Consumption Smoothing and Debtor Protections

Transcription

1 Consumption Smoothing and Debtor Protections Nathaniel Pattison 1 May 30, 2018 Abstract Restricting debt collection provides substantial consumption insurance to households. This paper evaluates the social insurance created by laws that protect households assets, both inside and outside of bankruptcy. First, I show that households are not fully insured; consumption declines by 3-5% upon non-bankruptcy default. Second, I use changes in states exemptions to estimate the consumption benefits and costs of additional asset protection. While higher exemptions do provide valuable insurance, the interest rate cost is large. Adopting a sufficient statistics formula from the literature, the estimates imply that the cost of additional exemption protection far exceeds what debtors are willing to pay. I then show that this large cost can be fully explained by the distortion to default decisions. JEL classification: D14, K35, G18 Keywords: social insurance; bankruptcy; consumer credit; consumption smoothing; sufficient statistic 1 Assistant Professor, Department of Economics, Southern Methodist University. npattison@smu.edu. See for the latest version. I thank Leora Friedberg, Richard Hynes, William Johnson, Daniel Millimet, Saltuk Ozerturk, John Pepper, Sarah Turner, and seminar participants at UVA, the Richmond Federal Reserve, S.S. Huebner Foundation Doctoral Colloquium, FDIC, CFPB, Bank of Canada, Clemson, SMU, and CUNY Baruch for helpful discussion and comments. I am grateful for financial support from the Bankard Fund for Political Economy and the Steer Family Endowed Fund.

2 1 Introduction Allowing borrowers to default on debt can be viewed as one of the largest social insurance programs in the United States. Around 10% of households have filed for bankruptcy and a greater percentage default on debt outside of bankruptcy. 1 Forgoing payment on debt frees financial resources, which can then be used to supplement existing social and private insurance or help smooth consumption over shocks that are not otherwise insured. The generosity of this default insurance is determined by the combination of state and federal laws that govern the collection of delinquent debts. A key law these households is the asset exemption, which protects specific property (e.g., home and vehicle equity) from seizure by unsecured creditors. The amount protected varies across states from less than $10,000 to more than $500,000. Higher exemptions help households smooth consumption in default, but lenders compensate for additional losses by reducing credit supply. While there exist several estimates of the costs of raising exemptions, 2 namely more costly credit, there is little empirical evidence of the consumption smoothing benefits of exemptions and their role as consumption insurance. 3 This paper helps to fill this gap by examining the consumption smoothing role of asset exemptions for households who default on consumer debts. There are three main contributions. First, using two separate empirical strategies, I provide estimates of the key determinants of the value of exemption protection: (i) the 1 Stavins (2000) reports that 8.5% of households have filed for bankruptcy, and more recently, Dobbie et al. (2016) reports that 15% of individuals have filed for bankruptcy based on their calculations in the Federal Reserve Bank of New York s Consumer Credit Panel/Equifax Data. Default without bankruptcy is more common. VISA reports that 55-60% of charge-offs occur without a bankruptcy filing (NBRC, 1997). 2 Papers estimating the effect of exemptions on interest rates include Gropp, Scholz and White (1997), Berkowitz and Hynes (1999), Berkowitz and White (2004), Berger, Cerqueiro and Penas (2011), and Severino and Brown (2017). 3 One exception is that Mahoney (2015) demonstrates that exemptions play a substantial role as an alternative form of health insurance. Also, Lehnert and Maki (2007) finds that exemptions slightly reduce state-level consumption volatility, but is unable to estimate the impact on specific families with delinquent debt. Focusing on families that have filed for bankruptcy, Filer and Fisher (2005) finds no significant impact of exemptions on consumption (Table 3 Column 2), though the small sample (131 bankruptcy filers in the PSID) limits the precision. 1

3 change in household consumption that occurs upon default and (ii) the effect exemptions on consumption in default. 4 Second, adapting the Baily-Chetty sufficient statistic formula of Dávila (2016) to incorporate these estimates, I evaluate the implications for welfare and optimal exemption policy. This normative analysis reveals that the interest rate cost of higher exemptions far exceeds what households are willing to pay. The policy implication is that lower exemptions would increase welfare. Finally, I provide an explanation for the high cost of exemption-generated insurance. I show that higher exemptions increase the default rate and, within a competitive model of lending, this default response explains 90% of the observed interest rate markup. The empirical strategy consists of two parts. First, I use the Panel Study of Income Dynamics to investigate the value of additional consumption smoothing for households that default on debt. The direction and magnitude of consumption changes upon default determine the value of consumption insurance (Baily, 1978, Gruber, 1997, Chetty, 2006). If households default to offset adverse shocks that cause declines in consumption, additional consumption insurance is valuable. 5 If, alternatively, households strategically default while consumption remains stable (Fay, Hurst and White, 2002), additional insurance has little value. Current evidence suggests additional protection for bankruptcy filers provides little value; Filer and Fisher (2005) find that, at the current level of protection, consumption increases by 8-13% during the year of a bankruptcy filing. Like Filer and Fisher (2005), I also examine consumption patterns around instances of default in the PSID, but I broaden the definition of default to include informal default, namely the occurrence of missed bill payments, debt collection calls, or judicial collections actions (e.g., garnishment or repossessions). This distinction is critical, as the results now suggest a clear consumption smoothing role for additional debtor protections, but only for default that occurs without a formal 4 I focus exclusively on consumer default. The relevant trade-offs for business bankruptcies and default are different, and exemptions for business debt can be viewed as a separate policy parameter. 5 There is evidence that individuals use default and bankruptcy to offset income and expense shocks such as unemployment (Keys, forthcoming), divorce (Lyons and Fisher, 2006), and health shocks (Himmelstein et al., 2005). 2

4 bankruptcy. Consumption falls by 3-5% upon default, and this is driven by default that occurs outside of the bankruptcy system. 6 Heterogeneity in the consumption drop suggests that, while targeted debtor protections can provide valuable insurance, additional exemption protection may have little value. The largest exemptions are for home equity, but the borrowers that benefit most from increasing these exemptions (homeowners and those with non-exempt equity) experience little to no change in consumption upon default. Having demonstrated a role for consumption insurance in default, the second part of the empirical strategy focuses on the benefits and costs of using asset exemptions to provide this insurance. Exemptions increase consumption by reducing the amount repaid in default. I directly estimate the effect of higher exemptions on the recovery rates of defaulted consumer debt, i.e., the amount that is eventually repaid by households after they default. Using data from Credit Union Call Reports for , I estimate difference-in-differences regressions using 57 within-state changes in exemption levels. 7 A 10% increase in exemptions reduces recovery rates on consumer debt by 32 basis points. Estimates from event study regressions show that the recovery rates in treatment and control states followed parallel trends and then diverged after an exemption increase. Using the same strategy, I find that interest rates on credit card and auto loans increase as exemptions rise, corroborating a set of papers finding higher interest rates in states with greater exemptions (Gropp, Scholz and White, 1997, Berkowitz and White, 2004, Berger, Cerqueiro and Penas, 2011, Severino and Brown, 2017). 8 Together, the estimates of the consumption smoothing benefit and interest rate cost of 6 The drop in consumption upon default and increase in consumption upon bankruptcy is also consistent with legal research that documents years of informal debt collection and financial struggle prior to most bankruptcy filings (Mann, 2007, Mann and Porter, 2010), as well as recent empirical research showing that access to bankruptcy alleviates financial distress (Dobbie and Song, 2015, Dobbie, Goldsmith- Pinkham and Yang, 2017). 7 Fedaseyeu (2015), who first used recovery data from credit unions in studying a different set of debt collection regulations, notes that unlike large commercial banks, credit union credit unions are often local lenders, so their financial information reflects state laws. An updated version of Fedaseyeu (2015) includes exemptions as a control in a robustness check, and the coefficient is consistent with the results of this paper. 8 The interest rate effect in this paper is very similar to that of Severino and Brown (2017) for unsecured personal loans, which is the only paper also using a difference-in-difference strategy. 3

5 increasing exemptions determine the price or markup of the exemption-generated default insurance, a critical determinant of optimal exemption policy. These estimates indicate that this insurance is expensive. For exemptions to increase expected consumption in default by $1, debtors must pay around $5 in higher interest rates. I then weigh the costs and benefits of exemption protection within the recent sufficient statistics model of Dávila (2016). Dávila (2016) applies the Baily-Chetty sufficient statistics approach to asset exemptions, focusing on their impact within formal bankruptcy, and derives a sufficient statistic for the optimal level of exemptions. Using illustrative benchmark values, Dávila (2016) finds that current exemptions are close to their optimal level, but also acknowledges the need for additional empirical work to generate improved estimates of key parameters. A primary contribution of this paper is to provide these estimates. My estimates indicate that exemptions are wellabove the optimal level. Strikingly, I find that while debtors are only willing to pay 5-25% over the actuarially fair rate for additional default insurance, exemptions generate insurance that is marked up 400%. Given the central role that the large markup of exemption-generated default insurance plays in the welfare analysis, I develop a second method to estimate the markup. Using the difference-in-difference strategy, I show that higher exemptions cause debtors to default more frequently. 9 Then, assuming a zero-profit condition on lenders, I infer the magnitude of the interest rate increase that is implied by the default response. This method is analogous to the sufficient statistic for optimal unemployment insurance, which infers the tax rate response of UI benefit increases from the elasticity of unemployment duration and the government s balanced budget constraint. The default response alone, when combined with the zero-profit model of lending, can explain 90% of the observed interest rate markup. This paper adds to the vast empirical literature examining the consumption smoothing benefits of the social safety net. The estimates of this paper suggest that additional debtor 9 I also show that bankruptcy filings rise when exemptions increase. Severino and Brown (2017), using credit report data, finds no overall change in default rates but some evidence of increases in delinquency among low-asset individuals. 4

6 protections, when applied broadly to all individuals with delinquent debt, generate less valuable insurance than other programs. A set of papers has estimated consumption drops upon job loss (7-10%), illness (11-14%), and the development of a disability with no workers compensation (30%). 10 Compared to these estimates, the average consumption drop upon default of 3-5% is small, suggesting that the shocks during times of default are, on average, less severe or more easily insured. Importantly, however, this consumption decline upon nonbankruptcy default differs sharply with the consumption increase of 8-13% upon bankruptcy found in Filer and Fisher (2005). This difference suggests that marginal increases in debtor protections are more valuable when the protection is offered outside of bankruptcy. Additionally, the estimates and approach of this paper complement the literature evaluating the impact within structural and macroeconomic models. Livshits (2015) provides a recent review and discussion of the dispersion of findings in the literature evaluating the welfare impact of default and bankruptcy policy, and several models focus on asset exemptions in particular (Athreya, 2006, Li and Sarte, 2006, Pavan, 2008, Mankart, 2014, Mitman, 2016, Hintermaier and Koeniger, 2016). The focus on exemptions also complements papers examining the benefits of access to bankruptcy more generally (Dobbie and Song, 2015, Dobbie, Goldsmith-Pinkham and Yang, 2017) and other debtor protection laws (Fedaseyeu, 2015). Another related subset of the literature studies the impact of exemptions on entrepreneurship (Fan and White, 2003, Cerqueiro and Penas, 2016), though the policy trade-offs differ from the consumer case examined in this paper. One omission from the welfare analysis, in this paper and the existing literature evaluating the overall welfare impact of exemptions, is the role of exemptions as health insurance. Mahoney (2015) shows that exemption protection plays an important role as a form of health insurance and that the protection also distorts health insurance decisions. A full welfare analysis would account for the costs and benefits of exemptions allowing default on medical debt, but quantifying the incidence and cost of 10 Gruber (1997), East and Kuka (2015), and Kroft and Notowidigdo (2016) estimate the consumption drop upon unemployment, Cochrane (1991) estimates consumption changes upon several shocks, including illness, and Bronchetti (2012) estimates the drop upon work-limiting disabilities. 5

