decreases mortality, suggesting that consumer bankruptcy benefits are an order of magnitude larger than previously estimated".

1.3 Data and Summary Statistics

1.3.1 Data Description

In order to address the impact of changes in bankruptcy protection on household debt, we collect and combine different data sources. The three main data sources include time series of state levels of protection under bankruptcy, and geographical distribution of household debt and interest rates information. In this section we describe this datasets in detail. The level of protection or exemptions represents the dollar amount of equity that the debtor is entitled to protect in the event of bankruptcy; it represents the amount 25 This is consistent with the corporate literature on bankruptcy reorganization which suggested that excessive creditor rights can lead to ex post inefficiencies in the form of a liquidation bias (Aghion et al. (1992); Hart et al. (1997); Stromberg (2000); Pulvino (1998); 26 and Povel (1999). Examples of papers showing the costs of increases in creditor rights include: Acharya et al. (2011); Acharya and Subramanian (2009); and Lilienfeld-Toal et al. (2012). 2 7 See White (2005) for a complete review of the literature.

24 of home equity and other personal assets that are protected. This information was manually extracted and compiled from many sources, from state bankruptcy codes to bankruptcy filing manual books2 8 We obtain level debt balances from the Federal Reserve Bank of New York Con- sumer Credit Panel/Equifax (CCP). This quarterly panel dataset is a 5% random sample of individuals in the US who have a credit history with Equifax and a social security number associated with their credit file. Debt data reported includes mortgage balances, home equity installment loans, and home equity lines of credit; auto loans, including loans from banks, savings and loan associations, credit unions, auto dealers and auto financing companies; and credit card debt: revolving accounts from banks, national credit companies, credit unions, and bankcard companies. The county level data is an aggregate of this information from 1999 to 2005 where, for privacy reasons, reporting is done only for counties with an estimated population of at least 10,000. This information is available for all debt types and the fraction of household with delinquency status of 90 days late is provided as well. The micro level data includes household level data of the debt variables described above, plus detailed information on credit card accounts and individual level delinquency status: current, 30 days late, 60 days late, 90 days late, 120 or more days late, and severely derogatory. The individual level data permits a unique insight into the ex post behavior of households, as we are able to track the delinquency behavior of consumers before they are affected by the change in protection2 9 We obtain interest rates from Rate-Watch. It provides historical rate and fee data from banks and credit unions across the country for a wide variety of banking products, such as CDs, checking, savings, money markets, promotional specials, auto loans, unsecured loans, and credit cards. They collect information at the branch-setters level by survey, and archive the information on a regular basis. For our purpose, interest rates for unsecure loans, credit cards, and mortgage loans are aggregate at the county level using branch-setter rate levels for the last quarter of each year to be consistent with the aggregate debt balances measure. We then use this detailed geographically dispersed measure of interest rates from 1999 to 2005 to analyze the supply response of changes in personal bankruptcy protection. County level income is measured as total wages and salary in a county according to the IRS; this data is available from 1999 to 2005. The house prices used in the regressions are obtained from the Federal Housing Finance Agency (FHFA) House Price Index (HPI) data at a state level. The FHFA house price index is a weighted, repeat-sales index and it measures average price changes in repeat sales or refinancing on the same properties. This information is obtained by reviewing repeat mortgage transactions on single-family properties whose mortgages have been purchased or securitized by Fannie Mae or Freddie Mac since January 1975. We use data on the state level index between 1999 and 2005. County based unemployment levels and unemployment rates are obtained using

28How to file for Chapter 7 Bankruptcy, Elias Renauer and Leonard Michon. Nolo editorial (1999-2009) 29See Lee and van der Klaauw (2010) for details on the sample design.

25 the Bureau of Labor Statistics Local Area estimates. Local Area Unemployment Statistics (LAUS) are available between 1976 and 2012 for approximately 7,300 areas that range from census regions and divisions to counties and county equivalent. We match the county equivalent data to the CCP data using Federal Information Processing Standard (FIPS) county unique identifiers. To look at the determinants of change in exemptions, we use four additional data sources: changes in state total medical expenses extracted from the National Health Expenditure Data, Centers for Medicare and Medicaid Services; state level changes in GDP and Personal Income from Bureau of Economic Analysis (BEA); bankruptcy filing statistics at the state level from the Statistics Division of the Administrative Office of the United States Courts30; and measures of political climates using the share of votes for the Democratic Party in the last House of Representatives election obtained from the Clerk of the House of Representatives (CHR). The net creation of sole proprietorships at a county level is obtained from Census non-employer statistics; we obtain the number of establishments for the period of 1999 to 2009 at the 2-digit NAICS level. In order to construct a measure of industries that use credit card as a source of capital, we look at the Survey of Business Owners (SBO) Public Use Microdata Sample (PUMS). The SBO PUMS was created using responses from the 2007 SBO and provides access to survey data at a more detailed level than that of the previously published SBO results. The SBO PUMS is designed to study entrepreneurial activity by surveying a random sample of businesses selected from a list of all firms operating during 2007 with receipts of $1,000 or more provided by the IRS. The survey provides business characteristics such as firm size, employer-paid benefits, minority- and women-ownership, access to capital, and firm age. For the purposes of this paper, we classified industries based on the "use of credit card as a start-up capital" for each firm and we group the answers to this question at the 2-digit NAICS industry level (the finest level available in the data) for firms established in 2007, and then focus specifically in 1-4 employee firms only.

1.3.2 Summary Statistics

Table 1.1 shows a description of our main variables; the sample spans from 1999 to 2005. The total debt balance in a county is 2.91 billion dollars. The level of credit card balance is 0.29 billion dollars. When looking at states that "eventually" change their level of protection during our sample period and compare them to states that never change their level of protection, the former holds 0.36 billion dollars on average, and the latter 0.22; however the difference is not statistically significant. The average debt growth in a county was 12.2%, and credit card debt growth during the same period experienced the same pattern, with a 7.6% average annual growth, with no significant difference between the "eventually" treated and the never treated group. The summary statistics seem to show that credit card balances are a small proportion of the average household balance sheet, as mortgage debt accounts for most of consumers' debt claim. However, it is important to point out that when

30 See http://www.uscourts.gov/Statistics/BankruptcyStatistics.aspx

26 compared in terms of monthly payments, this difference is much smaller, and arguably credit card debt is an important part of household budget and a relevant medium to relax budget constraint, allowing households to shift inter-temporal consumption (White 2007). The only strong significant difference between the two groups is seen in average house price growth. States which were never treated experienced a house price growth of 6.2% on average annually, and states which were eventually treated increased their house price growth by 8.8%. This difference is consistent with the fact that house prices are argued to be determinants of the changes in bankruptcy protection. However, we find in Table 1.6 that they have no predictive power in the changes in protection. Table 1.2 shows the description of the exemption levels and changes from 1999 to 2005. First, it is important to notice that bankruptcy exemption changes are quite common within our sample period; over the whole time there are 37 changes within 26 states. The average level of protection is around 73,000 dollars, and a median of 55,800 dollars, with most of the value coming from the homestead exemption (protection over homeowners' equity). The average change in protection is close to 40,000 dollars, with a median of 15,400 dollars, with some changes being very small and associated to inflation adjustments, and others being very substantial. Figure 1-2 shows the geographical dispersion of these changes.

1.4 Empirical Hypothesis

Changes in the level of asset protection in bankruptcy affects credit markets' equilibrium through demand and supply. In order to guide our empirical analysis we review the differences dimension through which increases in asset protection can affect the supply and demand of credit, and review the implications for our empirical exercise. Collateral channel. If markets are incomplete, the possibility of collateral pledg- ing enhances agents' debt capacity, as it gives the lender the option to repossess assets ex post, reducing the risk of borrowers, and easing borrowers' access to finance ex ante (Hart and Moore, 1994). In our case, the increase in protection diminishes the collateral value of assets, as it decreases the availability of assets to be seized by lenders, making the supply of credit less attractive; therefore reducing borrowers' access to credit. Insurance channel. In the presence of incomplete markets, increased protection also makes borrowing more attractive for risk-averse agents by improving risk-sharing. Effectively, the higher protection on the bad state of the world will incentivize risk- averse agents to take on leverage, increasing the demand for credit. Moral hazard channel. An increase in the level of protection might also foster borrowers' incentives to undertake riskier projects or over-borrowing, increasing the demand for credit, and the ability of lenders to distinguish the type of borrower that are they facing will define the supply response. Furthermore, according to Stiglitz and Weiss (1981), lenders' profit functions could set an upper limit to the increase in interest rates, leading to a decrease in the quantities due to the increase in borrower

27 risk. In summary, moral hazard increases the demand for credit, and in most cases, will reduce the supply of credit. Adverse selection channel. If the level of protection increases, more strategic defaulters with private information about their future income or propensity to default could participate in the markets, aiming to profit from the new borrowing conditions, increasing the riskiness of the pool of borrowers and also the demand for credit. Again the equilibrium response will be driven by lenders' ability to screen new borrowers. Therefore, the theoretical prediction is unclear, given that the net effect will depend on the relative magnitudes of the supply and demand response3 1 . Interest must weakly rise in equilibrium, independent of the prevailing force. If the supply demand dominates, quantities should go down, but if the demand effect dominates, quantities should go up. We attempt to distinguish between these channels empirically. It is plausible to imagine that in the presence of agency problems, a demand driven equilibrium takes place. In an extreme case, if the lender overestimates the quality of the pool of borrowers, the increase in protection would lead to an increase in quantities. However, in Appendix A we show that given very simple conditions, and without asymmetric information, we can observe a demand driven equilibrium where quantities and prices increase. This model of the credit market considers a risk-averse borrower who is financially constrained and a risk-neutral lender. The borrower has a stochastic income, and exogenous home equity that is realized in period 2. Only debt contracts are available. In case of default, the lender can seize the borrower's assets up to the exemption level dictated by law. The agents need to borrow in order to consume in period 1, while the interest rate is set such that the bank breaks even (zero profit). For a given interest rate, a risk-averse borrower will consume until a point where the marginal utility of consumption today is equal to the expected marginal utility in the future. Increased bankruptcy protection makes defaulting attractive to the borrower in more states of the world, and forces lenders to charge a higher interest rate to break even. The model shows that for a certain region with a given level of protection in bankruptcy, when the level of protection is increased, the agent will be willing to take on more debt despite the increase in interest rates. This happens when the marginal benefit from the increase in consumption at period 1 is greater than the loss of utility in the good state in period 2, due to the repayment of their debt claim; as in the bad state they are indifferent due to the protection level. Furthermore, if the marginal benefit is not enough to overcome the loss of consumption during the good state, we should see a decrease in quantities and increase in prices. Using exogenous variation on the level of protection, we aim to identify the type of equilibrium that rises after an increase in the level of consumer protection under bankruptcy. These results, which are highlighted by the model, are relevant as they show that the insurance channel in itself could lead to a demand driven credit market equilibrium shift, without the presence of moral hazard or adverse selection. Empirical Predictions The exposed theoretical framework allows us to sharpen our empirical exploration.

31 Figure 1-3 shows the possible outcomes in a simple demand and supply graph.

28 Based on the arguments above we have the following predictions. First, if the demand effect dominates, we should see an increase in quantities and prices. Furthermore, the increase in prices should be stronger for low-income borrowers, as the increase in risk-sharing (insurance channel) is more important for these borrowers, and they are also more likely to be under financial constraints. The effect should be stronger for homeowners, as the change in asset protection affects home-equity holding predominantly (see Table 1.2). The increase in bankruptcy protection does not directly affect secured debt, as the bankruptcy code only dis- charges unsecured debt. Therefore, we should see weaker or no effect on secured debt. Finally, if agency problems are an important driver of the increase in demand, we would expect to see a significant effect on ex post default, arguably driven by individuals who over-borrowed ex ante or invested in riskier projects. Second, if the supply effect dominates, we should see an increase in prices and a decrease in quantities. The rise in prices should be higher in places where the riskiness of the pool of borrowers, or the ex ante probability of defaults, increases more. The effect should also be stronger where the fundamental value of the ability to pledge assets is higher, and court enforcement of bankruptcy contracts is lower. Further, the effect should be stronger in areas where lenders have less information about their borrowers, as the dominance of the supply effect suggests that lenders are reducing the supply of credit more intensively. In the next section we show the empirical strategy we used to identify the equilibrium change: we find that the quantities and price effect is consistent with a stronger demand effect, and we describe the set of tests that we used to assure this finding, and the empirical test that attempts to distinguish between the different channels.

1.5 Empirical Strategy

Empirically identifying the actual effect of bankruptcy protection levels on household leverage is challenging, as these levels are correlated with unobservable borrower and lender characteristics, which might simultaneously affect credit availability and the level of protection. For example, on the one hand, states with a higher protection level may be states where households are less financially savvy and, as a result, are more willing to take on more debt; this in turn will lead to a positive correlation between debt and protection. On the other hand, if the level of protection correlates with better local economic conditions, people will be less financially constrained, potentially taking on less debt, and thus leading to a negative correlation between debt and protection levels. In this paper, we exploit exogenous variation in state level bankruptcy protection dollar amounts to identify the effect of this protection on household debt. We use different timing in the changes to exemption levels by state to identify how exemptions affect household leverage (there were a total of 37 changes in exemptions between 1999 and 2005) The proposed baseline specification is the following,

29 ADebtit = ai + at+ ppAProtectiont + FAXt + Eit (1)

Where ADebtit is the log change in either credit card debt, mortgage debt, auto loan debt, in a county i and year t .AProtectiont represents the log change in the level of Chapter 7 protection (homestead plus personal) in a state s and year t .aj is a county fixed effect, and at are year fixed effect.AXit represents a vector of county controls changes, such as county unemployment rate, log of house prices, and log of income in a county. We use the same specification in (1) to measure the effect of changes in protection on interest rates. To do so we replace the log change in debt, by changes in interest rates in percentage for mortgages, personal unsecure loans and credit cards. Since changes in protection vary at the state level, but debt balances and interest rates are observed at the county or individual level, the error term in equation (1) has a potentially time-varying state component. Following Bertrand, Duflo and Mullainathan (2004), the residuals are clustered by state. This allows for maximum flexibility in the variance-covariance matrix of residuals. It is also more general than state-year clustering, which would leave intact the possibility of serial correlation in the error term. If the measure of debt and the controls all display heterogeneous trends across counties, the most parsimonious treatment of these trends is to take first-differences, as in the equation above3 2 , with variables in differences; the presence of county fixed effects guarantees that differential county specific trends are controlled for in all variables. A first-differences specification is suitable in our case as it accommodates the repeated treatment present in our sample (in our sample period some states did change their level of protection more than once). The regressor 13p captures the changes in debt within the year as the level of protection increases. Additionally, the use of the amount of protection, i.e., intensity of treatment, guarantees that the main estimate is driven by big changes in the level of protection. Furthermore, we will conduct alternative specifications to show that our results are robust to the use of level specification, and to the use of alternative measures of the treatment effect. Effectively, we compare the change in the amount of debt between a county belonging to a state which increased the level of protection between t and t+1, with the amount of debt of a county belonging to a state in which the level of protection did not change during the same period. The two identifying assumptions are first, that the timing of the changes in the levels of protection are uncorrelated with determinants of household leverage; and second, that after controlling for observed time-varying characteristics, linear county trends, and time-invariant county characteristics, changes in the state level of protection will only affect the state which adopted the change, thus the only determinant of the difference in household debt across states is the exogenous change in the level of protection. We assess the first identifying assumption by looking at the correlation between

3 2 Paravisini (2008).

30 suspected determinants in the level of protection and changes in the levels of protection. Conventional wisdom attributes changes in the levels of bankruptcy protection to the gap between house prices and homestead exemption levels, as well as the cost of medical expenses. If our identification strategy is valid, changes in the measurable variables should be uncorrelated with changes in the level of protection, suggesting that the actual timing of the change is an exogenous shock to the credit demand and supply of credit in the affected regions. To assess the second identifying assumption, we need to rule out alternative hypotheses that could explain our results. First, changes in the level of protection could be correlated with state specific pre-existing trends that survive our controls, and thus our results are a reflection of this differential pre-trend rather than a result aris- ing from changes in the levels of protection. For example, states which increase their protection levels are states where economic conditions are booming in the period prior to the increase. We should expect that looking at the dynamic of the change, the inclusion of lags of the changes should have no effect on the coefficients and have no significant correlation with the levels of debt. A second alternative hypothesis is that there are state specific credit market trends that are correlated with the changes in protection that would explain our findings. For example, the areas where the level of protection increased were areas where all credit availability for all types was expanded. To meaningfully differentiate the impact of the change in the level of protection from these alternative hypotheses, we use the fact that personal bankruptcy laws allow households to renege only on unsecured debt, which implies that changes in personal bankruptcy laws will only directly affect unsecured debt. A third alternative hypothesis is that the observed increase in quantities is due to a contemporaneous decrease in prices that is correlated with the timing of the changes in bankruptcy protection. In other words, areas that increased the level of protection were areas where credit became cheaper. Using novel bank branch level data on credit rates for different types of credit, we can explore the effect of bankruptcy protection changes on interest rates; if interest rates are positively affected by the increase in the level of protection, it is less likely that our effect is driven by a relaxation of lending standards in credit markets. Local economic conditions could produce spurious effects due to geographical heterogeneity that is uncorrelated to changes in the levels of protection. To overcome this 33 endogeneity we compare neighboring county-pairs across state borders , but within the same income categories, using the following empirical specification:

ADebtipt = oi + aYipt + /pAProtectionst + FAXit + 6Ept (2)

Where ADebtipt is the log change in either credit card debt, mortgage debt, auto loan debt; in a county i, pair p and year t. AProtectionst represents the log change in the level of Chapter 7 protection (homestead plus personal) in a state s in year t. ai

3 3This methodology is similar to Heider and Ljungqvist (2013) and Dube et al. (2010)

31 is a county fixed effect, and aipt, is a dummy for each neighboring county pair for each year. Note that variables for county i maybe repeated for all pairs of which they are part. In this setup our estimate fp only uses debt variation within each neighboring county-pair across state borders. Our additional identifying assumption implies that the changes in protection are uncorrelated with the residual Eipt after controlling for observable characteristics, county fixed effects and county-pairs year fixed effect. We also assign counties to income buckets, and run the proposed specification only within county-pairs that are in the same income category. To attempt to identify the channel that is driving the demand effect we use individual level data to look at debt change, entry to the credit card market, and delinquency. We use the same specification (1) as for the county aggregates, but changing the dependent variable, and including in this case the zipcode level house prices, income, and county unemployment rates. The change in debt for each individual is estimated using log changes, and it therefore represents the change in debt for existing debtors. When looking at the number of accounts, our dependent variable is the difference between the number of credit cards in t -1 and t. Entry is defined in two ways as follows: opening the first credit card, which is a dummy equal to one if the household did not have a credit card in t-1, and have one or more credit cards in t. Alternatively, entry is defined as a dummy equal to one if the balance becomes positive between t and t -1. Both measures attempt to capture the entry of new borrowers to the credit card market. Finally, to measure delinquency, this is a dummy equal to one if household i is delinquent at time t, t+1, t+2, and t+3 respectively, and the regressions are estimated separately. Therefore, the estimated coefficient represents an intent-to-treat effect, as the same individual may be affected by the change in the levels of protection more than once during our sample period. Finally, we look at changes in the levels of self-employment to explore the effect on real outcomes. For this we use specification (1) but in this case, using the change in total county self-employment as a left hand side variable, or the change in self- employment in an industry and county between t and t-1.

1.6 Results and discussion

1.6.1 Bankruptcy Protection and Household Leverage and Interest Rates

We find that growth in bankruptcy protection leads to an increase in the level of credit card debt held by households (unsecured debt) between 1999 and 2005 (Table 1.3 ). Moreover, the increase in protection has no effect on other types of secured debt (auto and mortgage, Table 1.4 and 1.5)3.

3 4 The average effect is only present in the pre-bankruptcy reform period, when filing for bankruptcy was easier and cheaper (Table B8). If the cost of filing for bankruptcy increases enough, the effective protection is smaller, decreasing the ex ante benefit of increasing the amount of debt today. Considering that there is evidence that household bankruptcy filings are highly sensitive to

32 A possible concern may be that states which did not change the level of protection within our sample period are not a good control group, as they could be systematically different from the group which did opt to change their level of bankruptcy protection, and this would therefore invalidate our empirical inference. To overcome this concern, we replicated our main specification (Table 1.3 column 1), focusing only on the states in which changes in protection levels were implemented in our sample period (i.e. "eventually" treated, Table 1.3 column 6). In this case the main effects we estimate are basically unchanged, mitigating the endogeneity concern about the changes. Tables 1.10 and 1.11 replicates our main specification, but using interest rates changes as a dependent variable for personal unsecured loans, credit cards, and mortgage rates. The results show that the increase in bankruptcy protection leads to an increase in the level of interest rates for unsecured loans, but does not affect mortgage rates. These results suggest a demand driven credit market equilibrium, as we observe increases in quantities, and prices. Furthermore, in Table 1.6, columns 1 and 2, we look at the correlation between the levels of protection and contemporaneous and lag levels of determinants, which in a traditional view would be seen as driving the changes in the level of protection. Empirically, levels seems to be correlated with housing price and bankruptcy filing rates, which is consistent with evidence that cross-sectional variation in the level of protection is a state specific characteristic. Furthermore, Table 1.6, columns 3 to 6, looks at how changes in the levels of exemptions correlates with change in the determinants above, using an OLS estimation clustering standard errors at the state level, or running a linear probability model of the likelihood of change. In both cases, lag change in the candidates' determinants have no predictive power on changes in the level of protection. This is consistent with our identification assumption, that the timing of the changes is exogenous to characteristics which define the supply and demand of credit. While our results support the empirical strategy, there are alternative hypotheses that we need to rule out as explaining our results. First, changes in the level of protection could be correlated with pre-existing state specific trends that survive our controls, and thus our results are a reflection of these differential pre-trends rather than changes in the levels of protection. For example, states which increase their protection levels are states in which employment conditions are booming in the period prior to the change in protection levels. Table 1.7 looks at the effect of changes in protection when lags and leads of the changes are incorporated into the main specification; the first 4 columns show the specification without fixed effect, the second sets out with state fixed effect, and the last one with county fixed effect. These results show that our estimates are not affected by the inclusion of lag changes in the levels of protection, and that the coefficient in the lags is economically small and statistically insignificant 35 . Furthermore, the coefficients in the leads are increasing and statistically significant, especially for two periods after the change, which suggests

liquidity constraint (Gross et al., 2013), we should expect the effect to be weaker or nonexistent during the post period. 3 5 Considering that our exogenous variation is at the state level, we cannot control for state-time unobserved heterogeneity that is contemporaneous to our effect.

33 that there may be an overreaction of households to the changes in the first year and a long term effect that continues up to year two. Table 1.3 shows that the effect is concentrated in credit card debt (unsecured). This allows us to rule out the alternative explanation that our strategy is picking up state specific credit market trends that are correlated with the changes in protection and that can be confounded with our identified effect. Table 1.9 shows the effect is stronger in counties that are in the lowest tercile of the within state income distribution, monotonically decreasing as the level of income increases. It is expected that lower-income areas may be more affected by increases in protection, as the impact of the improvement in risk sharing should be more significant. Homeowner households should be more affected by the changes in the level of protection, as a big proportion of their protection comes from home equity protection. However, county level homeownership is correlated with income, so in order to gain a meaningful perspective on this variation, we look at the within income group variation on county level homeownership. Table 1.9 column 3 shows that the differential effect is aligned with the prediction, as the estimated coefficient for these particular areas almost triples with respect to the baseline specification. Following the same logic, we look at the within income group variation on bank concentration - a measure based on share of deposit holding at the branch level. Ta- ble 1.16column 2 shows that the effect is stronger in areas where markets are more concentrated, which is consistent with the Peterson and Rajan (1995) relationship lending model, where creditors are more likely to finance a credit constrained borrower when credit markets are concentrated because it is easier for these creditors to internalize the benefits of assisting these borrowers. Another alternative explanation of our finding is that the increase in quantities is due to a contemporaneous decrease in prices, which correlates with the timing of the changes in bankruptcy protection. In other words, areas which increased the level of protection were areas in which credit became cheaper. As mentioned above, Tables 1.10 and 1.11 show that the increase in bankruptcy protection leads to an increase in the level of interest rates for unsecure loans, not affecting mortgage rates. These results support our causal interpretation of the results, alleviating the concern that we are picking up a relaxation in the price of credit leading to an increase in quantities. Local economic conditions could produce spurious effects due to geographical heterogeneity that is uncorrelated with changes in the levels of protection. To overcome this endogeneity, we compare neighboring county-pairs within the same income bucket. Table 1.8 shows that when focusing on a county-pair in the same income bucket, the estimated results are very similar to the main specification. Moreover the effect is stronger when we concentrate on county-pairs in the lower end of the county income distribution.

1.6.2 Robustness Test We choose a first difference specification with county fixed effect to parsimoniously account for county level linear trends, and to account for multiples treatment for the

34 same state across time. However, in Table 1.1 Panel A, we show that our estimation is the same if we exclude county fixed effect, and change them by state level fixed effect or run debt levels on protection level with county fixed effect. In other words, our effect is invariant to the specific difference in difference specification. Table 1.1 shows how the effect changes with different measures of the treatment. We choose to use an intensity of treatment measure as our treatment; however, as Table 1.1 Panel A shows, our results are invariant to the use of only large changes, use of exemption dummies instead of the intensity of treatment, or if we restrict the analysis to only states which change their level of protection only once. Given the nature of our empirical strategy, as we argue before, time-varying changes at state levels may be omitted variables explaining our results; one candidate is the level of unemployment insurance in each state (Hsu et al., 2012). Table 1.15 shows that the inclusion of this variable has no impact on the estimated coefficient. The results are also robust to change, the depend variable for changes in debt to income, or percentage changes, or to replace the treatment only by the amount of homestead protection. Finally, all the results exclude DC, because within our sample period, this state changed the protection from a very low level to an unlimited level. If we include a time-varying dummy to account for this extreme change in the level of protection, Table 1.15 shows that it generates a decrease in the level of debt available to households, consistent with the empirical prediction of our model.

1.6.3 Magnitude of the effect In terms of magnitude, we find that the average county in our relevant period (1999- 2005) has a credit card balance of 290 million dollars, and the average increase in credit card debt is 7.6%. Our main estimate explains 10% of this balance growth. This magnitude represents the average treatment effect over the entire population. However, we believe that our effect is driven mostly by people close to financial distress, for whom the possibility of filing for bankruptcy is a real one. When we estimate the magnitude of the effect for the particular subgroup of areas, counties in the low-income tercile with higher homeownership percentage, we find that the effect now explains between 34% and 47% of the increases in their credit card balance. This heterogeneity is consistent with our interpretation that there is only one subset of people affected, e.g., homeowners within a county close to distress level on their credit cards. However, there is also the possibility that our estimates are biased downward (attenuation biased), due to measurement errors in our variables

1.6.4 Borrowers, Delinquency and Self-Employment Important remaining questions to address, include which households are expanding the amount of credit they hold, how they are doing so, and what their ex post conduct may be. Using individual level data to look at the ex ante and ex post behavior of households, first we replicate the county level results focusing on areas that are below the median county income. Table 1.12 Panel A shows that the effect of changes in protection is similar to those found when we focus on the lower end of the county

35 level distribution or county borders. When we focus on homeowners, defined as an individual for whom we observe home-related debt at some point between 1999 and 2005, the effect is stronger, which again is consistent with the county estimates (Table 1.12 Panel B). Furthermore, using detailed account information, we show in Table 1.12 columns 2-4, that changes in protection causally increase the number of credit cards per household; this increase is stronger among households that had ex ante credit card accounts. Even more interestingly, the increase in number of credit cards is stronger for households that also had a positive balance. This finding suggests that the credit expansion is due to existing borrowers acquiring more credit. Finally, Table 1.12 columns 5-6, show how changes in protection are uncorrelated with entry into the credit card market, defined as the time when a member of a household opens their first account, or as the time when their credit card balance goes from zero to positive. All these results provide evidence that in this sample, the effect is being driving by existing debtors expanding their current balance or their number of accounts, rather than new households entering the credit market. Focusing on the same sample, we explore their delinquency behavior up to three years after the increase in credit card usage induced by the change in protection. Three years is a long time frame when considering holdings on a credit card. Table 1.13 shows that within this sample there is no measurable increase in the level of delinquency; if anything, the probability of individuals becoming delinquent in the future decreases. If the households which are increasing their level of debt are over- borrowing, or taking on more risky projects, we would expect delinquency rates to increase. Although we cannot completely rule out an over-borrowing behavior, the results described are more consistent with risk-averse borrowers increasing their debt holding in response to the greater insurance received from the increase in protection. We show that areas which experienced an increase in the level of credit card debt also experienced an increase in the level of self-employment creation, specifically within industries that make more use of credit cards as start-up capital. Table 2.6 shows that, on average, the increase in self-employment is only positively correlated with the changes in the level of protection in low-income regions. Also, the estimated effect is stronger when we focus on industries for which credit card debt is an important source of financing (for example, construction or photography). It is important to point out that these outcome variables are only suggestive evidence of the real effect of the increase on the level of unsecured debt. Taking all this evidence together, the rise in credit card debt induced by the increase in the level of protection could have led to an increase in small business creation, and a decrease (or no increase) in the delinquency rates of unsecure creditors. The individual results seems to suggest that the channel driving the demand effect is consistent with a large impact from the insurance channel on existing borrowers, as we do not observe increases in the entry rates of new borrowers and ex post delinquencies within our micro level sample. Although this evidence is only suggestive, it highlights the important potential benefits of increasing the level of bankruptcy protection, especially for people in areas on the lower end of the wealth distribution, for which the insurance effect is more significant.

36 1.7 Conclusion

Overall, the evidence we present in this paper identifies the causal effect of the increase in the level of protection under personal bankruptcy on household leverage. We show that increases in the level of bankruptcy protection within our sample period, leads to an expansion in the levels of credit card debt that is stronger in counties that are in the lowest tercile of the within state income distribution, and monotonically decreasing as the level of income increases. Using micro level data we find that the expansion is concentrated among existing borrowers. This expansion is also correlated with an increase in small business creation, and seems to have no effect on counties' overall delinquency rates. These findings highlight the importance role that personal bankruptcy laws play as an insurance mechanism, providing down side protection especially for low-income regions. Therefore, the documented credit increase has important implications for our understanding of personal bankruptcy protection as a risk-sharing improving policy.

37 1.8 Bibliography

Acharya, Viral V., and Krishnamurthy V. Subramanian. Bankruptcy codes and innovation. Review of Financial Studies 22.12 (2009): 4949-4988. Acharya, Viral V., Rangarajan K. Sundaram, and Kose John. Cross-country variations in capital structures: The role of bankruptcy codes. Journal of Financial Intermediation 20.1 (2011): 25-54. Acharya, Viral V., Yakov Amihud, and Lubomir Litov. Creditor rights and corporate risk-taking. Journal of Financial Economics 102.1 (2011): 150-166. Adelino, Manuel, Antoinette Schoar, and Felipe Severino. House Prices, Collateral and Self-Employment. No. w18868. National Bureau of Economic Research (2013). Agarwal, Sumit, et al. Regulating Consumer Financial Products: Evidence from Credit Cards. No. w19484. National Bureau of Economic Research (2013). Aghion, Philippe, Oliver Hart, and John Moore. The economics of bankruptcy reform. Journal of Law Economics and Organization. 8 (1992): 523-546. Alvarez, Fernando, and Urban J. Jermann. Efficiency, equilibrium, and asset pricing with risk of default. Econometrica 68.4 (2000): 775-797. Baird, Douglas, and Robert Rasmussen. The End of Bankruptcy. Stanford Law Review 55 (2002). Benmelech, Efraim, and Nittai K. Bergman. Bankruptcy and the collateral channel. The Journal of Finance 66.2 (2011): 337-378. Berger, Allen N., Geraldo Cerqueiro, and Mara F. Penas. Does debtor protection really protect debtors? Evidence from the small business credit market. Journal of Banking & Finance 35.7 (2011): 1843-1857. Bernanke, Ben, and Mark Gertler. Agency costs, net worth, and business fluctu- ations. The American Economic Review (1989): 14-31. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan. How much should we trust differences-in-differences estimates? The Quarterly Journal of Economics 119.1 (2004): 249-275. Bolton, Patrick, and David S. Scharfstein. Optimal debt structure and the number of creditors. Journal of Political Economy (1996): 1-25. Cerqueiro, Geraldo, and Maria Penas. How does personal bankruptcy law affect start-ups? Working Paper (2011). Cerqueiro, Geraldo, Deepak Hegde, Maria Penas, and Robert Seamans. Personal Bankruptcy Law and Innovation. Working Paper (2013). Chaney, Thomas, David Sraer, and David Thesmar. The Collateral Channel: How did Real Estate Shocks Affect Corporate Investment. American Economic Review102 (6) (2012): 2381-2409. Chang, Tom, and Antoinette Schoar. Judge Specific Differences in Chapter 11 and Firm Outcomes. National Bureau of Economic Research (2013). Chatterji, Aaron K., and Robert C. Seamans. Entrepreneurial finance, credit cards, and race. Journal of Financial Economics 106.1 (2012): 182-195. Davydenko, Sergei A., and Julian R. Franks. Do bankruptcy codes matter? A study of defaults in France, Germany, and the UK. The Journal of Finance 63.2

38 (2008): 565-608. Djankov, Simeon, Caralee McLiesh, and Andrei Shleifer. Private credit in 129 countries. Journal of financial Economics 84.2 (2007): 299-329. Dobbie, Will, and Jae Song. Debt Relief and Debtor Outcomes:. Measuring the Effects of Consumer Bankruptcy Protection. Harvard University mimeo, (2013). Dube, Arindrajit, T. William Lester, and Michael Reich. Minimum wage effects across state borders: Estimates using contiguous counties. The Review of Economics and Statistics 92.4 (2010): 945-964. Elias, Stephen R. The New Bankruptcy: Will it Work for You?. Nolo, (2005). Fan, Wei, and Michelle J. White. Personal Bankruptcy and the Level of En- trepreneurial Activity Journal of Law and Economics 46.2 (2003): 543-567. Flynn, Ed, Gordon Bermant, and Karen Bakewell. A closer look at elderly chapter 7 debtors. American Bankruptcy Institute Journal. 21 (2002): 22-49. Gertner, Robert, and David Scharfstein. A Theory of Workouts and the Effects of Reorganization Law. The Journal of Finance 46.4 (1991): 1189-1222. Giannetti, Mariassunta. Do better institutions mitigate agency problems? Evi- dence from corporate finance choices. Journal of Financial and Quantitative Analysis 38.1 (2003): 185-212. Gropp, Reint, John Karl Scholz, and Michelle J. White. Personal bankruptcy and credit supply and demand. The Quarterly Journal of Economics 112.1 (1997): 217-251. Gross, David B., and Nicholas S. Souleles. Do liquidity constraints and interest rates matter for consumer behavior? Evidence from credit card data. The Quarterly Journal of Economics 117.1 (2002a): 149-185. Gross, David B., and Nicholas S. Souleles. An Empirical Analysis of Personal Bankruptcy and Delinquency. Review of Financial Studies, 15.1 (2002b) 319-347. Gross, Tal, Matthew J. Notowidigdo, and Jialan Wang. Liquidity constraints and consumer bankruptcy: Evidence from tax rebates. Review of Economics and Statistics (2012). Hart, Oliver and John Moore. A Theory of Debt Based on the Inalienability of Human Capital, The Quarterly Journal of Economics, 109.4 (1994) 841-79 Hart, Oliver, and John Moore. Foundations of incomplete contracts. The Review of Economic Studies 66.1 (1999): 115-138. Hart, Oliver, Rafael La Porta, Florencio Lopez-de-Silanes, and John Moore. A new bankruptcy procedure that uses multiple auctions. European Economic Review 41.3 (1997): 461-473. Haselmann, Rainer, Katharina Pistor, and Vikrant Vig. How law affects lending. Review of Financial Studies 23.2 (2010): 549-580. Heider, Florian, and Alexander Ljungqvist. As Certain As Debt And Taxes: Estimating The Tax Sensitivity Of Leverage From Exogenous State Tax Changes. No. w18263. National Bureau of Economic Research (2012). Hsu, Joanne, David Matsa, and Brian Melzer. Unemployment Insurance and Consumer Credit. Working Paper (2012). Hynes, Richard M., Anup Malani, and Eric A. Posner. The Political Economy of Property Exemption Laws. Journal of Law and Economics 47.1 (2004): 19-43.

39 Jagtiani, Julapa, and Wenli Li. Credit access and credit performance after consumer bankruptcy filing: new evidence. No. 13-24. (2013). Kehoe, Timothy J., and David K. Levine. Debt-constrained asset markets. The Review of Economic Studies 60.4 (1993): 865-888. Kiyotaki, Nobuhiro, and John Moore. Credit Chains. Journal of Political Econ- omy 105.21 (1997): 211-248. La Porta, Rafael, Florencio Lopez-de-Silanes, Andrej Shleifer, and Robert W. Vishny. Law and Finance. Journal of Political Economy 106 (1998): 1113-1155. Lee, Donghoon and Wilbert van der Klaauw. An Introduction to the FRBNY Consumer Credit Panel. Federal Reserve Bank of New York Staff Report no. 479. (2010) Lefgren, Lars, and Frank McIntyre. Explaining the Puzzle of Cross-State Differ- ences in Bankruptcy Rates. Journal of Law and Economics 52.2 (2009): 367-393. Levine, Ross. The legal environment, banks, and long-run economic growth. Jour- nal of Money, Credit and Banking (1998): 596-613. Li, Wenli, Michelle J. White, and Ning Zhu. Did Bankruptcy Reform Cause Mortgage Default to Rise?. No. w15968. National Bureau of Economic Research, 2010. Lilienfeld-Toal, Ulf von, Dilip Mookherjee, and Sujata Visaria. The Distributive Impact of Reforms in Credit Enforcement: Evidence from Indian Debt Recovery Tribunals. Econometrica 80.2 (2012): 497-558. Lin, Emily Y., and Michelle J. White. Bankruptcy and the market for mortgage and home improvement loans. Journal of Urban Economics 50.1 (2001): 138-162. Link, Frederick. The Economics of Personal Bankruptcy. MIT Thesis (2004) Martin, Nathalie, and Stewart Paley. JK Lasser's The New Bankruptcy Law and You. John Wiley & Sons, 2006.' Mian, Atif and Amir Sufi. 2011. House Prices, Home Equity-Based Borrowing, and the US Household Leverage Crisis. American Economic Review, 101 (2011): 2132-2156. Paravisini, Daniel. Local bank financial constraints and firm access to external finance. The Journal of Finance 63.5 (2008): 2161-2193. Petersen, Mitchell A., and Raghuram G. Rajan. The benefits of lending relationships: Evidence from small business data. The Journal of Finance 49.1 (1994): 3-37. Povel, Paul. Optimal soft or tough bankruptcy procedures. Journal of Law, Economics, and Organization 15.3 (1999): 659-684. Pulvino, Todd C. Do asset fire sales exist? An empirical investigation of commercial aircraft transactions. The Journal of Finance 53.3 (1998): 939-978. Qian, Jun, and Philip E. Strahan. How laws and institutions shape financial contracts: The case of bank loans. The Journal of Finance 62.6 (2007): 2803-2834. Rampini, Adriano A., and S. Viswanathan. Collateral, risk management, and the distribution of debt capacity. The Journal of Finance 65.6 (2010): 2293-2322. Stango, Victor, and Jonathan Zinman. Limited and varying consumer attention: Evidence from shocks to the salience of bank overdraft fees. No. w17028. National Bureau of Economic Research, 2011.

