Credit Aversion. Sean Hundtofte. February 15, 2018 Preliminary Draft

Transcription

1 Credit Aversion Sean Hundtofte February 15, 2018 Preliminary Draft I thank Daniel Paravisini and Donghoon Lee for fruitful discussions, Andreas Fuster, Michaela Pagel and participants at the NYC Household Finance Juniors Workshop for their comments. All errors are my own. The views expressed in this paper are solely my own and do not necessarily reflect those of the Federal Reserve Bank of New York, or the Federal Reserve System. Federal Reserve Bank of New York, New York, USA. sean.hundtofte@ny.frb.org, Phone

2 Abstract This paper studies the propensity of households to voluntarily terminate credit contracts and constrain their future borrowing. Using representative surveys of household beliefs and financial behavior, and a randomized sample of credit registry data from , I estimate that almost half (40%) of credit cards are voluntarily closed rather than kept open at a zero balance. Voluntary credit card closure is difficult to reconcile with a frictionless and rational model, as most cards (>95%) have no annual fee, and closure imposes an immediate cost on credit score in addition to lost consumption smoothing benefits. Yet I find that households are more likely to engage in credit aversion in response to negative employment and house price appreciation shocks, precisely when liquidity constraints are most likely to bind. I note that these findings are consistent with a pro-cyclical demand for credit: when economic news is bad, households appear less likely to borrow for consumption. Terminating a credit contract could act as a commitment to limit future (luxury) consumption.

3 1 Introduction How do U.S. households view their ability to borrow? Do they see an unsecured line of credit as ready to perform intertemporal smoothing, or do they worry about its potential to facilitate unnecessary temptation spending in the future (Bertaut et al., 2009). typically think of supply-side financial constraints, such as the requirement of collateral or a lack of appropriate institutions, as limiting borrowers capacity to smooth consumption, amplifying negative shocks to economic activity in a variety of contexts (Deaton, 1991; Kiyotaki and Moore, 1995; Zeldes, 1989). Indeed, access to credit is keenly monitored by central banks as a key financial condition for an economy s health. 1 But what if households deny themselves credit precisely when the outlook turns bleak and financial distress is more likely? Austerity characterised some fiscal responses to the Great Recession. 2 At the householdlevel, there were analagous accounts of voluntary credit card closures ( cutting up the credit card ) and a generational shift away from the traditional financial system. 3 We If agents propensity to consume or invest in risky assets varies by sentiment (Keynes, 1936; Schiller, 2000), household beliefs of appropriate credit usage could also similarly vary (Minsky, 1977). In this paper, I investigate the extent to which households constrain their access to credit by voluntarily closing credit cards. I measure self-imposed credit constraints in a 5% sample of all U.S. individuals credit registry records (the FRBNY s Consumer Credit Panel, or CCP) by observing whether an account is closed or kept open at a zero balance, with its option to borrow in the future still intact. I then compare the timing and borrower 1 The Federal Reserve Bank of New York for example runs a Credit Access Survey of households every 4 months in addition to polling lending officers directly. One of the 25 or more questions in the Credit Access Survey addresses whether a household has closed an account themselves. 2 See the UK, Eurozone for examples. The second definition of austerity in the Oxford English Dictionary (2017) is now Difficult economic conditions created by government measures to reduce public expenditure. 3 html; Chakrabarti et al. (2013) 1

4 characteristics associated with closure. To address omitted variables and possible supplyside confounds, I employ additional survey data on credit access, beliefs and financial literacy administered within a representative survey of U.S. households (the FRBNY s Survey of Consumer Expectations, or SCE). After documenting the propensity to close credit cards after bad news arrives, I then consider a number of potential explanations for the behaviour. I first estimate that approximately 40% of credit card accounts of low-risk borrowers are closed rather than simply kept open at a zero balance. Motivated by the stylized fact that credit card closures spike at the onset of the financial crisis, as illustrated by Figure 1 and noted in Campbell et al. (2016), I then exploit the panel nature of the dataset to test the hypothesis that credit aversion varies inversely with economic outlook holding time-period constant. I find closures are more likely in regions with negative unemployment shocks (using Bartik-style instruments) or negative House Price Appreciation (HPA). During the crisis, some U.S. banks close riskier zero balance credit card accounts even when the account is contractually performing. As a result, there is a potential supply confound. I address this by restricting all analysis using the CCP to prime borrowers (>700 FICO); by taking advantage of cross-sectional variation (bank-driven closures of zero balance cards were related to nationwide portfolio purchases); and by performing additional analysis restricting analysis to non-crisis years. If closures were supply-driven, we would expect to find an inverse relationship to credit score, but I find an economically insignificant (a statistically significant but small positive) relationship to credit score. Lastly, I replicate the CCP-based analysis with survey data, where households state when they voluntarily terminate a credit contract. In the separate survey data of the SCE, I confirm the economic magnitude of voluntary closures and again find that closures covary with beliefs of future negative shocks. Additionally, I find limited support for interpreting credit aversion as a beliefs-based mistake in terms of its impact on credit score: most people understand the direct cost of credit card 2