7 unpaid medical debt is challenging (Finkelstein, Mahoney and Notowidigdo, 2017). The paper proceeds as follows. Section 2 describes the role that exemptions play inside and outside of bankruptcy. Section 3 estimates the change in consumption that occurs upon default. Section 4 estimates the causal effect of changing exemptions on the recovery rate and interest rate. Finally, Section 5 calculates the welfare effect using these estimates and develops an alternative method for calculating the interest rate markup. Section 6 concludes by discussing limitations of the welfare analysis. 2 Institutional Background When debtors default, exemption laws protect specific assets from seizure by unsecured creditors. While federal exemptions are available, the large majority of states have opted out or set their own exemption laws alongside the federal exemptions. This generates substantial variation across states in the amount protected, shown in Figure 1. For example, for an unmarried debtor, Virginia exempts $5,000 in home equity and $6,000 in vehicle equity, while Texas exempts an unlimited amount of home equity and one vehicle per adult. These state exemptions can influence the debt collection process in both the formal bankruptcy system and outside of bankruptcy in the state courts. Inside of bankruptcy, almost all consumers file under Chapter 7 or Chapter In Chapter 7, which accounts for 70% of consumer bankruptcies, exemptions determine the debtor s non-exempt assets, which the court sells then transfers the proceeds to creditors. In Chapter 13, exemptions apply indirectly, since creditors must receive at least as much as they would have under a Chapter 7 liquidation. 12 Outside of bankruptcy, exemptions affect debt collection in two primary ways. First, the 11 Individuals are sometimes required to file under Chapter 11 if they fail the Chapter 7 means test and have debts that exceed Chapter 13 s debt limits, and some individuals choose Chapter 11 even though the other chapters are available. However, individual Chapter 11 bankruptcies account for only 0.15% of individual bankruptcy filings (2016). 12 In both chapters, secured creditors are first paid the collateral value, followed by other priority debts (e.g., domestic support obligations or taxes), so exemptions may have the largest impact general unsecured credit such as credit cards or unsecured loans. 6

8 majority of exemptions still protect debtors assets from the collection efforts of unsecured creditors in state court (Hynes and Posner, 2002, Gilles, 2006, Hynes, 2008, Dawsey, Hynes and Ausubel, 2013). If an unsecured creditor sues in state court, he can obtain a judgment allowing additional collection actions, including the right to seize non-exempt assets as payment. Second, asset exemptions determine the debtor s potential cost of filing for bankruptcy, which can influence informal negotiations between debtors and creditors. Thus, exemptions can affect debt collection directly through asset seizure, or with asset seizure operating indirectly as a threat. The main impact of exemptions may be through its role as a threat or determinant of bargaining power. 13 Assets are rarely seized in either bankruptcy (Flynn, Bermant and Hazard, 2003, Jiménez, 2009) or through the state courts (Hynes, 2008). There is, however, anecdotal and empirical evidence that exemptions influence negotiations between debtors and creditors outside of bankruptcy. A consumer guide advises delinquent debtors that when settling, the amount you offer to pay should be directly related to what the collector could seize... (NCLC, 2016). Similarly, creditors are more likely to accept partial payment if the debtor has few seizable assets (Finlay, 2010). Mahoney (2015) provides empirical support for the importance of these laws in the negotiation process of medical debt, showing that uninsured individuals with fewer seizable assets repay less of the debt. In Section 4, I provide direct evidence that increasing exemption levels reduces the amount that banks recover on delinquent consumer debt. These settlements, asset seizures (or the threat of seizure), and other collection efforts recover a nontrivial share of defaulted debt, particularly when done outside of the formal bankruptcy system. In 2013, the bankruptcy courts collected $3.2 billion from Chapter 7 asset cases, while third-party debt collection agencies alone, which excludes in-house collection, recovered over $55 billion (United States Trustees Program Annual Report, FY 2013 and (Ernst and Young, 2012)). 13 There is an analogy with criminal law, where the rules of criminal procedure influence plea bargains, although criminal trials are rare. 7

9 3 Changes in Consumption upon Default The change in consumption that occurs when a household defaults determines the value of default insurance. If individuals default in response to adverse events, such as job loss, divorce, or health shocks, then consumption may decline and debtor protections can provide valuable consumption insurance. Alternatively, if default is mostly strategic, rather than due to adverse events, consumption can be stable during times of default and there is no need for insurance. 14 This section uses the Panel Study of Income Dynamics (PSID) to examine consumption patterns around instances of default, including both formal bankruptcy and default outside of the bankruptcy system. This strategy follows a large literature using consumption changes to determine the value of additional social insurance (Cochrane, 1991, Gruber, 1997). 3.1 Data: Panel Study of Income Dynamics The PSID from contains information about instances of default and a measure of consumption. In 1996, the PSID asked families about financial distress that occurred between 1991 and Each family reports the year that they missed a bill payment, had a debt collector call, dealt with judicial collection actions (repossession, garnishment, lien), or filed for bankruptcy. 15 In the main analysis, I count the occurrence of any of these events as default, though in Appendix Table A1 I show similar estimates with a stricter definition of default. Since the goal is to estimate the change in consumption upon default, the unit of observation is an instance of default. The main defaulter sample consists of household heads that report an instance of default in some year t but not in year t 1. The same head can enter the sample multiple times as long as the instances of default are not in consecutive 14 A set of papers examine the degree to which bankruptcies are driven by adverse events and find evidence of both strategic and non-strategic bankruptcies (Fay, Hurst and White, 2002, Zhang, Sabarwal and Gan, 2015, Keys, forthcoming). 15 I use the PSID because these are the only years which include information about financial distress and an annual consumption measure. 8

10 years. 16 The measure of consumption available in the PSID during this period is each family s annual food expenditure. While focusing on food consumption seems limiting, Chetty (2006) shows that as long as agents make optimal consumption choices, the change in a single good is sufficient to calculate the value of additional insurance. Following Gruber (1997), I measure consumption as the sum of at-home food expenditure (including food stamps) and out-ofhome food expenditure, deflated by the corresponding component of the CPI for the month of the interview. I exclude households with imputed food consumption and households that report a change in food consumption over 300%. 17 I combine these data from the PSID with states asset exemption levels collected from historical state statutes and various editions of Elias, Renauer and Leonard ( ), a popular consumer bankruptcy guidebook. 18 For each individual, I sum the homestead and personal property exemptions available in the state during the year of default. If the household head is married and lives in a state that allows married couples to double their exemptions, I double the exemption value. I group individuals into exemption terciles based on the total amount of homestead and personal property exemptions available to them. 19 I refer to these groups as individuals living in low-exemption, mid-exemption, and highexemption states, although the exemption tercile is a function of both state and marital status. 16 I drop households that report default in 1991 since I do not know whether they defaulted in Following Zeldes (1989) and Gruber (1997), I drop households where log(c t /c t 1 ) > 1.1 or < 1.1 (4% of the sample). Including these households does not affect the results. 18 I collected data on homestead and property exemptions from editions of Elias, Renauer and Leonard ( ) and corrected the timing of the changes by referencing historical state statutes. I thank Jeffrey Traczynski for generously sharing data on exemptions from Traczynski (2011) for comparison. For states that allow individuals to choose between the state homestead exemption and the federal homestead exemption (available only in bankruptcy), I use whichever is higher. I code states with unlimited homestead exemptions as $550,000, the maximum exemption level among states without unlimited exemptions during the period I ignore lot size restrictions. I assume the filer is not a senior citizen. For personal property exemptions, I sum the wildcard, cash, and vehicle exemptions. 19 Low-exemption states have total exemptions less than $14,990, mid-exemption states range from $14,990-52,100, and high exemption states have total exemptions above $52,100 (including the unlimited exemption states). To allow for robustness checks on the sample of non-defaulters, I determine the tercile thresholds using the full PSID sample. 9

11 The first two columns of Table 1 report descriptive statistics for two groups: PSID respondents who never report an instance of financial distress (non-defaulters), and the analysis sample, which consists of the 1,144 instances of default (defaulters). The first row shows the average change in log food consumption. In the sample of non-defaulters, the average change in food consumption is 0.04%. In the defaulter sample, however, food consumption drops by 3.5%. One concern is that this drop may be due to shocks that change the food requirements of the family and are correlated with default, such as divorce or death of a spouse. The second row rules this out by showing that food needs, the PSID s measure of the household s food requirements based on household size and composition, does not change among defaulters. The third column shows that consumption increases by 5.7% during the year of a bankruptcy filing, consistent with the 8% increase found by Filer and Fisher (2005) using a slightly different sample. The remaining rows of Table 1 show that the sample of defaulters tends to be younger and are more likely to be female, non-white, unmarried and have more unsecured debt than non-defaulters. 3.2 Empirical Strategy The sample statistics in Table 1 show that borrowers experience an average drop in consumption of 3.5% upon default. The purpose of the additional analysis in this section is to investigate how the consumption drop varies with the exemption level and across individuals more or less likely to benefit from exemption protection. The sample consists of 1,144 instances of default, indexed by i. I estimate regressions of the form: logc i = α L exempt L i + α M exempt M i + α H exempt H i + δx i + ε i. (1) logc is the change in the log of consumption and exempt L, exempt M and exempt H are indicators for whether an individual is protected by low-, middle-, or high-exemption levels. X is a set of individual characteristics (discussed below) and ε is the error term. The X 10

12 variables, which include year fixed effects and other controls, are de-meaned, so the α coefficients capture the average drop in consumption upon default for borrowers in low-, middle, or high-exemption states. Since there is no constant, the significance of the α coefficient tests whether the average consumption change in low-, middle-, or high-exemption states is statistically different from zero. 20 Thus, if the average consumption changes (α L, α M, α H ) are negative and statistically significant, it indicates that consumption declines during times of default. The α coefficients do not capture the causal effect of exemptions on consumption smoothing. Instead, they capture the combined effect of exemptions and any other factors correlated with exemptions (e.g., social or private insurance) that affect changes in consumption upon default. It is this combined effect, not a causal effect, that is needed to determine the value of additional default insurance, as it reflects the extent to which borrowers remain imperfectly insured given existing methods of smoothing consumption in default. 21 I also investigate the role of observable differences by including controls for individual characteristics in some regressions. The individual characteristics X i include the year fixed effects, age, sex, years of education, an indicator for white, marital status, number of children, and the change in the log of the food needs of the family. In some specifications, I also add controls for the state unemployment rate and the log of state median income at the time of default. All results report standard errors clustered by state. 3.3 Results Table 2 reports the results from estimating specification (1) on the default sample. The key coefficients are α L, α M, and α H, which capture the average (log) consumption change upon default in low-, mid-, and high-exemption states. Column 1 includes only the exemption tercile indicators and year fixed effects. The estimate of α L indicates that, in the low- 20 Appendix Table A2 reports estimates from an alternate specification logc i = α + α M exempt M i + α H exempt H i + δx i + ε i. The α M and α H coefficients now test whether the drop in consumption among high- and mid-exemption states is statistically different from that in low-exemptions states (α). 21 Kroft and Notowidigdo (2016) highlights the same point for the case of unemployment benefits. 11

13 exemption states, consumption drops by an average of 5.1% and this drop is statistically different from zero. For comparison, the mean drop in consumption upon unemployment is 7-10% (Gruber, 1997, Kroft and Notowidigdo, 2016). The α M and α H show that the average drop in consumption is 3.4% in mid-exemption states and 1.9% in high-exemption states, though neither drop is statistically different from zero (nor from the estimate of α L ). That consumption drops upon default (at least in low-exemption states) demonstrates that some borrowers are not fully smoothing consumption over the shocks that cause default, so additional exemption insurance may be valuable. Again, the differences in the α estimates capture the combined effect of exemptions, compositional differences, and other factors correlated with exemptions. The remaining columns investigate the role of compositional differences and economic conditions in explaining these estimates. Column 2 adds controls for individual characteristics and column 3 adds controls for state-level economic conditions during the year that the default occurred. If differences in the observed characteristics of borrowers or economic conditions drive the results, the estimates of the average consumption drop across exemption terciles would converge after controlling for individual and economic characteristics. Instead, columns 2 and 3 show little change. The estimates in column 3 indicate that the average consumption drop is 5.0% in low-exemption states, 4.0% in mid-exemption states, and 1.5% in high-exemption states. Exemptions, which protect only certain types and amounts of assets, will not benefit all debtors equally. Figure 2 reports the average consumption drop for various subsets of defaulters. The first estimate shows the average drop in consumption among all debtors. The next two estimates show that the consumption drop is larger for renters (4.9%) than homeowners (1.7%) and those with non-exempt home equity (1.4%). If raising exemptions mostly protects homeowners and those with non-exempt equity, the small average declines in consumption among these groups would imply a low value for additional insurance. This is even more true for bankruptcy filers, who experience sizable increases in consumption 12