40 Stiglitz, Joseph E., and Andrew Weiss. Credit rationing in markets with imperfect information. The American Economic Review 71.3 (1981): 393-410. Stromberg, Per. Conflicts of interest and market illiquidity in bankruptcy auctions: Theory and tests. The Journal of Finance 55.6 (2000): 2641-2692. Vig, Vikrant. Access to collateral and corporate debt structure: Evidence from a natural experiment. The Journal of Finance 68.3 (2013): 881-928.

41 1.9 Appendix A. Model of Effect of Bankruptcy Protection on Household Borrowing

To explore the previous explanation, gain further insights into the effects of changes in the bankruptcy reforms on the supply of credit, and to guide the empirical analysis, we provide a simple model of the credit market where we abstract from considering the moral hazard and adverse selection behavior of borrowers. In our model, we highlight the effect of the increase of partial insurance provided by bankruptcy protection in the credit market equilibrium outcome, and how even in the absence of asymmetric information we could observe a demand effect. We do this using a two period model, where the agent needs to borrow in order to consume at period 1. Formally, the agent will consume c, at t=O and ci(s) at t=1, where s C {B, G} (good and bad states in t=1), with the correspondent probability {p, 1 - p} The agent is endowed with a wealth only at t=1, his wealth is a combination of home equity H (exogenous), and income y. For simplicity, assume that income follows a binomial distribution given by y(G) = W > 0 and y(B) = 0 . Exists a level of protection P (exogenously determinate) The agent's consumption will be given by

co = b

ci = y + H - Min{(1 + R)b, y + Max(H - P,0)} where R is endogenously determined

Agent's Maximization Problem

Given this setup, the agent will solve the following problem

V(b) = Max u(co) + E[u(c)]

Subject to the consumption above. Therefore, the agent's consumption in period 2 will be given by:

* No default, total repayment: ci = y + H - (1 + R)b

* Default and home-equity is not fully protected (H - P) > 0: c1 = P

* Default and home-equity is fully protected (H - P) < 0: ci = H

42 Bank's break even condition

It is given by

(1 +r)b = E[Min{(1 + R)b,y + Max(H - P,0)}] where r is the risk free rate (exogenous). The payoff for the bank are given by:

* No default, total repayment: b(1 + R)

" Default and home-equity is not fully protected: y + H - P

" Default and home-equity is fully protected: y

Consider a risk-averse agent, u(x) = ln(x), the solution of the problem above defines three regions as a function of the level of protection. Figure 1-4 illustrate the shape of the numerical solution using the following set of parameters r = 0.05, /3 = 0.925, p = 0.5, W = 5k. Fixed borrowing (between 0, P): There is no default; banks lend at a risk-free rate and the borrower demands a fixed quantity not related to the level of protection. Increase in borrowing (between P, P*): There is a probability of default greater than zero, interest rates go up, but quantities go up too. The agent's marginal utility of consumption at t = 0 is greater than the marginal cost in the good state, conditional on the level of protection on the bad state, that ensure a given level of consumption. Decrease in borrowing (between P*, P): The probability of default increases, and interest rates go up even more. Agents will decrease the equilibrium amount of debt with respect to the previous region, and the marginal cost in the good state overcomes the benefit of consumption today, given the level of protection in the bad state.

43 Figure 1-1: Debt Growth and Bankruptcy Filings

This figure plots the yearly number of non-business filings in the US from 1994 until 2012 extracted from the Statistics Division of the Administrative Office of the United States Courts, and the adjusted total revolving debt in the US extracted from the Federal Reserve Board of Governors Consumer Credit Report.

8 N4

1994 IM9 1998 20DO 2002 2004 2006 2008 2010 2012

Yearly non-businms fifins US - Consumer Revo"vn Debt US

44 Figure 1-2: States that Changed their Level of Bankruptcy Protection

This figure shows in dark the counties that were at some point treated between 1999 and 2005; "eventually" treated, in other words the level of bankruptcy protection changed at some point during that period. Lightly colored counties are the counties in which the level never changed, "never"' treated. Counties in gray represent counties for which FRBNY Consumer Credit Panel/Equifax did not provide information because their population was below 10,000 households during our sample period.

4.-

45 Figure 1-3: Iustration of Different Demand and Supply Responses

This figure uses supply and demand curves to illustrate possible net effects. Baseline Equilibrium is the initial equilibrium before the change. Increase in Price, No Increase in Q, show the effect when the supply response totally and perfectly upsets the demand increase. Increase in Price, Decrease in Q, show the effect when the supply response is stronger than the demand increase. Increase in Price, Increase in Q, show the effect when the demand effect dominates. P Da s Do So Do------= so So Pa -

Qo Qo=Qa Baseline Equilibrium Q Increase in Price, No Increase in Q Q

P Db Sb Dc Sc Do so Do so Pb

....- .- Pc ....---- -..-.- ..--..- Po PO

s c, QC Q Q Increase in Price, Decrease in Q Increase in Price, Increase in Q

46 Figure 1-4: Ilustration of a Solution of the Model This figure shows a stylized, schematic solution of the path obtained by solving numerically the model in Appendix A; the top figure shows the relationship between the debt amount and protection levels. The bottom figure shows the relationship between price and protection levels.

0 E

Crdi ExasoardtCnrcio ee fpoetoP

Credit Expansion Credit Contation Level of protection (P)

47 Table 1.1: Summary Statistics Data

All S ample Eventually Treated Never Treated N=1 5,519 N=7,091 N=8,428 Levels Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.

Debt to Income (DTI) 1.23 0.48 1.29 0.52 1.18 0.45 Mortgage Debt to Income (MTI) 0.90 0.45 0.97 0.49 0.84 0.41 Credit Card Debt to Income (CCTI) 0.16 0.04 0.16 0.04 0.17 0.05 Auto Loan Debt to Income (ATI) 0.17 0.06 0.16 0.06 0.17 0.07

County Total Debt (bil. USD) 2.89 10.51 3.93 13.95 2.01 6.18 County Mortgage Debt (bil. USD) 2.33 9.01 3.25 12.08 1.57 5.06 County Credit Card Debt (bil. USD) 0.29 0.83 0.36 1.03 0.22 0.61 County Auto Debt (bil. USD) 0.26 0.76 0.32 0.92 0.22 0.60

Pers. Unsec. Int. Rate (bp) 12.8 2.2 12.8 2.2 12.9 2.2 Credit Card hit. Rate (bp) 13.1 2.7 13.4 2.7 12.8 2.7 * 30 yr Fix. Mtg. Int. Rate (bp) 6.6 0.7 6.6 0.7 6.6 0.7

Mortgage Delinquency ('/ of pop) 1.5 1.3 1.5 1.2 1.6 1.3 Credit Card Deliquency (W of pop) 8.2 3.5 7.8 3.1 8.5 3.8 Auto Delinquency (W of pop) 2.4 1.5 2.3 1.4 2.4 1.5

County Household Population 100.306 269,477 123,735 331.573 80.594 200.934 IRS County Income (bil. USD) 1.90 5.56 2.46 6.85 1.43 4.11 * Unemployment Rate 5.32 1.90 5.35 1.87 5.30 1.93 No. of Bankruptcy Filing (1998) 604 2051 W of Owner Occupancy (2000 ) 73.35 7.84

N=13,302 N=6,078 N=7,224 Changes Mean Std. Dev. Mean Std. Dev. Mean Std. Dev.

DTI Change 0.099 0.113 0.101 0.109 0.098 0.116 MTI Change 0.115 0.149 0.115 0.145 0.115 0.151 CCTI Change 0.051 0.118 0.053 0.112 0.049 0.124 ATI Change 0.098 0.156 0.096 0.146 0.101 0.165

Total Debt Growth 0.122 0.091 0.123 0.089 0.122 0.092 Mortgage Debt Growth 0.133 0.120 0.133 0.119 0.133 0.120 Credit Card Debt Growth 0.076 0.099 0.078 0.093 0.075 0.104 Auto Debt Growth 0.117 0.125 0.115 0.118 0.119 0.130

Pers. Unsec. it. Rate Change (bp) -0.09 0.94 -0.12 0.93 -0.06 0.95 Credit Card JIt. Rate Change (bp) -0.75 1.88 -0.65 1.84 -0.84 1.91 30-yr Fix. Mtg. Int. Rate Change (bp) -0.34 0.50 -0.34 0.49 -0.33 0.51

Income Growth 0.033 0.053 0.032 0.054 0.033 0.052 Unemployment Rate Change 0.111 0.963 0.115 0.931 0.108 0.989 House Price Growth 0.075 0.046 0.088 0.050 0.062 0.037 ** *

Note. "All Sample" refers to all counties in the sample period. "Eventually Treated" refers to counties treated during the sample period, that is, states that changed their level of protection during the sample period. "Never Treated" refers to counties not treated during the sample period. County Debt (in bil. USD) for mortgage, credit card and auto loans, is obtained from the FRBNY Consumer Credit Panel/Equifax. IRS County Income (in bil. USD) is measured as total wages and salary in that county. Debt to Income is constructed using the two county measures described above. Personal unsecured, credit card, and 30-year fixed mortgage rates are constructed from branch-setter level rates from Rate-Watch. Delinquency rates for mortgage, credit card, and auto loans are from the FRBNY Consumer Credit Panel/Equifax, and represent the fraction of households that are 90+ days delinquent. County household population is the number of household per county and year in the FRBNY Consumer Credit Panel/Equifax. No. of Filings is the number of non-business filings in a county in 1998 from the American Court System. % of Owner Occupancy is the percentage of home ownership in a county in 2000 from the Census Bureau. For a complete description of the data sources see section 3.1. Data Description. House price growth is extracted from the Federal Housing Finance Agency (FHFA) House Price Index (HPI) data at a state level. The number of observations refers to the number of county-year observations. Almost all variables are available for every county (2,218), with the exception of interest rates, which are only available for (1232, 1323 and 1340 counties respectively). *, **, and *** denotes significance at the 10%, 5%, and 1% level cluster at the state level for the mean differences between "Eventually Treated" and "Never Treated" sample. The sample period is from 1999 to 2005.

48 Table 1.2: Summary Statistics Protection Level

5 7 5 9 5 All Sample Mean Std. Dev. p5 p 2 p50 p p

Protection Level 73,627 75,125 13,000 23,200 55,800 166,200 unlimited Homestead 63,932 73,356 7,500 20,000 40,000 150,000 unlimited Personal Assets 9,695 5,965 2,900 5,000 8,400 11,000 25,000 Unlimited States 7 No. of States 50

2 5 7 5 9 5 Eventually Treated Mean Std. Dev. p5 p p50 p p

Protection Level 85,655 86,100 11,000 32,300 51,000 110,300 390,000 Homestead 75,243 84,838 0,000 25,000 40,000 100,000 350,000 Personal Assets 10,411 6,061 3,000 7,200 9,100 11,000 25,000 No. of States 26

Protection Changes 38,841 52,992 2,000 3,250 15,400 50,000 200,000 No. of Changes 37

2 5 7 5 9 5 Never Treated Mean Std. Dev. p5 p p50 p p

Protection Level 56,922 52,366 14,400 20,700 57,700 586,000 unlimited Homestead 48,222 49,678 10,000 13,750 45,000 575,000 unlimited Personal Assets 8,700 5,705 2,900 4,800 6,300 12,300 42,000 No. of States 24 Note. "All Sample" refers to all counties in the sample period. "Eventually Treated" refers to counties treated during the sample period, that is states that changed their level of protection during the sample period. "Never Treated" refers to counties not treated during the sample period. Protection Level is the nominal value of household protection under Chapter 7. Homestead is the amount of home-equity protected under Chapter 7. Personal Assets, is the amount of assets protected under Chapter 7, such as, books, furniture, jewelry, etc. The exact description depends on the state. Unlimited States is the number of states with unlimited home-equity protection during our sample period. Protection Changes is constructed based on the yearly changes in the level of protection. Levels of protection and homestead are different at 10% between "Eventually Treated" and "Never Treated". The sample period is from 1999 to 2005.

49 Table 1.3: Effect of Bankruptcy Protection on Debt. Credit Card Debt

Changes Levels Level County State No Controls + Level on Level on Linear Linear Linear Level Inc-Year Eventually Changed Change Dummy Level Level Trend Trend Trend Controls Uep-Year Treated Once > 0.15 Treatment County FE State FE (1) (2) (3) (4) (5) (6) (7) (8) (10) (11) (7) ~(9)(1) 1) Protection 0.018 0.019** 0.018** 0.017** 0.017** 0.017** 0.022** 0.018** 0.012*** Growth s,t (0.008) (0.008) (0.007) (0.007) (0.008) (0.008) (0.009) (0.008) (0.004)

Protection 0.023** 0.027** Level s,t (0.011) (0.013)

Unemployment 0.002 0.003 0.002 0.000 0.003 0.002 0.002 0.002 0.002 Rate Change (0.002) (0.002) (0.002) (0.002) (0.003) (0.003) (0.002) (0.002) (0.002)

House Price -0.102 -0.109 -0.139*** -0.203** -0.183* -0.118 -0.049 -0.103 -0.105 Index Growth (0.086) (0.085) (0.037) (0.102) (0.099) (0.086) (0.108) (0.086) (0.083)

Incomue 0.079* 0.134*** 0.142*** 0.073* 0.088** 0.138* 0.081 0.079* 0.079* Growth (0.047) (0.041) (0.040) (0.041) (0.041) (0.077) (0.051) (0.047) (0.047)

Unemployment 0.005* 0.002 0.003 0.007* Rate (0.003) (0.003) (0.003) (0.004)

House Price 0.083*** 0.070** -0.166*** -0.263*** (0.031) (0.029) (0.042) (0.053)

Incomue 0.023 0.010 0.251*** 0.951*** (0.021) (0.020) (0.047) (0.006)

No. of Obs. 13,302 13,302 13.302 13,302 13,302 6,078 11,478 13,302 13,302 15,519 15,519 No. of Clusters 50 50 50 50 50 26 39 50 50 50 50 County FE y y y y y y y y State FE y Y Year FE y Y y y y y y y y y y R-Squared 0.30 0.29 0.28 0.30 0.31 0.29 0.30 0.30 0.30 Notes. This table shows the estimated coefficient following specification (1) of log changes to credit card debt on log changes in bankruptcy protection at the county level. FRBNY Consumer Debt county data is from the Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Protection Level is the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the income log change in county i at time t from IRS. Columns 1 and 2 show the result using county and state fixed effects respectively in the first difference specification. Column 3 shows state or county fixed the results if we exclude effect from specification (1). Column 4 shows the estimates including level of the controls. Column 5 shows the estimates including level controls groups times year and income and unemployment fixed effect, to allow for differential trends across states based on these observable characteristics. Column 6 shows the estimates for a regression that only uses states treated during the sample period, that is, states that changed level their of protection during the sample period. Column 7 shows the results if we only consider as treated state that changed once. Column 8 shows the estimates if we replace by zero changes below 0.15. Column 9 shows results if we replace the change with a dummy indicator that is one if the change is greater than zero. Columns 10 and 11 show the results of regression log levels of credit card debt on log levels of protection and including county and state fixed effect respectively. The sample period is from 1999 to 2005. *, * and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively. Table 1.4: Effect of Bankruptcy Protection on Debt. Mortgage Debt

Changes Levels Level County State No Controls + Level on Level on Linear Linear Linear Level Inc-Year Eventually Changed Change Dummy Level Level Trend Trend Trend Controls Uep-Year Treated Once > 0.15 Treatment County FE State FE (10) (11) (1) (2) (3) (4) (5) (6) (7) (8) (9) Protection 0.011 0.011 0.008 0.005 0.007 0.014 0.013 0.012 0.006 Growth s.t (0.012) (0.012) (0.015) (0.010) (0.010) (0.014) (0.013) (0.012) (0.007)

Protection 0.007 0.006 Level st (0.031) (0.026)

Unemployment -0.004 -0.003 -0.003 -0.004 0.000 -0.001 -0.005* -0.004 -0.004 Rate Change (0.003) (0.003) (0.003) (0.003) (0.003) (0.002) (0.003) (0.003) (0.003)

House Price 0.086 0.078 0.044 -0.378** -0.345** 0.128 0.046 0.086 0.084 Index Growth (0.161) (0.161) (0.079) (0.170) (0.174) (0.256) (0.209) (0.161) (0.161)

Income 0.114 0.185** 0.191** 0.039 0.060 0.208 0.125 0.114 0.114 Growth (0.107) (0.091) (0.091) (0.079) (0.081) (0.181) (0.114) (0.107) (0.107)

Uneniployient 0.000 -0.004 0.001 -0.055*** Rate (0.004) (0.004) (0.004) (0.007)

House Price 0.278*** 0.265*** 0.013 -0.223** (0.041) (0.041) (0.069) (0.089)

Income 0.133*** 0.105*** 0.319*** 1.123*** (0.039) (0.036) (0.067) (0.012)

No. of Obs. 13.302 13,302 13.302 13,302 13.302 6,078 11,478 13.302 13.302 15.519 15,519 No. of Clusters 50 50 50 50 50 26 39 50 50 50 50 County FE Y Y Y Y Y Y Y y y State FE Y Year FE Y Y Y Y Y Y Y Y Y Y Y 0.86 0.97 R-Squared 0.09 ().1() 0.08 0.11 0.13 0.11 0.09 0.09 0.09 is Notes. This table shows the estimated coefficient following specification (1) of log changes to mortgage debt on log changes in bankruptcy protection at the county level. Debt county data from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Protection Level is the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth, is the income log change in county i at time t from IRS. Columns 1 and 2 show the result using county and state fixed effects respectively in the first difference specification. Column 3 shows the results if we exclude state or county fixed effect from specification 1. Column 4 shows the estimates including level of the controls. Column 5 shows the estimates including level controls and income and unemployment groups times year fixed effect, to allow for differential trends across states based on these observable characteristics. Column 6 shows the estimates for a regression that only uses states treated during the sample period, that is, states that changed their level of protection during the sample period. Column 7 shows the results if we only consider as treated state that changed once. Column 8 shows the estimates if we replace by zero changes below 0.15. Column 9 shows results if we replace the change with a dummy indicator that is one if the change is greater than zero. Column 10 and 11, show the results of regression log levels of mortgage debt on log levels of protection and including county and state fixed effect respectively. The sample period is from 1999 to 2005. *, *, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively Table 1.5: Effect of Bankruptcy Protection on Debt. Auto Debt

Changes Levels Level County State No Controls + Level on Level on Linear Linear Linear Level Inc-Year Eventually Changed Change Dummy Level Level Trend Trend Trend Controls Uep-Year Treated Once > 0.15 Treatment County FE State FE (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Protection 0.009 0.009 0.009 0.009 0.013 0.009 0.009 0.010 0.002 Growth s,t (0.013) (0.013) (0.014) (0.012) (0.013) (0.012) (0.015) (0.013) (0.008)

Protection 0.000 0.007 Level s,t (0.024) (0.027)

Unemployment -0.005* -0.004 -0.005* -0.002 -0.005 -0.011*** -0.004 -0.005* -0.005* Rate Change (0.003) (0.003) (0.003) (0.003) (0.004) (0.003) (0.003) (0.003) (0.003)

House Price -0.005 -0.013 0.107** -0.104 -0.134 -0.230* 0.049 -0.005 -0.007 Index Growth (0.113) (0.113) (0.054) (0.124) (0.125) (0.118) (0.150) (0.113) (0.112)

Income 0.059 0.124*** 0.127*** 0.031 0.020 0.121*** 0.054 0.059 0.059 Growth (0.038) (0.032) (0.030) (0.032) (0.032) (0.043) (0.041) (0.038) (0.038)

Unemployment -0.011** -().009* -0.005 0.024*** Rate (0.005) (0.005) (0.004) (0.005)

House Price 0.009 0.033 0.107* 0.061 (0.043) (0.045) (0.055) (0.069)

Income 0.026 0.029 0.249*** 0.928*** (0.029) (0.030) (0.038) (0.1)08)

No. of Obs. 13,302 13,302 13,302 13,302 13,302 6,078 11,478 13,302 13,302 15,519 15,519 No. of Clusters 50 50 50 50 50 26 39 50 50 50 50 County FE Y Y Y Y Y Y Y Y State FE Y Y Year FE Y Y Y Y Y Y Y Y Y Y Y R-Squared 0.18 0.19 0.17 0.19 0.19 0.20 0.18 0.18 0.18 0.85 0.97

Notes. This table shows the estimated coefficient following specification (1) of log changes to auto debt on log changes in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Protection Level is the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the income log change in county i at time t from IRS. Columns 1 and 2 show the result using county and state fixed effects respectively in the first difference specification. Column 3 shows the results if we exclude state or county fixed effect from specification 1. Column 4 shows the estimates including level of the controls. Column 5 shows the estimates including level controls and income and unemployment groups times year fixed effect, to allow for differential trends across states based on these observable characteristics. Column 6 shows the estimates for a regression that only uses states treated during the sample period, that is, states that changed their level of protection during the sample period. Column 7 shows the results if we only consider as treated state that changed once. Column 8 shows the estimates if we replace by zero changes below 0.15. Column 9 shows results if we replace the change with a dummy indicator that is one if the change is greater than zero. Column 10 and 11 show the results of regression log levels of auto debt on log levels of protection and including county and state fixed effect respectively. The sample period is from 1999 to 2005. *, * and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively. Table 1.6: Determinants of Bankruptcy Protection Levels and Changes

Protection Level s,t Protection Growth s,t Protection Dummy s,t (1) (2) (3) (4) (5) (6) House Price/Growth st -3.900 -1.837*** -0.809** -0.537 -0.697 -0.858 (4.616) (0.671) (0.354) (0.572) (0.701) (0.789) House Price/Growth st-1 5.287 2.983*** 1.691*** 0.970 2.700*** 1.806* (4.503) (0.770) (0.619) (0.762) (0.776) (0.994)

Medical Exp./Growth st -3.332 0.836 -0.316 -1.150 -1.101 -2.380 (5.359) (1.001) (0.644) (0.821) (1.270) (1.834) Medical Exp./Growth st-i 4.635 0.348 -0.537 -1.805* -1.020 -2.274* (5.238) (1.106) (0.763) (1.001) (1.115) (1.287)

Unemp. Rate/Change s,t -0.023 0.028 0.005 0.002 0.027 0.026 (0.190) (0.036) (0.027) (0.033) (0.042) (0.048) Unermp. Rate/Change st-1 0.033 -0.081* -0.016 -0.008 -0.056 -0.058 (0.148) (0.042) (0.028) (0.032) (0.050) (0.065)

State Real GDP/Growth s,t 3.703 0.504 0.474 1.028 -1.665 -0.911 (4.464) (0.871) (0.668) (1.018) (1.034) (1.343) State Real GDP/Growth st-1 -6.950 -1.448 -0.277 0.425 -1.429 -0.547 (3.916) (0.742) (0.282) (0.457) (0.789) (0.802)

No. Filings/Growth st -0.299* 0.125* 0.030 -0.123 0.060* -0.114 (0.250) (0.039) (0.045) (0.098) (0.069) (0.098) No. Filings/Growth st-1 -0.482 0.194*** 0.053 -0.045 0.026 -0.080 (0.245) (0.072) (0.047) (0.071) (0.064) (0.090)

Political Climate st-1 0.045** -0.289*** 0.010 0.400 0.151 0.608 (1.509) (0.171) (0.161) (0.234) (0.151) (0.458)

Personal Income/Growth s,t 15.885* 1.077 1.554 0.996 3.264 3.190 (8.597) (1.257) (1.299) (1.928) (2.009) (2.399) Personal Income/Growth s,t-1 -13.235* -0.219* -0.720 -1.159 -0.525* -0.893 (9.202) (1.206) (0.929) (1.477) (1.849) (2.200)

No. of Obs. 350 350 300 300 300 300 State FE Y Y Y Year FE Y Y Y Y Y Y R2 0.13 0.12 0.07 0.22 0.13 0.25 Note. This table shows the estimated coefficient of regression of bankruptcy protection on contemporaneous and lag values of variables that could determinate the changes in protection levels. House Price s,t is the level or growth of house prices in state s at time t, from FHFA. Medical expenses is the level of growth in state's annual total medical expenses from the National Health Statistic. No. of Filings, is the number or change in the number of filings for non-business bankruptcies in a state. Political Climate s,t is defined as the share of democratic votes in the closer House of Representative election. State GDP and Personal Income are from BEA, and Unemployment Rate from BLS. Columns 1 and 2 show the coefficient of regressions of the protection level on levels of the explanatory variables using only year, and year and state fixed effect. Columns 3 and 4 show the coefficient of regressions of the growth in protection on growth of the explanatory variables using only year, and year and state fixed effect. Columns 5 and 6 show the coefficient of regressions of a dummy that is one if the growth in protection is greater than zero on the explanatory variables growth using only year, and year and state fixed effect. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level.

53 Table 1.7: Dynamics of the Change in Protection Levels on Credit Card Debt

1 Period 2 Periods No County County No County County Linear Trend Linear Trend Linear Trend Linear Trend Linear Trend Linear Trend (1) (2) (3) (4) (5) (6) Protection 0.001 -0.004 -0.005 Growth st-2 (0.019) (0.026) (0.025)

Protection -0.008 -0.010 -0.012 -0.007 -0.007 -0.010 Growth st-1 (0.008) (0.010) (0.009) (0.009) (0.015) (0.015)

Protection 0.018** 0.019** 0.016** 0.018** 0.022** 0.020** Growth st (0.007) (0.008) (0.007) (0.007) (0.009) (0.008)

Protection 0.002 0.006 0.006 0.003 0.010 0.010 Growth st+ (0.006) (0.008) (0.009) (0.006) (0.010) (0.011)

Protection 0.010** 0.016*** 0.016*** Growth st+2 (0.005) (0.005) (0.005)

Unemployment 0.002 0.002 0.000 0.002 0.002 0.001 Rate Change (0.002) (0.002) (0.002) (0.002) (0.002) (0.002)

House Price -0.139*** -0.108 -0.212** -0.142*** -0.120 -0.229** Index Growth (0.037) (0.085) (0.101) (0.037) (0.085) (0.100)

Income 0.143*** 0.080* 0.073* 0.143*** 0.080* 0.072* Growth (0.040) (0.047) (0.042) (0.040) (0.047) (0.041)

Unemployment 0.005* 0.004 Rate (0.003) (0.003)

House Price 0.085 0.086 (0.030) (0.029)

Income 0.024 0.025 (0.021) (0.021)

No. of Obs 13,302 13,302 13,302 13,302 13,302 13,302 No. of Clusters 50 50 50 50 50 50 County FE Y Y Y Y Year FE Y Y Y Y Y Y R-Squared 0.28 0.30 0.30 0.28 0.30 0.31 Note. This table shows the estimated coefficient following specification (1) of log changes to credit card debt on log changes in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the log change in income in county i at time t from IRS. Columns 1 and 4 show the without the inclusion of county fixed effects, including one lag and lead, and two lags and two leads. Columns 2 and 5 show the results with the inclusion of county fixed effect for including one lag and lead, and two lags and two leads, Columns 3 and 6 are the same than before but including level controls. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

54 Table 1.8: Local Business Conditions. Neighboring County-pairs across State Bor- ders. Credit Card Debt

All Equal Income Low Income County-Pairs County-Pairs County-Pairs State County State County State County Linear Liner Linear Liner Linear Liner Trend Trend Trend Trend Trend Trend (1) (2) (3) (4) (5) (6) Protection -0.006 -0.005 0.015 0.015* 0.099** 0.098** Growth s,t (0.011) (0.011) (0.010) (0.009) (0.046) (0.044)

Unemployment 0.003** 0.003** 0.002 0.001 0.002* 0.001** Rate Change (0.002) (0.002) (0.003) (0.003) (0.005) (0.005)

House Price -0.322** -0.317** -0.266 -0.261 -1.040* -1.037** Index Growth (0.157) (0.154) (0.178) (0.171) (0.550) (0.526)

Income 0.095*** 0.043 0.122* 0.066 0.121 0.102 Growth (0.024) (0.027) (0.071) (0.075) (0.125) (0.122)

No. of Obs 9,168 9,168 3,984 3,984 1,188 1,188 No. of Clusters 48 48 46 46 33 33 County FE Y Y Y State FE Y Y Y County-Pair-Year FE Y Y Y Y Y Y R-Squared 0.70 0.70 0.67 0.67 0.63 0.62

Note. This table shows the estimated coefficient following specification (2) of log changes in credit card debt on log changes in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the log change in income in county i at time t from IRS. Columns 1 and 2, show the estimates for state and county fixed effect for all neighboring county-pairs sample. Columns 3 and 4 show the results including state and county fixed effect for the sub-sample of neighboring county-pairs for which both counties are in the same income bucket. Columns 5 and 6 show estimates with state and county fixed effect for only the neighboring county-pairs in the same income bucket and in the lowest tercile of the income distribution. The sample period is from 1999 to 2005. *, *, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

55 Table 1.9: Heterogeneous Treatment of Bankruptcy Protection on Credit Card Debt: Income and Home ownership

Low Income Med Income High Income Income Home Home Home Ownership Ownership Ownership (1) (2) (3) (4) (5) (6) (7) Protection Growth s,t 0.007 0.028** 0.063*** 0.020** 0.029 0.006 0.014 (0.007) (0.011) (0.018) (0.010) (0.019) (0.006) (0.009)

Protection Growth st 0.022*** x Low Income (0.007)

Protection Growth s,t -0.050*** -0.012 -0.011 x Low Home Ownership (0.018) (0.025) (().009)

Protection Growth st 0.013** x Med Income (0.006)

Protection Growth st -0.049*** -0.014 -0.013 x Med Home Ownership (0.016) (0.019) (0.012)

Unemployment 0.003 0.005* 0.005* 0.002 0.002 0.002 0.002 Rate Change (0.002) (0.003) (0.003) (0.002) (0.002) (0.003) (0.003)

House Price -0.109 -0.015 -0.012 -0.099 -0.099 -0.208** -0.206** Index Growth (0.086) (0.094) (0.095) (0.098) (0.098) (0.092) (0.093)

Income 0.137*** 0.059** 0.057* 0.090*** 0.088*** 0.240*** 0.227*** Growth (0.040) (0.030) (0.031) (0.032) (0.028) (0.062) (0.064)

No. of Obs 13,302 4,536 4,536 4,422 4,422 4,344 4,344 No. of Clusters 50 50 50 50 50 50 50 State and Year FE Y Y Y Y Y Y Y R-Squared 0.29 0.24 0.24 0.29 0.30 0.46 0.48 Note. This table shows the estimated coefficient following a variation of specification (1) that incorporates interactions. Low/Med Income represents counties in the lowest/middle tercile of the within state income distribution. Low/Med Ownership represents counties in the lowest/middle tercile of the within income bucket distribution. Column 1 shows the result for the whole sample when interacted with income heterogeneity. Column 2 shows the result of specification (1) restricted to the low income counties. Column 3 shows the within low income heterogeneity in homeownership. Columns 4 to 7 replicates columns 2 and 3 for medium and high income levels. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

56 Table 1.10: Effect of Bankruptcy Protection on Interest Rates: Personal Unsecured Loans and Credit Cards

Personal Unsecured Loan Credit Card Debt Eventually County-Pairs Eventually County-Pairs St Linear Cty Linear Cty Linear Cty Linear St Linear Cty Linear St Linear Cty Linear Cty Linear Cty Linear St Linear Cty Linear Trend Trend Trend Trend Trend Trend Trend Trend Trend Trend Trend Trend (10) (11) (12) (1) (2) (3) (4) (5) (6) (7) (8) (9) Protection -0.260 0.584 Growth s.t-2 (0.395) (0.464)

Protection -0.022 0.083 Growth st-1 (0.274) (0.677)

Protection 0.389*** 0.415*** 0.373** 0.296* 0.755*** 0.820*** 0.007 0.147 -0.004 0.317 0.875* 0.775 Growth st (0.147) (0.144) (0.147) (0.170) (0.177) (0.157) (0.217) (0.183) (0.232) (0.229) (0.515) (0.573)

Protection -0.132 0.308* Growth s,t+1 (0.106) (0.166)

Protection -0.286 0.256 Growth s.t+2 (0.205) (0.273)

Unemploymient 0.003 0.001 -0.020 -0.009 0.106 0.084 -0.118 -0.103 -0.100 -0.086 -0.038 -0.059 Rate Change (0.046) (0.050) (0.073) (0.048) (0.103) (0.107) (0.089) (0.090) (0.096) (0.095) (0.151) (0.160)

House Price 4.938*** 4.812*** 4.363** 5.154*** -0.112 1.072 5.179 3.691 2.606 3.625 -5.857 -5.049 Index Growth (1.629) (1.623) (2.159) (1.607) (3.153) (3.315) (3.984) (3.895) (4.532) (4.014) (7.780) (8.608)

Income 0.198 0.182 0.551 0.203 2.299* 2.904 1.734*** 1.886*** 1.440 1.864*** -0.224 -0.868 Growth (0.268) (0.385) (0.622) (0.383) (1.255) (1.936) (0.558) (0.600) (0.973) (0.605) (4.195) (4.905)

No. of Obs 4.693 4.693 2,338 4.693 1,621 1,621 5.371 5.371 2.430 5,371 1.621 1,621 No. of Clusters 49 49 25 49 44 44 50 50 26 50 45 45 Y Y Cty and Year FE Y Y Y Y Y Y State and Year FE Y Y Y Y R-Squared 0.17 0.13 0.15 0.14 0.79 0.80 0.29 0.21 0.23 0.21 0.82 0.82

Note. This table shows the estimated coefficient following a variation of specification (1) of changes in interest rates (%) on changes in the level of protection. Personal Unsecured Loan and Credit Card Debt are county averages of the interest rates in a county for each type of credit. Columns 1 and 7 show the result using state fixed effect. Columns 2 and 8 show the estimates using county fixed effect. Columns 3 and 9 show the result restricting the sample to only the "eventually" treated sample. Columns 4 and 10 show the estimates looking at the dynamic effect of changes in protection on interest rates. Colums 5, 6, 11, and 12 show the results including state and county fixed effect for the sub-sample of neighboring county-pairs for which both counties are in the same income bucket. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively. Table 1.11: Effect of Bankruptcy Protection on Interest Rates: Mortagage Credit

3 Yr-ARM 15 Yr-Fixed 30 Yr-Fixed Eventually Eventually Eventually St Linear Cty Linear Cty Linear St Linear Cty Linear Cty Linear St Linear Cty Linear Cty Linear Trend Trend Trend Trend Trend Trend Trend Trend Trend (1) (2) (3) (4) (5) (6) (7) (8) (9) (7) Protection 0.037 0.053 0.041 0.014 0.019 0.005 0.026 0.029 0.027 Growth s,t (0.051) (0.062) (0.057) (0.041) (0.042) (0.035) (0.029) (0.030) (0.034)

Unemployment -0.066*** -0.100*** -0.048** -0.001 -0.002 -0.022 0.001 0.004 -0.040 Rate Change (0.031) (0.041) (0.026) (0.009) (0.011) (0.017) (0.019) (0.022) (0.017)

House Price 2.244*** 2.332*** 2.690** 0.009 0.045 0.637 -0.039 0.017 0.234 00 Index Growth (0.648) (0.677) (1.094) (0.319) (0.332) (0.403) (0.246) (0.252) (0.261)

Income -0.093 -0.191 -0.485 -0.003 -0.005 -0.136 -0.029 -0.034 -0.317*** Growth (0.228) (0.290) (0.374) (0.085) (0.118) (0.111) (0.107) (0.139) (0.115)

No. of Obs 3,919 3,919 1,945 5,723 5,723 2,802 5,533 5,533 2;732 No. of Clusters 47 47 24 50 50 26 49 49 25 Cty and Year FE Y Y y y y y State and Year FE Y y y R-Squared 0.85 0.85 0.85 0.87 0.86 0.87 0.86 0.85 0.87 Note. This table shows the estimated coefficient following a variation of specification (1) of changes in interest rates (%) in the level of protection. 3 Yr-ARM, 15 Yr-Fixed, 30 Yr-Fixed, averages of the interest rates in are county a county for each type of credit. Columns 1, 4, and 7 show the result using state fixed effect. Columns 2, 5 and 8, show the estimates using county fixed effect. Columns 3, 6 and 9, show the result restricting the sample to only the "eventually" treated sample. The sample period is from 1999 to 2005. *, *, and *** denotes significance the state level respectively. at the 10%, 5%, and 1% cluster at Table 1.12: Effect of Bankruptcy Protection on Debt. Number of Credit Cards and Entry

Panel A. All individuals Number of Credit Cards Entry A in A in N Credit Cards Credit Card A in A in N Credit Cards Conditional on n>0 Open First Balance Debt Balance N Credit Cards Conditional on n>0 & Balance >0 Credit Card Becomes >0 (5) (6) (1) (2) (3) (4) Protection 0.076*** 0.054*** 0.082*** 0.093*** 0.001 -0.002 Growth s.t (().009) (0.019) (0.026) (0.029) (0.003) (0.006)

Unenploynient 0.002 0.008** 0.008 0.009* 0.001 0.001 Rate Change (0.003) (0.004) (0.005) (0.005) (0.001) (0.002)

House Price -0.070* -0.050 -0.043 -0.039 -0.005 -0.031 Index Growth (0.041) (0.037) (0.049) (0.044) (0.008) (0.020)