5 closure on credit score. Only approximately 20% of a representative sample are mistaken in their understanding that closing a credit card is likely to have a negative effect on credit score. The similar economic magnitude and relationship to likely distress found in the SCE provides a clean replication of the CCP-based findings. While credit card closures provide a natural laboratory in which to examine the demand for credit, there are data challenges presented to analysis using the CCP. There is limited information at the individual level outside of credit data (e.g. FICO, debts and performance). Survey data addresses this; and I incorporate zipcode and census-tract averages for variables such as income and house prices. One friction not observed in the current dataset is which cards carry a fee. This could explain some voluntary closures, but is unlikely to explain all of them. In the US, if anything, fee cards are closed less often than no-fee cards. Fee cards represent only < 3% of all cards in in the CFPB database of credit card issuance (CFPB, 2013). Frictions such as mental accounting costs can explain the overall fact, but cannot explain closures varying inversely with economic outlook. Closing a typical credit card in the U.S. is suprising from a frictionless and rational viewpoint, as cards are typically costless and closure leads to a direct, albeit small, cost to one s credit score, and an expected cost from tighter liquidity constraints. Economic theory predicts that access to credit and its ability to smooth costs and consumption over time is important and beneficial. Mechanically, the costs and benefits associated with holding a credit card depend on its balance and credit limit (the maximum amount that a lender has approved borrowing up to). Major determinants of credit score construction are a borrower s overall utilisation ratio (balance divided by credit limit) the lower, the better; and the history a borrower has with paying an account on time the more accounts with long unblemished pay-histories, the better. On both these counts, closing a credit card presents a cost to FICO. In contrast, keeping a credit card open provides the benefit of assuring the opportunity to borrow up to a credit limit in the near future. 3

6 As expected, short-run costs to credit scores and delinquencies are associated with card-closure, albeit economically small in their magnitude. In favor of possibly interpreting closure as a commitment device, the short-run relationsihp to credit score is retraced over subsequent quarters, ending at a smaller positive point estimate 4 quarters later. Within-person analysis shows a similar relationship as the main analysis. In addition, I find some evidence in favour of households learning about liquidity constraints: individuals are less likely to close a card after previously closing one card if they missed a payment on one of their other debts (consistent with having encountered a shock and a subsequent binding liquidity constraint). This evidence is consistent with initially overconfident estimates by some households in the likelihood of liquidity constraints binding, consistent with an initial undervaluation of the future option to borrow. Relevant theory to explain credit aversion includes short-sighted doer/far-sighted planner (dual-self) models, with heterogeneity in self-control (Thaler and Shefrin, 1981). Keynes (1936) and Minsky (1992) describe how changes in outlook can lead to business cycles and pro-cyclical changes in what is an acceptable liability structure. Keeping up with the Jones s utility functions could also be relevant. Prelec and Loewenstein (1998) explore how the pain of paying, mental accounting, and financing method can induce debt aversion. Households could for instance only receive release from the burden of a debt s mental account by closing the actual account. Related empirical findings include the use of myopia and self-control issues to explain illiquid savings as a commitment device (Angeletos et al., 2001), and using impatience and sophistication to explain borrowing and repayment behaviour (Kuchler and Pagel, 2017; Meier and Sprenger, 2010). There is previous empirical evidence of a heterogeneous aversion to carrying debt balances even at a 0% interest rate, at least in a student loan setting (Cadena and Keys, 2013; Callender and Jackson, 2005). An additional motive for closure of a credit card by sophisticated consumers is that consumption behaviour may be less responsible or price-conscious when a credit card is used as the payment method 4