14 upon default (Filer and Fisher, 2005). 22 Overall, the figure shows that many of the groups likely to benefit from exemption increases - homeowners, those with non-exempt equity, and bankruptcy filers - experience smaller declines in consumption upon default, reflecting that they are already able to smooth consumption. Figure 3 provides additional evidence that these estimates capture declines in consumption associated with default by comparing the consumption changes upon defaulters to nondefaulters in the same states. The consumption changes among the group of non-defaulters are close to zero. Appendix Table A4 tests this more formally and finds a significant and similarly sized consumption change among defaulters relative when compared to non-defaulters in the same state and year. Additionally, borrowers may anticipate default and reduce consumption in advance. For example, Hendren (2017) shows that households reduce consumption in the year before a job loss. Appendix Table A5 shows that the largest changes in consumption occur in the year of default, with smaller changes in consumption (1.1%) in the year before default. These timing results also provide reassurance about concerns related to the ambiguity about the reference year of the food expenditure questions in the PSID (East and Kuka, 2015) or the fact that default dates were assigned retroactively from the 1996 survey. 23 Overall, this section shows a 3-5% decline in consumption that occurs upon default and is driven by non-bankruptcy default, particularly in low-exemption states. 4 The Consumption Smoothing Effects of Exemptions The previous section finds a drop in consumption upon default, implying a potential consumption insurance benefit. A complete analysis of exemptions must compare this benefit 22 In Appendix Table A3, I construct the PSID sample following Filer and Fisher (2005) and show that the difference between the consumption drop upon default and consumption increase upon bankruptcy is statistically significant and robust to using non-defaulters as a comparison group. 23 In most years, the PSID asks about food consumption in the average week, and the question is asked immediately after a question about food stamp use in the prior month. For this reason, I follow prior research in assuming that individuals report their consumption during the year of the interview (Zeldes, 1989, Gruber, 1997, East and Kuka, 2015). Also, the data on default was asked in 1996 and only contains the calendar year that default occurred, which will not correspond precisely to the timing of the PSID food expenditure questions in those years. 13

15 to the cost of the insurance generated by asset exemptions. While higher exemptions allow debtors to repay less of their debt (the insurance payout ), lenders compensate by raising interest rates (the insurance premium ). The cost of the insurance depends on the ratio of premiums to payouts. In this section, I use difference-in-differences and event study regressions to estimate the causal effect of exemptions on repayment rates in default and on consumer debt interest rates. Together, these estimates determine the cost of exemptiongenerated default insurance. 4.1 Data: Credit Union Call Reports I use state-level and individual credit union data on recovery rates and interest rates from Credit Union Call Reports. Each quarter, credit unions must submit a Call Report with financial information such as balance sheets and income statements. State-level aggregates from Credit Union Call Reports were first used to study collection in Fedaseyeu (2015), which examined how regulations on debt collectors, such as licensing requirements or criminal penalties, affect credit markets. I use similar state-level aggregates, but also use data at the level of individual credit unions. One advantage of using credit union data, as argued in Fedaseyeu (2015), is that credit unions are local lenders, so their lending practices reflect state laws. Over 92% of credit unions, as of 2013, had branches in only one state and over 98% had branches in two or fewer states. 24 A drawback, however, is that the lending practices of credit unions may differ from those of larger banks. I discuss and provide evidence for the external validity of the credit union estimates later in this Section. I use credit union data for the years I limit the sample to this period because two shocks, a major bankruptcy reform and a severe recession, likely weaken the impact of exemptions in the late 2000s and make it difficult to detect the impact of exemptions with a difference-in-difference strategy. 25 In particular, the bankruptcy reform generated large, 24 In the main analysis, I drop two major national credit unions, Navy Federal Credit Union and the Pentagon Federal Credit Union, from the sample. In Appendix Table B1, I show that the main results hold in the subsample of credit unions operating in only one state. 25 Appendix Table B2 compares the estimates from the period to the post-2007 estimates, 14

16 temporary spikes in bankruptcy filings and these changes correlate with asset exemption levels (Ashcraft, Dick and Morgan, 2007). Soon after, falling home prices erased a substantial amount of home equity, reducing the benefit of the largest exemption, the home equity exemption. In 2010, 55-65% of homeowners were completely protected by exemptions (Dobbie and Goldsmith-Pinkham, 2015). Finally, the reduction in credit supply and increase in unemployment during the Great Recession may weaken or alter the impact of exemptions and the pool of borrowers considering default. I use 4th-quarter Call Reports to construct annual recovery rates on charged-off non-real estate loans. A charge-off occurs when a creditor marks a debt as unlikely to be collected, typically after days of delinquency for consumer debts. 26 Recoveries reflect the amount collected after a debt has been charged-off, and can consist of post-charge-off payments by debtors or revenues from selling the charged-off debt (Furletti, 2003). Therefore, recoveries capture the amount that creditors ultimately collect on debt that is severely delinquent, including collections in and out of bankruptcy. Credit unions report total charge-offs and recoveries and real-estate charge-offs and recoveries separately. Exemptions matter most for unsecured credit (since they do not prevent the recovery of collateral), so I construct chargeoffs and recoveries of non-real estate loans for each credit union. 27 I use the individual credit union data as well as state-level aggregates. To form the state-level measure, I aggregate which are smaller and statistically insignificant. Similarly, Severino and Brown (2017) finds no effect of exemptions on interest rates in the post-bapcpa period. 26 Bank (FFIEC) regulatory accounting requirements state that revolving credit must be charged-off after 180 days of delinquency and installment loans after 120 days - Uniform Retail Credit Classification and Account Management Policy, 65 Fed. Reg (June 12, 2000). When loans are charged-off, issuers reverse the fees and finance charges on the loan in a process called purification (Furletti, 2003). Therefore, the charged-off amounts will reflect the unpaid principal (see NCUA 5300 CALL REPORT INSTRUCTIONS - June 2005). 27 These charge-offs are primarily unsecured consumer loans (e.g., credit cards) and the underwater portion of vehicle loans. Estimates based on the share of unsecured debt and non-real estate debt that is chargedoff suggest that at most 44% is auto loans. These numbers are obtained by multiplying the shares of unsecured (22%) and auto (78%) loans by their respective charge-off rates of 2.18% (for credit cards) and 0.56%. Since credit unions have the option of only charging off the difference between a loan and its collateral, auto loan charge-offs are likely a smaller share. Unsecured loans also include overdraft advances, but these likely make up a small share. As of 2005, total unfunded overdraft commitments made up approximately 8% of total unfunded commitments for unsecured loans (credit cards and personal loans). 15

17 non-real estate charge-offs and recoveries by state and measure the recovery rate in state j as aggregate recoveries divided by aggregate charge-offs. Credit Union Call Reports also include data on credit card interest rates. Each credit union reports the most common interest rate offered for credit cards and the total number of credit card loans. I also aggregate these interest rates to the state-level, weighting each credit union s interest rate by the number of outstanding credit card loans. Table 3 presents the descriptive statistics. The mean interest rate on credit card debt is 12.30% and the average recovery rate on charged-off non-real estate debt is 17.73%. 28 I combine the interest rates and recovery rates with data on exemption levels. To calculate the exemption level, I sum the homestead and property exemptions available to an unmarried bankruptcy filer under the age of 65 for each state and year, where personal property as defined in Section 3. Although some states allow doubling of exemptions for a married debtor, my specification uses the log of the exemption level and so the coefficient would not be affected by doubling. Between 1994 and 2004, there were 57 changes among 28 states, and the average (credit union membership-weighted) change is $10,507 or 17.5 log points. Appendix Figure B1 shows the number of exemptions changes in each state between 1994 and 2004, and Figure B2 shows the distribution and timing of changes in the exemption level. 4.2 Empirical Strategy I use the state-level data to estimate the effect of exemptions on recovery rates and interest rates. For state s at time t, the regressions are of the following form: y st = α + ηln(e st ) + X st β + δ s + τ t + u st. (2) 28 The recovery rate is similar to that of Visa, which reports that the average recovery rate on debt charged-off without a bankruptcy is 18% and 3% when a bankruptcy is filed (NBRC, 1997). 16

18 where ln(e st ) is the log of the exemption level. 29 Using the log of a state s exemption level allows the effect of exemptions to diminish as the exemption level rises, reflecting the fact that more debtors will be fully protected. 30 For example, an increase in Virginia s $5,000 homestead exemption would affect everyone with more than $5,000 in home equity, while an increase in Minnesota s $390,000 homestead exemption would affect the few with more than $390,000 in equity. Consistent with this, in Appendix Tables B3 and B4, I show that the estimated effect of a $1 increase in exemptions is much larger in states that have lower exemptions, and close to zero for states with high exemptions. The outcome variable y st is either the interest rate or the recovery rate in state s during year t. The coefficient η captures the effect of a 100 log point increase in a state s exemption level. The state controls, X st, contain the log of median income, the log of the home price index from the Federal Housing Finance Agency, and the state unemployment rate. I also include state fixed effects (δ s ) and year fixed effects (τ t ) in all specifications. The error term, u st, represents the unobserved state-year shocks that affect interest or recovery rates. Unlike the analysis in Section 3, I argue that these estimates reflect the causal effect of exemptions. The identifying assumption is the parallel trends assumption: in the absence of an exemption increase, interest rates and recovery rates in states that increase exemptions and in control states would have been parallel. I support this assumption in two ways. First, I argue that the changes in exemptions arise out of a political process that does not depend on states lending conditions. Several states exemptions and the federal bankruptcy exemptions are adjusted at predetermined intervals to adjust for inflation. Additionally, Severino and Brown (2017) examines a number of potential predictors of exemption changes, including house prices, state GDP, medical expenditures, the unemployment rate, the political climate, bankruptcy filings, and income growth. Only medical expenditure is found to be statistically 29 I focus on the effect of exemption increases because the exemption level is the parameter set by the policymaker. An alternative approach would be to calculate the financial benefit or cost of defaulting, which depends on households debts, assets, and the exemption level (Fay, Hurst and White, 2002, Mahoney, 2015). 30 Mankart (2014) highlights the point that exemption matter much more at lower levels, and their effect fades out quickly as exemption levels increase. 17

19 significant. Other important debtor protection laws, namely wage garnishment restrictions and statutes of limitations on debt, were stable over this period. Second, using an event study specification, I test whether trends in treatment and control states were parallel prior to an exemption increase. Since states have multiple exemption increases, a standard event study specification is not appropriate. Instead, I use a multiple event study framework, similar to those in Dube, Lester and Reich (2010) and Sandler and Sandler (2014), that allows for overlapping events within a state. I estimate the following regression for state s in year t: 31 5 y st = α + η k ln(e s,t k ) + X st β + δ s + τ t + u st. (3) k= 5,k 1 The one-period difference operator,, produces coefficients η k that represent the cumulative effect of a one log-point increase in the exemption level k years later. For example, if the log exemption increased by 0.5 in a state during 2000, η 3 captures the difference in bankruptcies in that state in 1997 ( ln(e s,t+3 ) = 0.5 when t = 1997), while η 3 captures the effect of that increase on state bankruptcies in The specification omits one-year lead term, so the coefficients capture differences in the outcome relative to differences that exist one year before an exemption increase. 32 The estimates provide evidence about the identification assumption by testing whether the trends were parallel prior to an exemption increase. If interest rates y st in the treatment and control states are similar prior to an exemption increase, then the coefficients η 5,..., η 1 will be close to zero. The coefficients η 0 through η 5 capture the cumulative effect of an exemption increase in the years after the increase. 31 To produce a balanced panel in this regression, I use exemption data from even though y st is only used from The 5-year lead difference is inclusive, in that it captures the difference in log(exemption) between 2015 and year t + 5. Similarly, the 5-year lag difference captures the difference in log(exemption) between year t 5 and If this were not done, the reference group would be both 1-year before an exemption increase and any year outside of the event window (Sandler and Sandler, 2014). 18