Incone 0.012 -0.048*** -0.017 0.001 -0.011*** -0.063*** Growth (0.016) (0.016) (0.018) (0.017) (0.004) (0.014)

N of Ohs 366,362 619,726 454,688 359,235 555,007 221,849 N of Clusters 40 40 40 40 40 :39 Y Y Y Y Y Y R-Squaredl 0.00 0.02 0.02 0.02 0.01 0.01

Panel B. Home owners Number of Credit Cards Entry A in A in N Credit Cards Credit Card A in A in N Credit Cards Conditional on n>O Open First Balance Debt Balance N Credit Cards Conditional on n>0 & Balance >0 Credit Card Becomes >0 (5) (6) (1) (2) (3) (4) Protection 0.102*** 0.081*** 0.103*** 0.115*** -0.002 -0.006 Growth s.) (0.014) (0.020) (0.024) (0.032) (0.003) (0.006)

Uneniployient 0.000 0.009* 0.009 0.009 0.000 0.001 Rate Change (0.004) (0.005) (0.006) (0.007) (0.001) (0.002)

House Price -0.088* -0.052 -0.045 -0.032 -0.003 -0.044 Index Growth (0.052) (0.057) (0.067) (0.071) (0.007) (0.029)

Incone 0.014 -0.036* -0.006 0.006 -0.005* -0.060*** Growth (0.017) (0.021) (0.024) (0.021) (0.003) (0.016)

N of Obs 210.863 304,005 248,955 205,458 291,353 10:3,854 N of Clusters 39 39 39 :39 39 37 Y Y Cty and Year FE Y Y Y Y 10.02 R-Squared 0()0 0.02 0.02 0.01 0.01 Note. This table shows the estimated coefficient following a variation of specification (1). Panel A uses all individuals in counties below the median income. Panel B restricts the sample to homeowners, defined as individuals for whom some home debt is observed during the sample period. Column 1 shows the estimated of log changes in individuals' credit card balance on log changes in the levels of bankruptcy protection. Column 2 shows the estimates of the effect of personal bankruptcy protection on the number of credit cards changes. Column 3 restricted the previous specification to borrowers with more than 0 credit card. Column 4 shows the estimates for individual with more than 0 credit cards and a positive balance. Column 5 shows the estimates for a linear probability model on the timing of opening the first card, in this case the dependent variable is one if the individual did not have a credit card at t-1, but has one at t. Column 6 shows the same linear probability model estimates, but defining entry based on the timing of going to a positive balance, in other words the variable is one if the individual did not have a positive balance at t-1 but has one at t. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

59 Table 1.13: Effect of Bankruptcy Protection on Credit Card Delinquency

Panel A. All individuals

90+ days 120+ days Severe t t+1 t+2 t+3 t t+1 t+2 t+3 t t+1 t+2 t+3 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Protection -0.001 -0.008** 0.000 0.004 -0.002 -0.009** -0.001 0.003 -0.003 -0.008** 0.001 0.002 Growth s,t (0.004) (0.004) (0.004) (0.003) (0.005) (0.004) (0.003) (0.004) (0.003) (0.004) (0.002) (0.002)

N of Obs 366,362 363,498 361,444 359,783 366,362 363,498 361,444 359,783 366,362 363,498 361,444 359,783 N of Clusters 40 40 40 40 40 40 40 40 40 40 40 40 Cty and Year FE Y Y Y Y Y Y Y Y Y Y Y Y Uep/Income/HP Controls Y Y Y Y Y Y Y Y Y Y Y Y R-Squared 0.02 0.02 0.02 0.01 0.02 0.02 0.02 0.01 0.02 0.02 0.02 0.01

Panel B. Home owners

90+ days 120+ days Severe t t+1 t+2 t+3 t t+1 t+2 t+3 t t+1 t+2 t+3 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Protection -0.004 -0.014*** -0.007 0.003 -0.003 -0.014*** -0.007 0.001 -0.001 -0.013*** -0.005 0.000 Growth s,t (0.004) (0.005) (0.007) (0.002) (0.003) 0.004) (0.006) (0.003) (0.003) (0.005) (0.004) (0.003)

N of Obs 210,863 209,878 209,173 208,616 210,863 209,878 209,173 208,616 210,863 209,878 209,173 208,616 N of Clusters 39 39 39 39 39 39 39 39 39 39 39 39 Cty and Year FE Y Y Y Y Y Y Y Y Y Y Y Y Uep/Income/HP Controls Y Y Y Y Y Y Y Y Y Y Y Y R-Squared 0.02 0.02 0.02 0.01 0.02 0.02 0.02 0.01 0.02 0.02 0.02 0.01 Note. This table shows the estimated coefficient following a variation of specification (1), where we replace the dependent variable for a dummy indicator that is equal to 1 if the person is delinquent at the specified time. Panel A uses all individuals in counties below the median income with a positive balance. Panel B restricts the sample to homeowners, defined as individuals for whom some home debt is observed during the sample period. Columns 1 to 4 show the estimates where delinquency is defined as being delinquent 90 days or more. Column 5 to 8 show the estimates where delinquency is defined as being delinquent 120 days or more. Columns 9 to 12 show the estimates where delinquency is defined as being severely delinquent. All regressions include controls. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively Table 1.14: Effect of Bankruptcy Protection on Self-Employment

Credit Card Credit Card Self Employment Startup > p50 Startup < p50 (1) (2) (3) (4) (5) (6) (7) Protection Gowth s,t 0.000 -0.003 -0.0l0(** -0.002 -0.014 -0.003 -0.007 (0.002) (0.003) (0.004) (0.007) (0.009) (0.002) (0.003)

Protection Gowth s,t 0.006** 0.012*** 0.024*** 0.005 x Low Income (0.003) (0.004) (0.007) (0.004)

Protection Gowth st 0.003 0.008*** 0.012** 0.006 x Med Incoie (0.002) (0.003) (0.005) (0.003)

Unemployment 0.001*** 0 .0 0 1 *** 0.01*** 0.001 0.001 0.001 0.001 Rate Change (0.000) (0.000) (0.001) (0.001) (0.001) (0.001) (0.001)

Honse Price 0.096*** 0.097*** 0.058** 0.057 0.056 0.059* 0.059 Index Growth (0.023) (0.022) (0.028) (0.035) (0.035) (0.033) (0.033)

Income 0.063*** 0.063*** 0.101*** 0.126*** 0.127*** 0.085*** 0.085 Growth (0.010) (0.009) (0.028) (0.037) (0.037) (0.025) (0.025)

Number of Observations 12,738 12,738 194,011 73,081 73,081 120,930 120,930 Number of Clusters 50 50 50 50 50 50 50 State FE Y Y State x 2-digit industry Y Y Y Y Y Year FE Y Y Y Y Y Y Y R-Squared 0.21 0.23 0.01 0.02 0.02 0.02 0.02 Note. This table shows the estimated coefficient following a variation of specification (1) of log changes in self- employment measures on log changes in the levels of protection. Column 1 shows the estimates for county self- employment aggregates. Column 2 shows the results for the effect interacted with income heterogeneity for aggregate self-employment. Column 3 shows the estimates interacted with low income using self-employment changes by industry and county. Column 4 and 5 show the estimates for industries that used the level of credit card debt as a start-up capital and Column 6 and 7 for industries that do not. The sample period is from 1999 to 2005. *, **, and denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

61 Table 1.15: Effect of Bankruptcy Protection on Credit Card Debt. Alternative Specifications

Event Event Event Event Event Event Baseline Baseline Unlimited Unlimited Unemp. Unemp. Debt to Debt to % Change % Change Homestead Homestead Cty Linear Cty Linear Change Change Insurance Insurance Income Income in Debt in Debt Only Only (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Protection 0.018** 0.017** 0.018** 0.017** 0.018** 0.017** 0.023** 0.020* 0.022** 0.020** 0.017*** 0.015** Gowth st (0.008) (0.008) (0.008) (0.008) (0.008) (0.008) (0.011) (0.011) (0.009) (0.010) (0.006) (0.006)

Unlimited -0.156*** -0.139*** Protection s,t (0.027) (0.027)

Unemployment 0.002 0.002 0.002 0.002 0.002 0.002 0.008*** 0.008*** 0.002 0.002 0.002 0.003 Rate Change (0.002) (0.003) (0.002) (0.003) (0.002) (0.003) (0.002) (0.003) (0.002) (0.003) (0.002) (0.003) House Price -0.102 -0.118 -0.103 -0.119 -0.099 -0.134 -0.210** -0.155 -0.106 -0.109 -0.111 -0.130 Index Growth (0.086) (0.086) (0.086) (0.086) (0.088) (0.091) (0.094) (0.111) (0.094) (0.097) (0.087) (0.091) Income 0.079* 0.138* 0.080* 0.139* 0.080* 0.138* 0.065* 0.111* 0.081* 0.141* Growth (0.047) (0.077) (0.047) (0.077) (0.047) (0.076) (0.037) (0.059) (0.047) (0.077)

N of Obs 13,302 6,078 13,308 6,084 13,302 6,078 13,302 6,078 13,302 6,078 13.140 5,916 N of Clusters 50 26 51 27 50 26 50 26 50 26 48 24 Cty and year FE Y Y Y Y Y Y Y Y Y Y Y R-Squared 0.30 0,29 0.30 0.29 0.30 0.29 0.22 Y 0.18 0.29 0.27 0.30 0.29 Note. This table shows the estimated coefficient following a variation of the specification (1). Columns 1 and 2 replicated the main results. Columns 3 and 4 show the result when unlimited change of DC is included as a dummy. Columns 5 and 6 show the results when controlling for level of unemployment insurance. Columns 7 and 8 replace the dependent variable for debt income and 10 replace the dependent to change. Columns 9 variable for percentage changes in level of debt, and Columns 11 and 12 show the result if changes in the level of protection are measured only as a home-equity protection. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively. Table 1.16: Other Heterogeneous Treatment of Bankruptcy Protection. Credit Card Debt

Low Inc Bank Total Credit Card Number Credit Card Baseline Concentration Debt/Income Debt/Income of Filing 90+ Delinq (1) (2) (3) (4) (5) (6) Protection Gowth st 0.028** 0.085*** 0.041** 0.043* 0.026** 0.048*** (0.011) (0.021) (0.018) (0.024) (0.012) (0.017)

Protection Gowth s,t -0.086*** -0.028 -0.005 -0.004 -0.042*** x Low Income (0.022) (0.021) (0.039) (0.026) (0.014)

Protection Gowth s,t -0.076*** -0.009 -0.040 0.010 -0.014 x Med Income (0.016) (0.031) (0.037) (0.019) (0.018)

Unemployment 0.005* 0.005* 0.005* 0.005* 0.005* 0.005** Rate Change (0.003) (0.003) (0.003) (0.003) (0.003) (0.002)

House Price -0.015 -0.012 -0.013 -0.018 -0.018 -0.013 Index Growth (0.094) (0.094) (0.095) (0.095) (0.094) (0.094)

Income 0.059** 0.060** 0.062** 0.101*** 0.064** 0.058** Growth (0.030) (0.030) (0.031) (0.031) (0.030) (0.029)

N of Obs 4,536 4,536 4,536 4,536 4,536 4,536 N of Clusters 50 50 50 50 50 50 State and year FE Y Y Y Y Y Y R-Squared 0.24 0.24 0.24 0.25 0.24 0.24 Note. This table shows the estimated coefficient following a variation of specification (1) that incorporate interactions, within low income counties. Low/Med represents counties in the lowest/middle tercile of the within state described variable distribution. Column 2 shows the result for bank concentration. Column 3 for the total debt to income heterogeneity. Column 4 for credit card debt to income. Column 5 for heterogeneity on the county level number of filing in 1998. Column 6, using credit card delinquency heterogeneity defined as delinquency in 1999. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

63 Table 1.17: Determinants of Bankruptcy Protection Levels and Changes. Eventually Treated

Protection Level s,t Protection Growth s,t Protection Dummy s,t (1) (2) (3) (4) (5) (6) House Price/Growth s,t -1.563 -2.224** -0.984** -0.699 -1.123 -1.503 (2.581) (0.970) (0.445) (0.776) (0.924) (1.135) House Price/Growth st-i 3.301 3.147*** 1.453* 0.648 2.087* 1.631 (2.676) (0.985) (0.778) (1.259) (1.152) (1.569)

Medical Exp./Growth s,t -1.237 0.027 -1.039 -1.533 -1.851 -3.300 (4.206) (1.611) (1.124) (1.604) (2.219) (3.542) Medical Exp./Growth st-1 0.670 0.863 -0.733 -3.089* -2.245 -4.823** (4.642) (2.067) (1.443) (1.590) (2.110) (2.277)

Unemp. Rate/Change s,t 0.150 0.059 0.016 0.010 0.100 0.100 (0.177) (0.068) (0.052) (0.065) (0.080) (0.101) Unemp. Rate/Change st-1 0.029 -0.093 0.000 0.008 -0.042 -0.077 (0.129) (0.071) (0.060) (0.069) (0.091) (0.129)

State Real GDP/Growth s,t -0.994 0.899 0.814 1.301 -2.145 -1.589 (5.869) (1.774) (1.185) (1.814) (1.858) (2.519) State Real GDP/Growth st-1 -2.495 -1.494 -0.241 0.391 -1.183 -1.055 (5.177) (1.210) (0.592) (0.807) (1.411) (1.419)

No. Filings/Growth s,t -0.284 0.073 0.004 -0.129 0.023 -0.073 (0.190) (0.051) (0.069) (0.134) (0.087) (0.126) No. Filings/Growth st-1 -0.268 0.158 0.035 -0.074 -0.030 -0.127 (0.159) (0.083) (0.057) (0.089) (0.083) (0.112)

Political Climate st-1 0.209* -0.123* 0.060 0.924 0.375 1.536 (1.547) (0.374) (0.266) (0.535) (0.211) (0.887)

Personal Income/Growth st 13.996* 2.387 2.838 2.147 7.406** 7.292 (7.586) (2.940) (2.642) (3.967) (3.512) (4.611) Personal Income/Growth st-1 -9.635 -0.875 -0.722 -0.613 -0.869 -0.545 (7.373) (2.035) (1.740) (2.809) (3.266) (3.806)

No. of Obs. 196 196 168 168 168 168 State FE Y Y Y Year FE Y Y Y Y Y Y R2 0.27 0.12 0.08 0.21 0.18 0.24 Note. This table shows the estimated coefficient of regression of bankruptcy protection on contemporaneous and lag values of variables that could determinate the changes in protection levels. House Price s,t is the level or growth of house prices in state s at time t, from FHFA. Medical expenses is the level of growth in state's annual total medical expenses from the National Health Statistic. No. of Filings, is the number or change in the number of filings for non-business bankruptcies in a state. Political Climate s,t is defined as the share of democratic votes in the closer House of Representative election. State GDP and Personal Income are from BEA, and Unemployment Rate from BLS. Columns 1 and 2 show the coefficient of regressions of the level protection on level of the explanatory variables using only year, and year and state fixed effect. Columns 3 and 4 show the coefficient of regressions of the growth in protection on growth of the explanatory variables using only year, and year and state fixed effect. Columns 5 and 6 show the coefficient of regressions of a dummy that is one if the growth in protection is greater than zero on the explanatory variables' growth using only year, and year and state fixed effect. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

64 Table 1.18: Dynamics of the Change in Protection. Mortgage Debt

1 Period 2 Periods No County County No County County Linear Trend Linear Trend Linear Trend Linear Trend Linear Trend Linear Trend (1) (2) (3) (4) (5) (6) Protection -0.024 -0.043* -0.054** Growth st-2 (0.017) (0.025) (0.026)

Protection 0.019 0.013 0.005 0.018 0.002 -0.006 Growth st-1 (0.013) (0.014) (0.012) (0.013) (0.017) (0.015)

Protection 0.007 0.011 0.005 0.006 0.005 -0.002 Growth st (0.016) (0.014) (0.012) (0.016) (0.014) (0.013)

Protection -0.009 -0.006 -0.004 -0.010 -0.012 -0.010 Growth st+1 (0.008) (0.009) (0.009) (0.008) (0.009) (0.010)

Protection -0.016* -0.014 -0.011 Growth s,t+2 (0.009) (0.011) (0.011)

Unemployment -0.003 -0.004 -0.004 -0.003 -0.004 -0.005* Rate Change (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)

House Price 0.046 0.092 -0.372** 0.049 0.092 -0.385** Index Growth (0.078) (0.163) (0.172) (0.075) (0.163) (0.173)

Income 0.190** 0.113 0.039 0.189** 0.114 0.040 Growth (().091) (0.107) (0.078) (0.091) (0.107) (0.078)

Unemployment 0.001 0.000 Rate (0.004) (0.004)

House Price 0.277 0.281 (0.040) (0.039)

Income 0.133 0.132 (0.039) (0.040)

No. of Obs 13,302 13,302 13,302 13,302 13,302 13,302 No. of Clusters 50 50 50 50 50 50 County FE Y Y Y Y Year FE Y Y Y Y Y Y R-Squared 0.09 0.09 0.11 0.09 0.09 0.12 Note. This table shows the estimated coefficient following specification (1) of log changes to mortgage debt on log changes in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the log change in income in county i at time t from IRS. Columns 1 and 4, show the without the inclusion of county fixed effects, including one lag and lead, and two lags and two leads. Columns 2 and 5 show the results with the inclusion of county fixed effect for including one lag and lead, and two lags and two leads, Columns 3 and 6 are the same than before but including level controls. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

65 Table 1.19: Dynamics of the Change in Protection. Auto Debt

1 Period 2 Periods No County County No County County Linear Trend Linear Trend Linear Trend Linear Trend Linear Trend Linear Trend (1) (2) (3) (4) (5) (6) Protection -0.022 -0.015 -0.020 Growth s,t-2 (0.019) (0.026) (0.028)

Protection -0.006 -0.004 -0.004 -0.005 -0.002 -0.003 Growth st-1 (0.013) (0.017) (0.017) (0.013) (0.017) (0.016)

Protection 0.008 0.006 0.007 0.009 0.010 0.010 Growth st (0.014) (0.011) (0.010) (0.014) (0.014) (0.013)

Protection -0.012* -0.011 -0.008 -0.011 -0.007 -0.004 Growth st+ (0.007) (0.010) (0.011) (0.007) (0.009) (0.011)

Protection 0.015 0.020* 0.022* Growth s,t+2 (0.011) (0.011) (0.012)

Unemployment -0.005* -0.005* -0.002 -0.005* -0.005* -0.002 Rate Change (0.003) (0.003) (0.003) (0.003) (0.003) (0.003)

House Price 0.110** 0.002 -0.097 0.105* -0.015 -0.127 Index Growth (0.053) (0.113) (0.124) (0.054) (0.114) (0.125)

Income 0.127*** 0.059 0.032 0.128*** 0.060 0.031 Growth (0.030) (0.038) (0.032) (0.030) (0.037) (0.032)

Unemployment -0.011** -0.012** Rate (0.005) (0.005)

House Price 0.009 0.012 (0.043) (0.042)

Income 0.025 0.026 (0.030) (0.029)

No. of Ohs 13,302 13,302 13,302 13,302 13,302 13,302 No. of Clusters 50 50 50 50 50 50 County FE Y Y Y Y Year FE Y Y Y Y Y Y R-Squared 0.17 0.18 0.19 0.17 0.18 0.19 Note. This table shows the estimated coefficient following specification (1) of log changes to auto debt on log changes in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the log change in income in county i at time t from IRS. Columns 1 and 4, show the without the inclusion of county fixed effects, including one lag and lead, and two lags and two leads. Columns 2 and 5 show the results with the inclusion of county fixed effect for including one lag and lead, and two lags and two leads, Columns 3 and 6 are the same than before but including level controls. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

66 Table 1.20: Local Business Conditions. Neighboring County-pairs across State Bor- ders. Mortgage Debt

All Equal Income Low Income County-Pairs County-Pairs County-Pairs State County State County State County Linear Liner Linear Liner Linear Liner Trend Trend Trend Trend Trend Trend (1) (2) (3) (4) (5) (6) Protection 0.006 0.007 0.006 0.006 0.051 0.051 Growth s,t (0.011) (0.011) (0.010) (0.010) (0.060) (0.058)

Unemployment -0.002 -0.002 0.001 0.000 -0.001 -0.001 Rate Change (0.005) (0.005) (0.005) (0.005) (0.008) (0.008)

House Price -0.116 -0.109 -0.050 -0.046 0.077 0.074 Index Growth (0.153) (0.150) (0.203) (0.196) (0.639) (0.617)

Income 0.089* 0.015 0.197*** 0.151* 0.160 0.177 Growth (0.054) (0.064) (0.074) (0.083) (0.115) (0.126)

No. of Obs 9,168 9,168 3,984 3,984 1,188 1,188 No. of Clusters 48 48 46 46 33 33 County FE Y Y Y State FE Y Y Y County-Pair-Year FE Y Y Y Y Y Y R-Squared 0.65 0.64 0.62 0.61 0.55 0.53 Note. This table shows the estimated coefficient following specification (2) of log changes in mortgage debt on log changes in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the log change in income in county i at time t from IRS. Columns 1 and 2 show the estimates for state and county fixed effect for all neighboring county-pairs sample. Columns 3 and 4 show the results including state and county fixed effect for the sub-sample of neighboring county-pairs for which both counties are in the same income bucket. Columns 5 and 6 show estimates with state and county fixed effect for only the neighboring county-pairs in the same income bucket and in the lowest tercile of the income distribution. The sample period is from 1999 to 2005. *, * and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

67 Table 1.21: Local Business Conditions. Neighboring County-pairs across State Bor- ders. Auto Debt

All Equal Income Low Income County-Pairs County-Pairs County-Pairs State County State County State County Linear Liner Linear Liner Linear Liner Trend Trend Trend Trend Trend Trend (1) (2) (3) (4) (5) (6) Protection 0.006 0.006 0.008 0.008 -0.018 -0.017 Growth s,t (0.010) (0.010) (0.014) (0.013) (0.050) (0.048)

Unemployment 0.000 0.000 -0.001 -0.001 -0.004 -0.003 Rate Change (0.004) (0.004) (0.005) (0.005) (0.006) (0.006)

House Price -0.079 -0.072 -0.275 -0.269 -0.381 -0.379 Index Growth (0.197) (0.193) (0.213) (0.206) (0.406) (0.389)

Income 0.143*** 0.062 0.295*** 0.239** 0.285* 0.279* Growth (0.049) (0.057) (0.102) (0.118) (0.160) (0.167)

No. of Obs 9,168 9,168 3,984 3,984 1,188 1,188 No. of Clusters 48 48 46 46 33 33 County FE Y Y Y State FE Y Y Y County-Pair-Year FE Y Y Y Y Y Y R-Squared 0.70 0.70 0.67 0.67 0.60 0.60 Note. rhis table shows the estimated coefficient following specification (2) of log changes in auto debt on log ch anges in bankruptcy protection at the county level. Debt county data is from the FRBNY Consumer Credit Panel/Equifax. Protection Growth is the log change in the level of protection in state s at time t. Unemployment rate change is the change in unemployment rate in county i at time t from BLS. House price growth is the log change in the FHFA state level index for state s at time t, and Income growth is the log change in income in county i at time t from IRS. Columns 1 and 2, show the estimates for state and county fixed effect for all neighboring county-pairs sample. Columns 3 and 4 show the results including state and county fixed effect for the sub-sample of neighboring county-pairs for which both counties are in the same income bucket. Columns 5 and 6 show estimates with state and county fixed effect for only the neighboring county-pairs in the same income bucket and in the lowest tercile of the income distribution. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

68 Table 1.22: Heterogeneous Treatment of Bankruptcy Protection: Income and Home- ownership. Mortgage Debt

Low Income Med Income High Income Income Home Home Home Ownership Ownership Ownership (1) (2) (3) (4) (5) (6) (7) Protection Growth st 0.018 0.011 0.012 0.006 0.006 0.012 0.019 (0.011) (0.013) (0.016) (0.016) (0.019) (0.010) (0.015)

Protection Growth s,t -0.005 x Low Income (0.014)

Protection Growth st 0.007 0.003 -0.016 x Low Home Ownership (0.024) (0.015) (0.014)

Protection Growth s,t -0.013 x Med Income (0.011)

Protection Growth s,t -0.010 -0.004 -0.001 x Med Home Ownership (0.016) (0.015) (0.015)

Unenployient -0.003 -0.003 -0.003 -0.001 -0.001 -0.007 -0.008 Bate Change (0.003) (0.004) (0.004) (0.003) (0.003) (0.008) (0.008)

House Price 0.078 0.070 0.070 0.137 0.137 0.042 0.041 Index Growth (0.161) (0.141) (0.141) (0.185) (0.185) (0.182) (0.182)

Income 0.189** 0.096** 0.096** 0.016 0.012 0.415*** 0.403*** Growth (0.089) (0.046) (0.045) (0.053) (0.052) (0.138) (0.143)

No. of Obs 13,302 4,536 4.536 4,422 4,422 4,344 4,344 No. of Clusters 50 50 50 50 50 50 50 State and Year FE Y Y Y Y Y Y Y R-Squared 0.11 0.08 0.08 0.10 0.11 0.29 0.31 Note. This table shows estimated coefficient a variation of specification (1) that incorporates interactions. Low/Med Income represents counties in the lowest/middle tercile of the within state income distribution. Low/Med Ownership represents counties in the lowest/middle tercile of the within income bucket distribution. Column 1 shows the result for the whole sample when interacted with income heterogeneity. Column 2 shows the result of specification (1) restricted to the low income counties. Column 3 shows the within low income heterogeneity in homeownership. Columns 4 to 7 replicates columns 2 and 3 for medium and high income levels. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

69 Table 1.23: Heterogeneous Treatment of Bankruptcy Protection: Income and Home- ownership. Auto Debt

Low Income Med Income High Income Income Home Home Home Ownership Ownership Ownership (1) (2) (3) 4) (5) (6) (7) Protection Growth s,t 0.000 0.032* 0.038* -0. 002 -0.003 -0.006 -0.023 (0.013) (0.019) (0.021) (0. 016) (0.027) (0.012) (0.016)

Protection Growth st 0.027 x Low Income (0.017)

Protection Growth s,t -0.020 0.006 0.021 x Low Home Ownership (0.030) (0.023) (0.017)

Protection Growth s,t 0.001 x Med Income (0.007)

Protection Growth s,t 0.008 -0.004 0.028** x Med Home Ownership (0.017) (0.025) (0.012)

Unemployment -0.005* -0.002 -0.002 -0.007** -0.007** -0.008 -0.009* Rate Change (0.003) (0.004) (0.004) (0.003) (0.003) (0.005) (0.005)

House Price -0.013 -0.114 -0.112 0.070 0.072 0.020 0.020 Index Growth (0.113) (0.146) (0.147) (0.116) (0.117) (0.105) (0.105)

Income 0.120*** 0.066 0.065 0.056* 0.059* 0.209*** 0.196*** Growth (0.031) (0.057) (0.054) (0.033) (0.031) (0.030) (0.030)

No. of Obs 13,302 4,536 4,536 4,422 4,422 4,344 4,344 No. of Clusters 50 50 50 50 50 50 50 State and Year FE Y Y Y Y Y Y Y R-Squared 0.19 0.12 0.13 0.20 0.20 0.34 0.36 Note. This table shows estimated coefficient following a variation of specification (1) that incorporates interactions. Low/Med Income represents counties in the lowest/middle tercile of the within state income distribution. Low/Med Ownership represents counties in the lowest/middle tercile of the within income bucket distribution. Column 1 shows the result for the whole sample when interacted with income heterogeneity. Column 2 shows the result of specification (1) restricted to the low income counties. Column 3 shows the within low income heterogeneity in homeownership. Columns 4 to 7 replicates columns 2 and 3 for medium and high income levels. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively.

70 Table 1.24: Effect of Bankruptcy Protection on County Delinquency Proportions

Credit Card Debt Mortgage Debt Auto Debt 1 year 2year 3 years 1 year 2year 3 years 1 year 2year 3 years Protection 0.088 0.020 0.072 -0.009 -0.018 -0.002 0.001 -0.021 -0.037 Growth s,t (0.204) (0.058) (0.065) (0.116) (0.043) (0.045) (0.165) (0.077) (0.049)

Unemployment 0.024 -0.018 -0.024 0.021 -0.008 0.001 0.077*** 0.012 0.010 Rate Change (0.034) (0.016) (0.018) (0.020) (0.014) (0.009) (0.021) (0.012) (0.009)

-1 House Price -1.910** -2.388 -3.404*** -1.245* -0.653*** 0.385*** 0.181 0.508 0.998* I, Index Growth (1.448) (1.185) (0.876) (0.576) (0.475) (0.659) (0.606) (0.455) (0.311)

Income -1.579** -1.130 -0.630*** -0.581* -0.650*** -0.401*** -0.387 -0.067 -0.132* Growth (0.695) (0.700) (0.180) (0.335) (0.232) (0.115) (0.255) (0.139) (0.076)

N of Obs 13,302 13,302 13,302 13,302 13,302 13,302 13,302 13,302 13,302 N of Clusters 50 50 50 50 50 50 50 50 50 county and year FE Y Y Y Y Y Y Y Y Y R-Squared 0.10 0.17 0.22 0.02 0.05 0.06 0.03 0.02 0.02 Note. This table shows the estimated coefficient following a variation of specification (1) that uses as a dependent variable the change in the fraction of delinquent households in each county, for each type of credit, for different periods: 1, 2, and 3 year annual changes. The sample period is from 1999 to 2005. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively. Table 1.25: Effect of Bankruptcy Protection on Debt After Bankruptcy Reform 2005

Credit Card Debt Mortgage Debt Auto Debt No Linear Cty Linear Cty Linear No Linear Cty Linear Cty Linear No Linear Cty Linear Cty Linear Trend Trend Trend Trend Trend Trend Trend Trend Trend

(1) (2) (3) (1) (2) (3) (1) (2) (3) Protection -0.002 -0.006 0.017** -0.002 0.007 0.011 -0.007 -0.003 0.013 Growth s,t (0.004) (0.006) (0.008) (0.005) (0.008) (0.013) (0.005) (0.005) (0.013)

Protection -0.021** -0.011 -0.022 Growth s,t x Post (0.009) (0.014) (0.014)

Unemployment -0.004** -0.007*** -0.001 -0.002 -0.006** -0.005** -0.007** -0.006** -0.007*** Rate Change (0.002) (0.002) (0.001) (0.003) (0.003) (0.002) (0.003) (0.003) (0.002) CA

House Price -0.254*** -0.139*** -0.197*** 0.065* 0.146*** 0.070** 0.166*** 0.125 0.161*** Index Growth (0.034) (0.038) (0.025) (0.036) (0.046) (0.033) (0.041) (0.082) (0.033)

Income 0.054 -0.174** 0.057* 0.455*** 0.160** 0.172* 0.420*** 0.323*** 0.123*** Growth (0.091) (0.076) (0.033) (0.087) (0.079) (0.092) (0.063) (0.053) (0.031)

N of Obs 8,868 8,868 22,170 8,868 8,868 22,170 8,868 8,868 22,170 N of Clusters 50 50 50 50 50 50 50 50 50 ety and year FE Y Y Y Y Y Y Y Y Y R-Squared 0.43 0.48 0.43 0.34 0.38 0.25 0.40 0.43 0.42

Note. This table shows the estimated following specification (1) but extending the sample, for each for each type of credit until 2009. Columns 1, in each type shows the estimates without county fixed effect. Columns 2, shows the estimates with fixed effect and Columns 3 shows the interaction with a post dummy equal to one for years greater or equal than 2006. The sample period is from 1999 to 2009. *, **, and *** denotes significance at the 10%, 5%, and 1% cluster at the state level respectively. Chapter 2

House Prices, Collateral and Self-Employment

2.1 Introduction

The boom-and-bust cycle of house prices over the past decade has featured promi- nently in explanations of the low unemployment during the surge in house prices and the high unemployment that followed the real-estate bust. The debate has focused on two primary explanations for the observed employment dynamics. One view is that consumers' use of their houses as "ATMs" drove demand and created employment during the surge in prices, so employment suffered when aggregate demand dropped because of household deleveraging and falling house prices (see, e.g., Mian and Sufi, 2011a; and Romer, 2011). The other view is that the increase in house prices and the rise in labor demand in the construction industry masked structural mismatches in the workforce caused by job losses in the manufacturing sector (see Charles, Hurst, and Notowidigdo, 2012; and Kocherlakota, 2010). Our paper documents an alternative channel that has received much less attention but significantly affects the dynamics of employment creation over the business cycle: the impact of the collateral lending channel, especially mortgage lending, on employment in small businesses. Seminal papers by Bernanke and Gertler (1989) and Kiyotaki and Moore (1997) and research since then suggest that improvements in collateral values ease credit constraints for borrowers and can have multiplier effects on economic growth. This collateral lending channel builds on the idea that information asymmetries between banks and firms can be alleviated more easily when collateral values are high, and therefore firms can have higher leverage (Rampini and Viswanathan, 2010), and that these problems are especially acute for small, more opaque firms (Gertler and Gilchrist, 1994; Kashyap, Stein, and Wilcox, 1993). Yet it has been difficult to cleanly identify the causal direction of the collateral effect empirically. The challenge is that, on the one hand, increased collateral values facili- tate lending but that, on the other hand, higher collateral values can be the result of improvements in economic conditions (e.g., lacoviello, 2005). This paper is the first to look directly at shocks to home values and consider the

73 impact these shocks have on employment in small firms relative to large firms. To identify the causal effect of higher house prices we instrument for the growth in prices between 2002 and 2007 using the elasticity measure developed by Saiz (2010). The measure uses exogenous geographic and regulatory constraints to housing supply to differentiate areas where an increase in housing demand translates into higher house prices and more collateral value (areas where it is hard to build - that is, in which the elasticity of the housing supply is low) or into higher volume of houses built (areas with high elasticity). By relying on exogenous restrictions on the expansion of housing volumes, we can identify the effect of high collateral values on employment in small businesses. This identification strategy is similar to Chaney, Sraer, and Thesmar (2012), who look at corporate investment decisions, and Mian and Sufi (2011b), who examine increases in consumption from household leverage. We show that during the housing price boom of 2002-2007, areas with rising house prices (and increased leverage) experienced a significantly bigger increase in small business starts and a rise in the number of people who were employed in establishments with fewer than ten employees compared to areas that did not see an increase in house prices. The same increase in employment cannot be found for large establishments in these same areas. In fact, the effect of home prices on job creation decreases monotonically with firm size. This asymmetric effect on small versus large holds only for instrumented house prices, which suggests that the non-instrumented part of the variation (the one that captures endogenous demand) chiefly impacts employment at larger firms. This asymmetry points to the interpretation of the collateral lending channel as an important driver of employment creation particularly for small firms, since large firms have access to other forms of financing and should be less affected by the collateral channel. To the extent that large firms are also affected by the increase in real estate values, our estimates may understate the effect of the collateral channel on total employment. Although the result above supports the importance of the collateral channel for small business creation, two alternative hypotheses must be ruled out as explaining our results. First, increases in housing prices can drive local demand for goods (Campbell and Cocco, 2007) and, consequently, employment at non-tradable industries (Mian and Sufi, 2011a). To the extent that small firms may be more sensitive to changes in demand (Kashyap and Stein, 1994), the asymmetry in the results could reflect increased consumer demand rather than use of the collateral lending channel. The second alternative hypothesis results from our use of housing and zoning restrictions for obtaining identification, because we rely on cross-sectional differences between high- and low-elasticity areas. These areas could also vary in other characteristics, such as the level of economic vitality. For example, not only could areas with low housing elasticity see higher home prices when demand for housing picks up - and therefore increased available collateral - but they could also be the areas where more investment opportunities become available. We devise a number of tests to differentiate the impact of the collateral lending channel from these alternative hypotheses. First, we verify that the results are not driven by changing industry composition: even within industries, areas with increasing home prices saw stronger employment growth in smaller establishments than areas

74 with stagnant prices. 1 Second, narrowing in on the importance of collateral for business financing, we look at variation across industries in the amount of start-up capital needed to set up a new firm. The minimal feasible scale of businesses differs across industries, and the availability of collateral matters more or less depending on that minimal scale. For example, some businesses, like home health-care services, can be started with small amounts of capital that could reasonably be financed through appreciation in home values. In contrast, many sectors within manufacturing, for example, require large amounts of capital and fixed investments; the capital needs in these areas are too high to be financed via individual loans against property. This strategy is similar to the approach used in Hurst and Lusardi (2004). Our results follow exactly the predicted pattern: when we repeat our regressions disaggregated by industries above and below median needs for start-up capital, we find that the effect of house price increases on the creation of employment in small establishments is especially strong among industries with lower capital needs. These results confirm that the collateral lending channel plays an important role in shaping employment dynamics. Borrowing against housing wealth allows people in areas with more rapid home price appreciation to start small businesses and drives the increase in employment at these small firms. Third, we confirm that our results are not driven by the non-tradable or construction sectors. As noted above, if the relation between increasing housing price and job creation in small firms were purely constrained to the non-tradable or construction sectors, one would be concerned that the results are driven not by changes in the collateral lending channel but by differences in local demand. However, our results are almost unchanged when we eliminate these sectors from the analysis, and they also hold for the manufacturing sector where products are easily tradable. The difference in employment creation between large and small firms is also particularly strong for industries in which firms report shipping goods across long distances. Our results are thus distinguished from the work of Mian and Sufi (2011a), which shows that areas where house prices increased most also exhibited an increase in unemployment in non-tradable industries due to deleveraging and lower demand in the aftermath of 2008. Any change in output in the low-elasticity areas must therefore be driven by changes on the input (production) side. This is the collateral lending channel. Last, we rule out that our results are driven by generally loosening credit standards in areas with rapid house price growth. The growth of small businesses could be caused not by better access to collateral but rather by easier access to other forms of credit because of banks' improved balance sheet position. We show that this is not the case. If anything, banks became increasingly more selective in credit approval in low-elasticity areas leading up to 2007. Using a calculation similar to that used in Mian and Sufi (2011a), we compute the approximate contribution of the collateral lending channel to changes in overall employment in the pre-crisis period, 2002-2007. Using this approach, we find that

'A similar relationship exists when we include proprietorships and unincorporated businesses in the regressions.