7 (Prelec and Simester, 2001). In a study examining the bigger picture of US household debt responses to employment shocks, Hundtofte and Pagel (2017) (HP hereafter) find an absence of increased borrowing using unsecured revolving accounts such as credit cards. On average, HP find that outstanding balances remain steady in the face of negative shocks even when households have slack in their credit limits. It appears the typical response to negative employment shocks is to smooth credit balances, rather than consumption. Collectively, the findings of these two papers beg the question of the usefulness of credit cards for consumption smoothing in response to transitory shocks, as there does not appear to be a counter-cyclical response in borrowing by US households. In this paper, I highlight and quantify a puzzle for modeling of intertemporal behaviour. Low-risk borrowers appear to constrain their own future access to credit, and this credit aversion is more likely to occur when recent economic news is bad. Credit aversion is puzzling as it imposes a cost on one s credit score and constrains consumption smoothing. Yet I find voluntary closures are more likely exactly when negative shocks are more likely. This household financial behaviour contrasts with that of sophisticated and well-advised corporations during the crisis, who draw on lines of credit for precautionary liquidity in response to negative news (Ivashina and Scharfstein, 2010). Multiple moments in the data pose challenges to a frictionless, rational view of household financial behaviour. Pre-emptive closure of a financially costless credit card could represent a household responsibly constraining their future (temptation) consumption. Such a belt-tightening form of debt aversion is consistent with models where changes in beliefs, or Animal Spirits, affect agents views on the appropriate use of credit for consumption. One policy implication for responses to contractionary economic events is that it might not be enough to ensure lenders continued willingness to lend, if borrowers are increasingly averse to using debt to fuel their consumption. 5

8 2 Data 2.1 Datasets The two main datasets I use are the Federal Reserve Bank of New York s Consumer Credit Panel (Lee and Van der Klaauw, 2010) and the Federal Reserve Bank of New York s Survey of Consumer Expectations (Armantier et al., 2013). The Consumer Credit Panel (CCP) is an anonymous longitudinal panel of individuals, comprising a 5% random sample of all individuals who have a credit report with Equifax. The quarterly sample starts in 1999Q1 and ends in 2017Q2. The data is described in detail in Lee and Van der Klaauw (2010). For the current analysis, I use the 0.1% sample, which provides information on approximately 250,000 randomly selected individuals each quarter. The CCP provides credit registry data on all debts monitored by one of the three main credit bureaus, in addition to public records (bankruptcy, death) and mobility (address changes) for any individual that is visible to the credit registry, e.g. excluding the young and those without reported debts. This panel dataset allows the econometrician to track all aspects of individuals financial liabilities, including bankruptcy and foreclosure, mortgage status, detailed delinquencies, various types of debt, with number of accounts and balances. Address information is available to the census block level. To supplement the analysis of the CCP, I obtain data on individual beliefs matched with credit events as provided by the Survey of Consumer Expectations (SCE). The SCE is a nationally representative, internet-based survey of a rotating panel of about 1,300 household heads. The survey has been conducted at a monthly frequency since June New respondents are drawn each month to match various demographic targets from the American Community Survey (ACS), and stay on the panel for twelve months. The SCE has high response rates: first-time respondents have a participation rate of about 60%; For repeat respondents the participation rate is about 90%. The survey contains a core monthly module on expectations about various macroeco- 6

9 nomic and household level variables, in addition to a module on credit access. Respondents are asked for their expectations of the state of the economy such as inflation, personal outcomes such as income growth, access to credit, their household s financial situation. In addition, the survey contains detailed demographic information about the respondents and their household. Previous work has shown SCE expectations are informative and meaningful in household financial decisions (Armantier et al. (2015) amongst others) 2.2 Data Challenges and Constructed Variables For the current analysis, I focus on the CCP designation of bank cards and ignore other revolving lines of credit such as home equity lines. Credit card activity in this dataset is observed at the individual-level, and contains data such as number of cards, total scheduled payments and balances but not features such as non-interest based annual fees. Analysis using the CCP is conditional on an individual paying off a credit card, 4 and examining whether one (or more) credit cards were closed that quarter. We can observe when the total number of credit cards an individual possesses declines from one quarter to the next. This ignores closures that are replaced by new credit cards, i.e. restricts our analysis to net drops in the number of credit cards. Individuals reported as deceased are dropped from analysis. I match individuals by zipcode to various additional features such as unemployment rates (BLS), median house prices and appreciation (CoreLogic), and 2000 income (BEA) for each county. Lastly, I construct Bartik-style shift shares of industry employment shocks based on Burea of Labor Statistics (BLS) county statistics. This approach uses changes in national employment levels of a county s various industries to instrument for changes in county level employment, with the intent to cleanse these changes of any local changes in demand. 4 Paying off a new credit card is defined as the number of bank cards held at a zero balance increasing by at least one, using the relevant CMA-defined variable for zero balance trade lines. The CMA definition is a subset of the CRTR definition so this will undercount some occurrences. 7