20 4.3 Results Table 4 reports the estimates from the difference-in-differences equation (2). The estimate in column 1 indicates that a 10% increase in the exemption level reduces recovery rates by 0.2 percentage points, or 20 basis points, though it is not statistically significant. Columns 2 and 3 add state-level economic controls and region-year fixed effects for the four Census regions. After controlling for economic conditions, the magnitude in column 2 indicates that a 10% increase in the exemption level reduces the recovery rate on charged-off debt by 36 basis points, significant at the 1%-level, and the magnitude is slightly larger after controlling for region year fixed effects in column 3. Put differently, moving from an exemption level of $15,000 (bottom terciles) to $60,000 (top tercile) would reduce recovery rates on chargedoff debt by 5 percentage points. Columns 4-7 show that the estimates are unchanged when using individual credit union data and including fixed effects for the individual credit unions. Thus, the effect represents changes that occur within an individual credit union and not a shift in credit across credit unions. Table 5 reports the estimates from the difference-in-differences equation (2) with the credit card interest rate as the dependent variable. The estimate in column 1 indicates that a 10% increase in the exemption level raises credit card interest rates by percentage points, or 4.2 basis points, and is statistically significant at the 1%-level. Put differently, moving from an exemption level of $15,000 (bottom terciles) to $60,000 (top tercile) would increase interest rates on charged-off debt by 0.6 percentage points. The results remain similar when economic, geographic, and individual credit union controls are added in columns 2-7. In Appendix Table B5, I show significant but smaller interest rate responses for auto loans, consistent with secured loans being less affected by exemption increases. Figure 4 plots the η t estimates and 95% confidence intervals from the event study specification in equation (3). The pattern of coefficients in these event study regressions lend credibility to the empirical design and show the longer-run effects of raising exemptions. For both interest rates and recovery rates, the coefficients η t 5,..., η t 1 are small and insignif- 19

21 icant, consistent with the parallel trends assumption. In the period t, when exemptions increase, interest rates rise and remain elevated for at least five years. At the same time, the recovery rate on charged-off loans falls sharply in period t and remains low over the next six years. In the Appendix, I report several robustness checks.table B1 shows that the results are similar if the sample is constructed from credit unions operating in only one state. Tables B3 and B4 show that the results are similar if the exemption level is included linearly and that the effects are driven by states with lower exemption levels. Table B6 shows that the results are largely unchanged if only the homestead exemption is used. Overall, these estimates demonstrated the costs and benefits of exemption generated default insurance. Borrowers repay less when they default, but must pay higher interest rates when they do not. In the next section, I show how these two effects can be compared to determine the cost of exemption-generated default insurance. One concern is whether these credit union estimates apply to the broader credit market. Between 1991 and 2004, credit unions issued 7-10% as much revolving credit as commercial banks and tend to have lower interest rates (12.3% vs. 14.7%) and recovery rates (17.2% vs. 19.3%). 33 The magnitude of these interest rate effects are very close to the estimate found for unsecured personal loans in Severino and Brown (2017), which uses a similar identification strategy with bank interest rates. The estimates are also in the lower end of the range of estimates found in papers using cross-sectional variation in exemptions (e.g., Gropp, Scholz and White (1997)). 34 That these estimate from credit unions are consistent with other results from commercial banks provides reassurance about their external validity. Additionally, in Appendix Table B7 and event study Figure B4, I estimate the main regressions using data from single-state commercial banks and find results similar to those from credit unions Source: Board of Governors of the Federal Reserve System, G.19 series - Consumer Credit. Revolving credit is mostly credit card loans. This number is slightly understated since the commercial bank figures include overdraft credit, but the credit union figures do not (J. Furletti and Ody, 2006). The interest rate and recovery rate data are author s calculations from aggregating commercial and Credit Union Call Reports and taking the mean annual recovery rate between 1994 and These numbers differ slightly from those in Table 3, which reports averages of the state-level recovery rates. 34 I provide a more detailed comparison with other estimates in Appendix A Since data for larger banks are only available at the bank-level, not the bank-state-level, the analysis 20

22 This suggests that the credit union estimates are, at least, applicable to smaller commercial banks. Moreover, the interest rate estimates for credit unions are very close to the magnitudes found in Severino and Brown (2017), which covers a similar period and uses data on rates from larger banks. 5 Calculating the Welfare Impact In this section, I use the sufficient statistic model of Dávila (2016) to map the estimates from Sections 3 and 4 into a statement about the welfare impact of raising exemptions. After introducing the sufficient statistics formula, I calculate the welfare impact of raising exemptions and then provide a new, alternative formula for the cost of raising exemptions. 5.1 Model of Default and Exemptions I use the baseline exemption model of Dávila (2016), which is a two-period model based on Eaton and Gersovitz (1981), and slightly modify it to incorporate the impact of exemptions on repayment in default. There are two periods, t = 0, 1, a single consumption good, and a unit measure of borrowers. At t = 0, income is certain and borrowers choose how much to borrow, B 0. At t = 1, borrowers receive an income draw y 1 from a distribution F ( ) with support [y 1, y 1 ] then decide whether to default or repay their debt. In deciding whether to default or repay, there is an endogenously determined optimal decision rule y where borrowers default if and only if y 1 < y. This income threshold y is chosen to satisfy C1 N (y ) = C1 D (y ) and it depends on the exemption level m and the amount borrowed B 0. Borrowers maximize: V (m) = y y1 max u(c 0) + u(c D C 0,{C 1 } y1,b 0,y 1 )df (y 1 ) + u(c1 N )df (y 1 ) (4) y 1 y cannot be conducted for the sample of larger commercial banks. Also, commercial bank interest rate data is not reported in call reports. 21

23 where C 0 = y 0 + B 0, C N 1 = y 1 (1 + r(m))b 0 C D 1 = (1 φ)y 1 s(m)b 0. In period 0, borrowers consume the income endowment, y 0, plus the amount borrowed, B 0. If y 1 > y in period 1, borrowers consume C1 N, which is the income that remains after repaying the debt. Otherwise, if y 1 < y in period 1, borrowers default and consume C1 D, which is income less the costs of default. The interest rate r(m) and the recovery rate upon default depend on the exemption level. 36 To capture this, I assume that defaulting borrowers repay a portion of their debts, s(m) < 1 that depends on the exemption level m and will be informed by the estimates in the empirical section. Debtors also pay a default cost that depends on their income φy 1, which can reflect other collection actions (e.g., wage garnishment), stigma, or reduced access to future credit. 37 V (m) captures the impact of exemptions on borrowers, but exemptions may also affect the balance sheets of lenders. Lenders profits are assumed to be unaffected by changes in the exemption level, so the change in consumer welfare captures the total welfare change. Later in this section, I will provide empirical support for this assumption by comparing the observed interest rate response to that implied by a zero-profit condition on lenders. The effect of a change in exemptions on social welfare is summarized by the following proposition: Proposition 1. Let c N and c D be average consumption in repayment and default. The welfare impact of using asset exemptions to provide $1 of additional expected consumption 36 Dávila (2016) allows the interest rate to also depend on the level of debt, B 0. Given that I only observe one interest rate per credit union, I omit this extension. 37 The major departure from the model of Dávila (2016) is in how exemptions affect consumption in default. This model incorporates the estimate of exemption on repayment, so CD 1 m original model assumes that a $1 exemption increase raises consumption by $1. = s (m), while the 22

24 during default is approximately 38 ( ) ( ) dw u dm (c D ) u (c N ) 1 (1 π) r (m) π s (m) 1. (5) Proof: See Appendix B. This formula weighs the value of additional default insurance against the cost of generating default insurance using asset exemptions. The first term on the right side of equation (11) is the ratio of marginal utilities in default and repayment. This ratio can be interpreted as the percentage markup over the actuarially fair rate that a borrower is willing to pay for additional default insurance. 39 The second term is the actual markup of the default insurance that is generated by increasing asset exemptions, i.e., the ratio of premium increases to payout increases. The numerator, (1 π)r (m), reflects the expected increase in interest payments (the insurance premium) and the denominator (π)s (m) reflects the expected reduction in payments made in default (the insurance payout). Baily (1978) and Chetty (2006) show that taking a Taylor expansion of u around c h yields the approximation u (c D ) u (c N ) 1 γ C C, where γ = u (c N ) u (c N ) c N is the coefficient of relative risk aversion evaluated at c N and C C to equation (11) yields the following formula: = c N c D c N. Applying the Baily-Chetty approximation dw dm γ C C ( ) (1 π) r (m) π s (m) 1 (6) Importantly, this formula does not require estimates of how exemption affect debt levels or default decisions. Moreover, Chetty (2006) shows that the logic of the envelope theorem underlying this formula holds in a much more general, dynamic setting with many other endogenous choices. Dávila (2016) shows that this formula for asset exemptions is robust to many extensions relevant for credit markets, including observed and unobserved heterogene- 38 The approximation requires that the third-order and above utility terms (e.g., u ) be small relative to the lower terms. 39 This markup interpretation and language follows Hendren (2013) and Hendren (2017). 23

25 ity. 5.2 Calculating the Welfare Impact The welfare formula in equation 6 can be evaluated with the empirical estimates of the decline in consumption upon default, the effect of exemptions on interest rates and recovery rates, and the probability of default. Following the empirical sections, the policy parameter is m = log(exemption), though the implications are similar when the linear exemption estimates in Table B3 are applied. 40 To calculate the welfare formula, I use the following empirical values: 41 C C = 0.05 r (m) = 0.36 s (m) = 3.17 π = There is uncertainty about the appropriate value for the coefficient of relative risk aversion over food consumption, γ, so I report the welfare gains for γ = 1, 3, 5. For these levels of risk aversion, debtors are willing to pay a markup of 5-25% over the actuarially fair rate, but the transfer generated by asset exemptions is marked up 405% over the actuarially fair rate. Consequently, at current exemption levels, increasing asset exemptions enough to raise the expected protection in default by $1 generates a welfare-reduction of $ per borrower. Put differently, doubling the exemption level reduces welfare by $40 per household. 42 Ultimately, the policy implication is that exemptions generate in- 40 The magnitude of the markup is robust to functional form. If I use the values from the linear exemption specification in Appendix Table B3 columns 2 and 5, the markup is 432% compared to 405% in the main analysis. 41 I use the mean charge-off rate, rather than the share of borrowers in default, because the benefits and costs are proportional to the amount of debt held. 42 For a household with $15,000 in unsecured debt, doubling exemptions raises expected protection in default by s π(15, 000) = $

26 surance that costs much more than what debtors are willing to pay. Consequently, welfare would increases if exemptions were lower (and this holds even in low-exemption states). 43 Given that the exemption levels in the United States vary from less than $10,000 to over $500,000, these estimates imply substantial differences in welfare across states. The result that lower exemptions increase welfare conflicts with the calibration exercise of Dávila (2016), which finds that current exemptions levels are near or slightly below the optimal level. 44 Two components differ significantly. First, this paper estimates the consumption drop upon default to be 3-5%, which more than halves the willingness to pay for insurance relative to the 10% value used in Dávila (2016). 45 Second, this paper s estimates imply a much smaller benefit from exemption-generated default insurance. Dávila (2016) assumes that a $1 exemption increase causes a $1 increase in consumption for bankruptcy filers with non-exempt assets. The estimates of this paper, in contrast, imply a transfer that is one-tenth as large, which leads to a much higher markup cost for exemption-generated default insurance (405% vs. 54%). 46 This small effect of exemptions is consistent with external evidence about exemption protection. Only 4-6% of bankruptcy filers have non-exempt assets (Flynn, Bermant and Hazard, 2003), and among this subset filers only around 10% would benefit from increases in the homestead exemption. 47 Relative to the values chosen in Dávila (2016), the estimates of this paper show that exemption insurance is less valuable (smaller willingness to pay) and costs more (higher markup), which ultimately implies that 43 While extrapolating from these estimates implies that optimal exemptions should be near zero, it is possible that the trade-offs changes as exemptions decrease. In particular, at current exemption levels, asset seizures are very rare. However, if seizures and foreclosures become more common at extremely low exemption levels, these events would introduce additional costs that are not captured by this welfare analysis. 44 The calibration values used in Dávila (2016) are sometimes chosen to demonstrate features of the model, and the author notes the need for improved estimates and additional empirical work. 45 This 10% decrease in consumption is based on Filer and Fisher (2005), but they find an 8-13% increase in consumption, not a decrease. 46 I provide the details of these calculations in Appendix A Between 2000 and 2013, only 10.8% of asset cases returned any funds to the debtor because they had non-exempt assets that were partially protected by exemptions (Chapter 7 Trustee Final Reports, ). Additionally, in a sample of cases with non-exempt assets, only 11% of cases had non-exempt equity in real estate of any kind, which can include home equity as well as unprotected equity in other property (Jiménez, 2009). Thus, only about 10% of asset cases, which make up 4-6% of all cases, would benefit from additional exemption protection. 25