75 the collateral channel accounts for 10-25% of the increase in employment in these years (depending on the specific assumptions about the reference group that best isolates the collateral effect), whereas the demand channel explains about 40% over the same period and the two effects are mutually non-overlapping. Interestingly, although the point estimate for the effect of the demand channel is large, the effect is noisily estimated for 2002-2007, so we cannot reject that there is no effect on employment of increased demand driven by higher house prices before the crisis. This is in stark contrast to the post-crisis period (2007-2009), when the drop in demand of over-leveraged areas shows up very strongly in the data (as documented in Mian and Sufi, 2011a). It is important to point out that these numbers provide rough approximations of the relative magnitudes of these two channels, but they ignore any general equilibrium effects in aggregation. When we consider the period after the financial crisis when house prices started to decline (2007-2009), we find that small firms experienced weaker employment declines than large firms in areas where the increase in house prices was stronger in the period before the crisis. This suggests that small firms that were created in low- elasticity areas during the time of increasing collateral values were more resilient than larger ones in those areas and did not immediately disappear when the crisis struck. This shows an interesting asymmetry in the mechanism behind the collateral lending channel - although it is a powerful channel in facilitating the creation of new small establishments, a contraction in the amount of available collateral does not lead to a disproportionate amount of destruction of employment in those small establishments. We are, however, cautious in interpreting our results for the post-2007 period. First, given the nature of our data, we cannot disentangle whether the relative persistence of jobs in small businesses is due to the survival of existing small businesses or a change in the entrance of newly started firms. Second, although the elasticity measure has a natural interpretation for positive housing demand shocks, we lack a good instrument for the house price drop. In fact, an increase in housing demand can translate into either higher house prices (inelastic areas) or an expansion of housing volume (elastic areas). However, on the downside, a drop in housing demand does not lead to the destruction of housing stock, and thus prices simply drop in both inelastic and elastic areas. So, instead of instrumenting for the price drop in the crisis period, we instead compare areas with large appreciation in the pre-crisis period (low elasticity) with those that had smaller house price increases - that is, the timing of the housing price changes remains 2002-2007, as in the rest of the analysis. Once the crisis hit, areas that experienced larger house price increases in the pre-crisis period were more leveraged (Mian and Sufi, 2011a, 2011b), so it should be harder for households to access collateral in these areas in the crisis. Our study builds on literature that shows that credit constraints at the household level matter for the creation of new businesses (Evans and Jovanovic, 1989; Holtz-Eakin, Joulfaian, and Rosen, 1994; Gentry and Hubbard, 2004; Cagetti and De Nardi, 2006), although some authors have argued that this relation is present only at the very top of wealth distribution (Hurst and Lusardi, 2004). At the same time, housing wealth in particular has been shown to be an important factor in the funding of business start-ups (see Fan and White, 2003; Fairlie and Krashinsky, 2012;

76 Fort, Haltiwanger, Jarmin, and Miranda, 2012; Kleiner, 2013; Corradin and Popov, 2013; and Schmalz, Sraer, and Thesmar, 2013, for France; and Black, De Meza, and Jeffreys, 1996; and Kleiner, 2013, for the United Kingdom). Previous work has also found that bank credit is an important source of financing for small businesses (Pe- tersen and Rajan, 1994; Robb and Robinson, 2012; Fracassi, Garmaise, Kogan, and Natividad, 2013) and that entrepreneurs often have to provide personal guarantees when they obtain financing (Berger and Udell, 1998). More recently, Greenstone and Mas (2012) use the sharp reduction in credit supply following the 2008 crisis, and the heterogeneity of this effect among banks, to show that a decrease in the origination of small business loans leads to a decrease in county employment and business formation during the period 2007-2009. The rest of the paper proceeds as follows: Section 2 describes our data and the empirical methodology. Section 3 discusses the results, and Section 4 concludes.

2.2 Data and Empirical Methodology

2.2.1 Data Description

We obtain employment growth from the County Business Patterns (CBP) data set published by the U.S. Census Bureau. The CBP data contain employment data by county, industry, and establishment size (measured in number of employees) between 1998 and 2010 as of March of the reported year. We use the data at the four-digit National American Industry Classification System (NAICS) level, broken down by county and establishment size, to construct our main dependent variable of interest: the employment growth by establishment size between 2002 and 2007. The breakdown of establishments by employee number allows us to differentially estimate the effect of housing price growth in the net creation of establishments of different sizes. 2 We use five establishment categories in our regressions that the Census Bureau commonly uses: establishments of one to four employees, five to nine, ten to 19, 20 to 49, and 50 or more. The CPB provides all but the final category. For establishments with 50 or more employees, the CBP has multiple categories, but if we were to use each one individually, it would add noise to our estimation because such large businesses become rare at the county level and even scarcer at the county and industry levels, which we need for some of the specifications discussed below. In order to create the category of establishments with more than 50 employees, we take the number of establishments in each category above 50 and multiply those by the midpoint of the category (for example, for the category of 100 to 249 employees, we multiply the number of establishments by 174.5), and then we add them all up at the country and industry levels.

2The data include only the number of establishments in each county, industry, and year by category of employment size (1-4 employees, 5-9, 10-19, etc.), not the total employment for each establishment category. As such to construct the employment in each bin we multiply the number of establishments by the middle point of each category. For example, to calculate the total employment of 1-4-employee establishments in a given industry, county, and year, we multiply the number of establishments by 2.5.

77 The housing prices used in the regressions come from the Federal Housing Finance Agency (FHFA) House Price Index (HPI) data at the Metropolitan Statistical Area (MSA) level. The FHFA HPI is a weighted, repeat-sales index, and it measures average price changes in repeat sales or refinancings on the same properties. We obtain this information by reviewing repeat mortgage transactions on single-family properties whose mortgages have been purchased or securitized by Fannie Mae or Freddie Mac since January 1975. We use data on the MSA-level index between 2002 and 2007. The use of MSA-level house prices is consistent with our identification strategy. To identify the causal effect of house prices on small business creation, we instrument house price growth between 2002 and 2007 with the measure of housing supply elasticity of Saiz (2010), which varies at the MSA level. The measure of the supply elasticity is constructed using geographical and local regulatory constraints to new construction. Areas where it is difficult to add new housing (due to geographic or regulatory restrictions) are classified as low elasticity and vice versa for areas where land is easily available. Low-elasticity areas correlate strongly with steeper house price growth in the years 2002-2007. This measure is available for 269 MSAs that we match to 776 counties using the correspondence between MSAs and counties for the year 1999 as provided by the Census Bureau.3 Although employment growth and our other controls are available for a much larger sample of counties, our regressions focus on the subset of counties for which we have the housing supply elasticity measure. An important measure for our analysis is the amount of capital needed to start a firm, since these investment requirements might affect how much a given industry depends on the housing collateral channel. To construct this variable we use the Survey of Business Owners (SBO) Public Use Microdata Sample (PUMS). The SBO PUMS was created using responses from the 2007 SBO and provides access to survey data at a more detailed level than that of previously published SBO results. The SBO PUMS is designed to study entrepreneurial activity by surveying a random sample of businesses selected from a list of all firms operating during 2007 with receipts of $1,000 or more provided by the IRS. The survey provides such business characteristics as firm size, employer-paid benefits, minority- and women-ownership, access to capital, and firm age. We focus here on the "Amount of start-up or acquisition capital" for each firm, and we group the answers to this question at the two-digit NAICS industry level (the finest level available in the data) for firms established in 2007. The classification is virtually identical if we use all years in the data or if we focus on firms with one tO four employees only. The median amount of capital needed to start a business in the data is $215 thousand. We follow Hurst and Lusardi (2004) and split industries above and below the median to measure the differential effect of the collateral channel on business creation for industries in the two groups. The average amount of capital needed by firms below the median is $132 thousand, whereas the average amount needed for industries above the median is $260 thousand (detailed amounts by two-digit NAICS sector are in Appendix Table2.14).

3 This correspondence is available at and for the New England Metropolitan Component Areas used by Saiz (2010).

78 Our classification of "non-tradable," "tradable," and "construction" industries at the four-digict NAICS level is obtained from Appendix Table 2 of Mian and Sufi (2011a). 4 Non-tradable codes are included largely in the 44 and 45 sectors (Retail Trade), as well as under 72 (Accommodation and Food Services). Construction industries include most codes under the Construction two-digit NAICS sector (23), as well as some subsectors in manufacturing, retail trade, and services that are directly connected to construction (e.g., 3273 - Cement and Concrete Products Manufactur- ing). Manufacturing industries include all 31-33 subsectors (Manufacturing), and in some specifications we restrict the sample to manufacturing industries that are also classified as "tradable" in Mian and Sufi (2011a) (i.e., those not in construction or in "other industries"). To address further the concern that the results might be driven by local demand, we construct a measure of the average distance that firms in an industry ship their goods similar to that used in Duranton, Morrow, and Turner (2013). These data are available from the 2007 Census Commodity Flow Survey, which reports the distance traveled by shipments of a sample of establishments in each three-digit NAICS manufacturing industry.5 The unit of observation in the census data is at the state and industry levels, so we construct a dollar-weighted average distance of shipments also for each state and industry individually. Summary statistics of the average distance shipped, as well as how often each industry appears in each decile, are shown in Appendix Table 2.13. We also use data on county-level births and deaths of establishments for each two-digit NAICS industry between 2002 and 2010 from the Census Statistics of U.S. Businesses (SUSB). Data on births and deaths of establishments is provided under the "Employment Change" section of SUSB, and it does not include a breakdown by establishment size at the county and industry levels, so we cannot use it as our main dataset. However, given that most establishment births are of a very small scale (Haltiwanger, Jarmin, and Miranda, 2011), we view the regressions performed on this data set as an important test of the mechanism in our main results. We compute the cumulative number of births and deaths between 2002 and 2007 for each county and industry as our dependent variable of interest and scale this number by the total number of establishments as of 2002 in the same county-industry cell. The net creation of sole proprietorships at a county level is obtained from two sources. We use both the yearly local area personal income and employment data from the Bureau of Economic Analysis (BEA and the census nonemployer statistics. From the BEA we use Non-Farm Proprietorship employment at the county level between 2002 and 2007 to estimate the growth of sole proprietorships in this period. From the census we obtain the number of establishments for the period 2002-2007 at the two-digit NAICS level. We use both sources of data in the regressions to ensure the robustness of our results. Unemployment and unemployment rate at the county level are obtained using

4The current version of the online appendix can be found here: http://faculty.chicagobooth.edu/amir.sufi/data-and-appendices/ 'The year 2007 is the first year in which the data is reported at the three-digit NAICS level (previous years included only commodity identifiers rather than industry data).

79 the Bureau of Labor Statistics Local Area estimates. Local Area Unemployment Statistics (LAUS) are available for approximately 7,300 areas that range from census regions and divisions to counties and county equivalents, and these data are available between 1976 and 2012. We match the county equivalent data to the CBP data using Federal Information Processing Standard (FIPS) county unique identifiers. The migrations data are extracted from the IRS county-to-county migration data series. The migration estimates are based on year-to-year address changes reported on individual income tax returns filed with the IRS. The data set presents migration patterns by county for the entire United States and is split by inflows - the number of new residents who moved to a county and where they migrated from - and outflows - the number of residents leaving a county and where they went.' We also compute net flows as inflows minus outflows, and we scale all figures by the number of nonmovers in the county. The data are available from 1991 through 2009 filling years. To better identify the effect of house prices on self-employment, we include a set of controls that capture some of the cross-sectional differences across counties. We use county-level information from the Census Bureau Summary Files for 2000 on: the number of households in a county; the natural logarithm of county-level population; the percentage of college-educated individuals, defined as the number of people over 25 with a bachelor degree or higher as a proportion of the total population over 25 years old; the percentage of employed people, defined as the employed population over the total population 16 years old or older; the share of the population in the workforce, defined as the total population in the civilian labor force over 16 years old divided by the total population 16 years old or older; the percentage of owner-occupied houses; and a measure of exposure of each county to imports from China, 7 and, therefore, better control for changes in investment opportunities in those counties.

2.2.2 Summary Statistics Panel A of Table 2.1 provides descriptive statistics for our data set: the first row shows total employment in 2002 for all counties in our sample, as well as the employment growth between 2002 and 2007 estimated from the CBP data. Our data include a total of 775 counties with nonmissing total employment data. We split the sample into counties above and below the median of the housing supply elasticity measure and show t-statistics (with standard errors clustered by MSA) for the difference in means between the two groups. We see that counties with low supply elasticity are larger but have similar unemployment rates in 2002 as those with high supply elasticity. The characteristics in 2002 from the census are broadly similar for the

6The data used to produce migration data products come from individual income tax returns filed before late September of each calendar year and represent between 95% and 98% of total annual filings. 7We construct the measure of competition from imports from China by multiplying the fraction of employment in each county and in each industry by the share of imported goods from China as a fraction of total domestic shipments in the industry in the United States. The variation is virtually the same if we instead use the growth in the weight of imports for each industry as a fraction of U.S. domestic shipments between 1998 and 2005. The import data at the industry level is obtained from Peter K. Schott's website: http://faculty.som.yale.edu/peterschott/subinternational.htm.

80 two groups, with the one exception being the percentage of college-educated people (somewhat higher in low-elasticity areas). Average household income is also higher in those counties, but the difference is economically small (about 10% of the mean). As expected, counties with a low elasticity of housing supply experienced much stronger growth in house prices than did counties with a high elasticity of supply (a "crude" version of the first stage in our regressions) and similarly experienced a much larger increase in average debt-to-income ratio (consistent with Mian and Sufi, 2011a). Panel B of Table 2.1 shows how employment is distributed across the different employment-size categories. The biggest firm category, 50 employees or more, accounts for 51.7% of employment in 2002, whereas the smallest category, 1-4 employees, accounts for 8.9%. Growth in employment is stronger among larger companies in the 2002-2007 period, especially among the industries that we classify as having low start-up capital needs.

2.2.3 Empirical Model

We test whether increases in real estate prices affect the growth in employment by facilitating the creation of small businesses (collateral channel). To differentiate the collateral channel from a pure (expansionary) demand shock, we look at the differential effect of home prices on the net creation of establishments in different size categories.8 Our identification relies on the idea that improved availability of collateral in the form of higher house prices can positively affect the creation of small businesses, whereas it is likely to have no effect on the creation of larger establishments since these firms cannot be started with capital that can be extracted from a house. We measure the availability of collateral to small business entrepreneurs by the growth in house prices in the area where the establishment is located. However, it is challenging to establish a causal link from the availability of collateral to the creation of small businesses, since there are many omitted variables that could simultaneously affect both the value of real estate collateral and the demand faced by small businesses, including changes in household income in the area and improvements in investment opportunities. To overcome this difficulty, we instrument for the changes in house prices during our period of interest (2002-2007) using the elasticity of housing supply by MSA (see Saiz, 2010). Our identification relies on the assumption that the elasticity of the housing supply only impacts employment creation at establishments of different sizes through its effect on house prices. The exclusion restriction is violated if housing supply elasticity is correlated with employment or business creation for reasons other than house price growth. Similar approaches have been used extensively in the recent literature (see, e.g., Mian and Sufi, 2011a, 2011b; Charles, Hurst, and Notowidigdo, 2012; and Robb and Robinson, 2012). Davidoff (2012) argues that the

8As we discuss in the data section, our data do not include changes in employment within establishments (i.e., along the intensive margin), so our measure of changes in employment relies on multiplying the number of establishments in each size category by the midpoint of the number of employees in each bin. It is thus equivalent to interpret our results in terms of number of employees or number of establishments.

81 supply elasticity measure does not capture the severity of the boom-and-bust bust cycle of the 2000s. In our setting we are concerned only with price increases between 2002 and 2007, and the supply elasticity measure developed by Saiz is a strong predictor of the increase in prices (i.e., there is no weak instruments problem). As we describe below, we also include specifications that include county fixed effects that should further mitigate concerns about the cross-sectional elasticity measure. We rely on two basic regression specifications for our analysis. The first specification aggregates data up to the level at which our instrument varies - that is, at the county-year establishment-size level. Each individual observation is the change between 2002 and 2007 of employees in a given county, year, and establishment size. We thus add up the number of employees in all industries in each establishment category and take the growth in total number of employees as the dependent variable. We then run two-stage least squares regressions of the type:

0 0 7 7 2- 6 A 2- Employmentij = o + 0 1 ,AHpF2-O + 0 2 1i + 3 1jAHP0 07 + -lXj + Ej

We index counties by j and establishment size categories by i. A 02-0 7 Employmentij is the change in employment for establishment size category i in county j between 2002 and 2007. Similarly, AHP02- 0 7 is the growth in housing prices at the county level for the same time period where, as we discuss above, we instrument for the growth in house prices using the housing supply elasticity of Saiz (2010). 1i is a set of dummy variables for each of the four included establishment categories (we omit the largest category of more than 50 employees). We then also include the product of the establishment size dummies and the growth in house prices, and 33 is the coefficient of interest in our regressions. In particular, the test we are interested in is whether the coefficient for the smallest establishments is larger (and positive) than those of the larger categories, which would confirm that house prices had a stronger impact on the creation of small establishments. Xj is a set of county-level controls that include the size of the county, the percentage of the population with a bachelor's degree or higher, the percentage of the population that is employed, the percentage of the population in the labor force, the percentage of owner-occupied houses, and the county share of China imports. Standard errors in this specification are heteroskedasticity robust and clustered at the MSA level (given that the variation in the instrument we use is at this level as well), and all regressions are weighted by the number of households in a county as of 2000, as in Mian and Sufi (2011a). The second specification disaggregates observations to the county, year, establishment size, and four-digit NAICS level, yielding a much larger number of observations than the specification above (since each county now appears multiple times for each industry). When using these disaggregated data we can include industry fixed effects in the regression, which allows us to control even further for common shocks (namely, nationwide demand shocks) to each four-digit industry. The coefficients in this case represent the differential impact that house prices have on establishments of different sizes within each industry. The specification becomes:

82 2 0 7 0 7 A 02- 7 Employentijz = a + O1AHP - + /21 + 031iAHPj2- + 'XY + Z + ij

in which z indexes the industries and lz is a set of indicator variables for each industry. The breakdown at the industry level allows us to address an important alternative hypothesis to the mechanism we identify- namely, that higher home prices caused increased demand, which then prompted the growth in new businesses. This type of demand story (as opposed to the collateral lending channel) comes in two versions. The first is that rising house prices lead to an increase in demand because households feel richer or have access to home equity. This channel is proposed in Mian and Sufi (2011a) to explain the drop in employment during the Great Recession of 2007-2009. A second version of the demand hypothesis is that increasing house prices may benefit certain industries more than others and that these industries happen to be composed of smaller establishments on average (i.e., a "composition" effect). We address these alternative demand hypotheses in several ways. First, by holding constant industry fixed effects we identify how employment in the smallest establishments reacts differently from that of large establishments within each four-digit NAICS industry. This addresses the composition effect described above. Second, as we have argued before, a pure local demand story should affect establishments of all sizes similarly, whereas the credit collateral channel is relevant mainly for small businesses. There is, however, still the possibility that smaller firms are more sensitive to local demand shocks than large firms. To see if this effect could explain our results, we exclude the most obvious candidate industries that might directly benefit from local demand shocks due to higher house prices- namely, those linked to construction and firms in the non-tradable sector as classified in Mian and Sufi (2011a), and we repeat our tests only for manufacturing firms, those that should be least affected by local demand shocks. As a robustness check to our results we also implement the approach in Chaney, Sraer, and Thesmar (2012) by constructing the product of the nationwide conventional mortgage rate (obtained from the Federal Reserve data website) with the local elasticity of housing supply measure. This provides time-varying shocks to the demand for housing - when mortgage rates drop more, the shock to demand for housing should be larger, consistent with Adelino, Schoar, and Severino (2012). This shock then translates into higher prices in areas with a low elasticity of housing supply than in places where it is easy to build. This specification uses a panel of yearly observations at the county level and includes county fixed effects, unlike the previous two specifications. As before, we run two-stage least squares regressions of the form:

1 1 AEmploymentijt = a + 01 IAHPJt + 02 lit + ( 3 1stAHPt + h'1j + Y2t + Eij

The instrument for house prices is the product of mortgage rates and housing elasticity, not just the elasticity measure as before. We include county fixed effects (I),

83 which absorbs all county-level controls included in the previous two specifications, as well as year fixed effects.'

2.3 Empirical Results

2.3.1 House Prices and Employment at Small Establishments

Our central hypothesis is that the availability of more valuable collateral (in our case through increased real estate prices) in the period before the financial crisis has an effect on the creation of small firms or on self-employment, since it provides individuals with easier access to start-up capital. As a result, we should see a sharper increase in self-employment and employment in small businesses in areas that had steeper housing price appreciation. We also expect this effect to be concentrated in firms in the smaller size categories, since large firms cannot finance themselves using home equity. This hypothesis is tested in Table 2.2, where we run two-stage least squares regressions of the growth in employment between 2002 and 2007 on five establishment size categories and their interaction with house price growth in the same period. The instrument for house price growth, as we discuss above, is the Saiz (2010) measure of housing supply elasticity. In the first column of Table 2.2 we show the first-stage regression of house price growth on the Saiz measure of housing supply elasticity to confirm the validity of the instrument. The coefficient of -0.09 means that a one standard deviation increase in elasticity of housing supply is associated with an 11.7 percentage point lower growth in prices (for an average house price growth of 33.9%). The F statistic on this regression is 14.5 (above the conventional threshold of 10 for evaluating weak instruments). This reflects that MSAs with a higher elasticity of supply experienced significantly lower housing price growth between 2002 and 2007, in line with previous literature. In Column 2 we run a regression of employment change between 2002 and 2007 on the change in house prices during the same period. In this regression we do not instrument the change in house prices in order to show the raw correlation between house prices and employment. The effect is positive and economically large. A one standard deviation increase in house prices is associated with an increase in total employment of 3.95% over this period, for an average growth in employment of 10.6%. In the simple weighted least squares regression we see no distinction between the effect of home prices on small and large establishments. This result highlights the need for an instrument for our dependent variable of interest, given the numerous factors that are likely to drive both employment creation and house prices (income growth, investment opportunities, etc.). In Column 3 of Table 2.2 we repeat the same regression but instrument the change in house prices with the Saiz measure for the elasticity of housing supply. We see that there is a positive but not significant causal relation between county-level employment change and house price growth on average, in contrast to the results in the

9 We do not rely on the panel specifications for most tests because mortgage rates did not experience large drops in the period we analyze. We effectively have one large shock to demand for housing in the period 2002-2007, and the first two specifications capture this fact more clearly.

84 previous column. However, when we look at the differential effect of instrumented housing price changes, the increase in home prices has a significant and large positive effect on the small establishments but no significant effect on employment growth for big establishments (50 or more employees). The coefficient on the interaction term between house price growth and the one-to-four-employee size category shows that a 1% increase in house prices translates into a 0.19% increase in employment at these establishments relative to the largest ones. This translates into an increase in employment of 5.3% for a one standard deviation change in house prices, for an average change in employment at the smallest establishments of 9.4% (the effects of a one standard deviation change in house prices for each size category are shown in Ap- pendix Table 2.12). Furthermore, the effect of collateral is decreasing monotonically with firm size. For firms with more than ten employees, the effect is indistinguishable from that of the very largest firms. This is consistent with the collateral channel of house price appreciation being an important mechanism for small firm creation, since the amount of collateral that is provided by real estate appreciation is not enough to start a larger firm. Also, these results suggest that the causal impact of house prices on employment growth in 2002-2007 did not work through increased demand, since in that case firms of all sizes (including the very large) should have been affected. One concern with the above specification could be that the change in house prices in areas with low Saiz housing elasticity induces a local demand shock that especially affects certain industries. If those industries are also, on average, disproportionately made up of smaller establishments, the result above might reflect a composition effect rather than the collateral channel, as we suggest. Although a number of factors would need to line up in a very specific way, we cannot rule this concern out on face value with the specifications in Table 2.2. In order to eliminate the alternative hypothesis about industry composition, we use our more disaggregated data, which provides data at the county, four-digit NAICS, and establishment size level. This allows us to hold industry fixed effects constant and test whether, conditional on an industry, the growth of small establishments is significantly stronger than that of large establishments in counties with greater increases in home prices. Intuitively, this specification asks whether within an industry the fraction of employment generated by small firms grows more quickly than that of large firms. This way we can confirm that the results are not a consequence of changing industry composition. The results for this specification are shown in Column 4 of Table 2.2. As before, we find that impact of house price changes (instrumented with the Saiz measure) is stronger for establishments with one to four employees when compared to the bigger firm categories. We again find that the effect is monotonically decreasing and not statistically significant beyond firms with ten or more employees. To be more specific about which industries show the strongest effects from the collateral channel, in Table 2.17 we show the three-digit NAICS industries that are not construction, manufacturing, non-tradable, and finance, insurance, and real estate, as well as the employment share in each size bin. The sample includes a variety of services and wholesale activities, with significant cross-sectional variation in the proportion of employees in the very small establishment size categories (from 26.3% of employment in one-to-four-employee establishments in the case of "NAICS 425 - Wholesale Electronic Markets and Agents and Brokers" to 0% in this category

85 for "NAICS 622 - Hospitals"). The third version of the instrumented regression is shown in Column 5 of Table 2.2, in which we use yearly observations on county-level employment and construct a time- varying instrument by taking the product of the average conventional mortgage rate in the United States and the Saiz elasticity measure. We then add county and year fixed effects to the regressions and run the specification described in Section 2.3, above. The results are very consistent with the two previous specifications, with the same monotonically decreasing effect of house prices on employment at establishments of increasing size. We run the robustness specifications with the time-varying instrument and county fixed effects to account for time-invariant differences across regions that could be correlated with elasticity and new business starts. The fact that the results are consistent with our main specification alleviates these concerns. To confirm that the effect we estimate runs through the collateral channel, we test whether our estimated effect is stronger in industries that have lower start-up capital needs. We expect this to be the case, given that the median amount of home debt at its peak in 2006 for all U.S. households was approximately $117 thousand (Mian and Sufi, 2011b) and that only a fraction of this amount would be available for use in starting a business. Also, Adelino, Schoar, and Severino (2012) show that the average value of a single family home during this period is approximately $309 thousand and that most families obtain an 80% loan-to-value (LTV) loan. Even accounting for the fact that most entrepreneurs are over age 35 and that almost half are over 45 (Robb and Robinson, 2012), and so we expect them to have built home equity relative to the initial 80% LTV, it is not plausible to finance a large amount of capital using home equity as collateral. Brown, Stein, and Zafar (2013) show that the average amount of home equity lines of credit (HELOC) in the boom period is $2,623, with a standard deviation of $13,672. This implies that even homeowners who are two standard deviations above the mean have less than $30 thousand in home equity loans. The paper also shows that the fourth quartile of homeowners in high house-price appreciation areas has about $8,500 in HELOC. These numbers suggest the range of funds that can be obtained from homes as collateral for starting a business. We split our sample of industries at the median amount of capital needed to start a firm to explore this source of variation. As we describe in Section 2, above, we obtain this information from the SBO PUMS by selecting the sample of new firms in each industry and averaging the amount of capital needed to start those firms. We show the results split by the amount of start-up capital needed in each industry in Columns 6-11 of Table 2.2. The results show that the effect of collateral on employment growth in small establishments is stronger for industries in which the amount of capital needed to start a firm is lower (the average amount of start-up capital for industries below the median is approximately $132 thousand). In fact, for this subset of industries the effect is statistically significantly different from that of the largest group even for establishments with up to 49 employees- that is, the causal effect of house prices extends to establishments other than the very smallest. When we include industry fixed effects, only the coefficient on the smallest establishments is statistically different from zero. For the group of industries that require more start-up

86 capital, the effect of house prices on employment is smaller and statistically significant only for the very smallest group both with and without fixed effects. These results confirm that job creation at small businesses in response to house prices changes is strongest in industries with low start-up capital needs that can reasonably be financed through loans on home equity. Notice that the assumption underlying these tests is that the contribution of housing as collateral is more likely to matter at the margin for firms that require low amounts of capital than for firms that require a lot of capital. In fact, for firms that require large amounts of capital, we expect entrepreneurs to seek out additional sources of capital, and housing collateral is unlikely to be as important for the decision to start a firm.

Effect After Removing Non-tradable Industries

In this subsection we document that our results are not driven by certain industries, in particular construction and non-tradable firms. One possible concern is that the increase in house prices led to a growth in demand for construction services or for local services (e.g., local retail or restaurants), resulting in an increase in new firms in these industries (e.g., more remodeling and new housing construction, more dry cleaners). This would be a consequence of increased demand rather than an effect through the collateral channel. We rerun our main specifications excluding all industries linked to the construction and non-tradable sectors as classified by Mian and Sufi (2011a), as well as Finance, Insurance, and Real Estate firms (NAICS 52 and 53). We report these results in Table 2.3. The first takeaway from Table 2.3 is that the direction and magnitude of the effects are virtually unchanged when we remove these sectors from the regressions. If the effect we measure were driven largely by a local demand shock (instead of the collateral channel), we would expect the coefficient to be significantly affected when we remove from the sample the sectors that are most sensitive to local demand (Columns 1-3 of Table 2.3). In the last two columns of Table 2.3 we limit the regressions to the manufacturing sector. These industries are the least likely to be affected by local demand. At the same time, however, they typically require significant start-up capital, which makes it harder to find the effect of the collateral channel using our experiment. Still, we find that small firms created more employment relative to large firms in period 2002- 2007 in areas where housing prices rose more (Columns 4 and 5 of Table 2.3). The effect is similar in magnitude for establishments of one to four employees, five to nine, and ten to 19, but it is statistically significant only at conventional levels for the smallest size category. We know that, on average, firms in the manufacturing sector lost jobs during this period, and the coefficient on the largest firms suggests that they lost more jobs in places where house prices rose more (coefficient is -0.16). When we combine this effect with the coefficient on the small firms, this implies that access to collateral allowed the smallest firms to preserve employment, whereas the largest firms were losing jobs during this period. This confirms that a simple demand-side story is not driving our results and confirms the importance of the collateral channel for the creation of smaller establishments in the period 2002-2007.

87 In Table 2.4 we perform an additional test for manufacturing industries. In this test, we split industries based on the average distance of shipments in each three-digit NAICS industry and state. This addresses further the concern that local demand shocks might be driving the results for manufacturing firms. Table 4 we show that the result for manufacturing shown in Table 2.3 is driven by firms in industries and states that ship goods across large distances. The median reported distance in the sample is 600 miles, so firms that report shipping goods over more than 600 miles are unlikely to make decisions as a function of local demand shocks (details on the distances shipped by firms in each industry and state are in Appendix Table ??). One possible concern with the test using distances is that small firms in a given sector could be very different from large firms, so the small firms in those industries could depend more on local demand. Although we do not have shipment data by firm, in Table A7 we consider the relation between the reported distance shipped in a given state and industry cell and the share of small businesses in that cell. We use the same distance measure from before and separately compute the share of employment in establishments that have 50 or more employees for each state and three-digit NAICS manufacturing industry. Then, for each industry, we compute the average (over all states) of the distance shipped, as well as the average share of employees in firms that have 50 or more employees. Finally, for each state and industry observation, we compute the deviation from the industry mean for both measures and classify observations into deciles based on these deviations. 0 The takeaway from this table is that there is no visible relation between the distance shipped and the share of employees at large firms versus small firms. In particular, there is a lot of heterogeneity across industries in the fraction of small firms and the distance shipped. This should mitigate the concern that a strong positive relation between firm size and distance shipped might explain the results in the last two columns of Table 2.4. Our measure of growth of establishments by size category does not allow us to observe the creation and destruction of establishments directly, so in a separate set of regressions shown in Table A8 we use the Statistics of U.S. Businesses from the census to look at births and deaths of establishments at the two-digit NAICS industry level. The disadvantage of this data set is that it does not include the breakdown of establishments by employment size. Given that an overwhelming percentage of new businesses are very small (Haltiwanger, Jarmin, and Miranda, 2011; Robb and Robinson, 2012), this robustness test directly speaks to the validity of our main results. We find that births of establishments are very strongly affected by increasing house prices instrumented with the elasticity of housing supply and that the result holds when we consider the net creation of establishments (i.e., births minus deaths), and the coefficient is unchanged when we include two-digit NAICS fixed effects (the finest industry category available in this data set at a county level).

10 So, state-industry observations that are in the first decile of the distance are those that ship goods at short distances relative to the industry average. Similarly, those in the first decile of the share of employment at large firms, are state-industry observations that have few employees in large firms relative to the industry average.

88 Magnitude of the Collateral Effect Relative to Previous Work

One way to give a rough estimate of the importance of the collateral lending channel is to compare the magnitude of the employment gains that can be attributed to this channel to those that can be assigned to the demand channel shown in Mian and Sufi (2011a). To do so, we follow the same calculation used in that paper to aggregate the effect across all counties. The authors compute the effect of debt- to-income (DTI) ratios as of the beginning of the crisis on the employment change between 2007 and 2009 in non-tradable industries." These are the industries that are most likely to be affected by a drop in local demand due to overleveraged households. They aggregate this effect by computing the predicted change in employment in non- tradable industries and then extrapolating this effect to the rest of the economy. We perform essentially the same calculations for the period 2002-2007 to establish a benchmark employment effect that can be attributed to the demand channel. We start by obtaining the effect of a change in house prices on employment in the non- tradable industries at a county level for the 2002-2007 period. That regression is shown in Table 2.5 in Column 3. If we aggregate in the same way as described above (where the baseline employment is now as of 2002), we obtain an increase in employment in the non-tradable sector of 451.8 thousand jobs, which, given a share of employment in this sector of 18.4% in 2002, translates into a predicted total job gain due to increased aggregate demand of 2.452 million jobs. This is about 40% of the jobs created in the private sector in the 660 counties used for the calculation. The confidence interval for this estimation is very large and includes zero, which opens the possibility that the aggregate demand effect for the period before the crisis may actually be quite small. This is in sharp contrast to the estimates obtained by Mian and Sufi (2011a) for the years after the crisis, where the same regression yields very strong effects for the drop in demand on non-tradable employment. We now turn to the calculation of the magnitude of the collateral channel over the same period. We rely on the differential impact of house prices on employment creation at small firms relative to firms with 50 or more employees, and we focus on the specifications in which we exclude non-tradable industries and construction (Table 2.3, Column 2). We again first compute predicted county-level employment gains for these industries (relative to the 10th percentile county) and then we aggregate to all counties. When we do that, we obtain an estimated total job gain in firms with fewer than 50 employees relative to those with 50 or more employees of 1.698 million jobs in all counties, or 27.8% of jobs created between 2002 and 2007 in this period. If we restrict our attention to the specification where the demand explanation for our results is the least plausible - that is, the manufacturing sector and, in particular, firms in industries and states where the shipment distance is largest (Column 6 of Table 2.4), the same computation would yield an estimate of 676 thousand jobs, or about 11% of jobs created in this period and subset of counties. Section Al of the

"Using county-level debt-to-income ratio or the run-up in house prices between 2002 and 2007 as the independent variable (as we do in this paper) yields virtually the same results, as counties with high debt-to-income by the end of this period are also the ones that experienced large increases in home values.

89 appendix describes the calculation we perform in more detail. The magnitude we estimate above is a lower bound for the total importance of collateral for job creation for two reasons. First, our data do not allow us to track firms over time, so if a firm grows to become very large, we do not attribute the employment creation of that firm to our effect (it would be in the 50+ category that we use as our baseline). Second, we are focusing on the importance of this channel for very small firms. This ignores the role that collateral value plays for larger firms, as pointed out in Chaney, Sraer, and Thesmar (2012), Cvijanovic (2013), and Chakraborty, Goldstein, and MacKinlay (2013). Last, this exercise is useful as a comparison to previous work and not as a proper calibration of the importance of the collateral effect for the whole economy. In extending the effect that we observe for a subset of firms and industries in individual counties to the whole economy, we ignore general equilibrium effects that could potentially be important.

2.3.2 Sole Proprietorships

We now expand our analysis to include the creation of businesses without employees, also called sole proprietorships or nonemployer businesses. Table 2.6 shows the effect of housing price growth on net creation of proprietorships relative to all the establishment categories listed in the previous tables using the Saiz measure to instrument for exogenous movements in housing price changes. The first column in this table uses employment data on sole proprietorships from the BEA, while the last three columns rely on census data on nonemployer establishments (which includes information on the two-digit NAICS sector in which the establishment operates). The coefficient on housing price growth in Column 1 interacted with the sole proprietorship category is significantly different from that on the largest establishments and close in magnitude to that on the 1-4-employee category. In Column 2 we use census data and find a smaller coefficient on the sole proprietorships, and we cannot distinguish that coefficient from the others in the regression. In the last two columns we again split the sample by the amount of capital needed to start a business in a given industry, as discussed above. We find that the effect of home prices on the net creation of sole proprietorships is stronger in industries with low start-up capital needs, which is in line with our findings for the other size categories. Note, however, that the difference between the coefficients in the two specifications (below and above median capital needs) is not statistically significant.

2.3.3 Crisis Period (2007-2009) One question that remains regarding the business establishments created as a consequence of the increasing value of collateral during the rise in house prices is whether these establishments were then eliminated after the housing bubble burst. In this section we try to distinguish whether these newly created businesses were particularly fragile and were disproportionately affected by the crisis or, alternatively, whether they behaved like the rest of the firms in the economy.

90 Our data do not allow us to track individual establishments, so we cannot know whether the specific firms created in the 2002-2007 period survived the crisis. We can, however, test whether small establishments in general were more or less likely to downsize or disappear in the crisis. That is, we can assess whether employment loss was stronger at larger or smaller firms during the crisis in counties where the increase in house prices had been stronger in the precrisis period (which are also the most leveraged counties, as shown in Mian and Sufi, 2011a). We run those regressions in Table 2.7. The results show that employment loss was either similar across large and small establishments or, if anything, was worse at large firms (in the specifications without industry fixed effects) in counties where house prices rose more. This suggests that, at least as a group, small firms were no more likely to destroy jobs as a consequence of the increased leverage accumulated during the precrisis period. This is consistent with the findings of Mian and Sufi (2011a) regarding non-tradable industries for this period.

2.3.4 Migration Our final consideration is the effect of house price changes on the net migration of people in and out of each county. We measure net migration as the difference between inflows and outflows of individuals at the county level. Table 2.8 shows county-level regressions of county-to-county Net Migration, as well also Inflows and Outflows separately, on house prices changes instrumented with the Saiz measure and the same county-level controls as the previous tables. The results on migration show no significant effect of the (instrumented) change in house prices on net migration. This masks stronger results when we break down the results by inflows and outflows. Indeed, counties that experience higher growth in house prices had larger outflows that were offset in part by somewhat bigger inflows of people at the same time. This alleviates the concern that low-elasticity counties experience high growth in demand due to large in-migration. If anything, the results seem to suggest the opposite. Of course, we cannot observe who is entering and who is migrating out of each county, so we cannot address the more detailed question of whether entrepreneurs were moving in as other individuals were moving out, but the aggregate trends suggest stronger outflows than inflows in the high-appreciation areas.