27 lower exemptions would increase welfare. Appendix Table B8 shows that this policy implication is not sensitive to reasonable variation in the estimates of the components (e.g., within the 95% confidence intervals). 5.3 The Role of the Default Distortion Given the central role that this large markup plays in the welfare analysis, this section uses an alternative method of calculating the markup. Rather than relying on the observed interest rate increase, I calculate the markup using the interest rate response that is implied by a zero-profit restriction on lenders and the magnitude of the default rate distortion. 48 First, I estimate the effect of exemptions on default rates for credit cards. Table 6 reports difference-in-differences estimates from equation (2) with the credit card charge-off rate as the outcome. The credit card charge-off rate is not available before 1998, so the sample covers After including controls for economic conditions, the point estimate in column 2 indicates that a 10% increase in asset exemptions raises the credit card charge-off rate by percentage points, a 1.6% increase. That is, as exemptions increase, borrowers chargeoff a larger share of their debt (possibly due to both moral hazard and adverse selection into borrowing). The estimates are similar using the individual credit union data and credit union fixed effects (columns 4-6), with a 10% increase causing a percentage point increase in the charge-off rate. Appendix Figure B6 reports the event study estimates for this specification. Additionally, in Appendix Table B9 and event study Figure B5, I show a similar response for another measure of default - Chapter 7 bankruptcy filings. If lenders are competitive and risk-neutral, then the returns from lending satisfy a zeroprofit condition (1 π(m))(1 + r(m)) + π(m)s(m) = (1 + r), where π is the probability of default, r is the interest rate, s(m) is the recovery rate on defaulted debt, and r is the risk-free rate of return. 49 Differentiating this zero-profit condition with respect to m gives 48 This is analogous to the sufficient statistic for optimal unemployment insurance, which uses the government s balanced budget constraint and the behavioral response of unemployment duration to infer the response of the tax rate. 49 For simplicity, I assume the income cost of default φy 1 reflects stigma or access to future credit, so it 26

28 the following alternative formula for the markup: (1 π) r (m) π s (m) 1 = (m)[1 + r s] π πs (m) (7) That is, the markup can be expressed as the losses due to the behavioral distortion to default decisions relative to the losses from the mechanical effect on recovery rates. This version of the formula clarifies that the markup of exemption-generated default insurance is due to the behavioral response of default decisions. It also provides a second method of calculating the markup. Using the estimates of π, s and sample means for r, s, and π in equation 7, the implied markup is 366%. 50 This default distortion therefore explains 90% of the 405% markup obtained by using the observed interest rate increase. This is important for several reasons. First, the magnitude of the default distortion provides a plausible explanation for the observed interest rate markup. Second, it supports the critical assumption that lenders profits are unaffected by exemption increases; it appears that the additional losses are passed on to borrowers. Finally, it supports the policy implications two alternative methods of calculating the markup both suggest that exemptions create insurance that costs far more than the 5-25% markup that debtors are willing to pay. 6 Conclusion In this paper, I estimate the consumption smoothing benefits and costs of the default insurance provided by asset exemptions. I find that consumption falls when debtors default outside of bankruptcy, so there is potentially a consumption smoothing role for debtor protections. Exemptions, however, are an expensive means of providing this protection. While exemptions do create default insurance, this insurance is marked up 400% over the actuaridoes not enter the zero-profit condition. A similar formula holds if φy 1 is paid to lenders. 50 Specifically, I set π = 0.27, s = 3.17, and the sample means from Table 3. 27

29 ally fair rate, while debtors are only willing to pay a 5-25% markup. As a result, a sufficient statistic analysis indicates that welfare would increases substantially if states lowered exemption levels. This welfare analysis captures the main trade-off in raising exemptions, but omitted costs and externalities could potentially alter the trade-offs. First, exemptions may also affect the set of loan contracts offered or loan denial rates, and this could make increasing asset exemptions even more costly. Second, the welfare analysis assumes that consumers make financial decisions optimally, but behavioral biases are present in many household financial decisions (see Zinman (2014) and Zinman (2015) for an overview). If consumers do not make decisions optimally, the analysis in this paper could either overstate or understate the welfare gains of exemptions, depending on the specific behavioral biases of borrowers. Third, debtor protections exist alongside many other forms of social and private insurance programs and protections also influence the collection of non-consumer debts, such as medical debt. There is evidence that some of these programs interact, as consumers view health insurance, unemployment insurance, and default or bankruptcy as substitutes (Gross and Notowidigdo, 2011, Hsu, Matsa and Melzer, Forthcoming, Mahoney, 2015). Changes in exemption policy may reduce or exacerbate externalities in other social insurance programs, and this paper ignores these effects. The interaction of debtor protections and social insurance programs is important since debtor protections affect consumers ability to self-insure through credit markets. Additionally, there is some evidence that, by reducing foreclosures, exemption protections had positive macroeconomic effects during the Great Recession (Dobbie and Goldsmith-Pinkham, 2015). Alternative responses of lenders, behavioral biases in borrowing, and these externalities on other forms of insurance or default decisions are three avenues for future research on the welfare impact of debtor protection laws. 28

30 References Ashcraft, Adam B, Astrid Andrea Dick, and Donald P Morgan The Bankruptcy Abuse Prevention and Consumer Protection Act: Means-Testing or Mean Spirited? Athreya, Kartik Fresh Start or Head Start? Uniform Bankruptcy Exemptions and Welfare. Journal of Economic Dynamics and Control, 30(11): Baily, Martin Neil Some Aspects of Optimal Unemployment Insurance. Journal of Public Economics, 10(3): Berger, Allen N, Geraldo Cerqueiro, and María F Penas Does Debtor Protection Really Protect Debtors? Evidence from the Small Business Credit Market. Journal of Banking & Finance, 35(7): Berkowitz, Jeremy, and Michelle J White Bankruptcy and Small Firms Access to Credit. RAND Journal of Economics, 35(1): Berkowitz, Jeremy, and Richard Hynes Bankruptcy Exemptions and the Market for Mortgage Loans. The Journal of Law and Economics, 42(2): Bronchetti, Erin Todd Workers Compensation and Consumption Smoothing. Journal of Public Economics, 96(5): Cerqueiro, Geraldo, and María Fabiana Penas How Does Personal Bankruptcy Law Affect Startups? The Review of Financial Studies, Chetty, Raj A General Formula for the Optimal Level of Social Insurance. Journal of Public Economics, 90(10): Cochrane, John H A Simple Test of Consumption Insurance. Journal of Political Economy, Dávila, Eduardo Using Elasticities to Derive Optimal Bankruptcy Exemptions. Unpublished. Dawsey, Amanda E, Richard M Hynes, and Lawrence M Ausubel The Regulation of Non-Judicial Debt Collection and the Consumer s Choice Among Repayment, Bankruptcy and Informal Bankruptcy. American Bankruptcy Law Journal. Dobbie, Will, and Jae Song Debt Relief and Debtor Outcomes: Measuring the Effects of Consumer Bankruptcy Protection. American Economic Review, 105(3): Dobbie, Will, and Paul Goldsmith-Pinkham Debtor Protections and the Great Recession. Dobbie, Will, Paul Goldsmith-Pinkham, and Crystal S Yang Consumer bankruptcy and financial health. Review of Economics and Statistics, 99(5):

31 Dobbie, Will, Paul Goldsmith-Pinkham, Neale Mahoney, and Jae Song Bad Credit, No Problem? Credit and Labor Market Consequences of Bad Credit Reports. National Bureau of Economic Research. 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): East, Chloe N, and Elira Kuka Reexamining the Consumption Smoothing Benefits of Unemployment Insurance. Journal of Public Economics, 132: Eaton, Jonathan, and Mark Gersovitz Debt with Potential Repudiation: Theoretical and Empirical Analysis. The Review of Economic Studies, 48(2): Elias, Stephen, Albin Renauer, and Robin Leonard How to File for Bankruptcy. Berkeley: Nolo Press. Ernst and Young The Impact of Third-Party Debt Collection on the National and State Economies. Fan, Wei, and Michelle J White Personal Bankruptcy and the Level of Entrepreneurial Activity. Journal of law and economics, 46(2): Fay, Scott, Erik Hurst, and Michelle J White The Household Bankruptcy Decision. American Economic Review, 92(3): Fedaseyeu, Viktar Debt Collection Agencies and the Supply of Consumer Credit. Filer, Larry H, and Jonathan D Fisher The Consumption Effects Associated with Filing for Personal Pankruptcy. Southern Economic Journal, Finkelstein, Amy, Neale Mahoney, and Matthew J. Notowidigdo What Does (Formal) Health Insurance Do, And For Whom? prepared for Annual Review of Economics. Finlay, Steven The Management of Consumer Credit: Theory and Practice. Springer. Flynn, Ed, Gordon Bermant, and Suzanne Hazard Chapter 7 Asset Cases Part II. American Bankruptcy Institute Journal. Furletti, Mark Credit Card Pricing Developments and Their Disclosure. Federal Reserve Bank of Philadelphia Payment Cards Center Discussion Paper,, (03-02). Gilles, Stephen G The judgment-proof society. Wash. & Lee L. Rev., 63: 603. Gropp, Reint, John Karl Scholz, and Michelle J White Personal Bankruptcy and Credit Supply and Demand. The Quarterly Journal of Economics, 112(1):

32 Gross, Tal, and Matthew J Notowidigdo Health Insurance and the Consumer Bankruptcy Decision: Evidence from Expansions of Medicaid. Journal of Public Economics, 95(7): Gruber, Jonathan The Consumption Smoothing Benefits of Unemployment Insurance. The American Economic Review, 87(1): Hendren, Nathaniel Private information and insurance rejections. Econometrica, 81(5): Hendren, Nathaniel Knowledge of Future Job Loss and Implications for Unemployment Insurance. American Economic Review, 107(7): Himmelstein, David U., Elizabeth Warren, Deborah Thorne, and Steffie J. Woolhandler Illness and Injury as Contributors to Bankruptcy. Health affairs (Project Hope), W5 63. Hintermaier, Thomas, and Winfried Koeniger Debt Portfolios and Homestead Exemptions. American Economic Journal: Macroeconomics, 8(4): Hsu, Joanne W, David A Matsa, and Brian T Melzer. Forthcoming. Positive Externalities of Social Insurance: Unemployment Insurance and Consumer Credit. American Economic Review. Hynes, Richard M Broke But Not Bankrupt: Consumer Debt Collection In State Courts. Fla. L. Rev., 60: 1. Hynes, Richard M, and Eric A Posner The Law and Economics of Consumer Finance. American Law and Economics Review, 4(1): J. Furletti, Mark, and Christopher Ody Measuring U.S. Credit Card Borrowing: An Analysis of the G.19 s Estimate of Consumer Revolving Credit. Jiménez, Dalié The Distribution of Assets in Consumer Chapter 7 Bankruptcy Cases. American Bankruptcy Law Journal, 83: 795. Keys, Benjamin J. forthcoming. The Credit Market Consequences of Job Displacement. The Review of Economics and Statistics, 0(ja): null. Kroft, Kory, and Matthew J Notowidigdo Should Unemployment Insurance Vary with the Unemployment Rate? Theory and Evidence. Review of Economics Studies. Lehnert, Andreas, and Dean Maki Consumption, debt, and portfolio choice: Testing the effects of bankruptcy law. In Household Credit Usage Springer. Livshits, Igor Recent Developments in Consumer Credit and Default Literature. Journal of Economic Surveys, 29(4):

33 Li, Wenli, and Pierre-Daniel Sarte US Consumer Bankruptcy Choice: The Importance of General Equilibrium Effects. Journal of Monetary Economics, 53(3): Lyons, Angela C, and Jonathan Fisher Gender Differences in Debt Repayment Problems after Divorce. Journal of Consumer Affairs, 40(2): Mahoney, Neale Bankruptcy as Implicit Health Insurance. The American Economic Review, 105(2): Mankart, Jochen The (Un-) Importance of Chapter 7 Wealth Exemption Levels. Journal of Economic Dynamics and Control, 38: Mann, Ronald J Bankruptcy Reform and the Sweat Box of Credit Card Debt. U. Ill. L. Rev., 375. Mann, Ronald J, and Katherine Porter Saving up for Bankruptcy. Georgetown Law Journal, 98: 289. Mitman, Kurt Macroeconomic Effects of Bankruptcy and Foreclosure Policies. American Economic Review, 106(8): NBRC Bankruptcy: The Next Twenty Years. Final Report of the Bankruptcy Review Commission. NCLC Guide to Surviving Debt. National Consumer Law Center. Pavan, Marina Consumer Durables and Risky Borrowing: The Effects of Bankruptcy Protection. Journal of Monetary Economics, 55(8): Sandler, Danielle H, and Ryan Sandler Multiple Event Studies in Public Finance: A Simulation Study with Applications. Journal of Economic and Social Management. Severino, Felipe, and Meta Brown Personal bankruptcy protection and household debt. Stavins, Joanna Credit Card Borrowing, Delinquency, and Personal Bankruptcy. New England Economic Review, 15. Traczynski, Jeff Divorce Rates and Bankruptcy Exemption Levels in the US. Journal of Law and Economics. Zeldes, Stephen P Consumption and Liquidity Constraints: An Empirical Investigation. The Journal of Political Economy, Zhang, Shuoxun, Tarun Sabarwal, and Li Gan Strategic or nonstrategic: The role of financial benefit in bankruptcy. Economic Inquiry, 53(2):

34 Zinman, Jonathan Consumer Credit: Too Much or Too Little (or Just Right)? Journal of Legal Studies, 43(3): 9. Zinman, Jonathan Household Debt: Facts, Puzzles, Theories, and Policies. Annual Review of Economics, 7(1):

35 Table 1: Consumption Sample (PSID) Non-Defaulters Defaulters Bankruptcy Filers log consumption log food needs Food consumption (1990$) 4,815 4,346 4,697 Age Female Years of education White Number of children Married Unsecured debt (1990$) 3,504 3,827 4,703 Mortgage debt (1990$) 24,986 14,171 15,007 Observations 20,717 1, This table displays means for a sample of individuals who never reported defaulting (non-defaulters), and the defaulter sample, which consists of the observations of defaulters during the period of default. Column 3 shows the means for the sample of bankruptcy filers, as in Filer and Fisher (2005). Following Filer and Fisher (2005), the sample in Column 3 includes individuals who were not heads of their household at the time of default. Source: PSID

36 Table 2: The Consumption Drop Upon Default Default sample Baseline Controls State Controls (1) (2) (3) Low-exemption states (α L ) *** *** *** (0.014) (0.014) (0.015) Mid-exemption states (α M ) * * (0.023) (0.021) (0.023) High-exemption states (α H ) (0.018) (0.018) (0.019) Log(median income) (0.080) Unemp. rate (0.012) Observations 1,144 1,144 1,144 Year FE X X X Demographic controls X X This table reports results from the regression in equation (1) estimated on the sample of default instances. Demographic controls consist of age, sex, years of education, an indicator for white, marital status, number of children, and the change in the log of the food needs of the family, which is a function of family size and age computed by the PSID. The control variables are de-meaned and there is no constant, so the coefficients on the exemption terciles represent the average drop in consumption among defaulters in those states. Standard errors are clustered by state. Source: PSID

37 Table 3: Credit Union Summary Statistics Variable Mean Std. Dev. Min. Max. N Interest rate Recovery rate, non-real estate debt Charge-off rate, credit cards Charge-off rate, non-real estate debt Exemption level (1990$) 33,222 40, , This table shows descriptive statistics of the state-year level credit union data. Observations are weighted by the credit union membership in that state-year. The sample size is smaller for credit card charge-offs because that data is only available from The exemption statistics exclude the 7 states with unlimited exemptions. Source: NCUA Call Reports; Exemptions are from Elias, Renauer and Leonard ( ) and state statutes 36

38 Table 4: The Effect of Exemptions on Recovery Rates Dependent Variable: Recovery Rate on Non-Real Estate Debt State Aggregate Data Individual Credit Union Data (1) (2) (3) (4) (5) (6) (7) Log(exemption) *** *** * *** ** ** (1.648) (1.262) (1.527) (1.525) (1.073) (1.246) (1.270) Log(median income) (4.735) (4.700) (4.225) (4.157) (3.617) Unemp. rate *** ** *** *** *** (0.608) (0.592) (0.551) (0.565) (0.509) Log(house price index) 11.95*** 10.19*** 12.76*** 12.21*** 12.27*** (2.897) (2.359) (2.928) (2.813) (2.393) Observations ,731 52,731 52,731 51,108 Year FE X X X X X X X State FE X X X X X Region-year FE X Credit Union FE X X Drop Rec. rate > 1 X This table reports regression results from estimating equation (2). Standard errors clustered at the state-level are in parentheses. Columns 1-3 show estimates from the state-level aggregates, with observations weighted by credit union membership. Columns 4-7 use individual credit union data. The sample of credit unions is restricted to those with a positive amount of credit card loans. Some credit unions (less than 0.5% of the weighted sample) report recovery rates or charge-off rates over 100% due to timing issues (recoveries can be from previous years charge-offs, while charge-offs are only from the current year) or reporting errors. To reduce the influence of these outliers, I truncate the recovery rates at 100% in columns 4-5, and drop observations with recovery rates over 100% in column 6. Observations in columns 4-7 are weighted by the amount of credit card debt. Source: NCUA Call Reports 37

39 Table 5: The Effect of Exemptions on Interest Rates Dependent Variable: Credit Card Interest Rate State Aggregate Data Individual Credit Union Data (1) (2) (3) (4) (5) (6) (7) Log(exemption) 0.415*** 0.448*** 0.357** 0.306** 0.343** 0.363** 0.360** (0.129) (0.125) (0.158) (0.141) (0.146) (0.142) (0.144) Log(median income) * * (0.511) (0.540) (0.487) (0.508) (0.513) Unemp. rate (0.0448) (0.0524) (0.0542) (0.0526) (0.0534) Log(house price index) (0.278) (0.400) (0.299) (0.329) (0.331) Observations ,731 52,731 52,731 51,108 Year FE X X X X X X X State FE X X X X X Region-year FE X Credit Union FE X X Drop Rec. rate > 1 X This table reports regression results from estimating equation (2). Standard errors clustered at the state-level are in parentheses. Columns 1-3 show estimates from the state-level aggregates, with observations weighted by credit union membership. Columns 4-7 use individual credit union data. The sample of credit unions is restricted to those with a positive amount of credit card loans. Column 6 drops observations with recovery rates or charge-off rates over 100%. Observations in columns 4-7 are weighted by the amount of credit card debt. Source: NCUA Call Reports 38

40 Table 6: The Effect of Exemptions on Credit Card Charge-off Rates Dependent Variable: Credit Card Charge-off Rate State Aggregate Data Individual Credit Union Data (1) (2) (3) (4) (5) (6) (7) Log(exemption) *** 0.397*** *** 0.268** 0.270** (0.169) (0.0900) (0.112) (0.152) (0.0757) (0.108) (0.109) Log(median income) * * (0.386) (0.385) (0.345) (0.511) (0.511) Unemp. rate 0.101*** *** 0.127*** 0.142*** 0.142*** (0.0358) (0.0328) (0.0360) (0.0455) (0.0456) Log(house price index) *** *** *** *** *** (0.224) (0.276) (0.168) (0.167) (0.167) Observations ,399 33,399 33,399 32,697 Year FE X X X X X X X State FE X X X X X Region-year FE X Credit Union FE X X Drop Rec. rate > 1 X This table reports regression results from estimating equation (2). Standard errors clustered at the state-level are in parentheses. Columns 1-3 show estimates from the state-level aggregates, with observations weighted by credit union membership. Columns 4-7 use individual credit union data. The sample of credit unions is restricted to those with a positive amount of credit card loans. Some credit unions (less than 0.5% of the weighted sample) report recovery rates or charge-off rates over 100%. This is due to timing issues (recoveries can be from previous years charge-offs, while charge-offs are only from the current year) or reporting errors. To reduce the influence of these outliers, columns 4-5 truncate the recovery rates at 100%, and column 6 drops observations with recovery rates over 100%. Observations in columns 4-7 are weighted by the amount of credit card debt. Source: NCUA Call Reports 39

41 Figure 1: State Exemption Levels (2004) Exemption levels are from historical state statutes and various editions of Elias, Renauer and Leonard ( ). The exemption level is the sum of the home and non-home exemptions. 40

42 Figure 2: Heterogeneity in the Consumption Change upon Default This figure presents the mean and 95% confidence intervals of the change in consumption upon default for subsamples of defaulters. Non-exempt HE shows the drop for homeowners with non-exempt equity. For comparison, I also report the estimates and confidence intervals for bankruptcy filers from Filer and Fisher (2005) Table 4 column 1 as Bank (FF2005). The means in this figure are shown in Table A1. Source: PSID

43 Figure 3: Change in Consumption by Exemption Tercile This figure presents the mean consumption drop upon default and 95% confidence intervals for defaulters living in low-, mid-, and high-exemption states. It also presents the average consumption change for non-defaulters living in those states. The estimated coefficients and 95% confidence intervals are from the regression in specification (1). The Repayer results present the estimated coefficients from the same regression estimated on the sample of individuals who never report financial distress (non-defaulters). Both regressions include only year fixed effects and an indicator for whether the respondent lives in a low-, mid-, or high- exemption state. Low-exemption states have total exemptions less than $14,990, mid-exemption states range from $14,990-52,100, and high exemption states have total exemptions above $52,100 (including the unlimited exemption states). Source: PSID

44 (a) Recovery Rates on Charged-Off Loans (b) Credit Card Interest Rates Figure 4: Annual Effects of Exemption Increases in Year t The cumulative effect of a 100 log point increase in asset exemptions in period t, estimated from the distributed lag model in equation (3). The sample period is , with exemption data used from to allow for 5 leads and lags for each observation. The dotted lines show 95% confidence intervals for standard errors clustered at the state-level. Source: Bankruptcy Filings from the Administrative Office of the U.S. Courts 43

45 Online Appendices A Comparison with Other Estimates in Literature A.1 The Interest Rate Effect Other estimates of the interest rate effect are similar to or larger than the estimate in this paper. There are a variety of samples, specifications, and loan types used in the literature, so to make the estimates comparable, I consider the effect of moving from a state with a $5,000 exemption to one with a $50,000 exemption. The estimates in this paper predict such a change would result in an 80 basis point increase on credit card interest rates. Using a sample of 310 auto loans rates in the 1981 Survey of Consumer Finances, Gropp, Scholz and White (1997) report that such a change would result in a 230 basis point increase for the average borrower. Berkowitz and White (2004), using a sample of non-corporate small business loans, would predict a 225 basis point increase. Berger, Cerqueiro and Penas (2011), using a sample of corporate small business loans, predict a 23 basis point increase. The other paper using panel variation, Severino and Brown (2017), finds effects on unsecured loan (not credit card) interest rates from Ratewatch.com that are extremely close to the estimates of this paper. In summary, despite using different data, empirical strategies, and loan types, four of the five other papers providing estimates of the impact of exemptions on interest rates find estimates that are similar to or larger than the effect that I estimate. Using one of these larger estimates for the interest rate effect would strengthen the policy conclusion that lower exemptions would increase welfare. A.2 Calculating the Markup with the Values of Dávila (2016) To explain the different policy implications of this paper and Dávila (2016), I show how the chosen values of Dávila (2016) would alter the markup of exemption-generated insurance 44