2.3.5 Credit Conditions and Elasticity of Housing Supply One possible concern with the instrument we use is that the behavior of lenders in high- and low-elasticity areas during our time frame was different. Specifically, if it became easier to obtain credit in low-elasticity areas relative to high-elasticity areas during our sample period for reasons unrelated to collateral availability, and if this drove the creation of new businesses, this would violate the exclusion restriction for our instrument. One mechanism for such an effect would be that banks might become laxer on all their credit decisions because of the improvement on the quality of their mortgage portfolio due to higher house prices. Although the evidence points

91 to commercial lending having become more difficult in places where house prices boomed (Chakraborty, Goldstein, and MacKinlay, 2013), making it unlikely that small business credit provision became easier because of stronger mortgage portfolios, we wish to address this concern directly. To test whether such an effect is plausible, we use data on denial rates of mortgage applications from HMDA. The underlying assumption is that the cross-sectional variation on the looseness of credit conditions should be positively correlated with the same variation for mortgage credit, especially given that the reason why credit might have become laxer is the fact that house prices increased. We consider the number of applications that are denied by financial institutions as a proportion of the total loan applications in a county and in a year." Using the yearly estimates we compute the proportional change in denial rates between 2002 and 2007. We focus on loans used for purchasing homes because they are less sensitive to the issue of relationship lending and/or private lender information about the borrower and therefore should better reflect the loosening of credit conditions. Panel A of Table 2.9 shows that credit conditions tightened rather than loosened in low-elasticity areas (those below median elasticity in the sample) when we use this measure of credit supply. Denial rates increased by about 2% in counties with low elasticity of housing supply, whereas they go down in high-elasticity areas by 1% - that is, credit loosened in those areas. The difference between the two types of counties is statistically significant at the 1% level. In addition, total volume of applications decreases by 1% in low-elasticity areas in comparison to the 10% increase in the high-elasticity areas. We formally test these differences in a regression framework using a continuous elasticity measure as our independent variable. Panel B of Table 2.9 shows the results. Consistent with the summary statistics of Panel A, we find that lower elasticity is associated with higher denial rates of loan applications, and these results are robust to different specification and controls. Although the regressions condition on the applicant pool (and so the denial ratc could mask riskier borrowers applying for loans), we control for the debt-to-income in these regressions to account for changes in applicant types. Overall, this result allows us to rule out the concern that our instrument is picking up changes in the way that lenders granted credit instead of access to credit through an increase in collateral values.

2.4 Conclusion

Overall, the evidence we present identifies the causal effect of rising house prices in the creation of new small firms. Increased access collateral allowed individuals to start small businesses or to become self-employed. We conjecture that without access to this collateral in the form of real estate assets, many individuals would not have

1 2 Volume of applications is calculated as the sum of all loans that are originated plus applications that are approved but not accepted, applications denied by the financial institution, and loans purchased by the financial institution itself.

92 made the transition to starting a new business or self-employment. Our study is in line with recent survey evidence from the NY Fed" that shows that: (i) access to capital is the top growth challenge for small firms in 2013; (ii) the most cited reason for not receiving credit is insufficient collateral; and (iii) that the most used form of collateral for small businesses is personal real estate (in line also with the findings of Kleiner, 2013). This implies that the effect we uncover is a collateral effect and not the result of changing household risk-aversion due to increased wealth (as suggested by Kihlstrom and Laffont, 1979). We show that the effect of house prices is concentrated in small firms only and has no causal effect on employment at large firms. Importantly, our results also hold when we exclude industries that are most likely to be affected by local demand shocks and when we restrict our attention to manufacturing industries. The effect of house prices is also stronger in industries where the amount of capital needed to start a new firm is lower, consistent with the hypothesis that housing serves as collateral but is not sufficient to fund large capital needs. Our results on the collateral effect on the upside (2002-2007) and after the crisis hit, paired with the results on the effect of demand on job creation, suggest an interesting asymmetry of these effects. Collateral was particularly important in explaining job creation when more collateral became available, but we observe no significant destruction when collateral became scarce. This is consistent with a "bright side" of bubbles (as suggested in Caballero, Farhi, and Hammour, 2006, although the effect we emphasize is quite different). On the other hand, a drop in demand is a strong predictor of employment loss, but a similar shock on the upside (at least in the recent experience) does not seem as powerful in predicting where jobs will be created.

2.5 Bibliography

Adelino, M., Schoar, A., Severino, F., 2012. Credit supply and house prices: evidence from mortgage market segmentation. NBER Working Paper No. 17832. Berger, A., Udell, G. F., 1998. The economics of small business finance: the roles of private equity and debt markets in the financial growth cycle. Journal of Banking and Finance 22, 613-673. Bernanke, B. S., Gertler, M., 1989. Agency costs, net worth, and business fluctu- ations. American Economic Review 79, 14-31. Black, J., De Meza, D., Jeffreys, D., 1996. House prices, the supply of collateral and the enterprise economy. Economic Journal, 60-75. Brown, M., Stein, S., Zafar, B., 2013. The impact of housing markets on consumer debt: credit report evidence from 1999 to 2012. Federal Reserve Bank of New York Working Paper. Caballero, R. J., Farhi, E., Hammour, M. L., 2006. Speculative growth: hints from the US economy. American Economic Review, 1159-1192.

1 3 Small Business Credit Survey, May 2013, Federal Reserve Bank of New York, available at http://www.newyorkfed.org/smallbusiness/2013/.

93 Cagetti, M., De Nardi, M., 2006. Entrepreneurship, frictions, and wealth. Journal of Political Economy 114, Campbell, J. Y., Cocco, J. F., 2007. How do house prices affect consumption? Evidence from micro data. Journal of Monetary Economics 54, Chakraborty, I., Goldstein, I., MacKinlay, A., 2013. Do asset price bubbles have negative real effects? Unpublished working paper. Chaney, T., Sraer, D., Thesmar, D., 2012. The collateral channel: how real estate shocks affect corporate investment. American Economic Review 102, 2381-2409. Charles, K. K., Hurst, E., Notowidigdo, M. J., 2012. Manufacturing busts, housing booms, and declining employment: a structural explanation. Unpublished working paper. Corradin, S., Popov, A., 2013. House prices, home equity and entrepreneurships. ECB Working Paper. Cvijanovic, D., 2013. Real estate prices and firm capital structure. Unpublished working paper. Davidoff, T., 2012. Supply elasticity and the housing cycle of the 2000s. Real Estate Economics, forthcoming. Duranton, G., Morrow, P., Turner, M. A., 2013. Roads and trade: evidence from the US. CEPR Discussion Papers 9393. Evans, D. S., Jovanovic, B., 1989. An estimated model of entrepreneurial choice under liquidity constraints. Journal of Political Economy 97, 808-827. Fairlie, R. W., Krashinsky, H. A., 2012. Liquidity constraints, household wealth, and entrepreneurship revisited. Review of Income and Wealth 59, 279-306. Fan, W., White, M. J., 2003. Personal bankruptcy and the level of entrepreneurial activity. Journal of Law and Economics 46. Fort, T., Haltiwanger, J. C., Jarmin, R. S., Miranda, J., 2012. How firms respond to business cycles: the role of firm age and firm size. Unpublished working paper. Fracassi, C., Garmaise, M. J., Kogan, S., Natividad, G., 2013. How much does credit matter for entrepreneurial success in the United States? Unpublished working paper. Gentry, W. M., Hubbard, R. G., 2005. Success taxes, entrepreneurial entry, and innovation. In: Innovation Policy and the Economy, Volume 5. MIT Press, Cam- bridge, pp. 87-108. Gertler, M., Gilchrist, S., 1994. Monetary policy, business cycles, and the behavior of small manufacturing firms. Quarterly Journal of Economics 109, 309-340. Greenstone, M., Mas, A., 2012. Do credit market shocks affect the real economy? Quasi-experimental evidence from the Great Recession and "normal" economic times. Unpublished working paper. Haltiwanger, J. C., Jarmin, R. S., Miranda, J., 2011. Who creates jobs? Small vs. large vs. young. Unpublished working paper. Holtz-Eakin, D., Joulfaian, D., Rosen, H. S., 1994. Entrepreneurial decisions and liquidity constraints. NBER Working Paper No. 4526. Hurst, E., Lusardi, A., 2004. Liquidity constraints, household wealth, and entrepreneurship. Journal of Political Economy 112, 201-248.

94 Iacoviello, M., 2005. House prices, borrowing constraints, and monetary policy in the business cycle. American Economic Review 95, 739-761. Kashyap, A. K., Stein, J. C., 1994. Monetary policy and bank lending. In: Monetary Policy, University of Chicago Press, Chicago, pp. 221-261. Kashyap, A. K., Stein, J. C., Wilcox, D. W., 1993. Monetary policy and credit conditions: evidence from the composition of external finance. American Economic Review 83, 78-98. Kihlstrom, R. E., Laffont, J.-J., 1979. A general equilibrium entrepreneurial theory of firm formation based on risk aversion. Journal of Political Economy 719-748. Kiyotaki, N., Moore, J., 1997. Credit cycles. Journal of Political Economy 105, 211-248. Kleiner, K., 2013. How real estate drives the economy: an investigation of small firm balance sheet shocks on employment. Unpublished working paper. Kocherlakota, N., 2010. Inside the FOMC. President's speech, Marquette, MI, August 17, 2010. Mian, A., Sufi, A., 2011a. What explains high enemployment? The aggregate demand channel. NBER Working Paper No. . Mian, A., Sufi, A., 2011b. House prices, home equity based borrowing, and the U.S. household leverage crisis. American Economic Review 101, 2132-2156. Petersen, M. A., Rajan, R. G., 1994. The benefits of lending relationships: evidence from small business data. Journal of Finance 49, 3-37. Rampini, A. A., Viswanathan, S., 2010. Collateral, risk management, and the distribution of debt capacity. Journal of Finance 65, 2293-2322. Robb, A. M., Robinson, D. T., 2012. The capital structure decisions of new firms. Review of Financial Studies, forthcoming. Romer, C. D., 2011. Dear Ben: It's time for your Volcker moment. New York Times, October 29, 2011. Saiz, A., 2010. The geographic determinants of housing supply. Quarterly Journal of Economics 125, 1253-1296. Schmalz, M., Sraer, D., Thesmar, D., 2013. Housing collateral and entrepreneurship. Unpublished working paper.

95 Table 2.1: Summary Statistics

Panel A

All Counties High Elasticity Low Elasticity Total Employment (2002) 113,918 69,057 157,523 45,454 33,228 63,286 (238,831) (129,569) (304,041) Unemployment Rate (2002, percent) 5.4 5.3 5.4 5.3 5.2 5.4 (1.5) (1.5) (1.4) Number of Households (2000, thousands) 100.2 59.3 139.8 46.2 34.2 66.4 (188.1) (92.6) (241.4) Growth in Total Employment (02-07, percent) 10.6 10.2 11.0 8.2 7.5 8.9 (15.8) (16.9) (14.5) Growth in DTI (02-07, percent) 51.8 36.6 66.3 42.6 34.9 58.3 (36.4) (23.0) (40.7) Growth in Income (02-07. percent) 27.6 27.2 28.0 23.9 23.0 24.5 (21.1) (24.2) (17.6) Growth in House Prices (02-07, percent) 33.9 23.5 43.7 26.8 19.4 40.9 (21.1) (14.3) (21.9) Change in Unemployment Rate (02-07, percent) -0.9 -0.7 -1.0 -0.8 -0.5 -1.0 (1.0) (0.9) (1.0) Number of Counties 775 382 393

Panel B

1-4 Emp 5-9 Emp 10-19 Emp 20-49 Emp 50+ Emp Emp. in All Sectors Total 9,101 9,122 12,819 21,466 72,939 Growth (02-07) 9.4 8.0 12.5 10.6 13.3 Percentage of Total 8.9 9.0 12.1 18.3 51.7 Emp. in Firms P50 of Start-Up Capital Total 2.866 3,542 5,454 10,433 32,975 Growth (02-07) 6.9 4.4 13.1 9.6 9.3 Percentage of Total 5.8 7.4 11.7 20.5 54.6 Note. Panel A reports summary statistics for all counties in the sample in Column 1, and Columns 2 and 3 show the summary statistics for counties above and below the median elasticity of housing supply in the sample. For each variable we show the pooled average, median (italicized) and standard deviation (in parenthesis). The last column shows the t-statistic for the difference in means of the two groups, adjusted for clustering at the Metropolitan Statistical Area level. Total Employment refers to the total number of employees in a county in thousands across all establishment sizes and industries using the County Business Patterns data as of 2002. Unemployment Rate is shown in percentage and comes from the Bureau of Labor Statistics Local Area statistics in 2002. Percent College Educated is the percentage of the population with a college degree, Percent Employed is the percentage of the labor force that is employed, Workforce as a Percentage of Population is the share of the population in the workforce, and Percent of Homes Owner-occupied is the percentage of homes that are owner-occupied (i.e., not rental properties). Average Household Income is the total income in a county divided by the number of households as of 2002 and Growth in Income is the percentage change in income in a county between 2002 and 2007. Change in DTI is the percentage change in debt to income ratio in the same period. The debt to income ratio is estimated using county level household debt data from the New York Fed-Equifax and income is computed using IRS county-level information. Growth in House Prices is the percentage change in house prices between 2002 and 2007 at the MSA level from the Federal Housing Finance Agency. Panel B shows the Total Employment in 2002 in thousands, Employment Growth between 2002 and 2007 in percentage points, and the percentage of Total Employment for each establishment size for all firms, as well as split by the start-up amount of capital needed to start a firm.

96 Table 2.2: Employment Growth, Firm Size, and House Price Appreciation

First Stage All Industries All Industries Start-up Capital Start-up Capital (WLS) (IV) < P50 > P50 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Housing Supply Elasticity -0.09*** (0.02)

Growth in House Prices 0.19*** 0.05 -0.06 0.02 -0.01 -0.04 -0.11*** 0.06 -0.07 0.10** (0.04) (0.06) (0.10) (0.03) (0.07) (0.13) (0.04) (0.07) (0.10) (0.04)

Growth in House Prices * 1-4 Employees 0.03 0.20*** 0.26** 0.16*** 0.33*** 0.32** 0.31*** 0.14** 0.18** 0.10* (0.03) (0.05) (0.09) (0.05) (0.07) (0.12) (0.06) (0.06) (0.09) (0.06)

Growth in House Prices * 5-9 Employees -0.02 0.08** 0.17 0.00 0.19*** 0.14 0.17*** 0.04 0.19** -0.10* (0.03) (0.04) (0.10) (0.04) (0.05) (0.15) (0.05) (0.06) (0.08) (0.05)

Growth in House Prices 10-19 Employees -0.02 0.01 0.06 -0.05 0.14*** 0.02 (.10* -0.07 0.09 -0.12** (0.02) (0.04) (0.09) (0.04) (0.05) (0.12) (0.05) (0.06) (0.09) (0.05)

Growth in House Prices * 20-49 Employees 0.01 0.00 0.07 -0.07** 0.13*** 0.10 0.06 -0.07 0.02 -0.14*** (0.02) (0.04) (0.07) (0.03) (0.05) (0.10) (0.05) (0.05) (0.08) (0.04)

Log of the Population 0.00 -0.02*** -0.02*** -0.04*** -0.03*** -0.05*** -0.02*** -0.04*** (0.03) (0.01) (0.01) (0.01) (0.01) ((1.01) (0.01) (0.01)

Percent College Educated 0.00 0.00** 0.00** 0.00 0.00 0.00 0.00** 0.00 -- I (0.00) (0.00) ((1.))) (1.1 ) (0.))) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) -o.01*** 0.00 0.00 0.00 0.00 0.00 0.00** 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -0.69 -1.09*** -1.11*** -0.86*** -1.16*** -1.00*** -1.08*** -0.72*** (0.63) (0.19) (0.19) (0.22) (0.20) (0.25) (0.20) (0.21)

Percent of Homes Owner-occupied 0.00 0.00** 0.00** 0.00 0.00** 0.00 0.00 0.00*

((1.00) (1.1)0) (1.1 )) ((.1) (0.0) (1.1)) (11.111) (11.1)

China Import Share in County (2005) 0.10 0.09 0.12 -0.08 0.33 0.08 -0.01 -0.19 (0.91) (0.23) (0.23) (0.32) (0.26) (0.38) (0.22) (0.30)

4-Digit Industry Fixed Effects - - Y Y - Y County Fixed Effects Y Y Y Number of Observations 731 3,653 3.653 373.576 21,962 3,653 196,027 21.954 3,651 177,549 21,949 R2 0.30 0.27 0.22 0.30 0.02 0.21 0.39 0.00 0.14 0.10 0.03 The table shows two-stage least squares regressions of employment growth on house price growth instrumented with the elasticity of housing supply, indicator variables for each establishment size (not shown in the table) and interactions of house price growth with the size of establishments. All regressions are weighted by the number of households in a county as of 2000. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data. Growth in House prices is the percentage change between 2002 and 2007, and each interaction is with a dummy indicator for the size of the establishment. Column 1 shows the first stage regression of the change in house prices between 2002 and 2007 on the Saiz elasticity measure. Columns 2 through 5 "All Industries" shows the results for the whole sample of firms, first the weighted least squares results, then the IV at a county level, the IV results at a county and industry level and then the IV results using yearly observations and the interaction of the elasticity measure with the conventional mortgage rates as the instrument. Columns 6 through 11 show the coefficients split by the start-up capital amount (above and below the median) also at the county, at the county and industry level, and at the county level with yearly observations. The omitted category refers to establishments with 50 or more employees. All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. Controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, **, *** indicate statistical significance at 10, 5, and 1% levels, respectively. Table 2.3: Employment Growth and House Prices: Excluding Construction, Non-Tradable, and Finance Industries and Con- sidering Manufacturing Only

Drop Drop Const. Drop Const., Manufacturing Manufacturing (Tradable) Construction and Non-Trad. Non-Trad. and F.I.R.E. (Tradable) Growth in House Prices -0.09 -0.12 -0.14 -0.17 -0.16 (0.10) (0.10) (0.10) (0.11) (0.12)

Growth in House Prices * 1-4 Employees 0.27*** 0.32*** 0.35*** 0.13* 0.15* (0.09) (0.09) (0.10) (0.07) (0.09)

Growth in House Prices * 5-9 Employees 0.19* 0.21* 0.24** 0.12 0.10 (0.10) (0.11) (0.11) (0.08) (0.09)

Growth in House Prices 10-19 Employees 0.08 0.12 0.12 0.11 0.16 (0.09) (0.09) (0.09) (0.11) (0.11)

Growth in House Prices * 20-49 Employees 0.08 0.12* 0.11* 0.01 -0.05 (0.06) (0.06) (0.06) (0.12) (0.09)

Log of the Population -0.04*** -0.04*** -0.04*** -0.02** -0.02* (0.01) (0.01) (0.01) (0.01) (0.01)

Percent College Educated 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (1.00) (0.00) (0.00)

Workforce as a Percentage of Population -0.88*** -0.84*** -0.84*** -0.64** -0.66** (0.22) (0.23) (0.24) (0.29) (0.30)

Percent of Homes Owner-occupied 0.00 0.00 0.00* 0.00* 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

China Import Share in County (2005) -0.11 -0.23 -0.28 -0.88* -1.24** (0.34) (0.36) (0.36) (0.50) (0.56) Controls Y Y Y Y Y 4-Digit Industry Fixed Effects Y Y Y Y Y Number of Observations 325,349 264,901 242,510 55,345 44,649 R2 0.29 0.30 0.31 0.02 0.02 Growth HP * 1-4 E. = Growth HP * 5-9 E. 0.04** 0.02** 0.02** 0.95 0.48 Growth HP * 1-4 E. = Growth HP * 10-19 E. 0.00*** ).00*** 0.00*** 0.85 0.91 Growth HP * 1-4 E. = Growth HP * 20-49 E. 0.00*** 0.00*** 0.00*** 0.33 0.10* The table shows two stage least squares regressions of employment growth on house price growth instrumented with the elasticity of housing supply, indicator variables for each establishment size (not shown in the table) and interactions of house price growth with the size of establishments. Each observation is at a county, 4-digit NAICS industry, and establishment size level. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data. Growth in House prices is the percentage change between 2002 and 2007, and each interaction is with a dummy indicator for the size of the establishment. All regressions include 4-digit industry fixed effects. Column 1 shows the results when we exclude construction industries, column 2 excludes both construction and non-tradable industries, column 3 also excludes finance, insurance and real estate-related industries (NAICS codes 52 and 53), column 4 includes only manufacturing industries (NAICS 31 to 33) and column 5 has manufacturing industries that are classified as "tradable" in Mian and Sufi (2011a). All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner-occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, *, * denote statistical significance at the 10, 5, and 1% levels, respectively. Table 2.4: Breakdown of Manufacturing Industries by Distance Shipped

Manufacturing Manufacturing Dist. Shipped P50 Growth in House Prices -0.11 -0.29** (0.17) (0.14)

Growth in House Prices * 1-4 Employees 0.07 0.21** (0.14) (0.09)

Growth in House Prices * 5-9 Employees 0.11 0.20** (0.17) (0.09)

Growth in House Prices * 10-19 Employees -0.03 0.24** (0.17) (0.11)

Growth in House Prices * 20-49 Employees 0.06 0.04 (0.30) (0.12)

Log of the Population -0.02 -0.02* (0.02) (0.01)

Percent College Educated 0.00 0.00 (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 (0.00) (0.00)

Workforce as a Percentage of Population -0.42 -0.58* (0.36) (0.32)

Percent of Homes Owner-occupied 0.00 0.00* (0.00) (0.00)

China Import Share in County (2005) -0.29 -1.21** (0.45) (0.58) Controls Y Y 4-Digit Industry Fixed Effects Y Y Number of Observations 27,599 27,294 R2 0.02 0.02 Growth HP * 1-4 E. = Growth HP * 5-9 E. 0.82 0.90 Growth HP * 1-4 E. = Growth HP * 10-19 E. 0.59 0.77 Growth HP * 1-4 E. = Growth HP * 20-49 E. 0.96 0.13 The table shows two-stage least squares regressions of employment growth on house price growth instrumented with the elasticity of housing supply, indicator variables for each establishment size (not shown in the table) and interactions of house price growth with the size of establishments. Each observation is at a county, 4 digit NAICS industry, and establishment size level. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data for manufacturing industries (NAICS codes 31 to 33). Growth in House prices is the percentage change between 2002 and 2007, and each interaction is with a dummy indicator for the size of the establishment. All regressions include 4 digit NAICS fixed effects. The table splits industries and states based on the median of the shipment distance distribution (about 600 miles). Data for distance shipped is from the Census Commodity Flow Survey for 2007 and represents a dollar weighted average of shipment distance calculated at the 3 digit NAICS and state of origin level. All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, **' *** denote statistical significance at the 10, 5, and 1% levels, respectively.

99 Table 2.5: Employment and House Price Appreciation across Industry Types

First Stage All Industries Non-Tradable Tradable Construction Others Housing Supply Elasticity (0.02)

Growth in House Prices 0.09 0.10 -0.01 0.32*** 0.06 (0.06) (0.07) (0.11) (0.08) (0.06)

Log of the Population 0.00 -0.02** -0.01 -0.02** -0.02* -0.03 (0.03) (0.01) (0.01) (0.01) (0.01) (0.01)

Percent College Educated 0.00 0.00* 0.00** 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) -0.01*** 0.00 0.00* 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -0.69 -1.15*** -1.13*** -0.82 -0.83** -1.35 (0.63) (0.23) (0.28) (0.51) (0.37) (0.24)

Percent of Homes Owner-occupied 0.00 0.00** 0.00 0.00** 0.00(** 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

China Import Share in County (2005) 0.10 -0.23 0.42 -1.94*** -0.52 0.42 (0.91) (0.28) (0.32) (0.47) (0.42) (0.32)

Number of Observations 731 731 731 730 731 731 R2 0.30 0.24 0.18 0.10 0.30 0.21 The table shows two stage least squares regressions at a county level of employment growth on house price growth between 2002 and 2007. Each observation is at a county level. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data. Industry type definitions follow Mian and Sufi (2011a). All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, **, *** denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table 2.6: Proprietorships and House Price Appreciation

BEA Census Start-up Capital Start-up Capital Data Data < P50 (Census) > P50 (Census) Growth in House Prices 0.02 0.03 -0.04 0.05 (0.06) (0.06) (0.07) (0.07)

Growth in House Prices * Proprietorships 0.14* 0.06 0.12* 0.08 (0.07) (0.06) (0.06) (0.08)

Growth in House Prices * 1-4 Employees 0.20*** 0.20*** 0.33*** 0.14** (0.05) (0.05) (0.07) (0.06)

Growth in House Prices * 5-9 Employees 0.08** 0.08** 0.19*** 0.04 (0.04) (0.04) (0.05) (0.06)

Growth in House Prices * 10-19 Employees 0.01 0.01 0.14*** -0.07 (0.04) (0.04) (0.05) (0.06)

Growth in House Prices * 20-49 Employees 0.00 0.00 0.13** -0.07 (0.04) (0.04) (0.05) (0.05)

Log of the Population -0.02** -0.02** -0.02*** -0.02** (0.01) (0.01) (0.01) (0.01)

Percent College Educated 0.00** 0.00* 0.00 0.00** (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -1.02*** -1.16*** -1.21*** -1.13*** (0.19) (0.20) (0.21) (0.21)

Percent of Homes Owner-occupied 0.00** 0.00** 0.00** 0.00* (0.00) (0.00) (0.00) (0.00)

China Import Share in County (2005) 0.02 0.03 0.18 -0.02 (0.22) (0.23) (0.24) (0.23)

Number of Observations 4,381 4,384 4,384 4,382 R2 0.48 0.38 0.31 0.28

The table shows two-stage least squares regressions at a county level of employment growth on house price growth, indicator variables for each establishment size (not shown in the table) and interactions of house price growth with the size of establishments. Proprietorships are establishments with zero employees. Each observation is at a county and establishment size level. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data except in the case of proprietorships. The data on growth in proprietorships is obtained from the Bureau of Economic Analysis in the first column and from the Census in Columns 2 to 4. All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner-occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, *, *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

101 Table 2.7: Employment Growth, Firm Size, and House Price Appreciation, Crisis Period (2007-2009)

All Industries All Industries Start-up Capital Start-up Capital (WLS) (IV) < P50 (IV) > P50 (IV) Growth in House Prices -0.04* -0.12*** -0.13*** -0.14*** (0.02) (0.03) (0.04) (0.04)

Growth in House Prices * 1-4 Employees 0.04** 0.10*** 0.11*** 0.13*** (0.02) (0.03) (0.04) (0.05)

Growth in House Prices * 5-9 Employees 0.01 0.05* 0.05* 0.09 (0.02) (0.03) (0.03) (0.05)

Growth in House Prices * 10-19 Employees 0.00 0.06* 0.07** 0.09** (0.02) (0.03) (0.03) (0.04)

Growth in House Prices * 20-49 Employees -0.02 0.02 0.00 0.07 (0.02) (0.03) (0.03) (0.05)

Log of the Population -0.01*** 0.00** 0.00* -0.01*** (0.00) (0.00) (0.00) (0.00)

Percent College Educated 0.00*** 0.00*** 0.00*** 0.00*** (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00* 0.00*** 0.00*** 0.00*** (0.00) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -0.25*** -0.26*** -0.26*** -0.25*** (0.07) (0.06) (0.07) (0.07)

Percent of Homes Owner-occupied 0.00*** 0.00*** 0.00*** 0.00*** (0.00) (0.00) (0.00) (0.00)

China Import Share in County (2005) 0.12* 0.14* 0.25*** 0.06 (0.07) (0.08) (0.09) (0.08)

Number of Observations 3,654 3,654 3.651 3,653 R2 0.16 0.12 0.08 0.13 The table shows two-stage least squares regressions of employment growth between 2007 and 2009 on house price growth for the previous 5 years (2002-2007), indicator variables for each establishment size (not shown in the table) and interactions of house price growth with the size of establishments. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2007 and 2009 estimated using County Business Patterns (CBP) data. Growth in House prices is the percentage change between 2002 and 2007, and each interaction is with a dummy indicator for the size of the establishment. Columns 1 and 2, All Industries, shows the results for the whole sample of firms (first the weighted least squares results and then the IV), Columns 3 to 6 show the coefficients split by the startup capital amount. The omitted category refers to firms with 50 or more employees. The first column for each sample of industries is aggregated at the county and establishment size level, whereas the second column is at the county, establishment size and industry level, and includes industry fixed effects. All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, **, *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

102 Table 2.8: Total Employment, Unemployment, and Migration

Total Employment Unemp. Unemp. Rate Net Migration Inflows Outflows Growth in House Prices 0.09 -0.20 -1.29** -0.16 0.19 0.34** (0.06) (0.14) (0.66) (0.12) (0.12) (0.17)

Log of the Population -0.02*** -0.01 0.03 0.00 -0.07*** -0.07*** (0.01) (0.02) (0.10) (0.01) (0.01) (0.01)

Percent College Educated 0.00** -0.01***-0.03*** 0.00 0.01*** 0.00*** (0.00) (0.00) (0.01) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 0.04** 0.00 0.00 0.00 (0.00) (0.00) (0.02) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -1.15*** -0.13 3.94 -0.01 -0.63* -0.62** (0.23) (0.52) (2.67) (0.19) (0.34) (0.26)

Percent of Homes Owner-occupied 0.00** 0.00*** 0.03*** 0.00** .0** -0.01*** (0.00) (0.00) (0.01) (0.00) (0.00) (0.00)

China Import Share in County (2005) -0.23 -0.60 -4.76 0.19 -1.08*** -1.27*** (0.28) (0.64) (3.65) (0.29) (0.28) (0.44)

Number of Observations 731 721 721 731 731 731 R2 0.24 0.26 0.33 0.41 0.18 The table shows two stage least squares regressions at a county level of the net migration on house price growth between 2002 and 2007. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Net Migration, Inflows and Outflows are obtained from the IRS county to county migration data series. Net Migration is calculated by county using inflows of taxpayers minus outflow of taxpayers in a year as a proportion of non migrants (i.e. people that filed in the same county in t-1 and t). For each dependent variable the first column shows the results for the regressions without controls, and the second column shows the coefficients controlling for log of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, *, * denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table 2.9: Denial Rates Panel A Low Elasticity High Elasticity Difference Denial Rate (2002) 0.12 0.14 Change in Denial Rate (02-07) 0.02 -0.01 0.03*** (0.06) (0.05) Volume (2002) 9,454 3,811 Panel B Volume per Household (2002) 0.07 0.06 Change in Volume (02-07) -0.01 0.10 .-0.11*** (0.27) (0.22) Number of Counties 394 382 Denial Rates Volume Elasticity -0.03*** -0.01*** -0.01*** 0.07** -0.01 0.02 (0.00) (0.00) (0.00) (0.03) (0.02) (0.02)

Debt to Income (2002) 0.11*** -0.01 -0.57*** -0.13 (0.02) (0.04) (0.11) (0.21)

Changre in Debt to Income (02-07) 0.02* 0.06*** -0.26*** -0.29** (0.01) (0.01) (0.05) (0.10)

Log of the Population 0.02*** 0.02*** -0.05** -0.08** (0.00) (0.00) (0.02) (0.03)

Percent College Educated 0.00*** 0.00*** 0.01** 0.00 (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00*** -0.01** 0.00 (0.00) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -0.15* -0.08 -1.05** -1.10* (0.08) (0.10) (0.44) (0.61)

Percent of Homes Owner-occupied 0.00* 0.00 -0.01*** -0.01*** (0.00) (0.00) (0.00) (0.00)

China Import Share in County (2005) -0.39*** -0.49*** -0.12 0.47 (0.11) (0.11) (0.66) (0.90)

DTI data NY Fed / IRS HMDA NY Fed / IRS HMDA Number of Observations 776 763 774 776 763 774 P2 0130 0.58 0.55 0.09 0.42 026 The table shows the relation between mortgage denial rates and mortgage volume at a county level and the elasticity of housing supply. Total application volume is calculated as the sum of all loans that are originated plus applications that are approved but not accepted, applications denied by the financial institution and loans purchased by the financial institution itself in each county and year, all scaled by the total number of households in a county as of 2000. Denial rates are computed as the proportion of applications denied by the financial institution over total volume in each county and year. All the data is extracted from HMDA LAR records. Panel A shows the average denial rates and average volume in 2002 and 2007, as well as the change in these variables during this period for counties above and below the median elasticity of housing supply in the sample. Panel B shows OLS regressions of the change in denial rate the change in total volume of applications on housing supply elasticity as a continuous variable and controls (debt to income level and changes, the natural logarithm of the population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, the percentage of homes that are owner occupied). All regressions are weighted by the number of households as of 2000. *, **, *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

104 2.6 Appendix. Calculating the magnitude of the collateral effect

We follow the same calculation as Mian and Sufi (2011a) to aggregate the collateral effect across all counties in the data. We start with the differential impact of house prices on employment creation at small firms relative to firms with 50 or more employees, and we focus on the specifications where we exclude non-tradable industries and construction (Table 2.3, Column 2). We first compute predicted county-level employment gains for each establishment size bins in this subset of industries (relative to the 10th percentile county), and then we aggregate to all counties. Below we describe each step in detail. First, we compute the county-level predicted change in employment in each establishment size category by multiplying the regression coefficient by the change in house prices between 2002 and 2007 in each county. We then subtract the predicted change in the 10th percentile county in the change in house prices (to avoid being affected by outliers at the bottom of the distribution). Second, we multiply the predicted county-level change in employment in each establishment size bin by the employment in that size bin in each county as of the beginning of the period (2002) to obtain a predicted change in employment in terms of numbers of workers for each county and establishment size. Third, we sum up the predicted changes across all counties and establishment size bins to obtain an economy-wide predicted change due to the collateral channel in the subset of industries in our preferred specification. Fourth, and last, we divide the number of employees obtained in step 3 by the share of the economy made up by the industries included in the specification (for example, 70.8% of employment is in the industries included in Table 2.3, Column 2). As an illustration of the calculations, we can take the regression coefficient of 0.315 for size bin 1-4 employees from Column 2 in Table 2.3. Given a change in house prices of 0.12 in the 10th percentile county, this yields a predicted employment change in this size bin in the subset of industries in this regression (all except non-tradable and construction) for the county in the 10th percentile growth in house prices of 3.8% more than for the size bin 50 and more employees. If we take another county that has a change in house prices at the median (0.267) the predicted change in that county for this subset of industries is 0.267*0.315=8.4%. Subtracting the predicted employment change in the 10th percentile county yields 4.6% predicted change in employment in the smallest establishment size bin in this county for this subset of industries. We would then multiply this change by the number of employees in this establishment size bin in this county and in this subset of industries. When we obtain a total number of employees by county and bin category, we sum across the four smallest categories and divide by the share of the economy that is made up by the industries included in each specification. We estimate a total job gain in firms with fewer than 50 employees relative to those with 50 or more employees of 1.698 million jobs in all counties, or 27.8% of jobs created between 2002 and 2007. This is composed of 600 thousand employees in 1-4 employee establishments, 488 thousand employees in the 5-9 category, 291 thousand

105 for the 10-19 employee bin, and 319 thousand for the bin with 20-49 employees. If we restrict our attention to the specification where the demand explanation for our results is the least plausible - that is, the manufacturing sector and, in particular, firms in industries and states where the shipment distance is largest (Column 6 of Table 2.4), the same computation would yield an estimate of 676 thousand jobs, or about 11% of jobs created in this period and subset of counties.

106 Table 2.10: Employment Growth, Firm Size, and House Price Appreciation: Individ- ual Industries by Firm Size

1-4 Emp 5-9 Emp 10-19 Emp 20-49 Emp 50+ Emp Growth in House Prices 0.13*** 0.11** 0.05 -0.02 0.03 (0.05) (0.05) (0.05) (0.08) (0.12)

Log of the Population -0.03*** -0.06*** -0.06*** -0.04*** -0.06*** (0.01) (0.01) (0.01) (0.02) (0.02)

Percent College Educated 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

Percent of Potential Worker Population -0.75*** -1.16*** -0.83*** -0.58* -0.99** (0.20) (0.18) (0.21) (0.31) (0.44)

Percent of Homes Owner-occupied 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

4-Digit Industry Fixed Effects Y Y Y Y Y Number of Observations 110,069 80,915 71,947 61,427 50,381 R-Square 0.34 0.37 0.37 0.34 0.27 The table shows two-stage least squares regressions at a county level of employment growth on house price growth split by size of establishment. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data. Growth in House prices is the percentage change between 2002 and 2007, and each interaction is a dummy indicator for the size of the establishment. All regressions include 4 digit industry fixed effect and control for log of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce and the percentage of homes that are owner occupied. We drop the top and bottom one percentile of the change in employment in each county, industry and establishment category. Standard errors are in parenthesis and are clustered by MSA. *, **, * denote statistical significance at the 10%, 5%, and 1% levels, respectively.

107 Table 2.11: Robustness Test: Difference between High and Low Start-up Capital

1-4 Emp 5-9 Emp 10-19 Emp 20-49 Emp 50+ Emp Growth in House Prices 0.23*** 0.11* 0.03 0.03 0.01 (0.06) (0.06) (0.06) (0.09) (0.13)

Growth in HP * High Startup Capital -0.21*** 0.00 0.05 -0.11 0.03 (0.05) (0.06) (0.06) (0.07) (0.09)

Log of the Population -0.03*** -0.06*** -0.06*** -0.04*** -0.06*** (0.01) (0.01) (0.01) (0.02) (0.02)

Percent College Educated 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

Percent of Potential Worker Population -0.75*** -1.16*** -0.82*** -0.59* -0.99** (0.20) (0.18) (0.21) (0.31) (0.44)

Percent of Homes Owner-occupied 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00)

4-Digit Industry Fixed Effects Y Y Y Y Y Number of Observations 110,069 80,915 71,947 61,427 50,381 R2 0.34 0.37 0.37 0.34 0.27 The table shows two-stage least squares regressions at a county level of employment growth on house price growth split by size of establishment and interacted with a High Startup Capital indicator (indicator itself not shown). High Startup Capital is defined as 4 digit industries for which the amount of capital to start the firm is higher than the median for all industries. All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Employment growth is the percentage change in employment between 2002 and 2007 estimated using County Business Patterns (CBP) data. Growth in House prices is the percentage change between 2002 and 2007, and each interaction is a dummy indicator for the size of the establishment. All regressions include 4 digit industry fixed effect and control for log of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. We drop the top and bottom one percentile of the change in employment in each county, industry and establishment category. Standard errors are in parenthesis and are clustered by MSA. *, *, *** denote statistical significance at the 10%, 5%, and 1% levels, respectively.