46 implied by the formula used in this paper: Markup = ( ) (1 π) r (m)b 0 1, π s (m)b 0 where π is the probability of default, r is the effect of exemptions on interest rates, s is the effect of exemptions on recovery rates, and B 0 is the level of debt. This formula is the ratio of the expected increase in interest rates to the expected increase in payments in default. The key conceptual differences in Dávila (2016) are that (i) only bankruptcy filers with nonexempt assets benefit from exemption protection, and (ii) a $1 increase in exemptions raises the consumption of these filers by exactly $1. Consequently, the corresponding formula is Markup alt = ( ) (1 π) r (m)b 0 1, (8) π m 1 where π is the probability of bankruptcy, π m is the probability of filing bankruptcy and having non-exempt assets, and B 0 is the level of debt. This formula still compares the total increase in interest rate payments to the total reduction in payments in default. Using the values of Dávila (2016), π = 0.008, π m = , and r (m) = The value of B 0 implied by the chosen leverage ratio of 8.4% is $4, Evaluating formula (8) with these values implies a markup of 54%, much lower than the 400% markup found in this paper. This difference due to differences in the values for r. The interest rate estimates in this paper imply values very similar to the value used in Dávila (2016). 52 Instead, the difference is caused by the estimated benefit of increasing exemptions. In Dávila (2016), a $1 increase in exemptions generates a transfer of $0.0008, while using the estimated recovery rate implies a transfer that is only one-tenth as large I obtain this estimate by multiplying the 8.4% leverage ratio by median household income in 2016 of $59,039 (U.S. Census). Technically, this value should be lower as the leverage ratio is relative to disposable income. 52 The main estimate in this paper implies that increasing exemptions from $5,000 to $50,000 increases interest rates by 80 basis points, which implies that a $1 exemption increase raises interest rates by , similar to the calibrated value of in Dávila (2016). 53 To get the effect of a $1 increase in exemptions, I divide the estimated effect of a 100% increase in 45

47 The markup using the values of Dávila (2016) is 54%. This is approximately equal to the willingness to pay for default insurance γ C, which equals 50% when evaluated using C the baseline values in Dávila (2016) of γ = 5 and C C = 0.1. Since, with these values, the willingness to pay approximately equals the markup cost, the policy implication of Dávila (2016) is that exemptions are close to their optimal value. This differs sharply from the willingness to pay (5-25%) and markup cost (400%) that are implied by the estimates of this paper. B Theory B.1 Proof of Proposition 1 In this appendix, I derive the welfare formula in equation (11): ( ) ( ) dw u dm (c D ) u (c N ) 1 (1 π) r (m) π s (m) 1. The derivation follows Dávila (2016), which also shows that the formula holds under a variety of extensions relevant to credit markets. Borrowers maximize: V (m) = y y1 max u(c 0) + u(c D C 0,{C 1 } y1,b 0,y 1 )df (y 1 ) + u(c1 N )df (y 1 ) y 1 y exemptions (πs (m)b 0 = , 949) then divide by an average exemption level of $50,000 in order to convert the elasticity to the effect of a $1 increase. The transfer is even smaller if the linear exemption estimates of the effect on recoveries are used (Table B3). That the benefit is roughly onetenth as large in this paper is also consistent with the differences between the two calculated markups (406% vs. 54%). 46

48 where C 0 = y 0 + B 0, C N 1 = y 1 (1 + r(m))b 0 C D 1 = (1 φ)y 1 s(m)b 0. With 0 < φ < 1, the optimal default rule is to default if y 1 < y, where y satisfies C N 1 (y ) = C D 1 (y ). The first order condition for borrowing is y u (C 0 ) = u (C1 D )s(m)df (y 1 ) + y 1 y1 y u (C N 1 )r(m)df (y 1 ), where the optimal default rule eliminates the terms related to dy db 0. From the borrower s problem, dv dm = [ y y 1 u (C1 D ) s(m) m B 0dF (y 1 ) y y1 y y1 u (C1 N ) r(m) m B 0dF (y 1 ) ] u (C N )r(m)df (y 1 ) u (C 0 ) u (C1 D )s(m)df (y 1 ) } y 1 {{ y 1 } =0 [ U(C D 1 (y ))f(y ) U(C1 N (y ))f(y ) ] } {{ } =0 dy dm. The second and third lines are zero due to the first order condition for borrowing and the optimal default rule. This is an application of the envelope theorem. If individuals are already optimizing, then changes in borrowing decisions (line 2) or default decisions (line 3) in response to an exemption change have no first-order impact on welfare. As shown in Chetty (2006), this logic generalizes to a large set of endogenous choices and constraints that are potentially affected by the policy parameter. db 0 dm 47

49 Let the average marginal utility of consumption while in default and repayment be Eu (c D ) = Eu (c N ) = y y 1 u (C1 D )df (y 1 ) y y 1 df (y 1 ) y1 u (C N y 1 )df (y 1 ) y1. df (y y 1 ) As shown in Chetty (2006), if the third- and higher-order terms of u are small (u 0), then the average marginal utility of consumption in default (or repayment) is approximately equal to the marginal utility of consumption at the average consumption in default (or repayment): Eu (c D ) u (c D ) Eu (c N ) u (c N ). Using these approximations, the welfare impact of an increase in exemptions is dv dm πu (c D )s (m)b 0 (1 π)u (c N )r (m)b 0, (9) where π = y df (y 1 ) is the probability of default. To obtain the money-metric measure of y 1 the welfare gain, normalize the effect of exemptions by the marginal utility of an additional dollar in repayment, u (c N ): dv (m)/dm u (c N ) [( ) ( )] u (c D ) u (c N ) 1 (1 π) r (m) π s (m) 1 ( πs (m)b 0 ). Define T = ( πs (m)b 0 ), which is the expected increase in consumption in default from a one unit increase in m. Using the Baily-Chetty approximation, u (c D ) u (c N ) 1 γ C C, where γ = u (c N ) c u (c N ) N is the coefficient of relative risk aversion evaluated at c N and C C = c N c D c N. 48

50 Defining the welfare metric as dv (m)/dm normalized by u (c N )T gives the welfare formula dw dm γ C C ( (1 π) π r (m) s (m) 1 ). B.2 Differences with Model of Dávila (2016) In the model of Dávila (2016), borrowers similarly maximize: V (m) = max U(C 0) + β C 0,{C 1 } y1,b 0,y [ y y 1 U(C D 1 )df (y 1 ) + y1 y U(C N 1 )df (y 1 ) ] but C 0 = y 0 + q 0 (B 0, m)b 0, C N 1 = y 1 B 0 C D 1 = min{y 1, m}. The key conceptual difference is in how exemptions affect consumption in default. Dávila (2016) assumes that C D 1 = min{y 1, m}, so a $1 exemption increase raises consumption by $1, but only for bankruptcy filers with non-exempt assets (y 1 > m). In contrast, I assume that exemptions affect all defaulters, so that C D 1 = (1 φ)y 1 s(m)b 0, and incorporate the empirical estimates of s (m). The other differences are minor and allow a simpler mapping from the empirical estimates to the model. First, rather than focusing on the price of debt charged in period 0 q 0 (B 0, m) as in Dávila (2016), I focus on the interest rate charged in period 1. This difference is motivated by the fact that I estimate the impact on interest rates, which allows for comparison with other estimates of the interest rate effect in the literature. Additionally, because I observe only a single interest rate, and not an interest rate schedule, I do not allow the interest rate to depend on the amount of debt. Reverting these few changes will result in the baseline model and sufficient statistic of Dávila (2016). Therefore, as discussed in the text, the differences in the policy implications result from the 49

51 new estimates applied within the model, rather than substantial changes to the model itself. B.3 An Equivalent Formula from an Effort Choice Model In this section, I show that the same welfare formula can be obtained from an alternative model where individuals choose effort to avoid default. This is one example of the formula being a sufficient statistic; there are multiple underlying models that generate the same formula. There are two periods, t = 0, 1, and a single consumption good. In the first period, income is certain and borrowers choose how much to borrow, B 0. Borrowers also exert costly effort, e, that determines the probability that they will repay their debts in the second period. This effort choice reflects actions that borrowers can take to increase their ability to repay debt, such as purchasing insurance against shocks or adjusting work effort. The units of effort, e, are normalized so that e is equal to the probability of repayment in the second period. The cost of effort is given by a convex function f(e). In the second period, borrowers enter one of two states: the default (low state) or repayment (high state). Assignment to the default state can be viewed as an income or expense shock that makes repayment infeasible. If borrowers enter the default state, they earn income y D 1 and repay some share of their debt at the rate s < 1. This recovery rate s reflects that creditors collect some portion of what is owed even when borrowers are unwilling or unable to repay the full amount. If borrowers enter the repayment state, they earn income y N 1 and repay their full debt plus interest at the rate r > 0. Through the response of creditors, exemptions also influence the interest rate r(m). Borrowers take the interest rate r(m) and the recovery rate s(m) as exogenous and choose their effort and debt to maximize state-independent utility. Indirect utility, V (m), written as a function of the exemption level m, is equal to the utility from consumption in period 0 plus the expected utility in period 1, minus the cost of effort. V (m) = max e,b 0 u(c 0 ) + { eu(c N 1 ) + (1 e)u(c D 1 ) } f(e) (10) 50

52 where C 0 = y 0 + B 0, C N 1 = y N 1 (1 + r(m))b 0 C D 1 = y D 1 s(m)b 0. In period 0, borrowers consume the income endowment, y 0, plus the amount borrowed, B 0. In period 1, borrowers enter the repayment state with probability e and consume income y N 1 less the amount it takes to repay the debt plus interest (1 + r(m))b 0. Alternatively, borrowers default with probability (1 e) and consume income y D 1 less the portion of the debt that repaid in default s(m)b 0. I assume the regularity conditions that generate an interior solution with borrowers holding a positive level of debt. The change in expected utility from increasing the exemption level is obtained by differentiating V (m) with respect to m. Using the envelope conditions, dv (m) dm = { er (m)u (C N 1 )B 0 (1 e)s (m)u (C D 1 )B 0 }. This welfare change is in units of utility. To obtain a money-metric measure of the welfare gain, I normalize the effect of exemptions by the marginal utility of an additional dollar in the repayment state u (C N 1 ), so that dw (m) dm dv (m)/dm =. u (C1 N ) [( ) ( dw (m) u dm = (C1 D ) u (C1 N ) 1 e )] r (m) 1 e s (m) 1 T, (11) where T = (1 e)s (m)d > 0 is a scaling factor equal to the total amount of money transferred to those in default. Dividing both sides by T and denoting the probability of repayment e as (1 π) produces the welfare formula of equation (11). 51

53 Appendix: Changes in Consumption Upon Default 52

54 Table A1: Heterogeneity in the Consumption Drop Consumption Drop: Average N Defaulters ,144 Renters Homeowners Non-exempt home equity Non-exempt assets Severe Defaulters Bankruptcy (FF2005).081 This table reports the mean drop in consumption for subsamples of defaulters. Non-exempt HE is the subsample of homeowners that are not fully protected by exemptions. Severe defaulters are those who report a more serious type of financial distress (debt collection actions, judicial actions, or bankruptcy). For comparison, Bankruptcy (FF2005) shows the mean consumption increase from Filer and Fisher (2005) Table 4 column 1. Source: PSID

55 Table A2: The Consumption Drop Upon Default Default sample Average Baseline Controls State Controls (1) (2) (3) (4) Constant (α) *** *** *** *** (0.011) (0.014) (0.014) (0.015) Mid-exemption states (α M ) (0.028) (0.026) (0.028) High-exemption states (α H ) (0.023) (0.023) (0.026) Log(median income) (0.080) Unemp. rate (0.012) Observations 1,144 1,144 1,144 1,144 Year FE X X X X Demographic controls X X This table reports regression results from the regression logc i = α + α M exempt M i + α H exempt H i + δx i + ε i. Demographic controls consist of age, sex, years of education, an indicator for white, marital status, number of children, and the change in the log of the food needs of the family, which is a function of family size and age computed by the PSID. The control variables are de-meaned and there is no constant, so the coefficients on the exemption interactions represent the average drop in consumption among defaulters in those states. Standard errors are clustered by state. Source: PSID

56 Table A3: Default vs. Bankruptcy Defaulters Sample Pooled Sample (1) (2) (3) (4) (5) Default *** *** ** *** ** (0.019) (0.022) (0.022) (0.013) (0.013) Bankruptcy 0.110** 0.109** 0.110** 0.106** 0.096* (0.053) (0.054) (0.054) (0.049) (0.049) Log(median income) (0.076) Unemp. rate (0.009) Observations 1,658 1,658 1,658 24,667 24,667 Year FE X X X X X State FE X X Demographic controls X X X I construct the sample following Filer and Fisher (2005), which uses the PSID and includes individuals with consecutive defaults, non-heads of household, and consumption changes over 300%. Default is an indicator for default and includes bankruptcy, so the coefficient on bankruptcy represents the difference between bankruptcy and non-bankruptcy default. Columns 1-3 report estimates from the regression in equation (1) estimated on the sample of default instances, but adds an indicator for a formal bankruptcy filing. The control variables are de-meaned and there is no constant, so the coefficient on Default represents the average drop in consumption among defaulters. Columns 4-5 pool the sample of defaulters and non-defaulters, include state year fixed effects, and report coefficients for the exemption level interacted with an indicator for default and an indicator for bankruptcy. Demographic controls consist of age, sex, years of education, an indicator for white, marital status, number of children, and the change in the log of the food needs of the family, which is a function of family size and age computed by the PSID. State controls include the log of median income and the unemployment rate. Standard errors are clustered by state. Source: PSID

57 Table A4: Non-Defaulters as a Comparison Group Non-Defaulters Sample Pooled Sample Baseline Controls State Controls Baseline Controls (1) (2) (3) (4) (5) Low-exemption (α L ) * (0.003) (0.003) (0.003) Mid-exemption (α M ) (0.004) (0.004) (0.004) High-exemption (α H ) (0.003) (0.003) (0.003) Low-exemption Default *** ** (0.015) (0.015) Mid-exemption Default * (0.022) (0.022) High-exemption Default (0.018) (0.018) State controls X Demographic controls X X X Year FE X X X X X State Year FE X X Observations 20,717 20,717 20,717 21,861 21,861 Columns 1-3 estimate equation (1) on the sample of non-defaulters. Columns 4-5 pool the sample of defaulters and non-defaulters to estimate logc i = α L exempt L i Default i + α M exempt M i Default i + α H exempt H i Default i + δx i + τ s(i),t(i) + ε i, where τ s(i),t(i) represents the set of state year fixed effects. There are multiple years for each household, so i indexes household-year observations. Demographic controls consist of age, sex, years of education, an indicator for white, marital status, number of children, and the change in the log of the food needs of the family, which is a function of family size and age computed by the PSID. State controls include the log of median income and the unemployment rate. Standard errors are clustered by state. Source: PSID

58 Table A5: Leads and Lags of the Consumption Change Full sample of defaulters Defaulters with a lead and lag Period relative to default t-2 to t-1 t-1 to t t to t+1 t-2 to t-1 t-1 to t t to t+1 (1) (2) (3) (4) (5) (6) Change in consumption *** ** (0.013) (0.011) (0.014) (0.014) (0.012) (0.014) Observations 1,030 1,144 1,130 1,017 1,017 1,017 This table reports estimates from a regression of leads and lags of the log change in consumption on a constant and the de-meaned demographic controls. The constant captures the mean change in consumption. Columns 1-3 estimate the equation on the full sample of defaulters. Columns 4-6 estimate the equation on the subsample of defaulters for which a lead and lag of the consumption change are available. All specifications include year fixed effects and demographic controls for age, sex, years of education, an indicator for white, marital status, number of children, and the change in the log of the food needs of the family. Additionally, state-year level controls for the log of median income and the unemployment rate are included. Standard errors are clustered by state. Source: PSID

59 Appendix: The Consumption Smoothing Effect of Exemptions 58

60 Table B1: Estimates from Sample of One-State Credit Unions Credit card interest rates Recovery rates on non-real estate debt (1) (2) (3) (4) (5) (6) Log(exemption) 0.466** 0.467*** 0.366* ** ** (0.175) (0.163) (0.184) (1.363) (1.190) (1.282) Log(median income) (0.494) (0.512) (5.796) (5.591) Unemp. rate *** ** (0.0485) (0.0552) (0.530) (0.534) Log(house price index) *** 13.31*** (0.364) (0.536) (3.457) (3.070) Observations State and year FE X X X X X X Region-year FE X X *** p<0.01, ** p<0.05, * p<0.1 Estimates are from specification (2), but the sample of credit unions is restricted to those with branches in only one state. Credit union call reports available from the NCUA begin including branch locations in I use the 2013 data, which include branch locations for 99.97% of credit unions (compared with 95.29% in the 2010 data). 92.9% of credit unions have branches in only one state, and 98.2% have branches in two or fewer states. Observations are weighted by credit union membership. Standard errors clustered at the state-level are in parentheses. Source: NCUA Call Reports 59

61 Table B2: Comparing Pre- and Post-Period Estimates CU Rec. Rate Bank Rec. Rate CU Interest Rate (cc) (1) (2) (3) (4) (5) (6) Log(exemption) *** *** *** (1.262) (2.131) (1.757) (3.160) (0.125) (0.148) Log(median income) (4.735) (7.855) (8.503) (18.17) (0.511) (0.543) Unemp. rate *** (0.608) (0.874) (1.479) (1.319) (0.0448) (0.0376) Log(house price index) 11.95*** * (2.897) (7.109) (4.754) (18.57) (0.278) (0.821) Observations State and year FE X X X X X X This table compares estimates from specification 2 for the pre-bapcpa period and the post-bapcpa period Observations are weighted by credit union membership. Standard errors clustered at the state-level are in parentheses. Source: NCUA Call Reports, FFIEC Call Reports 60

62 Table B3: Linear Exemptions Credit card interest rates Recovery rates on non-real estate debt (1) (2) (3) (4) (5) (6) Exemption ($1,000s) *** *** * *** *** ( ) ( ) ( ) ( ) ( ) ( ) Log(median income) (0.519) (0.548) (4.942) (5.083) Unemp. rate *** * (0.0471) (0.0553) (0.631) (0.682) Log(house price index) *** 10.14*** (0.281) (0.409) (3.087) (2.475) Observations State and year FE X X X X X X Region-year FE X X Estimates are from specification (2), but including exemption linearly instead of as log(exemption). Observations are at the state-year level and weighted by credit union membership. Standard errors clustered at the state-level are in parentheses. Source: NCUA Call Reports 61

63 Table B4: Linear Exemptions - Heterogeneity by Exemption Level Credit card interest rates Recovery rates on non-real estate debt (1) (2) (3) (4) (5) (6) Exemption ($1,000s) *** *** *** ** *** *** (0.0104) (0.0103) (0.0102) (0.180) (0.104) (0.126) High Exemption *** *** *** 0.356* 0.339*** 0.325** (0.0102) (0.0103) ( ) (0.178) (0.104) (0.125) Log(median income) (0.513) (0.544) (4.618) (4.836) Unemp. rate *** ** (0.0420) (0.0509) (0.567) (0.592) Log(house price index) *** 9.555*** (0.284) (0.407) (2.751) (2.419) Observations State and year FE X X X X X X Region-year FE X X Estimates are from specification (2), but including exemption linearly instead of as log(exemption). High is an indicator that is constant within each state and equals 1 if the state s average exemption level from is above the median average exemption level. Observations are at the stateyear level and weighted by credit union membership. Standard errors clustered at the state-level are in parentheses. Source: NCUA Call Reports 62

64 Table B5: Impact of Exemptions on Credit Card and Auto Loan Interest Rates Credit cards New auto loans Used auto loans (1) (2) (3) (4) (5) (6) Log(exemption) 0.415*** 0.448*** * 0.172* 0.191** (0.129) (0.125) (0.0918) (0.102) (0.0862) (0.0835) Log(median income) (0.511) (0.330) (0.309) Unemp. rate ** *** (0.0448) (0.0278) (0.0297) Log(house price index) (0.278) (0.207) (0.212) Observations State and year FE X X X X X X This table reports regression results from estimating equation (2) with state-level interest rate data. Observations are weighted by credit union membership. Standard errors clustered at the state-level are in parentheses. Source: NCUA Call Reports 63

65 Table B6: Estimates using Homestead Exemptions Only Credit card interest rates Recovery rates on non-real estate debt (1) (2) (3) (4) (5) (6) Log(home exemptions) 0.376*** 0.393*** 0.305* *** *** (0.130) (0.130) (0.159) (1.936) (1.257) (1.460) Log(median income) (0.524) (0.543) (4.851) (4.886) Unemp. rate *** * (0.0454) (0.0538) (0.604) (0.595) Log(house price index) *** 9.289*** (0.283) (0.412) (2.900) (2.389) Observations State and year FE X X X X X X Region-year FE X X *** p<0.01, ** p<0.05, * p<0.1 Estimates are from specification (2) with the (log) homestead exemption used as the main independent variable. Observations are weighted by credit union membership. Maryland and Delaware, which had no homestead exemption from , are excluded from the sample. Standard errors clustered at the state-level are in parentheses. Source: NCUA Call Reports 64

66 Table B7: Estimates from Commercial Bank Call Reports Dependent Variable: Recovery Rate on Charged-Off Consumer Debt State Aggregate Data Individual Bank Data (1) (2) (3) (4) (5) (6) (7) Log(exemption) * *** ** * * (2.069) (1.728) (2.397) (1.994) (1.709) (1.198) (1.159) Log(median income) (7.544) (7.599) (6.237) (6.061) (4.576) Unemp. rate ** * (1.539) (1.080) (1.430) (1.043) (0.778) Log(house price index) 9.738** *** 6.924** 9.325*** (4.782) (5.518) (3.408) (3.379) (3.301) Observations ,373 73,373 73,373 66,927 Year FE X X X X X X X State FE X X X X X Region-year FE X Bank FE X X Drop Rec. rate > 1 X This table replicates Table 4 using data from commercial bank call reports. The sample of commercial banks included is restricted to those with branches in only one state. I also drop banks with holding companies that have branches in multiple states. To match the credit union analysis, I define consumer debt as the sum of credit cards and other household loans (single payment, installment, student, and revolving plans other than credit cards). Observations are weighted by the average amount of consumer debt outstanding per state (Columns 1-3) or per bank (Columns 4-7). Standard errors clustered at the state-level are in parentheses. Source: FFIEC Call Reports 65

67 Table B8: Sensitivity of the Welfare Impact Parameter Assigned Value Lower 95% CI Upper 95% CI Values Needed for Welfare Gain Risk aversion γ > 81 c Consumption drop c > 1.35 Interest rate change r (m) < Recovery rate change s (m) < Probability of default π > 0.09 Assigned value reports the value used in the welfare calculations. The lower and upper bounds of the 95% confidence intervals are shown for the estimated parameters. The final column shows the range of values for each parameter that would generate a welfare gain from increasing exemptions, holding other parameters constant at their assigned value. 66

68 Table B9: Exemptions and Bankruptcy Filings Dependent Variable: Bankruptcy filings per 1,000 Chapter 7 Chapter 13 (1) (2) (3) (4) (5) (6) Log(exemption) ** 0.594*** (0.283) (0.215) (0.211) (0.105) (0.0965) (0.120) Log(median income) (0.530) (0.590) (0.339) (0.382) Unemp. rate 0.153** (0.0703) (0.0881) (0.0450) (0.0502) Log(house price index) *** *** *** *** (0.421) (0.445) (0.348) (0.347) Observations Year FE X X X X X X State FE X X X X X X Region-year FE X X This table shows the effect of exemptions on Chapter 7 and Chapter 13 bankruptcy filing rates per 1,000 people. Standard errors clustered at the state-level are in parentheses. Source: Bankruptcy Filings from the Administrative Office of the U.S. Courts 67

69 Figure B1: Number of Exemption Changes ( ) Source: Exemptions are from Elias, Renauer and Leonard ( ) and state statutes. 68

70 (a) Size of Exemption Changes (b) Number of Exemption Changes Figure B2: Distributions of the size and number of changes in homestead exemptions from Source: Exemptions are from Elias, Renauer and Leonard ( ) and state statutes. 69