108 Table 2.12: Effect of One Standard Deviation Change in the Independent Variable

1-4 Emp 5-9 Emp 10-19 Emp 20-49 Eip 50+ Emp Employment in All Sectors Effect of 1 sigma change in HP 5.2 2.7 1.3 1.1 1.1 Growth (02-07) 9.4 8.0 12.5 10.6 13.3 Employment as of 2002 9,101 9,122 12.819 21,466 72,939 Employment in Firms P50 of Start-Up Capital Effect of 1 sigrma change in HP 4.2 2.1 -0.1 -0.2 1.3 Growth (02-07) 6.6 4.3 13.0 9.4 9.3 Employment as of 2002 2,888 3,556 5,468 10,453 33,018 The table show effect of one standard deviation change in house prices on employment for different establishment sizes.

109 Table 2.13: Dollar-weighted Average Distance Shipped in Manufacturing (miles) Panel A: Summary Statistics

Industry x State Industry Average 630.2 651.7 Std. Dev. 368.4 218.3 Percentiles: 1% 25.0 168.9 25% 378.1 559.3 50% 600.8 620.4 75% 817.7 831.7 99% 1,789.2 1,021.3 Number of Observations 950 21

Panel B: Deciles of NAICS and State Dollar-weighted Average Distance Measure

Industry-State Deciles NAICS Description 1 2 3 4 5 6 7 8 9 10 311 Food Manuf. 1 2 7 10 13 2 6 4 4 1 312 Beverage & Tobacco Product Manuf. 15 16 8 3 2 1 2 313 Textile Mills 2 1 4 4 3 6 5 8 3 314 Textile Product Mills 3 2 8 2 3 4 4 4 6 7 315 Apparel Manuf. 1 1 2 1 3 4 5 3 4 316 Leather & Allied Product Manuf. 1 2 3 2 2 3 11 321 Wood Product Manuf. 8 12 13 4 4 3 2 3 1 322 Paper Manuf. 2 3 7 9 6 8 6 3 3 1 323 Printing & Related Support Activities 5 11 5 13 5 2 6 1 1 1 324 Petroleum & Coal Products Manuf. 27 10 4 2 1 1 325 Chemical Manuf. 1 1 2 11 9 8 4 6 7 326 Plastics & Rubber Products Manuf. 1 1 3 7 8 12 8 8 2 327 Nonmetallic Mineral Product Manuf. 16 20 12 3 331 Primary Metal Manuf. 2 4 9 8 7 2 5 5 4 332 Fabricated Metal Product Manuf. 3 2 3 11 10 7 7 2 6 333 Machinery Manuf. 1 1 1 7 7 12 10 9 334 Computer & Electronic Product Manuf. 3 1 1 5 5 5 10 3 15 335 Electrical Eq., App., & Component Manuf. 2 1 2 5 5 6 15 10 336 Transportation Equipment Manuf. 2 4 1 3 6 6 2 10 4 9 337 Furniture & Related Product Manuf. 5 2 8 11 6 3 7 3 2 1 339 Miscellaneous Manuf. 2 1 5 9 13 9 10

The table shows the dollar weighted distance of shipments for 3 digit NAICS manufacturing industries. Data is obtained from the 2007 Commodity Flow Survey. The first column of Panel A shows the weighted average distance for each industry and state, and the second column aggregates the distances shipped at the 3 digit NAICS level. Panel B shows the frequency with which each industry appears in each state x industry decile.

110 Table 2.14: Detail on Average Start-up Amount by 2-digit NAICS Sector

Average Start-Up Above/Below Industry NAICS2 Amount (USD) Median Agriculture, Forestry, Fishing and Hunting 11 146,033 0 Mining, Quarrying, and Oil and Gas Extraction 21 673,609 1 Utilities 22 601,149 1 Construction 23 78,372 0 Manufacturing 31 363,166 1 Wholesale Trade 42 188,085 0 Retail Trade 44 216,302 1 Transportation and Warehousing 48 131,893 0 Information 51 236,126 1 Finance and Insurance 52 203,799 0 1 Real Estate and Rental and Leasing 53 220,691 0 Professional, Scientific, and Technical Services 54 87,879 1 Management of Companies and Enterprises 55 488,681 0 Admin. and Supp. and Waste Mgnt and Remediation Svcs 56 91,278 Educational Services 61 156,893 0 0 Health Care and Social Assistance 62 214,889 1 Arts, Entertainment, and Recreation 71 218,061 1 Accommodation and Food Services 72 273,186 0 Other Services (except Public Administration) 81 161,995 The table shows the average startup amount by 2 digit NAICS sector used in Tables 2 and 3 in the paper. Data is from the Survey of Business Owners (SBO) Public Use Microdata Sample (PUMS) using responses to the question about "Amount of startup or acquisition capital" for each firm with employees in the 2007 survey year.

111 Table 2.15: Distance Shipped and Share of Employees at Large Establishments

Industry-Demeaned Fraction of Employees in Industry-Demeaned Distance Deciles > 50 Employee Establishments (2002), Deciles 1 2 3 4 5 6 7 8 9 10 1 10 7 6 3 2 3 2 5 9 10 2 15 12 6 3 5 10 5 13 12 16 3 11 9 5 10 12 10 6 9 12 11 4 5 7 13 11 10 12 11 13 8 9 5 8 10 10 11 10 13 17 5 8 7 6 5 9 9 9 14 7 17 15 8 6 7 9 15 12 17 6 9 12 4 6 7 8 6 9 12 14 14 7 5 15 7 10 9 8 9 11 10 10 12 11 8 9 5 10 16 5 9 4 9 10 6 6 13 11 This table uses the distance measure at the state and 3 digit NAICS manufacturing industry from the 2007 Census Commodity Flow Survey, and also the share of employment in establishments that have more than 50 employees for each state and 3 digit NAICS manufacturing industry. For each industry, we compute the average distance shipped, as well as the average share of employees in firms that have more than 50 employees. Finally, for each state and industry observation, we compute the deviation from the industry mean for both measures and classify observations into deciles based on these deviations.

112 Table 2.16: House Price Growth and Creation of Establishments

Births of Est. Deaths of Est. Net Creation of Est. Births, Capital < P50 Births, Capital > P50 (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Growth in House Prices 0.46*** 0.46*** 0.31*** 0.28*** 0.16** 0.18*** 0.57*** 0.43*** 0.32*** 0.50*** (0.12) (0.12) (0.07) (0.08) (0.06) (0.06) (0.13) (0.14) (0.11) (0.13)

Log of the Population -0.01 -0.01 0.00 0.01 -0.01* -0.02*** -0.01 -0.01 0.00 -0.01 (0.01) (0.02) (0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.01) (0.02)

Percent College Educated 0.01* 0.00 0.00* 0.00 0.00* 0.00 0.00 0.00 0.01** 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

Percent Employed (2000 Census) 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

Workforce as a Percentage of Population -2.34*** -1.78** -1.06** -0.65 -1.28*** -1.13*** -2.43*** -2.17** -2.17*** -1.35* -I (0.67) (0.79) (0.40) (0.49) (0.29) (0.33) (0.71) (0.88) (0.63) (0.77)

Percent of Homes Owner-occupied 0.00* 0.00* 0.00 0.00 0.00** 0.00** 0.00* 0.01** 0.00 0.00 (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00) (0.00)

China Import Share in County (2005) -0.62 -0.45 -0.46 -0.60 -0.16 0.16 -0.58 -0.24 -0.69 -0.68 (0.57) (0.67) (0.35) (0.40) (0.29) (0.35) (0.64) (0.61) (0.49) (0.85)

2-Digit NAICS Fixed Effects - Y Y Y - Y Y Number of Observations 731 13,482 731 13.482 731 13,482 731 7,167 731 6,315 R2 0.29 0.20 0.21 0.22 0.31 0.16 0.29 0.20 0.27 0.20

The table shows two stage least squares regressions of establishment births and deaths on house price growth instrumented with the elasticity of housing supply. Each observation is at a county level for the regressions without sector fixed effects (odd numbered columns) and at a county and 2 digit NAICS industry level whenever we include fixed effects (even numbered columns). All regressions are weighted by the number of households in a county as of 2000. House Price Growth is instrumented using the Saiz (2010) measure of elasticity of housing supply at an MSA level. Births and deaths of establishments come from the Census Statistics of U.S. Businesses and are summed between 2002 and 2007 and scaled by the number of establishments in a county as of 2002. Growth in House prices is the percentage change between 2002 and 2007, and each interaction is with a dummy indicator for the size of the establishment. Columns 1 and 2 shows the results for births of establishments, Columns 3 and 4 show results for disappearance of establishments and Columns 5 and 6 use the net creation of establishments as the dependent variable. The final four columns split the sample by the amount of capital necessary for starting a business and show results for establishment births. All regressions control for the natural logarithm of population, the percentage of the population with a college degree, the percentage of the labor force that is employed, the share of the population in the workforce, and the percentage of homes that are owner occupied. All controls are at a county level for the year 2000 and are obtained using Census Bureau Data Summary Files. Standard errors are in parenthesis and are clustered by MSA. *, *, * denote statistical significance at the 10%, 5%, and 1% levels, respectively. Table 2.17: List of 3-digit NAICS Industries Excluding Non-tradables, Manufacturing, F.I.R.E., and Construction

NAICS Description 1-4 Emp. 5-9 Emp. 10-19 Emp. 20-49 Emp. 50+ Emp. 113 Forestry and Logging 19.8% 2.6% 7.3% 4.3% 65.9% 114 Fishing, Hunting and Trapping 12.7% 7.4% 6.9% 8.2% 64.9% 115 Support Activities for Agriculture and Forestry 17.9% 8.0% 8.2% 13.7% 52.2% 213 Support Activities for Mining 5.3% 3.2% 5.7% 10.8% 75.1% 221 Utilities 1.5% 1.5% 2.4% 8.1% 86.6% 237 Heavy and Civil Engineering Construction 5.4% 5.2% 8.1% 17.1% 64.2% 423 Merchant Wholesalers, Durable Goods 7.8% 9.3% 14.6% 22.7% 45.6% 424 Merchant Wholesalers, Nondurable Goods 6.8% 6.3% 9.5% 16.1% 61.3% 425 Wholesale Electronic Markets and Agents and Brokers 26.3% 12.1% 11.9% 12.6% 37.1% 454 Nonstore Retailers 12.3% 9.4% 12.6% 16.6% 49.2% 481 Air Transp. 0.9% 0.6% 1.5% 4.3% 92.6% 483 Water Transp. 3.0% 2.1% 3.8%, 8.4% 82.7% 484 Truck Transp. 9.7% 5.5% 9.3% 18.8% 56.7% 485 Transit and Ground Passenger Transp. 4.3% 2.8% 4.6% 13.9% 74.3% 486 Pipeline Transp. 3.7% 3.9% 11.9% 16.0% 64.5% 487 Scenic and Sightseeing Transp. 19.3% 6.8% 8.1% 16.3% 49.6% 488 Support Activities for Transp. 7.6% 7.4% 10.2% 17.0% 57.8% 492 Couriers and Messengers 2.6% 1.6% 2.8% 7.0% 86.0% 493 Warehousing and Storage 2.8% 3.4% 6.9% 17.8% 69.3% 511 Publishing Ind. (except Internet) 3.2% 3.0% 5.2% 10.2% 78.4% 512 Motion Picture and Sound Recording Ind. 12.1% 4.5% 6.9% 18.7% 57.7% 515 Broadcasting (except Internet) 2.4% 2.3% 4.8% 13.4% 77.1% 516 Internet Publishing and Broadcasting 7.8% 5.0% 6.4% 12.7% 68.1% 517 Telecomnmunications 4.1% 3.6% 6.4% 11.0% 74.9% 518 ISPs, Web Search, and Data Processing 5.9% 3.2% 5.3% 10.9% 74.7% 519 Other Information Serv. 7.6% 8.0% 13.3% 21.6% 49.5% 541 Professional, Scientific, and Technical Serv. 16.8% 9.0% 11.1% 14.9% 48.2% 551 Management of Companies and Enterprises 1.3% 1.5% 2.8% 7.2% 87.2% 561 Administrative and Support Serv. 5.9% 4.1% 5.9% 11.2% 73.0% 562 Waste Management and Remediation Serv. 5.8% 6.1% 10.6% 22.2% 55.3% 611 Educational Serv. 3.2% 2.9% 5.2% 12.0% 76.7% 621 Ambulatory Health Care Serv. 10.6% 14.5% 16.5% 19.2% 39.3% 622 Hospitals 0.0% 0.0% 0.0% 0.0% 100.0% 623 Nursing and Residential Care Facilities 1.2% 3.0% 5.5% 9.4% 80.9% 624 Social Assistance 5.3% 6.8% 15.9% 28.8% 43.2% 711 Performing Arts, Spectator Sports, and Related Ind. 18.2% 5.3% 5.9% 9.3% 61.3% 712 Museums, Historical Sites, and Similar Institutions 4.7% 4.4% 6.5% 12.4% 71.9% 713 Amusement, Gambling, and Recreation Ind. 4.8% 4.7% 8.3% 20.5% 61.6% 721 Accommodation 2.3% 2.1%, 7.5% 16.5% 71.6% 811 Repair and Maintenance 23.1% 22.1% 20.5% 18.9% 15.4% 812 Personal and Laundry Serv. 19.7% 19.5% 21.4% 18.8% 20.6% 813 Religious, Grantmaking, Civic Org. 11.8% 13.3% 15.2% 22.4% 37.3%

The table shows the 3 digit NAICS codes, as well as the proportion of employees in each establishment size category and the total number of employees in each industry in our sample of counties. Chapter 3

Credit Supply and House Prices: Evidence from Mortgage Market Segmentation

3.1 Introduction

One of the central debates in finance focuses on the impact of the cost of funding on the level of asset prices (see, e.g., Brunnermeier, Eisenbach and Sannikov, 2012). A salient recent example is the US housing market: many observers of the 2008 financial crisis have proposed that reduced cost of credit was the central factor fueling the increase in housing prices as well as the subsequent reversal (Hubbard and Mayer, 2008; Mayer, 2011). Others have argued that cheaper credit alone cannot explain the bubble (Glaeser, Gottlieb, and Gyourko, 2010) and that other factors must have also been at play, including a reduction in collateral constraints (Favilukis, Ludvigson, and Van Nieuwerburgh, 2010; Khandani, Lo, and Merton, 2009), financial innovation (Mian and Sufi, 2009; Calomiris, 2009; Pavlov and Wachter, 2011), or market sentiment and expectations about future appreciation (Shiller, 2008). The key difficulty in measuring the effect of the cost of credit on the price of housing is establishing the direction of causality between cost of funding and house price growth: On the one hand, cheaper credit is likely to reduce borrower financing constraints and increase total demand for housing, which in turn would lead to higher prices. On the other hand, however, credit conditions in general might be responding to expectations of stronger housing demand and, as a consequence, higher house prices. In this latter scenario, cheaper credit is not the driver of house price increases, but a byproduct of increased demand for housing, since housing as collateral becomes more valuable. As we see in the existing literature, it has been very difficult to separate these two effects. 1 In this paper, we develop a new instrument that uses annual changes in the con-

'A recent paper by Favara and Imbs (2012) uses branching deregulation in the 1990s to identify the causal link between credit supply and house prices and finds that states where there is deregulation subsequently experience larger house price increases.

115 forming loan limit (CLL) as exogenous variation in the cost of credit, which allows us to provide clean estimates of the effect of lower cost of credit on house prices. The CLL determines the maximum size of a mortgage that can be purchased or securitized by Fannie Mae or Freddie Mac. Mortgages below the CLL therefore have lower interest rates compared to jumbo loans (loans that are above the CLL), since the former benefit from implicit (and since 2008, explicit) government support for Fannie Mae and Freddie Mac. The difference in interest rates between conforming loans and jumbo loans has been estimated to be up to 24 basis points.2 . In addition, Loutskina and Strahan (2009, 2011) show that more borrowers are able to access mortgages below the conforming loan limit than above, which suggests that not only the cost of credit it lower below the CLL, but also access to credit itself might be easier. The underlying idea of our identification strategy is that changes in the conforming loan limit (CLL) from one year to the next are exogenous to local housing markets and the local economy, since this change is based on the national average appreciation in house prices. That means that, in a given year, a house just above the CLL threshold has to be financed by an expensive jumbo loan, while the next year the equivalent house can be financed via cheaper conforming loan. Our empirical approach involves comparing transactions that can be financed more easily using a conforming loan, and houses that are more expensive so that buyers need to obtain larger (jumbo) loans to maintain the same loan-to-value ratio. We track transactions in the price range just above and below the CLL in the year that the limit is in effect and compare them to the subsequent year, once the limit is raised and houses just above the CLL become eligible for conforming loans. This setup enables us to cleanly identify the effect of the cost of credit and control for any overall trends in house prices. The threshold that we use to define houses that are "cheap" to finance with a conforming loan in a given year is obtained by dividing the conforming loan limit by 0.8.' By construction, buyers of houses with a price below this threshold can get a conforming loan that makes up 80 percent of the price of the house, whereas if the price of the house is above 125 percent of the CLL, it can no longer be financed at 80 percent with a conforming loan. Loans with a loan-to-value (LTV) ratio below 80 are associated with more attractive terms, and conforming loans above 80 percent require private mortgage insurance in order to qualify for purchase by Fannie Mae or Freddie Mac (Green and Wachter, 2005). Above this price threshold, borrowers either finance their home with an 80 percent first mortgage using a jumbo loan (i.e. a loan above the CLL) at a higher interest rate, or, if they want to take advantage of the lower interest rate below the CLL, they have to use savings or alternative forms of financing to make a larger down payment. Importantly, our sample includes all transactions in this price range independent of financing choice of each borrower. This allows us to eliminate any bias due to the endogenous choice of financing of a

2 See for example McKenzie (2002), Ambrose, LaCour-Little, and Sanders (2004), Sherlund (2008), Kaufman (2012), or DeFusco and Paciorek (2013) 3 Kaufman (2012) uses this threshold for appraisal values to study the effect of the conforming status of a loan on its cost and contract structure. Loutskina and Strahan (2013) follow our approach and use changes in the CLL interacted with regional constraints to look at financial integration and the propagation of shocks.

116 specific transaction. An example of such a bias would be that richer people who can afford to put more money down might also purchase houses that are more expensive based on (unobservable) quality dimensions. Our instrument eliminates this type of concern. We first document that the conforming loan limit (CLL) impacts borrowers' choice of financing. The data shows that the norm in the mortgage market during this period was to borrow at an LTV of exactly 0.8 (on average 60 percent of transactions are at an LTV of 0.8). However, for houses that transact just above 125 percent of the CLL, a much larger fraction of purchases are at an LTV below 0.8, since many borrowers choose to take out a mortgage to exactly max out the conforming loan limit. Borrowers that buy houses with a price above the threshold have a higher funding cost than borrowers who buy houses at a price below 125 percent of the CLL, since they either have to take a jumbo loan or use a conforming loan and finance the rest of the house price with other forms of financing. In our main analysis, we measure the causal effect of cheaper credit on house prices instrumented via the change in the conforming loan limit from one year to the next. We run differences-in-differences regressions in which we compare transactions just above and just below the threshold of 125 percent of the CLL in the year that the limit is in effect, and in the subsequent year when all of the transactions can obtain an 80 percent conforming loan.4 We use three different dependent variables to capture the value of a property: (1) the value per square foot; (2) the residuals of log of house prices from a hedonic regression using a large set of controls for the underlying characteristics of the house, and (3) the residuals of the value per square foot from similar hedonic regressions. 5 We find that transactions just above 125 percent of the CLL, i.e. in the "high cost" group of borrowers, are made at lower values per square foot than those for the unconstrained group. We see a 1.16 dollar discount per square foot for a mean value per square foot of 220 dollars (i.e., about 53 basis points of the average house value). This difference is reduced to 0.65 dollars per square foot (30 basis points) after we control for house characteristics, suggesting that part of the effect we find can be accounted for by differences in the observable quality of houses above and below the threshold. These effects are significantly different from those we obtain when we use "placebo" loan limits elsewhere in the distribution, which confirms that we are picking up a cost of credit effect of the CLL. The effect is smaller (and often insignificant) in the second half of our sample (2002-2005), which is the period when

4This is the case for all years between 1998 and 2005. For example, the CLL in 1999 is USD 240, 000, which gives a threshold of USD 240, 000/0.8 = 300, 000 for this year. This means that in the regression for 1999, we include houses priced at between 290, 000 and 310, 000 in the years of 1999 (the year the CLL is in effect) and 2000. The CLL in the year 2000 was raised to 252, 700, so the new threshold for that year is 315,875. Clearly, all the houses we included in the analysis for 1999 can be financed at 80 percent with a conforming loan in the year 2000. 'We run the hedonic regressions by year and by metropolitan statistical area (of which we have 10) and we use the set of controls available from deeds registries data, which includes common variables such as number of rooms and number of bedrooms, but also detail on the type of heating, architectural type, building type, among many others (we discuss these controls in more detail in Section 3.3.2).

117 jumbo loans became cheaper and easier to obtain (partly due to the increased ease with which they could be securitized) and also when second lien mortgages became widely available (see Figure 3-5 ). Both these effects reinforce the idea that when the CLL was more important in the earlier part of the sample, its impact is also more significant in our estimates.

Given our estimate for the change in house prices due to changes in credit conditions, we can compute the semi-elasticity of prices to differences in interest rates in the region close to the threshold. We use the differences in interest rates estimated in the prior literature of 10 to 24 basis points between conforming and jumbo loans as our measure of the cost differential for buyers above and below the threshold. We obtain local elasticity estimates that range from a low end of 1.2 to an upper range of 9.1 depending on the period and the exact estimate for the interest rate differential between jumbo and conforming loans that we use for our calculations. These elasticity estimates are at the lower end of what has been previously found in the literature, and they imply that the 55

We next investigate the cross-sectional heterogeneity of our elasticity estimates by focusing on whether the effect of cheaper credit is stronger when buyers face other types of constraints at the same time, as proxied for by lower income. Specifically, we interact the changes in the CLL with whether a zip code and year is below the 10th percentile of the income growth distribution for each individual regression. The point estimate for these areas shows that value per square foot is 2.50 dollars higher in the year that a house becomes eligible to be financed with a conforming loan. This is more than double the size of the average elasticity that we found in the overall sample, suggesting that cheaper credit may have had a disproportional impact on economically more depressed households and regions.

We show that our results are not driven by a subsidy effect that provides a focal point to draw in more bidders. First, there is no visible bunching in the number of transactions just around the threshold of CLL divided by 0.8, suggesting that the supply of housing does not react strongly to the CLL. We also do not find that there is bunching in the number of unobserved bidders for homes around the CLL, which we measure as the share of borrowers that apply for loans but ultimately either withdraw or do not use the loans they are approved for. If the CLL served simply as a focal point for home sales, we should expect more bidders for homes that are eligible for conforming loans. Instead, we find that our measure of the share of failed bids is lower, not higher, for borrowers that borrow up to the CLL. The fact that there is neither a significant jump in the quantity of transactions nor in our proxy of failed offers for homes suggests that the effect we find on prices is more consistent with a cost of credit interpretation.

The rest of the paper is structured as follows: Section 3.2 discusses related literature and the user cost model. Section 3.3 describes our data and the identification strategy. In Section 3.4, we lay out the regressions results and robustness checks of our main analysis. Section 3.5 discusses the findings and concludes.

118 3.2 The User Cost Model

In this paper, we are interested in estimating the impact of changes in the cost of credit on the price of housing. The existing literature has focused on different versions of the user-cost model of Poterba (1984) to draw conclusions about the role of interest rates and other costs of owning for house prices. In this model, agents are indifferent between owning and renting if the housing market is in equilibrium, where the mortgage interest rate is the main determinant of the cost of owning. The existing literature shows that different assumptions yield very different conclusions about the role of interest rates in driving the cost of housing and highlight why our estimate of the impact of the cost of credit on prices is an important contribution to this debate. We follow the notation in Glaeser, Gottlieb, and Gyourko (2010) to describe the basic elements of the user cost model. Renting a property involves paying rent equal to Rt in each period. Owning a property, on the other hand, includes making a down- payment 0 that is a proportion of the price of the house Pt and obtaining a mortgage that is rolled over each period, such that principal is never paid down completely. The borrower pays interest on the mortgage at a rate rt that is deflated by the relevant tax rate #, as well as property taxes and maintenance costs equal to T that both grow at a rate g. The model assumes that individuals have a private discount factor of pt. If we assume that market interest rates and private discount rates are constant and equal to each other, we can write the indifference condition for users as: = (1 -#)r - g+ (3.1) Pt

This is shown in Glaeser et al (2010) and is similar to what is presented in Hubbard and Mayer (2008) as well as a simplified version of the user cost in Himmelberg, Mayer, and Sinai (2005). If the assumptions of this model hold, changes in the user cost (the right-hand side of the equation) should lead to changes in the price to rent ratio. For example, if the user cost is 5 percent, then the price of a house should be about 20 times its market rent. In such a world, a drop of 1 percentage point in mortgage rates would lead to a decline of (1 - r) in the user cost, or 0.75 if we assume a marginal tax rate of 25 percent. The price to rent ratio would then be 23.5, an increase in the price of 17.5 percent. This is the magnitude of the elasticities proposed in Himmelberg et al (2005), and in Hubbard and Mayer (2008). Glaeser, Gottlieb, and Gyourko (2010) dispute some of the simplifying assumptions in the model above, and show that a more realistic model can produce much lower elasticities of prices to interest rates. In particular, if private discount rates are not the same as market rates, changes in interest rates wont alter the way users discount future expected house price appreciation. Glaeser et al (2010) show that this change alone can reduce the elasticity to just 8, instead of the initial 17.5. Other mechanisms through which the elasticity could be substantially reduced include mean reverting interest rates, which means borrowers anticipate having to sell a home at a time when rates are higher, or the possibility of prepaying a mortgage. Our econo-

119 metric approach allows us to more carefully identify the magnitude of the change in house prices due to changes in the average cost of financing, since we look at exogenous movements in the cost of capital for home buyers. Our empirical results provide local estimates for the numerator of the elasticity calculation. In Section 3.4.4, we discuss the range of elasticities that are consistent with our results.

3.3 Data and Methodology

The dataset we use in this paper contains all the ownership transfers of residential properties available in deeds and assessors records over 11 years, from 1998 to 2008, and seventy-four counties in ten metropolitan statistical areas (MSAs) - Boston, Chicago, DC, Denver, Las Vegas, Los Angeles, Miami, New York, San Diego, and San Francisco. We limit our attention to transactions of single-family houses, which account for the large majority (approximately 78 percent) of all observations. Each observation in the data contains the date of the transaction, the amount for which a house was sold, the size of the first mortgage, and an extensive set of variables about the property itself. These characteristics include the property address, interior square footage, lot size, number of bedrooms, number of bathrooms, total rooms, house age, type of house (single-family house or condo), renovation status, and date of renovation. Additional characteristics include the availability of a fireplace, parking, the architectural and structural style of the building, the type of construction, exterior material, availability of heating or cooling, heating and cooling mechanism, type of roof, view, attic, basement, and garage. We describe the procedure for cleaning the raw data received from Dataquick in the Appendix to the paper.

3.3.1 Summary Statistics

The dataset that we use for this paper contains 3.98 million transactions of single- family houses that are summarized in Tables 3.1 and 3.2.6 We can see in Panel A of Table 3.1 that the average transaction value is 309 thousand dollars with a standard deviation of 124 thousand dollars. The average size of the houses is 1,735 sqft, and the houses have, on average, 3 bedrooms and 2 bathrooms. The average loan to value is 0.81 (including only the first mortgage for each transaction), and the median LTV is 80 percent. The average value per square foot is 194 dollars with a standard deviation of 92 dollars per square foot (first row of Panel B). Table 3.1 also shows the summary statistics for the sample we use in the regressions in the final three columns. For the regression sample, the average price for each house is higher than in the whole dataset (at 371 thousand dollars, compared to 309 thousand in the first column). This is consistent with the fact that the conforming loan limit was set to cover substantially more than 50 percent of the mortgages made every year (Acharya, Richardson, Nieuwerburgh, White, 2011). These houses are also, on average, larger and have more bedrooms and bathrooms than the whole dataset.

6 Please see the Appendix for a detailed description of the procedure for cleaning the data initially obtained from Dataquick and how we arrive at the 3.98 million observations.

120 Panel A of Table 3.2 shows marked differences in the summary statistics for each of the ten MSAs included in our data. The table shows that San Francisco is the metropolitan area with the highest valuation, with an average house price of 384 thousand dollars. Denver and Las Vegas represent the areas with the lowest valuation, with an average of approximately 250 and 262 thousand dollars respectively. When we compare values per square foot, we get a similar picture, namely San Francisco is the area with the highest valuation with an average of 266 dollars per square foot, and Las Vegas is the area with the lowest valuation with an average of 137 dollars per square foot. Table 3.2 Panel B shows the evolution of prices through time. Here we see the increase in house prices from an average of 240 thousand dollars in 1998 to a peak of 366 thousand dollars in 2006, as well as the increase in the volume of transactions over the same period. The increase in prices and volume is linked to an increase in volatility. The standard deviation of the transactions increased from 102 thousand dollars in 1998 to 122 thousand dollars in 2006. A similar pattern can be observed for the value per square foot measure, where standard deviation is 51 dollars per square foot in 1998, and increases to 106 dollars per square foot in 2006. Finally, the loan to value average (including only the first mortgage) is stable both across MSAs and through time at around 0.8.

3.3.2 Hedonic Regression

One of the advantages of using deeds registry data is the richness of the information provided on the property characteristics, which allows us to account for price differences between houses that can be attributed to observable features. Specifically, we will be able to assess whether the price impact we observe due to the changes in the conforming loan limit can be attributed to differences in the quality of the houses, or whether these differences are there even after accounting for quality. In order to distinguish between these two explanations, we estimate hedonic regressions of value per square foot and log of house price on a number of house characteristics, and estimate the residuals for each of these two left-hand side variables (which we denote by LHSi). Specifically, we estimate the following regressions by MSA and by year:

LHS, = -yo + PX + monthi + zipcodei + Ej We use both the logarithm of the price of a transaction as well as the value per square foot as our dependent variables. By estimating these regressions by year and by MSA, we allow the coefficients on the characteristics to vary along these two dimensions. We also use monthly indicator variables to account for seasonality in the housing market, as well as zip code fixed effects. The set of controls Xi is a similar set of controls to that used in Campbell, Giglio, and Pathak (2010) with some additional characteristics. The controls include square footage, high and low square footage dummies, the size of the lot, number of bedrooms and bathrooms, and a number of indicators for interior and exterior house characteristics (eg. fireplace, style of the

121 building, etc.). We describe which variables are included, as well as the detail of the construction of each variable, in the Appendix to the paper. The estimated R2 of each of these regressions (80 in total for each of the two left-hand side variable-10 MSAs in 8 years) is between 40 and 60 percent for the price of the transaction, and 50 to 70 percent when we use value per square foot as a dependent variable. Summary statistics for the residuals from the hedonic regressions for the whole sample are shown in Panel B of Table 3.1. The average residuals are, by construction, zero. The standard deviation of the errors is about 42 dollars per square foot, and 0.17 thousand dollars for the log of the price of the house. The hedonic regressions are estimated on the whole dataset of transactions (the 3.98 million observations mentioned above), so when we restrict our attention to the regression sample, the average error no longer has to be zero. Indeed, for the regression sample, the average residual from the hedonic regressions for the value per square foot is positive at 5.3 dollars, and the average error for the log of transaction value of the house is 0.05 dollars (last three columns of Panel B of Table 3.1). The standard deviation of the residuals for the regression sample is similar in magnitude to what we obtain for all the transactions.

3.3.3 Empirical Approach Identification Strategy

To identify the effect of changes in credit conditions on house prices, we restrict our analysis to two groups of buyers who all buy houses in a tight price range, but differ in the financing available to them. The sample for our regressions is made up of houses that transact in a band around 125 percent of each year's conforming loan limit, as well as houses in the subsequent year in the same price range. Specifically, we divide houses into two groups: houses below the threshold of 125 percent of the year's CLL (i.e. transactions that fall between 125 percent of CLL and 125 percent of CLL minus USD 10,000) and houses above that threshold that transact between 125 percent of CLL and 125 percent of CLL+10, 000. By construction, in the year that the conforming loan limit is in effect, houses above the threshold of 125 percent of the CLL cannot be financed at 80 percent using a conforming loan, whereas the houses below the threshold can be financed. Thus, home buyers that bid for houses priced above 125 percent of CLL cannot finance a full 80 percent of the transaction with the cheaper and more easily available conforming loans. In the subsequent year, the CLL is raised and both groups of transactions can be financed at 80 percent 7 with a conforming loan. Our sample includes all transactions in this price range, independent of the mortgage choice made by each buyer. This way, our estimates are not biased by the endogeneity of the choice of financing of each specific transaction. The identification strategy is best understood through an example. Consider the

7 While this was no longer true for the years after 2006, in all cases between 1998 and 2005, the limit increases enough from year to year to make up 80 percent of the price of the transactions we have in the sample.

122 year 1999: In that year, the conforming loan limit (CLL) for single-family houses was USD 240, 000. The corresponding threshold for house prices that we use for this year is 300, 000 (240, 000/0.8 or, equivalently, 1.25 * 240, 000). In this year, the group of houses "above the threshold" have prices between USD 300, 000 and USD (300, 000 + 10, 000) = 310, 000 and houses "below the threshold" have a transaction price between USD (300, 000 - 10, 000) = 290, 000 and USD 300, 000 (those that transact at exactly USD 300, 000 are included in this second group). For the purposes of our main regressions, we track these two groups of houses from 1999 to 2000, where 1999 is the year in which the CLL is in effect and 2000 is the year in which all these transactions could be bought using a conforming loan at a full 80 percent LTV. In fact, the CLL changed in 2000 to USD 252, 700, so the threshold of 125 percent of CLL was now USD 315,875 and even our "above the threshold" group for 1999 is now eligible to get an 80 percent LTV conforming loan.

One important assumption in our analysis is that borrowers in the group "above the threshold" of 125 percent are constrained in their choice of financing. In order to stay at an LTV of 0.8, they have to take a jumbo loan and these have been found to be more expensive by between 10 and 24 basis points relative to conforming loans (McKenzie, 2002; Ambrose, LaCour-Little, and Sanders, 2004; Sherlund, 2008; Kaufman, 2012; DeFusco and Paciorek, 2013). Alternatively, they can also borrow up to the CLL and then cover the rest of the house price with savings or other funding, which means having a first mortgage LTV of less than 80 percent. This additional source of funding is likely substantially more expensive relative to the conforming mortgage rate. For some borrowers, this may, in fact, be the only option, as they may be excluded from the jumbo market altogether because of more careful screening of jumbo loans done by originating banks (Loutskina and Strahan, 2009, 2011). Whether they choose a jumbo loan or they make up the difference using other sources of financing, these borrowers have a higher average cost of capital than the buyers below the threshold.

As Figure 3-1 shows, the most frequent choice on the part of borrowers is to have an LTV of exactly 80 percent (that is, the large mass along the diagonal of the figure). The main exception to this rule occurs exactly at the conforming loan limit, where a significant mass of borrowers chooses an LTV below 0.8 by sticking to a conforming loan (in 2000 the limit was USD 252,700, and in 2004 it was 333,7000). The data shows that in the year in which the CLL is in effect, about 45 percent of the houses below the threshold in our sample are bought with an LTV of exactly 80 percent, whereas for houses above this boundary just 19 percent of borrowers pick 80 percent LTVs (which for these transactions means using a jumbo loan). Additionally, on average 55 percent of the transactions just above the threshold are financed using a conforming loan, which means having an LTV lower than 80 percent. These borrowers end up with an LTV of 77-79.5 percent, which is a very infrequent choice anywhere else in the distribution. Again, these borrowers might have a lower LTV because they choose to stay below the CLL due to the cost of the loan, or because they are excluded from the jumbo market altogether. Whatever the reason, this group of borrowers is "constrained" in the set of options available for financing their house.

123 Empirical Specification

Our main regressions estimate the size of the effect of the constraint imposed by the conforming loan limit on the valuation of transactions made just above the threshold of 125 percent of the CLL. We run differences-in-differences regressions year-by-year with one indicator variable for houses priced above the conforming loan limit divided by 0.8, another indicator for the year in which the CLL is in effect, and an interaction of these two indicator variables. We also include ZIP code fixed effects in all regressions, so our estimates do not reflect differences between neighborhoods, but rather variation within zip codes. The sample for each year-by-year regression includes houses within a USD 10,000 band around the conforming loan limit in the year in which the limit is in force, as well as the subsequent year. This implies that the "Above the Threshold" indicator variable takes a value of 1 if the price at which a house transacts is greater than 125 percent of the conforming loan limit of a certain year, and less than that amount plus 10,000 dollars. This same variable is a 0 for transactions between 125 percent of the CLL and 125 percent of the CLL minus 10,000 dollars. The "Year CLL" indicator variable is a 1 in the year in which the CLL is in effect for each regression, and a 0 in the subsequent year. We use a tight band around the threshold so that all transactions in the year after the limit is in effect are eligible for an 80 percent LTV conforming loan. We thus have a group of transactions that is "easy to finance and another one that is "hard to finance in the year that the limit is in effect, but all transactions in the sample are "easy to finance once the limit is raised.' We run regressions of the following form:

Valuation measurei = N + /11AboveThreshold + /3 2lYearCLL±

/31Above ThresholdxYear-CLL + YZIP + Ei We estimate this regression for each year between 1998 and 2005. We cannot include 2006 and 2007 in our estimates because the conforming loan limit did not change after 2006 in our data (house prices dropped and the administration left the 9 limit unchanged). After we obtain 01, /2, and 03 for all 8 years (1998-2005), we estimate Fama-MacBeth averages (Fama and MacBeth, 1973) of these coefficients and obtain the standard errors of this average by using the standard deviation of the estimated coefficients and dividing it by the square root of the number of coeffi-

8 An alternative way to run our test would be to compare the year in which each limit is in effect with the previous year, when all transactions in this range would be above the threshold for that year. The results for this alternative specification are reported in the Appendix. 9 We do not run our analysis on the changes that were made to conforming loan limits in 2008 in high-cost areas as part of the Housing and Economic Recovery Act of 2008 for two reasons: First, the limit was chosen by the government, as opposed to being mechanically related to previous limits, so this introduces the possibility that the "jumbo-conforming" program was designed to assist specific areas and thus would be endogenous to expected future appreciation. Second, to the best of our knowledge, there is no empirical evidence that the program had any discernible impact on the cost of funding of mortgages that were made between the old limit of USD 417,000 and the new, higher limits.

124 cients. We test the robustness of our results to serial correlation in the error term by constructing Newey-West standard errors, and all the results are unchanged. We should point out that our approach is not a regression discontinuity design, but rather differences-in-differences for each pair of years. There are a couple of reasons for this: First, the threshold that we use does not imply a sharp discontinuity in the ease of financing a home. For a house just one dollar above the threshold, a homebuyer only has to come up with one additional dollar of equity (and still obtain a conforming mortgage), which means the total cost of financing the house is almost unchanged. As we move progressively away from the threshold, transactions become harder to finance. For our differences-in-differences estimator to be valid, all we need is that houses above the threshold are somewhat harder to finance, though not necessarily discontinuously so. The second reason for not using a regression discontinuity design is that in the year that the limit is in effect, homebuyers choose to buy houses above or below the threshold, i.e. the position with respect to the limit is not exogenous. On the contrary, our differences-in-differences specification uses the exogenous change in the conforming loan limit to compare a group of transactions that are above the limit in a year, but below in the next with a group of transaction that are always below the limit, achieving a clean identification of the effect of credit availability on house prices. Our estimation strategy allows us to estimate the causal effect of changes in the cost of credit on the valuation of houses. Since house price levels differ across the various states of the United States, the change in the CLL affects different parts of the housing stock across areas depending on the price level of the area. Using this instrument we can account for the possibility that there are differential growth rates within the distribution of house types across the country. For example, one concern would be that middle class families might buy a certain type of house and, at the same time, have a different income growth from other parts of the population. Our instrument allows us to rule this out, because the same "type" of house will have different prices depending on where it is located in the country. Finally, we can rule out that selection effects are driving our results: one could worry that buyers of houses "above the threshold" in the year that the conforming loan limit is in effect are different along some unobservable dimensions from the other buyers. Several features of our analysis make selection an unlikely explanation of the results. First, for a selection hypothesis to be a true alternative to our explanation, it would have to involve arguments other than cost of credit to explain why buyers were different above and below the threshold. Second, these "special" buyers would both have to be better able to deal with the higher cost of credit (potentially because they are wealthier or have higher income) and bargain harder for houses. It is unclear why wealthier borrowers should pay less for a similar house than poorer borrowers. If wealthier people bought higher quality houses and we did not observe these differences, these unobservable characteristics would create bias in the opposite direction. Third, our identification strategy would require that the selection effect change each year parallel to the change in the size of the conforming loan limit, which is very unlikely. Lastly, to further alleviate any concerns about selection, we run our main

125 regressions excluding borrowers that choose LTVs below 80 percent in the year that the CLL is in effect. If selection was the explanation of the results, these transactions should be by "wealthy" borrowers driving the results. We find that the results do not change materially when we exclude this subset of transactions.

Differences in Financing Choices

As we pointed out above, the equivalent to a first stage in our empirical strategy is to show that the changes in the conforming loan limit have a significant effect on the financing choices of borrowers. In Figure 3-1 we can see the importance of both the 80 percent LTV rule, as well as the conforming loan limit, in determining financing choices for the whole distribution of transactions. In Figure 3-2 we focus on the groups of transactions that we include in the regressions. The first panel tracks transactions up to USD 10,000 below 125 percent of the conforming loan limit in each year, whereas the second panel includes transactions up to USD 10,000 above the threshold. We show the total number of transactions (for all years between 1998 and 2006) in each month during the year prior to the limit being in effect, in the year that the limit is valid, and in the subsequent year. We also break down the transactions by the choice of LTV - the transactions at the bottom of each panel have an LTV below 75 percent, the second group includes transactions with an LTV between 75 percent and 79.5 percent, the third has transactions with LTV=80 percent, and the top group has all the transactions with an LTV above 80.1 percent. The main message from Figure 3-2 is that in the year that the CLL is in effect, the composition of financing choices by borrowers differs very significantly, with the 80 percent group becoming very prominent for the transactions below 125 percent of the CLL, whereas it is small for the transactions above the threshold. At the same time, the borrowers who stick with a conforming loan and buy houses above 125 percent of the CLL become an important fraction of all borrowers (they have an LTV between 75 and 79.5 percent).10 In the year after the limit is in effect, the choice of LTV across the two groups becomes indistinguishable. In Table 3.3, we present the effect of the changes in the conforming loan limit on the financing choices made by the borrowers included in the sample of our main regressions. In this table, we are verifying what we see in the pictures, namely that borrowers on average end up with lower LTVs when they buy houses above the threshold of 125 percent of CLL. We find that LTVs are, on average, 0.3 to 0.7 percentage points lower for the group of transactions that happen above the threshold of 125 percent of the CLL in the year that the limit is in effect. This effect is statistically and economically significant given how little variation there is in the modal choice of LTV of borrowers. The second panel on Table 3.3 shows that borrowers also obtain,

'OThe first picture for the group below 125 percent of the CLL also shows a noticeable fraction of borrowers with an LTV between 75 and 79.5 percent in the year before the CLL is in effect. This is because these transactions were not eligible for a conforming loan at an 80 percent LTV in the year before the new limit was in effect and were, in general, just slightly above that threshold. This is thus a reflection of the same phenomenon we see for the group above 125 percent of the CLL in the year that the new limit is in place.

126 on average, smaller loans in the year that the limit is in effect and when the price of the house is above the threshold. The difference in log loan amount is, on average, 0.0056 to 0.0088 dollars, and based upon the findings in our main results, we conjecture that it is the fact that borrowers obtain smaller first mortgages that leads to the difference of approximately 1.16 dollars per square foot (for an average value per square foot of 220 dollars).

Differences in the Number of Transactions

There are several reasons to expect quantities to change due to differential cost of credit, including different levels of down-payment (Stein, 1995) or sellers waiting for buyers to obtain better credit conditions (Genesove and Mayer, 1997). In fact, unless the supply elasticity of houses is very low (or zero), we expect the price effect due to a change in the demand for housing to be accompanied by a change in the number of transactions. As discussed in Section 3.3.3, we do not use a regression discontinuity approach to address the question of the change in the quantity of transactions. Figure 3-3 confirms that this would produce no significant result. This figure shows the number of transactions relative to the threshold in each year. The figure is centered at 0, i.e. the transactions at exactly 125 percent of the CLL. The figure shows that there is no discontinuity in the number of transactions above and below the threshold. Given that a regression discontinuity would not be appropriate in our setting, we use a setup similar to our main regressions to look for changes in the number of transactions above and below the threshold. We consider the difference in the share of transactions in our sample that fall above and below the threshold in the year that the limit is in effect and in the subsequent year in a differences-in-differences setup. This test is equivalent to a T-test for the mean of the variable "Above Threshold" that compares the average of this variable in the year that the limit is in effect and in the subsequent year. If our instrument affects the quantity of transactions, we should see an increase in the share of observations above the threshold when the limit is raised, as credit becomes cheaper for those transactions. We show in Table 3.4 that this test reveals no changes in the share of transactions above and below the threshold for the first part of our sample (1998-2001), and that there is a statistically significant effect for the second part of the sample. This translates into a share of transactions above the threshold approximately 60 basis points lower in the year that the conforming loan limit is in effect during the period 2002-2005. This regression shows that cheaper credit provided by conforming loans is reflected only on house prices in the first part of our sample, and that in the second part of the sample, it impacts both quantities and prices, i.e. local supply elasticity of houses seems to have been higher in the second part of the sample. This, along with the reasons we give in Section 3.4.1 on the availability of second liens and jumbo loans, may help explain why the effect we find on prices is smaller relative to the earlier years (when the quantity response is not there).

127 3.4 Cost of Credit and House Prices 3.4.1 Main Regression Results

We present the results for our canonical specification in Table 3.5. This table presents Fama-MacBeth coefficients from year-by-year regressions, as described before in Sec- tion 3.3.3. The coefficient of interest in Panel A of Table 3.5 is that on the interaction variable, and it shows that houses above the threshold of CLL/0.8 transacted at a value per square foot that was lower by about 1.16 dollars in the year that the CLL was in effect. The results are stronger for the first half of the sample, where the point estimate is -1.55 dollars per square foot for this set of transactions. The other coefficients on the regressions for value per square foot are consistent with what we know about house prices over this period. First, houses that are above the threshold of 125 percent of CLL (i.e. the more expensive houses in the regression sample) are associated with a higher average value per square foot. In unreported analyses, we find that more expensive houses are generally associated with a higher value per square foot (i.e. price rises quicker than house size in the whole distribution of transactions), and here we find that this is also the case for the regression sample. Also, the "Year CLL" dummy variable is associated with a strong negative effect, reflecting the strong increase in house valuations that we saw in this period in the US. Given that the year in which the CLL is in effect is always the "pre" year in the regressions, we expect those transactions, on average, to be associated with a lower value per square foot. In Panels B and C we use the residuals from the regressions we described in Section 3.3.2 as the dependent variable to account for differences in quality between houses. The results are qualitatively and quantitatively very similar to the ones we present in Panel A. In Panel B we are using the residuals of a regression of log of house price on a set of characteristics, and we find a point estimate of -0.0017 that translates to residual being lower by 620 dollars for houses above the threshold of 125 percent of the CLL when the CLL binds, considering an average transaction value of 371,340 dollars. This suggests that transactions that cannot be financed at 80 percent with conforming loans are made at lower prices even after we control for a rich set of house characteristics. 11 Similarly in Panel C of Table 3.5, we confirm that even when we use the value per square foot as a dependent variable but control for house quality, the interaction term is significant and economically large even though the point estimate of 0.65 dollars for houses above the threshold is slightly lower than the results in Panel A where we do not adjust for house quality. The difference between the point estimate of 1.16 dollars of Panel A and 0.65 dollars in this specification indicates that houses above the limit are of somewhat worse quality than those below the limit in the year that the limit is in effect. We also show that the estimated effect of the conforming loan limit on house prices is stronger in the first half of the sample than in the second half. This result holds

"In the Appendix we show that the results are unchanged if we include the characteristics as controls in the regressions, as opposed to running the regressions with the hedonic residuals.

128 for all three left-hand side variables. This is in line with our expectations, given that borrowers had easier access to second lien loans after 2002 (we show the evolution of the use of second liens in Figure 1 of the Appendix). Additionally, more borrowers use jumbo loans, which may reflect a reduction of the cost differential of this type of loan relative to conforming loans, and an increase in the ease of access to this type of loan, possibly driven by an increased ease of securitization of these loans. Finally, in the Appendix we show the robustness of our results to serial correlation in the error term by constructing Newey-West standard errors, and all the results are unchanged

3.4.2 Credit Supply and Income

We now turn to how the effect of credit supply on house prices changes with the growth in income in a zip code. To do this, we obtain data on zip code level average household income each year from 2000 to 2007 from Melissa Data. 2 We create a new variable that is a "1" if a zip code has negative nominal average income growth from one year to the next, and "0" otherwise. We then run similar regressions to what we did before (year-by-year), adding an interaction between our previous variables and this new zip code level "Negative Income Growth" variable. Looking at the coefficient on the triple interaction term (negative income growth, the year that the CLL is in effect, and being above 125 percent of the CLL) allows us to identify how the effect of credit supply differs in times of positive and negative income growth. Our hypothesis is that the effect of credit supply is stronger in times of negative income growth, as households in a certain zip code are more likely to be constrained and there is likely to be less competition for housing, which increases the probability that a seller sells to a constrained buyer. We show the results for these regressions in Table 3.6. In the first column of Table 3.6 we repeat our main regressions for the period 2001-2005 only, as this is the period for which we were able to construct the income growth indicator variable. The results are consistent with those in Table 3.5. In the second column of Table 3.6 we show Fama-MacBeth coefficients from the regressions with the income growth interaction term. The triple interaction terms show that the effect of credit supply on value per square foot is significantly stronger in zip codes and years that are below the 10th percentile of income growth for the individual regression. The point estimate shows that value per square foot is 1.55 dollars lower in the year that the conforming loan limit is in effect for houses above 125 percent of the limit when income growth is low in a zip code. We also find that the main effect from our regressions in Table 3.5 is quantitatively similar to before, implying that the simple inclusion of ZIP code level income does not change any of our main results. In the Appendix we plot the distribution of value per square foot for ZIP codes of different income levels. Those pictures also suggest that the distribution of value per square foot is affected by the conforming loan limit in ZIP codes in the lowest quartile of the income distribution. In particular, the average value per square foot is monotonically increasing for up to conforming loan limit threshold, and from this

1 2 Melissa Data obtains this data from the IRS and provides it in an easy-to-read format.

129 point onwards the distribution becomes flat. This pattern is not visible for zip codes with higher median incomes.

3.4.3 Robustness and Refinements Differential House Price Trends

We want to rule out that our results are driven by differences in secular trends between houses above and below the threshold of CLL/0.8. Specifically, if more expensive houses have, on average, lower house price growth from one year to the next relative to less expensive houses, we might obtain the results reported in Table 3.5, but we might also obtain similar results for samples with transactions above and below other arbitrary thresholds. In order to address whether the effect that we find is indeed the product of the true conforming loan limits and not due to different trends along the distribution of houses, we run the same regressions described in Section 3.3.3 for "placebo" loan limits. We do this by shifting the true conforming loan limit in USD 10,000 steps from the true value each year. We start at CLL-100,000 and move 20 steps until we reach CLL+100,000. For each of these 21 tests, we first define the "shift" relative to the true conforming loan limits, and then we change the limits for all years by that amount. For example, when we are changing all the limits by -20,000, this means that the "placebo" limit for 1999 is 220,000 dollars instead of the true 240,000 dollars, the "placebo" limit for 2000 is 232,700 instead of 252,700, and so on. We then run the same year-by-year regressions and produce Fama-MacBeth coefficients for each of the 20 alternative "placebo" values for the CLL. The results from this exercise are shown in Table 3.7. The table shows that the coefficients of interest we obtain for all three dependent variables (values per square foot, residuals from the transaction amounts, and residuals of values per square foot) are systematically among the lowest of all obtained with the 20 "placebo" trials (the ranking is given in the last two rows of the table). The coefficient on the value per square foot measure is the lowest of the 21 trials whether we use the whole sample, or whether we limit our attention a sample of transactions that all have an LTV between 0.5 and 0.8.13 When we use the whole sample and the two residual measures from the hedonic regressions as the left-hand side variables in the regressions, the coefficients for the true conforming loan limits are the second and third lowest. In the restricted sample with LTVs between 0.5 and 0.8, these two measures produce the second lowest and the lowest coefficient out of the 21 trials. If we limit our attention to placebo limits that are below the true limits (i.e. the top half of Table 3.7), all our measures produce the lowest coefficients out of those trials. We consider these to be true "placebos", because all the transactions used for those regressions are, by construction, below the "eligibility" criteria of 125 percent of the true conforming loan limit both in the year that the limit is in effect, and in the

13 We discuss this subsample in more detail and show the equivalent to our Table 3.5 for this sample in the Appendix

130 subsequent year. As such, these transactions should not have any changes in credit availability from one year to the next. When we compute the standard deviation of those coefficients, we find that the coefficients using value per square foot as the dependent variable are statistically significantly different from the average of the other coefficients at a 5 percent level in both the whole sample and in the restricted sample with LTV between 0.5 and 0.8. T-statistics for these tests are shown in the fourth row of Table 3.7. When we use the value per square foot residual measure as a left-hand side variable, the coefficient has a t-statistic of 1.77 in the whole sample, and above 2.37 in the restricted sample. Finally, the coefficient from the regression that uses the residual from the log of house price hedonic regression as a left-hand side variable is not significantly different from the average of the other coefficients, as the t-statistics are between 1.0 and 1.2 in both the whole sample and in the restricted sample. The fact that the results are directionally the same when using all three left-hand side variables, and that there is no "placebo" limit that consistently produces results that are as strong as the ones from the true limit, further confirms that our coefficients are not obtained by pure chance.

Selection Into Treatment

As discussed in the introduction, there can be at least two alternative mechanisms for the effect of the conforming loan limits on house valuation. The first mechanism is that cheaper credit around the threshold leads to an increase in the demand for houses of a certain type, which then leads to higher valuation of these houses (or, conversely, higher cost of credit reduces the demand for houses above the threshold in the year that the limit is in effect). The alternative mechanism is that different credit conditions above and below the threshold attract a type of buyer in the year that the limit is in effect that is both better able to deal with the higher cost of funding (possibly because of higher wealth or income), and is a more effective negotiator than other "typical" buyers. This would still mean that our results are driven by credit conditions being different above and below the threshold, but it would be a different mechanism for our results. This selection effect results from the fact that borrowers can choose the level of their LTV. If all borrowers mechanically had to use an LTV of 80 percent, there would not be any possibility for selection. To understand whether the aforementioned form of selection is important, we divide transactions that are just above the cut off for being eligible for a CLL at 80 percent in a given year into two groups: (1) transactions that nevertheless use a conforming loan and therefore choose to have an LTV below 80 percent (making up the difference with other forms of financing), and (2) transactions that use a jumbo loan with an 80 percent LTV, which means they do not get a conforming loan. The first group isolates the set of borrowers where selection could be an issue. These borrowers might be optimizing around the CLL threshold and could therefore have other unobservable differences from the rest of the borrowers. For example, these "special" buyers could have more wealth or higher income and thus might also differ in other unobservables such as their ability to bargain. By excluding the group of

131 home buyers who choose this type of financing, we can test if these are driving our results, i.e. whether they alone buy cheaper houses. As an aside, it is ex ante not clear why those borrowers would buy cheaper houses (based on value per square foot). The fact that they are wealthier would usually lead us to believe that the omitted variable bias goes in the other direction, i.e. they buy houses with higher unobservable quality. The following regressions show that this group of borrowers does not drive our results. To test the importance of the selection effect, we run differences-in-differences regressions excluding each of the two groups described above at a time (in the year that the limit is in effect) and construct Fama-MacBeth coefficients, as we did in Table 3.5. The results are shown in Table 3.8. We find that results do not change much when we exclude the jumbo loans or when we exclude the conforming loans, which implies that our main results are not being driven solely by either one of these groups of transactions. The statistical significance of the results is similar, and the magnitude of the coefficients sometimes is larger for one group and other times for the other, depending on the left-hand side measure we use. Overall, the results point in the same direction for both sets of regressions. This robustness test shows that the effect of credit conditions on house prices in our setting is not likely to be driven solely by selection of different buyers in our "treated" group. If this were the case, we would expect the borrowers that pick a conforming loan and end up with an LTV below 80 percent to be the ones driving our main result. The fact that we also see similar results when we exclude this subgroup increases the likelihood of our alternative explanation, namely that differential cost of credit changes demand for housing, and that this shift in demand for housing drives the change in house valuation. In the Appendix we show that our results are stable if we use a 5,000 dollar band around the threshold of CLL/0.8 instead of the 10,000, which suggests that the difference in the cost of credit is likely to be similar for these two sets of buyers relative to buyers below the threshold. This is further evidence that the result is not driven solely by buyers who choose to obtain a conforming mortgage and put up additional equity from other sources. Finally, we also show in the Appendix that the effect of the CLL is similar for the first 9 months of the year and for the last three months, indicating that borrowers do not behave differently after the limit for the subsequent year has been defined by the administration.

Constraints to Housing Supply

To understand whether the effect of credit supply is amplified by the inability of housing supply to adjust quickly to demand, we divide zip codes into high and low house supply elasticity according to the measure in Saiz (2010). If the supply of housing were perfectly elastic and able to adjust quickly to an increase in demand for houses, the effect on prices should not be there. In this test, we find that the constraint imposed by the conforming loan limit is stronger in zip codes located in more inelastic, metropolitan, statistical areas (MSAs) according to the Saiz measure (Table 3.9). This result is in line with what we expect and with previous literature (e.g. Mian and Sufi, 2009), namely that cheaper credit will feed through to house

132 prices more frequently in regions where the supply of houses cannot adjust as easily. We are cautious to interpret this result, however, because we have limited cross- sectional variation in the elasticity measure in our data. In fact, all of the MSAs in our sample are above the median elasticity found in Saiz (2010) for the whole country, and 7 of the 10 MSAs are in the top 20 percent of MSAs with the least elasticity in the nation.

3.4.4 Economic Magnitude of the Effect

As we discuss in Section 3.2, there is significant disagreement as to what the magnitude of the elasticity of house prices to interest rates is, as changes to the way a standard user cost model is specified can produce vastly different estimates. To understand the magnitude of our estimated effect, we compute the semi-elasticity of house prices to interest rates, calculated as the percentage change in prices divided by the change in interest rates. The change in the CLL gives us an unbiased local estimate of the numerator of this semi-elasticity. To obtain an estimate of the denominator, we use the differential in interest rates between jumbo and conforming loans estimated in the prior literature. Table 3.10 shows that the change in house prices around the CLL ranges from 30 to 91 basis points. We obtain the low of 30 basis points when we use the residuals from the hedonic regressions of value per square foot as the dependent variable and include the whole time period (1998 to 2006).4 The high end of the estimate (91 basis points) comes from the specification where we constrain the period to 1998-2001 and use the raw value per square foot as the dependent variable. We exclude our estimates for the period 2002 to 2005 since we know that the CLL was less important during that time. There is an extensive literature that provides estimates of the jumbo-conforming spread, see McKenzie (2002), Ambrose, LaCour-Little, and Sanders (2004), Sherlund (2008), Kaufman (2012) and DeFusco and Paciorek (2013). The most common estimates that have been found across all the papers range from a low of 10 basis points to a high of 24 basis points. 15 If we divide our estimated range of house price changes by the range in the jumbo-conforming spread, we obtain estimates for the elasticity of house prices to interest rates that vary between 1.2 and 9.1 (Table 3.10). While these estimates are local in nature, i.e. they do not use the full distribution of housing transactions in the data nor do they take into account general equilibrium effects, this is the first unbiased estimate of this semi-elasticity in the literature and the results are at the lower end of the estimates that have been proposed previously (see, for example, Glaeser, Gottlieb, and Gyourko, 2010). In fact, given our data, it is hard

14 The point estimate in the regressions is 0.65 dollars from Panel C in Table 3.5, and we scale that by the average value per square foot for the sample to obtain 30 basis point changes in value per square foot. 15The paper by Kaufman (2012) obtains an estimate of 10 basis points by using a regression discontinuity approach on the access to conforming loans around the threshold of CLL/0.8 in appraisal values. This estimate is particularly relevant for our purposes given that it explores the part of the distribution of homes that we also consider.

133 to justify estimates above 10 without making very aggressive assumptions about the cost differential above and below the threshold. The prior calculation is our preferred method of obtaining an estimate of the elasticity. However, we can obtain an alternative estimate of the elasticity by considering borrowers who choose to obtain a conforming loan of less than 80 percent LTV above the threshold. This means they put up additional equity which either has to be financed through a third party loan or through savings. On average, given the range of transactions used in the regressions, these borrowers put up an additional USD 5,000. If we assume that the cost of the additional equity is 5 percentage points or more above the conforming mortgage rate, this is equivalent to a spread of 6-8 basis points in the total cost of financing for these borrowers relative to those who buy a house below the threshold. This translates into an elasticity of between 4.4 and 11.4, depending on the house price effect we use from our regressions. The assumption for the spread of 5 percentage points over the conforming mortgage rate is not high if we consider that many people use a jumbo loan even very close to the threshold of the CLL, indicating that the cost of additional equity is, at least for some borrowers, very substantial. The fact that we see borrowers stick with a conforming loan and put up additional equity above the threshold may, in fact, be an indication that they are excluded from the jumbo market altogether, rather than evidence that this is a cheaper option. As Loutskina and Strahan (2009, 2011) show, jumbo loans are associated with more careful screening of borrowers, which may mean that many households simply could not use an 80 percent LTV above the threshold of 125 percent of the CLL even if they were looking to do so. Another way of assessing the economic importance of the effect we find is by comparing the dollar amount of savings through lower interest rates and the house price differential. Assume a loan of USD 300,000, which is approximately the conforming loan limit midway through our sample (2002). If we use the upper end of the jumbo- conforming spread of 24 basis points, we calculate a cost difference of USD 720 in the first year of the life of the loan. The present value of the cost difference over 30 years is USD 8,557 assuming a 6 percent discount rate. If we use the lower end of the jumbo-conforming spread that has been estimated (10 basis points), this cost difference is USD 3,604. Our estimated effect of the conforming loan is a price difference of USD 0.65-1.16 per square foot for an average size of a house of 1,935 square feet. This translates into a USD 1,256-2,244 difference in the price of the house. Thus, for each dollar of savings in the present value of interest costs, home values increase by about 25-60 cents (always less than 1 dollar). One possible concern with our estimation is that home buyers might expect the conforming loan limit to rise in the subsequent year and would thus refinance their loan shortly after obtaining it. If refinancing were frictionless, buying a house above the threshold would cost 10-24 basis points more than the conforming loan rate for only one year, because borrowers who took a jumbo loan would immediately refinance into a conforming loan in the following year (once the limit was raised). This would imply a very high elasticity of house prices to interest rates, as the difference in the effective interest rate over the life of the loan paid by a borrower who took a conforming loan and one who took a jumbo loan would be very small. However,

134 this analysis misses the transaction costs of refinancing, and the estimates of these transaction costs that have been found in the literature are very large. A paper by Stanton (1995) finds that transaction costs for mortgage prepayment are around 30 to 50 percent of the remaining principal balance of a mortgage. These transaction costs include both explicit monetary costs (about one-sixth of the total costs) and non-monetary prepayment costs (the remaining five-sixths). A more recent paper by Downing, Stanton and Wallace (2005) produced a lower, but still substantial, average transaction cost of refinancing of 11.5 percent of face value. The bottom line from both these studies is clear - transaction costs are too high for the jumbo conforming spread alone to significantly change the prepayment behavior of borrowers. In other words, the benefit from obtaining lower interest rates by refinancing to a conforming loan in a year or two are too small to overcome the transaction costs of refinancing.

3.5 Conclusion

In this paper we use the exogenous changes in the annual level of the conforming loan limit as an instrument for lower cost of funding. We find that a home that becomes eligible for cheaper mortgages due to an increase in the CLL has, on average, a 1.16 dollar higher value per square foot compared to a house that is just above the threshold that allows it to be financed with a conforming loan at 80 percent loan to value. The magnitude of the difference that we find is economically important given the average value per square foot of houses that transact around the CLL of 220 dollars, which means that a 1.16 dollar increase constitutes almost a 0.45 percent increase in prices. Under our assumptions for the interest rate differential for transactions above and below the threshold, this corresponds to a semi-elasticity of prices to interest rates of less than 10. Another way of stating our results is to say that the interest rate subsidy granted by the GSEs and, ultimately, the taxpayer, does not fully benefit the buyers of homes and, instead, partially accrues to the sellers of homes in the form of higher house prices. Also, the results suggest that mortgages are being supplied in a competitive fashion, and that originating banks are not appropriating the mortgage subsidy provided by the GSEs. In addition, we see that the CLL constitutes a first order factor in how houses are financed: there is a significant fraction of borrowers who choose an LTV below 80 percent, between 77 and 79.5 percent, in order to stay below the conforming loan limit. These borrowers either were unable to get a jumbo loan, or are trying to take advantage of the lower interest rate of a conforming loan. But, as a result, many borrowers end up holding a larger fraction of equity in their house than most other borrowers. These results are stronger in the earlier part of our sample when borrowers were less likely to have access to other forms of financing, such as second liens, and when the interest rate differential between jumbo loans and conforming loans was larger. After 2004 in particular, we see that the vast majority of borrowers even above the threshold of 125 percent of the CLL choose an LTV of 80 percent, which supports the idea that access to jumbo loans and other forms of financing became much easier in the

135 second half of the sample. At the same time, the house price impact of the conforming loan limit is also smaller in this time period. This suggests that those houses which were previously just out of reach of being financed by a conforming loan at 80 percent could now be bid up in price since people had easier access to jumbo loans and other forms of finance. The results are also stronger in ZIP codes with the lowest income growth, usually negative, and also in areas with lower elasticity of housing supply. While we can only estimate a local treatment effect around the CLL, this presents a first test of the exogenous effect of cheaper mortgage loans on house prices. We estimate an elasticity of house prices to interest rates that is below 10, implying that the drop in mortgage rates cannot account for the increase in house prices between 2000 and 2006. However, we do show that those credit conditions matter for the formation of prices. Our results do not support a view that credit market conditions purely respond to housing demand, but point instead to a directional effect that easier credit supply leads to an increase in house prices.

136 3.6 Bibliography

Acharya, V., Richardson, M., Nieuwerburgh, S. V., and White, L. J. (2010) Guar- anteed to Fail: Fannie Mae, Freddie Mac, and the Debacle of Mortgage Finance. Princeton University Press, March 2011. Ambrose, B. W., LaCour-Little, M., and Sanders A.B. (2004) The Effect of Con- forming Loan Status on Mortgage Yield Spreads: A Loan Level Analysis. Real Estate Economics, Vol. 32, No. 4, 541-569. Brunnermeier, Markus K., Eisenbach, T., and Sannikov, Y. (2012) Macroeco- nomics with Financial Frictions: A Survey. Working Paper. Calomiris, C. W.. (2009). Financial Innovation, Regulation, and Reform. Cato Journal (29) p. 6 5 . Campbell, J.Y., Giglio, S., and Pathak, P. (2010) Forced Sales and House Prices. American Economic Review, Forthcoming. DeFusco, A. A., and Paciorek, A. (2013) The Interest Rate Elasticity of Mortgage Demand: Evidence From Bunching at the Conforming Loan Limit. Working Paper. Downing, C., Stanton, R., and Wallace, N. (2005) An Empirical Test of a Two- Factor Mortgage Valuation Model: How Much Do House Prices Matter? Real Estate Economics, Vol. 33, Issue 4, 681-710. Fama, E. F., and MacBeth, J. D. (1973) Risk, Return, and Equilibrium: Empirical Tests. Journal of Political Economy, Vol. 81, No. 3, 607-636. Favara, G., and Imbs, J. (2011) Credit Supply and the Price of Housing. CEPR Discussion Paper, No. 8129. FavilukisJ., Ludvigson, S.C., and Nieuwerburgh, S. V. (2010) The Macroeconomic Effects of Housing Wealth, Housing Finance, and Limited Risk-Sharing in General Equilibrium. NBER Working Paper, No. 15988. Genesove, D. and Mayer, C. J. (1997) Equity and Time to Sale in the Real Estate Market. American Economic Review, Vol. 87, No. 3. (Jun, 1997), 255-269. Glaeser, E. L, Gottlieb, J., and Gyourko, J. (2010) Can Cheap Credit Explain the Housing Boom. NBER Working Paper, No. 16230. Green, R. K., and Wachter, S. M. (2005) The American Mortgage in Historical and International Context. Journal of Economic Perspectives, Vol. 19, No. 4, 93-114. Himmelberg, C., Mayer, C., and Sinai, T. (2005) Assessing High House Prices: Bubbles, Fundamentals and Misperceptions. Journal of Economic Perspectives, Vol. 19(4), 67-92. Hubbard, G., and Mayer, C. (2008) House Prices, Interest Rates, and the Mort- gage Market Meltdown. Columbia Business School Working Paper. Kaufman, A. (2012) What do Fannie and Freddie do? Unpublished Manuscript. Khandani, A. E., Lo, A.W., and Merton, R.C. (2009) Systemic Risk and the Refinancing Ratchet Effect. NBER Working Paper,No. 15362. Loutskina, E., and Strahan, P. (2009) Securitization and the Declining Impact of Bank Financial Condition on Loan Supply: Evidence from Mortgage Originations. Journal of Finance, 64(2), 861-922.

137 Loutskina, E., and Strahan, P. (2011) Informed and Uninformed Investment in Housing: The Downside of Diversification. Review of FinancialStudies, 24(5), 1447- 80. Mayer, C. (2011) Housing Bubbles: A Survey. Annual Review of Economics, 3:55977. McKenzie, J.A. (2002) A Reconsideration of the Jumbo/Non-jumbo Mortgage Rate Differential. Journal of Real Estate Finance and Economics, Vol. 25, No. 2-3, 197-213. Mian, A., and Sufi, A. (2009) The Consequences of Mortgage Credit Expansion: Evidence from the U.S. Mortgage Default Crisis. Quarterly Journal of Economics, Vol. 124, No. 4, 1449-1496. Pavlov, A., and Wachter, S. (2011) Subprime Lending and Real Estate Prices. Real Estate Economics, 39: 117 Poterba, J. (1984) Tax Subsidies to Owner-occupied Housing: An Asset-Market Approach. Quarterly Journal of Economics, Vol. 99(4), 729-52. Saiz, A. (2010) The Geographic Determinants of Housing Supply. Quarterly Jour- nal of Economics, 125(3): 1253-1296. Sherlund, S.M. (2008) The Jumbo-Conforming Spread: A Semiparametric Ap- proach. Finance and Economics Discussion Series Working Paper, 2008-01. Stanton, R. (1995) Rational Prepayment and the Valuation of Mortgage-Backed Securities Review of FinancialStudies, Vol. 8, No. 3, 677-708. Stein, J.C. (1995) Prices and Trading Volume in the Housing Market: A Model with Down-Payment Effects. Quarterly Journal of Economics, Vol. 100, No. 2, 379-406.

138 Figure 3-1: Transaction-Loan Value Surface

Note: This figure shows the frequency of transactions at each house price-loan value combination for the year 2000 and 2004, and the 10 MSAs covered in our data, where both house prices and loan values were binned at USD 10,000 intervals. The mass of transactions on the diagonal have a loan to value of approximately 0.8.

(a) 2000

4500

4000

3500

2500

I'm

3500

1000

Transaction Value Loan Value

(b) 2004

SODO 40M IiLIL

Transaction Value LOWn Value

139 Figure 3-2: Borrower Composition for the Regression Sample

Note: This figure shows the number of transactions by month for transactions within USD 10,000 of the threshold of 125 percent of CLL. Transactions below and above this threshold are tracked from the year prior to the CLL being in effect to the year after the CLL is lifted to its new value. We break down transactions by LTV range to show the differences that emerge between houses above and below 125 percent of the CLL.

(a) Transactions below 125 percent of CLL

3500 (A) (B) (C) CLI year-i CLL year 3000 CLLyear+ 1

12500

2000

1500

Li -02

0 1 2 3 4 5 6 7 8 9 101112131415161718192021222324252627282930313233343536 months

E LTV<75 m7580

(b) Transactions above 125 percent of CLL

3500

3 0 0 0 CLL year -1 CLLyear i CLL year +1

2500

2000

1500

500

1 2 3 4 5 6 7 8 9 1011121314Im1617181920212223242526272829303132333435 Months

EILTV<75 *7580

140 Figure 3-3: Frequency of Transactions as Percentage of CLL Threshold

Note: This figure shows the frequency of transactions by their distance to the threshold of 125 percent of the conforming loan limit. The vertical red line is the threshold and the transactions for all years are centered around that value. The x-axis is represented as one minus the transaction value as a percentage of each year's threshold of 125 percent of the conforming loan limit (e.g. if the threshold is 200,000, a transaction of 150,000 will appear as -25 percent).

0 03

C C 0D

I0 03 0 $1 *@@ O C vfI %

E z 0% .met0406.

R -100 -50 0 50 100 Transaction Value as Percentage of 1.25CLL

141 Figure 3-4: Share of Unused Mortgage Applications Note: The horizontal axis indicates the difference between loan amounts and the conforming loan limit as a percentage of the conforming loan limit. The share of unused mortgages is constructed from HMDA as the number of "withdrawn" or "unused" mortgage applications as a percentage of total applications. We aggregate these proportions into 1% bins and each dot in the figure represents the share of unused mortgages for each bin. We also plot third degree polynomials ( to the left and right of the conforming loan limit) as well as 95% confidence intervals (dashed lines). Data extracted from HMDA, 1998-2006.

-A *~ Soa a a ~-a o -a-'. S

10 U S 0~ U r 0~

-50 -45 40 -35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35 40 45 50 Distance from the conforming loan limit (%)

142 Table 3.1: Summary Statistics

Panel A. House Characteristics. All Transactions N=3,983,575 Regression Sample N=262,671 Mean Std. Dev. Median Mean Std. Dev. Median Transaction Value (USD 1000) 308.52 123.93 286.00 371.34 54.92 380.00 Loan to value 0.81 0.15 0.80 0.76 0.13 0.80 House Size (sqft) 1,735 672 1,592 1,935 701 1,816 Lot Size (sqft) 10,197 15,495 6,700 11,734 17,923 7,203 Number of rooms 6.84 1.60 7.00 7.23 1.61 7.00 Number of bedrooms 3.20 0.78 3.00 3.33 0.78 3.00 Number of bathrooms 1.93 1.03 2.00 2.11 1.07 2.00 House age (years) 35.40 27.70 34.00 34.74 27.40 34.00

Panel B. House Valuation. All Transactions N=3,983,575 Regression Sample N=262,671 Mean Std. Dev. Median Mean Std. Dev. Median Value per sqft (USD/sqft) 193.59 91.60 172.03 219.63 93.37 200.20 Value per sqft residual (USD/sqft 0.00 42.30 -0.95 5.29 44.26 3.43 Log of transaction value residual (USD) 0.00 0.17 0.01 0.05 0.14 0.04

Note: Panel A shows the descriptive statistics for all transactions in our data from 1998 to 2008. The data was extracted from deeds records by Dataquick. Panel B shows the different valuation measures we use in the regression analysis. Value per sqft is the transaction amount divided by the size of the house measured in square feet. Both the residual measures are obtained from hedonic regressions run by year and by metropolitan area of value per sqft and transaction value on a set of detailed house characteristics. We give more information on the construction of the residuals in Section 2, Data and Methodology.

143 Table 3.2: Summary Statistics by Geography and Year

Panel A. Geographic Distribution

MSA Transaction Value Value per sqft Loan to Value N Obs Mean Std. Dev Mean Std. Dev Mean Std. Dev Boston 279,261 320.29 112.40 197.67 73.81 0.78 0.16 Chicago 377,031 262.41 108.15 174.37 68.63 0.81 0.15 DC 396,211 329.95 126.16 186.97 85.93 0.82 0.14 Denver 397,293 250.22 94.93 155.84 49.28 0.83 0.15 Las Vegas 345,219 262.24 102.87 136.62 45.38 0.82 0.14 Los Angeles 725,897 332.28 129.71 231.29 108.35 0.81 0.13 Miarni 483,541 270.10 111.74 144.80 57.04 0.81 0.14 New York 487,104 341.00 121.13 221.25 92.55 0.78 0.17 San Diego 219,489 353.14 124.63 222.18 94.86 0.79 0.14 San Francisco 272,529 383.59 123.74 266.47 109.26 0.79 0.13 Total 3,983.575 308.52 123.93 193.59 91.60 0.81 0.15

Panel B. Distribution By Year and Thresholds

Year Thresholds Transaction Value Value per sqft Loan to Value N Obs House Price Conf. Loan Mean Std. Dev Mean Std. Dev Mean Std. Dev 1998 134,200 283,938 227,150 239.78 102.07 133.84 50.59 0.81 0.15 1999 350,827 300,000 240,000 246.38 104.88 139.33 54.03 0.81 0.15 2000 354,071 315,875 252,700 257.67 109.21 149.65 61.64 0.81 0.16 2001 365,814 343,750 275,000 265.16 108.82 156.74 63.81 0.82 0.15 2002 397,527 375,875 300,700 283.79 114.34 171.06 71.85 0.81 0.15 2003 423,939 403,375 322,700 303.37 118.32 187.40 80.05 0.81 0.15 2004 525,407 417,125 333,700 331.81 121.20 212.65 90.51 0.79 0.14 2005 475,723 449,563 359,650 357.51 121.71 237.24 100.72 0.78 0.13 2006 376,182 521,250 417,000 366.27 121.89 247.02 105.50 0.79 0.13 2007 293,329 521,250 417,000 359.24 122.53 237.79 101.57 0.82 0.14 2008 286,556 521,250 417,000 325.11 119.84 206.92 91.62 0.84 0.15 Total 3,983,575 308.52 123.93 193.59 91.60 0.81 0.15

Note: This table uses all the deed registry data on house transactions for 10 MSAs. Panel A shows the mean and standard deviation by city of (i) house price, (ii) value per sqft and (iii) loan to value. Panel B the mean and standard deviation by year for the same three variables.

144 Table 3.3: Verification of the Impact of the CLL on Financing Choices

Panel A: Loan to Value

All years 1998-2001 2002-2005 Above Threshold -0.004*** -0.006*** -0.002*** (0.001) (0.002) (0.001) Year CLL -0.008*** -0.005** -0.011*** (0.002) (0.002) (0.001) Above Threshold x -0.004*** -0.004* -0.003* Year CLL (0.001) (0.002) (0.002) No. Obs. 242,753 100,870 141,883

Panel B: Log Loan Amount

All years 1998-2001 2002-2005 Above Threshold 0.023*** 0.024*** 0.021*** (0.002) (0.003) (0.001) Year CLL -0.013*** -0.009*** -0.017*** (0.002) (0.003) (0.003) Above Threshold x -0.006** -0.007* -0.005 Year CLL (0.002) (0.004) (0.003) No. Obs. 242,753 100,870 141,883

Note: This table shows Fama MacBeth coefficients computed from year by year regressions that use two measures of financing choice as the dependent variable in each of the two panels. The sample includes all transactions within USD 10,000 of each year's conforming loan limit, as well as transactions of the same amount in the subsequent year. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect.

145 Table 3.4: Impact of CLL on Number of Transactions

All years 1998-2001 2002-2005 Year CLL -0.003*** 0.000 -0.006*** (0.000) (0.001) (0.001) No. Obs. 262,671 109,496 153,175

Note: This table shows Fama MacBeth coefficients computed from year by year regressions that use a dummy variable for whether a transaction happens above the threshold of 125 percent of the CLL as the dependent variable. The sample includes all transactions within USD 10,000 of each year's conforming loan limit, as well as transactions of the same amount in the subsequent year. Year CLL is the year in which the conforming loan limit is in effect. Zip Codes fixed effects are included on each regression

146 Table 3.5: Effect of the CLL on House Valuation Measures

Panel A: Value Per Square Foot

All years 1998-2001 2002-2005 Above Threshold 1.261** 1.669*** 0.852 (0.494) (0.573) (0.836) Year CLL -22.869*** -14.851*** -30.886*** (4.047) (2.314) (5.314) Above Threshold x -1.162*** -1.553*** -0.771** Year CLL (0.264) (0.297) (0.369) No. Obs. 262,671 109,496 153,175

Panel B: Log of Transaction Value Residual from Hedonic Regressions

All years 1998-2001 2002-2005 Above Threshold 0.0129*** 0.0154*** 0.0104*** (0.0013) (0.0015) (0.0009) Year CLL 0.0387*** 0.0356*** 0.0417*** (0.0041) (0.0047) (0.0072) Above Threshold x -0.0017** -0.0020 -0.0013*** Year CLL (0.0008) (0.0015) (0.0004) No. Obs. 251,431 103,535 147,896

Panel C: Value Per Square Foot Residual from Hedonic Regressions

All years 1998-2001 2002-2005 Above Threshold 1.733*** 2.060*** 1.407** (0.360) (0.425) (0.595) Year CLL 4.103*** 3.935*** 4.270*** (0.644) (0.495) (1.293) Above Threshold x -0.651*** -0.940*** -0.362 Year CLL (0.238) (0.351) (0.291) No. Obs. 251,764 103,709 148,055

Note: This table shows Fama MacBeth coefficients computed from year by year regressions that use three alternative measures of valuation as the dependent variable in each of the three panels. The hedonic regressions that produce the residuals for panels B and C are described in Section 3.3.2. The sample for each year's regression includes all transactions within +/- USD 10,000 of that year's conforming loan limit, as well as transactions in the same band in the subsequent year. All year by year regressions include ZIP code fixed effects. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect.

147 Table 3.6: Effect of the CLL on House Valuation in Different Income Growth Areas

Panel A: Value Per Square Foot

2001-2005 2001-2005 Above Threshold 0.731 0.601 (0.667) (0.638) Year CLL -28.869*** -29.364*** (4.706) (4.510) Above Threshold x -0.846*** -0.953*** Year CLL (0.257) (0.210) Above Threshold x -1.548** Year CLL x Low Inc. Growth (0.652) No. Obs. 179,828 179,828

Panel B: Log of Transaction Value Residual from Hedonic Regressions

2001-2005 2001-2005 Above Threshold 0.0109*** 0.0108*** (0.0008) (0.0009) Year CLL 0.0418*** 0.0439*** (0.0056) (0.0057) Above Threshold x -0.0016*** -0.0022*** Year CLL (0.0003) (0.0006) Above Threshold x -0.0018 Year CLL x Low Inc. Growth (0.0051) No. Obs. 173,347 173,347

Panel C: Value Per Square Foot Residual from Hedonic Regressions

2001-2005 2001-2005 Above Threshold 1.396*** 1.347*** (0.453) (0.412) Year CLL 4.314*** 4.806*** (1.017) (1.072) Above Threshold x -0.504** -0.750*** Year CLL (0.250) (0.158) Above Threshold x -0.319 Year CLL x Low Inc. Growth (0.651) No. Obs. 173,550 173,550

Note: This table shows Fama MacBeth coefficients computed from year by year regressions that use three alternative measures of valuation as the dependent variable in each of the three panels. The sample for each year's regression includes all transactions within +/- USD 10,000 of that year's conforming loan limit, as well as transactions in the same band in the subsequent year. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect. This specification interacts the diff-in-diff specification with a dummy variable that uses changes in income at a zipcode level as proxy for good and bad times. Specifically, the dummy is 1 if the changes in the average zipcode income are below the 10th percentile of each particular diff-in-diff regression and 0 otherwise. We use tax income data at zipcode level available from 2000-2006, which restricted our sample to 2001-2005

148 Table 3.7: Placebo Test for Coefficient of Interest

All Transactions 0.5

True CLL -1.162 -0.002 -0.651 -1.257 -0.002 -0.931

Placebo 0.045 0.001 0.222 -0.107 0.000 0.110 (0.467) (0.002) (0.494) (-0.107) (0.002) (0.440)

T-Statistic 2.586 1.206 1.770 2.626 1.009 2.365

CLL Rank 1 4 2 1 3 1 CLL Rank 1 2 1 1 1 1 below only

Note: This table shows the average and standard deviation (in parenthesis) of a series of 20 placebo tests we perform by shifting the conforming loan limit in USD 10,000 intervals from CLL-100,000 until CLL+100,000 (i.e. the limits of all years are first changed by -100,000, then by -90,000, etc.). The first row shows the coefficients when we use the true conforming loan limit. We use the placebo loan limits to run year-by-year regressions and form Fama-MacBeth coefficients like those in Table 3.5 for each set of "false" loan limits. The t-statistic is for the difference between the coefficients when we use the true conforming loan limit and the average of all the other coefficients, using the standard deviation given by the 20 trials. The three dependent variables are the same we use in Table 3.5. The coefficient of interest is on the interaction between our "above threshold" variable and the year in which the conforming loan limit is in effect. As in the previous tables, the sample for each year's regression includes transactions within +/- USD 10,000 of that year's CLL, as well as transactions in the same band in the subsequent year. The first three columns include all such transactions, whereas in the last three columns the sample is constrained to transactions with an LTV between 0.5 and 0.8. All year by year regressions include ZIP code fixed effects. The last two rows show the ranking of the coefficient when we use the true CLL, first for all placebo limits and then when we only consider the placebo tests below the true CLL.

149 Table 3.8: Effect of the CLL on the Valuation of Different Groups of Transactions

Panel A: Value Per Square Foot

Keeping Conforming Keeping Jumbo All years 1998-2001 2002-2005 All years 1998-2001 2002-2005 Above Threshold 0.939** 1.580*** 0.297 0.868* 1.530*** 0.207 (0.472) (0.568) (0.666) (0.481) (0.545) (0.701) Year CLL -24.539*** -15.953*** -33.126*** -24.874*** -16.040*** -33.708*** (4.351) (2.564) (5.712) (4.454) (2.596) (5.813) Above Threshold x -0.967** -1.314** -0.621 -2.177*** -2.618** -1.736** Year CLL (0.416) (0.572) (0.634) (0.639) (1.119) (0.724) No. Obs. 177,227 72,048 105,179 160,342 62,905 97,437

Panel B: Log of Transaction Value Residual from Hedonic Regressions

Keeping Conforming Keeping Jumbo All years 1998-2001 2002-2005 All years 1998-2001 2002-2005 Above Threshold 0.0117*** 0.0145*** 0.0090*** 0.0119*** 0.0146*** 0.0091*** (0.0014) (0.0018) (0.0007) (0.0013) (0.0016) (0.0008) Year CLL 0.0367*** 0.0335*** 0.0398*** 0.0370*** 0.0337*** 0.0402*** (0.0038) (0.0041) (0.0067) (0.0039) (0.0042) (0.0068) Above Threshold x -0.0027** -0.0019 -0.0034*** 0.0004 -0.0020 0.0028** Year CLL (0.0011) (0.0022) (0.0009) (0.0015) (0.0025) (0.0012) No. Obs. 170,808 68,719 102,089 154,848 60,114 94,734

Panel C: Value Per Square Foot Residual from Hedonic Regressions

Keeping Conforming Keeping Jumbo All years 1998-2001 2002-2005 All years 1998-2001 2002-2005 Above Threshold 1.573*** 1.947*** 1.199*** 1.583*** 1.991*** 1.175** (0.290) (0.357) (0.414) (0.308) (0.333) (0.470) Year CLL 3.514*** 3.485*** 3.543*** 3.529*** 3.552*** 3.507*** (0.579) (0.431) (1.175) (0.573) (0.409) (1.168) Above Threshold x -1.399*** -1.216** -1.583*** 0.225 -0.462 0.911** Year CLL (0.344) (0.535) (0.493) (0.418) (0.536) (0.464) No. Obs. 170,946 68,790 102,156 154,949 60,165 94,784

Note: This table shows Fama Macbeth coefficients computed from year by year regressions that use three alternative measures of valuation as the dependent variable in each of the three panels. The hedonic regressions that produce the residuals for panels B and C are described in Section 3.3.2. The sample for each year's regression includes transactions within +/- USD 10,000 of that year's conforming loan limit. All year by year regressions include ZIP code fixed effects. We divide the transactions that happen at a price above 125 percent of a year's CLL in the year that the limit is in effect into two groups: those with a conforming loan and those with a jumbo loan. We then run the same regressions including just one of these two groups at a time. The first three columns include the transactions with a conforming loan and the last three columns include transactions with a jumbo loan. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect.

150 Table 3.9: Effect of the CLL on House Valuation in Low Supply Elasticity Areas ( Elasticity<1)

Panel A: Value Per Square Foot

All All 1998-2001 1998-2001 2002-2005 2002-2005 Above Threshold 1.261** 1.221 1.669*** 3.069*** 0.852 -0.628 (0.494) (0.799) (0.573) (0.374) (0.836) (0.749) Year CLL -22.869*** -15.282*** -14.851*** -8.015*** -30.886*** -22.550*** (4.047) (3.920) (2.314) (0.843) (5.314) (5.981) Above Threshold x -1.162*** -0.430 -1.553*** -2.100** -0.771** 1.239 Year CLL (0.264) (0.831) (0.297) (0.817) (0.369) (0.832) Above Threshold x -0.870 0.726 -2.466** Year CLL x Low Elasticity (0.977) (1.332) (0.992) No. Obs. 262,671 262,671 109,496 109,496 153,175 153,175

Panel B: Log of Transaction Value Residual from Hedonic Regressions

All All 1998-2001 1998-2001 2002-2005 2002-2005 Above Threshold 0.0129*** 0.0106*** 0.0154*** 0.0182*** 0.0104*** 0.0030* (0.0013) (0.0030) (0.0015) (0.0012) (0.0009) (0.0016) Year CLL (I).0387*** 0.0263*** 0.0356*** 0.0306*** 0.0417*** 0.0219*** (0.0041) (0.0037) (0.0047) (0.0044) (0.0072) (0.0055) Above Threshold x 0.0017** 0.0008 -0.0020 -0.0018 -0.0013*** 0.0033* Year CLL (0.0008) (0.0022) (0.0015) (0.0037) (0.0004) (0.0018) Above Threshold x -0.0032 -0.0002 -0.0063*** Year CLL x Low Elasticity (0.002) (0.004) (0.0020) No. Obs. 251,431 251,431 103,535 103,535 147,896 147,896

Panel C: Value Per Square Foot Residual from Hedonic Regressions

All All 1998-2001 1998-2001 2002-2005 2002-2005 Above Threshold 1.733*** 1. 338** 2.060*** 2.623*** 1.407** 0.054 (0.360) (( .524) (0.425) (0.278) (0.595) (0.319) Year CLL 4.103*** 1. 811** 3.935*** 3.316*** 4.270*** 0.305 (0.644) (C.716) (0.495) (0.270) (1.293) (0.898) Above Threshold x -0.651*** -).503 -0.940*** -1.620*** -0.362 0.615 Year CLL (0.238) (C.546) (0.351) (0.306) (0.291) (0.684) Above Threshold x ).241 0.843 -1.325 Year CLL x Low Elasticity (2).740) (0.744) (1.104) No. Obs. 251,764 2C1,764 103,709 103,709 148,055 148,055

Note: In this case the dummy is 1 for low elasticity places. For this specification that corresponde to the lowest MSA ( Miami, San Francisco, San Diego, Los Angeles, New York, Chicago and Boston). The areas with elasticity higher than 1 are Las Vegas, Denver and DC

151 Table 3.10: Elasticity Estimates

Jumbo-Conforming Spread A House Prices in bp Min (10 bp) Max (24 bp) Max: 91.2 9.1 3.8 Min: 29.7 3.0 1.2

Note: This table shows elasticity calculations for different scenarios of both the house price increase estimated in the regressions and the interest rate differential implied for transactions above and below the threshold of 125 percent of the conforming loan limit. We use the jumbo-conforming spread in interest rates as the denominator in the elasticity calculation.

152 3.7 Appendix A. Robustness and Refinements - Additional Tests

3.7.1 Restrict LTV Choices

We want to test that our estimates are not driven by borrowers with very unusual LTV levels, namely those with LTV below 50 percent and above 80 percent. Borrowers with those choices of LTV are likely to either have access to abundant equity to put up when buying a home, or to be very constrained and need a very high LTV. By limiting our sample to include only borrowers who choose a first lien LTV between 50 and 80 percent, we capture the transactions that should be most affected by the conforming loan limit. In particular, this subsample includes the group of borrowers that end up with an LTV between 77 percent and 79.5 percent in the year that the CLL is in effect because they stick with a conforming loan, even though their house costs more than 125 percent of the CLL. This choice of LTV is very common for the "Above the Threshold" group of borrowers in the year that the limit is in effect, but very infrequent everywhere else in the distribution of transactions. Also, this subsample includes all the borrowers that choose an 80 percent LTV, the most frequent choice in the data. This means getting a jumbo loan for transactions "Above the Threshold" and a conforming loan for transactions below that threshold. Finally, the transactions that are excluded from this sample should be least affected by the conforming loan limit, either because their LTVs are very low, in which case they are never affected by the limit anyway, or alternatively, because they have high LTVs and thus obtain jumbo loans in the year in which the limit is in effect whether the price of the transactions is above or below the 125 percent of the CLL threshold. Table 3.12 shows the results for Fama-MacBeth coefficients from year-by-year regressions, much like we described in the Main Results section of the paper, except using only transactions with an LTV between 0.5 and 0.8. The results are quantitatively similar to those we obtain for the whole sample, which means that our main results are not being driven by very low or very high LTVs. This reinforces our interpretation that our main results are caused by the CLL and not some other spurious factor. The magnitude of the coefficients is very similar to the ones in the previous table, but we lose statistical significance for the coefficient of interest when we use the "Value Residual" measure as the left-hand side measure.

3.7.2 Different Bands

Table 3.14 shows that the result is very stable as we move away from the threshold of CLL/0.8. In fact, the point estimates are indistinguishable from each other whether we use a band of USD 5,000 or USD 10,000, which suggests that the difference in the cost of credit is likely to be similar for these two sets of buyers relative to buyers below the threshold. This is further evidence that the result is not driven solely by buyers who choose to obtain a conforming mortgage and put up additional equity from other sources.

153 3.7.3 Timing of the Control Group We run an additional robustness test in which, instead of comparing the year in which the limit is in effect with the subsequent year, we compare it to the previous year. In this way, we are comparing houses that are never eligible for an 80 percent conforming loan (those above the threshold) to transactions that initially are not eligible, but become eligible once the limit changes. The research design is the same as before, but we shift the window of analysis back one year. Table 3.13 shows the Fama-MacBeth coefficients for this specification. The point estimates are smaller than the ones in Table 3.13, but they are in the same direction and remain statistically significant for the first years in the sample.

3.7.4 Pos-October Effect

One concern with our tests is that the conforming loan limit is announced in or around October of each year, which might mean that the anticipation of a raise of the conforming loan limit would confound our results. In order to address this issue, we interact our main effect with the last three months of the year, to see if the coefficients are being driven by this time period. Table 3.15 shows the results for this specification, and we see that the estimates for the effect are the same for the last three months of the year as they are for the first nine. The main effect is almost unchanged.

3.7.5 Value per Square Foot by ZIP Code Income In Figure 3-6, we split ZIP codes by their median income in order to consider the effect of the conforming loan limit on the distribution of value per square foot on the whole sample of transactions. We plot the average value per square foot as a function of the distance of each transaction to the threshold of 125 percent of the CLL. We can see that for the ZIP codes in the lowest quartile of the income distribution, the average value per square foot is monotonically increasing for up to conforming loan limit threshold, and from this point onwards the distribution becomes flat. This pattern is not visible for zip codes with higher median incomes, where the distribution seems monotonically increasing both below and above the threshold.

154 3.8 Appendix B. Data Manipulation

3.8.1 Data Cleaning

In order to clean the raw data received from Dataquick, we perform the following modifications to the data:

Table 3.11: Data Cleaning Description

Criterion Deleted Observations Remaining Observations Initial data 11,884,730 Transaction value equal to zero 1,365,429 10,519,301 Missing zipcode 18,766 10,500,535 Missing square feet 1,509,732 8,990,803 Mislabeled year 5 8,990,798 First loan greater than transaction value 353,552 8,637,246 House of less than 500 square feet 47,059 8,590,187 Transaction greater than 1,2 MM and smaller than 30 M 381,786 8,208,401 Company owned observation based on Dataquick flag 451,295 7,757,106 Company owned obs based on owner/seller/buyer information 746,754 7,010,352 Simple duplicated transactions 0 7,010,352 Value per square feet yearly outliers 142,079 6,868,273 Same property, date and buyer/seller information 11,577 6,856,696 Same property, and (late and no seller information 364 6,856,332 Same property, (late and transaction value 41,855 6,814,477 Same property, date and A sell to B and B sell to C 22,258 6,792,219 Special Transaction, based on Dataquick flag 609 6,791,610 Same property and (late, multiple sales in a day 248 6,791,362 Clean data 6,791,362 Ifemove single-family houses 1,751,670 5,039,692 Transaction greater than 600 M and smaller than 130 M 1,056,117 3,983,575 Whole sample for hedonic regressions 3,983,575 Transactions outside the 10k band for each year 3,742,840 240,735 Transactions used twice ( treatment in year t and control in +21,936 262,671 year t+1 Regression sample 262,671

Note: This table enumerates the steps taken in the data cleaning process and gives the number of observations that are dropped in each step, as well as the remaining observations after each step.

Table 3.11 shows the number of observations deleted in each step of the data preparation and a basic description of the criterion used to drop those observations from the sample. In the following paragraphs, we categorize each step and describe the criteria we used in detail, providing additional information about the data construction. We start with 11,884,730 observations.

Missing observations and outliers

We drop records with missing transaction value, house size, zip code, property unique identifier, or mislabeled year.

155 - We drop a record if the house size is smaller than 500 square feet, as well as records with transaction values smaller than three thousand and greater than one million and two hundred thousand dollars.

- Value per square foot outliers per year: We drop observations that are above the ninety-ninth percentile for the value per square foot variable or below the first percentile each year.

Company owned observations

- We drop observations that Dataquick identifies as being bought by a corpora- tion.

- Company owned observations based on owner/seller/buyer information: If the owner, seller, or buyer names contain LLC, CORP, or LTD, the observation is removed from the sample.

Duplicate transactions

Simple duplicated transactions: Remove records for which all the property information is the same.

Same property, date, and buyer/seller information: Drop observations that are duplicated based on transaction value, date, and buyer/seller information.

Same property and date, no seller information: Drop observations for which the property unique identifier and date are the same and have no seller information.

Same property, date, and transaction value: Drop observations for which property unique identifier, date, and transaction value are the same.

Same property and date and A sells to B and B sells to C: If person A sells to B and B sells to C in the same date, we keep the most recent transaction.

Special transaction, based on Dataquick flag: This flag allows us to identify records that are not actual transactions. For example, if a transaction was only an ownership transfer without a cash transfer, this field is populated, allowing us to delete this transaction.

Same property and date, multiple sales in a day: If a property is sold more than twice during the same day, we keep only one transaction. Additional information

We merge the Metropolitan Statistical Area (MSA) classification obtained from the Census Bureau definition, using FIPS unique code identifier by county16 . 6 1 FIPS county code is a five-digit Federal Information Processing Standard (FIPS) code which uniquely identifies counties and county equivalents in the United States, certain U.S. possessions, and certain freely associated states. The first two digits are the FIPS state code and the last three are the county code within the state or possession.

156 Change the second lien amount to missing if the first loan amount is equal to the second loan amount, or if the second loan amount is greater than the transaction value.

- Change the second lien amount to missing if combined loan to value (CLT) is greater than two and loan to value (LTV) is equal to one.

- Change house age to missing if house age, calculated using transaction year minus year built, is smaller than zero.

This procedure gives us our clean sample with 6,791,362.

Whole Sample for Hedonic Regression Sample

- We further restricted the sample for the hedonic regressions to transactions that are between one hundred and thirty thousand and six hundred thousand dollars. This selection aims to avoid that the estimates from the hedonic regression be driven by transactions that are far from the region of interest.

This gives us our whole sample with 3,983,575 observations that are summarized in the Summary Statistics section of the paper.

Regression Sample

Non-single-family houses: Our identification strategy relies on the change in the conforming loan limit for single-family houses, therefore, we restrict our attention to this type of house.

Transactions outside the USD 10,000 band for each year: Based on the threshold value for each year that we describe in the Identification Strategy subsection, we define a relevant transaction band around that threshold. For example, in 1999 the house threshold (1.25 of the conforming loan limit) is 300,000 dollars. Therefore, we keep records with transaction values between 290,000 and 310,000 dollars that happened between 1999 and 2000. This subsample will be the sample used to run the differences-in-differences specification using the 1999 threshold. For years when transaction bands overlapped, transaction will be treatment in year t and controls in year t+1, and therefore used twice in the empirical strategy

This gives us our regression sample with 262,671 observations

3.8.2 Variable Construction

In this appendix, we describe in more detail the variables used in the hedonic regressions. The hedonic regressions use two left-hand side variables: value per square foot and price of each transaction. As we pointed out when we describe the hedonic

157 regression in the paper (Section 3.2), we use a similar set of controls as those used in Campbell, Giglio, and Pathak (2010), and we add a few more characteristics. The variables we use are interior square feet (linearly, high and low square feet dummies), lot size, bedrooms, bathrooms, total rooms, house age (linearly and squared), type of house, an indicator for whether the house was renovated, an indicator for fireplace and parking, indicators for style of building (architectural style and structural style), and additional indicators for type of construction, exterior material, heating and cooling, heating and cooling mechanism, type of roof, view, attic, basement, and garage. While interior square feet, lot size, and age are included as continuous variables, all the other controls are included as indicator variables.

Type of house: This variable is 1 if the house is a single-family house and 0 if it is a condo or a multifamily property.

Bedrooms: This characteristic is divided into four categories (dummies): one bedroom, two bedrooms, three bedrooms, and more than three bedrooms.

Bathrooms: This characteristic is divided into four categories: one bathroom, one and a half bathrooms, two bathrooms, and more than two bathrooms.

Rooms: This characteristic is divided into five categories (dummies): one room, two rooms, three rooms, four rooms, and more than four rooms.

Building Shape, Architectural Code, Structural Code, Exterior Material, Con- struction Code, Roof Code, View Code: These characteristics were divided based on the numeric categorization of the original field. For example, construction code was divided into 10 different categories that indicated the material used on the framework of the building. In this case, we created 10 dummies based on this categorization.

Heating and cooling: This information was divided into four categories: only heating, only cooling, both heating and cooling, and heating-cooling information missing. The last variable was created to avoid dropping transactions for which the information was not available.

Heating and cooling type: These characteristics were divided based on the numeric categorization of the original field. In this case, they discriminate the type of cooling or heating system that is being used in the house.

158 - Garage and Garage Carport: A dummy is created to account for houses that have garage surface greater than 50 square feet. For those transactions without the information, a missing dummy is created for this category. Finally, we used additional information to create a dummy that indicates if the houses have a garage carport or not.

- Renovation: This variable accounts for the number of years since the last renovation. Based on this continuous variable, five categories (dummies) are defined: missing renovation if the renovation date is missing or renovation period is negative, last renovation in less than 10 years, renovated between 10 and 20 years, renovated between 20 and 30 years , and last renovation in more than or equal to 30 years.

Attic: This characteristic is accounted for using a dummy for houses with an attic greater than 50 square feet, and another dummy to account for missing information about the attic in the houses.

Basement Finished and Unfinished: For the finished basement information, we created a dummy for houses with basement size greater than 100 square feet, and another dummy to account for missing information about the finished basement. The same procedure is used to incorporate the information about unfinished basement.

We use both the price of a transaction as well as the value per square foot as our dependent variables. By estimating these regressions by year and by Metropolitan Statistical Areas (MSA), we allow the coefficients on the characteristics to vary along these two dimensions. We included monthly indicator variables to account for seasonality in the housing market, as well as zip code fixed effects. The set of controls Xi is composed of all the variables described above, but in the case of the value per square foot regression, we exclude the interior square feet continuous variables.

LHS, = -yo + IX, + monthi + zipcodej + Ei When a record is missing the interior square feet, the lot size, the number of bedrooms or bathrooms, or information on a houses age, we do not include this observation in the hedonic regressions. This explains the difference between the number of observations for the value per square foot hedonic regressions (where we exclude interior square footage) and the transaction value residual in our main regression results.

159 Figure 3-5: Fraction of Transactions with a Second Lien Loan by Year

Note: This figure shows the average fraction of transactions with a second lien loan by year for the whole sample and the restricted sample used in the regression. Years 2007 and 2008 are excluded from the regression sample because there was no change on the conforming loan limits on those years.

0 0-

XtD -

U--

0 -'- 1998 1999 2000 2001 2002 2003 2004 2005 2006 whole sample restricted sample

160 Figure 3-6: Value per Square Foot by House Value and by ZIP Code Income

Note: This figure shows the average value per square foot plotted against the value of the house. We split ZIP codes into quartiles according to their median income, where 1 includes the ZIP codes in the lowest income quartile and 4 includes the ZIP codes with the highest median income. We use the average of the median yearly income over the whole sample to place ZIP codes into the quartiles. The x-axis is represented as one minus the transaction value as a percentage of each year's threshold of 125 percent of the conforming loan limit (e.g. if the threshold is 200,000, a transaction of 150,000 will appear as -25 percent). The vertical red line is the threshold and the transactions for all years are centered around that value.

1 2 A- OR A"-- W V - -- W4PW *W -'04

0 LL..0. 700-l 3 4

S -04 0 woprl"- --

-17100 - 0 so 16o -ibo 70/-50 0 !b 10 Transaction Value as Percentage of 1.25CLL

161 Figure 3-7: Income as a Percentage of CLL Threshold

Note: The horizontal axis indicates the difference between loan amounts and the conforming loan limit as a percentage of the conforming loan limit. The figure plots average mortgage applicant income computed from HMDA mortgage applications. We aggregate these proportions into 1% bins and each dot in the figure represents the share of unused mortgages for each bin. We also plot third degree polynomials (to the left and right of the conforming loan limit) as well as 95% confidence intervals (dashed lines). Data extracted from HMDA, 1998-2006.

~0 C

0 6 E ; O

I I 1 i I I r - -50 -45 -40 -35 -30 -25 -20 -15 -10 -5 0 5 10 15 20 25 30 35 40 45 50 Distance from the conforming loan limit (%)

162 Table 3.12: Effect of the CLL on House Valuation Measures, Constrained Sample (0.5

Panel A: Value Per Square Foot

All years 1998-2001 2002-2005 Above Threshold 0.956** 1.584*** 0.328 (0.462) (0.556) (0.650) Year CLL -24.627*** -15.935*** -33.319*** (4.386) (2.576) (5.726) Above Threshold x -1.257*** -1.610** -0.904 Year CLL (0.422) (0.646) (0.576) No. Obs. 190,450 75,304 115,146

Panel B: Log of Transaction Value Residual from Hedonic Regressions

All years 1998-2001 2002-2005 Above Threshold 0.0118*** 0.0145*** 0.0090*** (0.0014) (0.0017) (0.0007) Year CLL 0.0367*** 0.0335*** 0.0398*** (0.0038) (0.0040) (0.0066) Above Threshold x -0.0017 -0.0019 -0.0015* Year CLL (0.0011) (0.0022) (0.0008) No. Obs. 183,643 71,843 111,800

Panel C: Value Per Square Foot Residual from Hedonic Regressions

All years 1998-2001 2002-2005 Above Threshold 1.565*** 1.958*** 1.172*** (0.298) (0.356) (0.431) Year CLL 3.431*** 3.470*** 3.392*** (0.550) (0.417) (1.113) Above Threshold x -0.931*** -1.085*** -0.777** Year CLL (0.260) (0.413) (0.360) No. Obs. 183,789 71,917 111,872

Note: This table shows Fama Macbeth coefficients computed from year by year regressions that use three alternative measures of valuation as the dependent variable in each of the three panels. The hedonic regressions that produce the residuals for panels B and C are described in Section 3.2. The sample for each year's regression includes transactions within +/- USD 10,000 of that year's conforming loan limit, as well as transactions in the same band in the subsequent year. Unlike the main regression table in the paper, the sample for these regressions is constrained to transactions with an LTV between 0.5 and 0.8. All year by year regressions include ZIP code fixed effects. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect.

163 Table 3.13: Effect of CLL on Valuation Measures - Alternative Timing of the Control Group

Panel A: Value Per Square Foot

All Transactions 0.5

Panel B: Transaction Value Residual from Hedonic Regressions

All Transactions 0.5

Panel C: Value Per Square Foot Residual from Hedonic Regressions

All Transactions 0.5

Note: Table shows Fama McBeth coefficients computed from year by year regressions that use three alternative measures of valuation as the dependent variable in each of the three panels. The sample includes all transactions within USD 10,000 of each year's conforming loan limit, as well as transactions of the same amount in the previous year (unlike the previous tables where we use the subsequent year). In this table we include the results for all transactions, as well as those for the sample that is restricted to having an LTV between 0.5 and 0.8. Below the Threshold refers to transactions up to USD 10,000 below the conforming loan limit at year t divided by 0.8 (i.e. the transactions that were "eligible" to be bought with a conforming loan at a full 80 percent LTV in year t , but were "ineligible" in year t-1) and Pre-Year CLL is the previous year in which the conforming loan limit is in effect. This specification makes the interaction coefficient directly comparable to the main regression on signs and magnitudes.

164 Table 3.14: Effect of the CLL on Valuation - Alternative Bands

Panel A: Value Per Square Foot

10K Ok to 5K 5K to 1OK Above Threshold 1.261** 0.969 1.406*** (0.494) (0.722) (0.544) Year CLL -22.869*** -23.008*** -23.194*** (4.047) (3.988) (4.177) Above Threshold x -1.162*** -1.064* -1.181** Year CLL (0.264) (0.556) (0.581) No. Obs. 262,671 134,117 128,554

Panel B: Log of Transaction Value Residual from Hedonic Regressions

10K Ok to 5K 5K to 10K Above Threshold 0.0129** 0.0071 0.0180*** (0.0013) (0.0019) (0.0013) Year CLL 0.0387*** 0.0384*** 0.0389*** (0.0041) (0.0045) (0.0038) Above Threshold x -0.0017*** -0.0015* -0.0023** Year CLL (0.0008) (0.0011) (0.0016) No. Obs. 251,431 128,429 123,002

Panel C: Value Per Square Foot Residual from Hedonic Regressions

10K Ok to 5K 5K to 1OK Above Threshold 1.733*** 1.255* 2.110*** (0.360) (0.700) (0.387) Year CLL 4.103*** 4.052*** 3.946*** (0.644) (0.678) (0.763) Above Threshold x -0.651*** -0.712 -0.623*** Year CLL (0.238) (0.508) (0.238) No. Obs. 251,764 128,601 123,163

165 Table 3.15: Effect of CLL on Valuation: Post October

Panel A: Value Per Square Foot

1998-2005 1998-2005 Above Threshold 1.261** 1.039* 0.000 (0.625) (0.531) Year CLL -22.869*** -23.460*** 0.000 (5.119) (5.079) Above Threshold x -1.162*** -1.086*** Year CLL (0.334) (0.393) Above Threshold x -0.213 Year CLL x Post October (1.031) No. Obs. 262,671 262,671

Panel B: Log of Transaction Value Residual from Hedonic Regressions

1998-2005 1998-2005 Above Threshold 0.0129*** 0.0132*** (0.0016) (0.0014) Year CLL 0.0387*** 0.0398*** (0.0052) (0.0056) Above Threshold x -0.0017* -0.0027** Year CLL (0.0010) (0.0013) Above Threshold x 0.0033 Year CLL x Post October (0.0027) No. Obs. 251,431 251,431

Panel C: Value Per Square Foot Residual from Hedonic Regressions

1998-2005 1998-2005 Above Threshold 1.733*** 1.751*** (0.456) (0.407) Year CLL 4.103*** 4.176*** (0.815) (0.813) Above Threshold x -0.651** -0.696** Year CLL (0.301) (0.277) Above Threshold x 0.031 Year CLL x Post October (0.805) No. Obs. 251,764 251,764

Note: This table shows Fama MacBeth coefficients computed from year by year regressions that use three alternative measures of valuation as the dependent variable in each of the three panels. The sample for each year's regression includes all transactions within +/- USD 10,000 of that year's conforming loan limit, as well as transactions in the same band in the subsequent year. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect. This specification interacts the diff-in-diff specification with a dummy variable that is 1 in October, November and December of each year.

166 Table 3.16: Effect of the CLL on House Valuation with In-Sample Controls

Panel A: Value Per Square Foot

All years 1998-2001 2002-2005 Above Threshold 2.926*** 3.272*** 2.581*** (0.366) (0.416) (0.612) Year CLL -15.158*** -9.681*** -20.634*** (2.706) (1.206) (3.567) Above Threshold x -0.771** -1.211*** -0.332 Year CLL (0.299) (0.428) (0.327) No. Obs. 251,764 103,709 148,055

Panel B: Log of Transaction Value

All years 1998-2001 2002-2005 Above Threshold 0.0281*** 0.0323*** 0.0239*** (0.0018) (0.0011) (0.0011) Year CLL -0.0004*** -0.0005*** -0.0004*** (0.0001) (0.0001) (0.0001) Above Threshold x 0.0000 -0.0001 0.0001 Year CLL (0.0000) (0.0001) (0.0001) No. Obs. 251,431 103,535 147,896

Note: This table shows Fama MacBeth coefficients computed from year by year regressions that use two alternative measures of valuation as the dependent variable in each of the two panels. Instead of using residuals from a hedonic regression, all characteristics of the houses are included as controls within the estimation sample. The sample for each year's regression includes all transactions within +/- USD 10,000 of that year's conforming loan limit, as well as transactions in the same band in the subsequent year. All year by year regressions include ZIP code fixed effects. Above the Threshold refers to transactions up to USD 10,000 above the conforming loan limit divided by 0.8 (i.e. the transactions that were "ineligible" to be bought with a conforming loan at a full 80 percent LTV) and Year CLL is the year in which the conforming loan limit is in effect.

167