10 2.3 Sample Summary Table 1 presents summary statistics of the representative sample of US households in the CCP (column 1), and the sample used for analysis conditional on the proxy for having recently paid off a credit card (column 2). The average age is approximately 50 and credit score 690, the average 12 month house price appreciation is 4%, unemployment rate is 6%, average per-capita income for that county is approximately $30k. The sample of observations where a credit card is paid off is slightly older, wealthier, higher credit score, and most substantially have higher debt burdens (almost double the average in entire dataset - $100k versus $60k). 3 Results The summary statistics of Table 1 show that on average, after paying at least one credit card down to a zero balance, high credit score households close their credit cards 40% of the time. Turning to look at covariates of these closures, Table 2 finds that closure is more likely according to age, wealth and credit score. The greater the typical value of a home in a county, the less likely a borrower engages in credit aversion. The statistically significant and positive relationship to credit score, while small, is inconsistent with the approximations of voluntary credit card closures picking up lender-driven closures. Other estimated relationships are smaller: holding all other variables constant, for each ten years of age, a household is 0.5% less likely to close a zero balance card; for each $100k in median home value approximately 0.3% less likely; for $10k of avg. per-capita income 0.5% less likely; and for a 1% greater overall unemployment rate, 3% more likely to close a card. Introducing a dummy for the crisis period of , the point estimate of the dummy indicates an additional 8% likelihood of credit card closure over that specific time period and other estimated relationships weaken. Given an extension of credit by a lender, some households might ruthlessly optimise 8

11 while others apply a heuristic or other bias in their management of household debt. The positive correlations with age, which is likely correlated with experience, and wealth and credit score shown in Table 2, provide initial suggestive evidence that credit aversion might not arise owing to a lack of financial sophistication. In untabulated analysis I also find households are more likely to engage in credit aversion the greater amount of outstanding credit card debt they owe, and the larger the household (for an additional family member, there is a 2% greater propensity to close when number of household members enters the estimation linearly). Table 3 estimates that voluntary closures are more likely when proxies for changes in household beliefs recent house price trends, and changes in employment using plausiblyexogenous (Bartik-style) shocks are negative. The economic magnitudes are large - albeit twice as large for changes in employment rate as for recent changes in house prices. Over the entire time period, one standard deviation in trailing 12 month HPA is equivalent to a 1.4% greater likelihood of closing a credit card, while one standard deviation in unemployment rate changes is equivalent to a 2.5% greater likelihood in card-closure. All specifications control for general changes in propensity over time, likely sapping some of the explanatory variation in the underlying relationship if it is driven by changes in outlook. Columns 2 and 4 introduce a fixed effect for years in the depth of the crisis, Even after sapping a lot of the variation that would be useful to test for a procylical demand for credit, the cross-sectional relationship between negative employment shocks and voluntary credit card closure persists. The relationship to trailing house prices is more significantly weakened, consistent with a model of belief-driven closures and the period being a time of house-price driven negative news. A potential concern could be that the extensive margin of zero balance credit cards decreases in response to negative shocks, while the amount of closures remains fixed in some way, to lead to the observed relationship. Figure 1 previously illustrated that both the volume of closures and ratio (as a percent of closed and zero balance cards) increases in 9

12 the time series in response to bad news. When examining the extensive margin of paying down a credit card (whether subsequently keeping it at zero balance or closing) in the cross-section, I find that the likelihood either does not materially change according to HPA or increases according to unemployment shocks (Details in Appendix Table I.I). For the remainder of this analysis, for purposes of brevity, I restrict myself to examining changes according to trailing house prices. Results are similar when examining the stronger relationship to unemployment shocks. 3.1 Addressing Supply-side Confounds In all analysis, I restrict the sample to high credit quality, prime borrowers to address the issue that lenders could selectively choose risky accounts to close in some pre-emptive fashion. Banks did not generally have regular approaches to close zero balance unsecured revolving accounts in response to local geographic economic conditions such as house prices prior to the crisis. If we thought a significant portion of the observed credit card closures were driven by lenders monitoring risk, this should be captured in the estimated coefficient on lagged credit scores, but Table 2 estimated that closures were positively related to lagged credit scores. To fully separate out any potential for supply-driven closures I turn to surveys of household behaviour, and before that repeat analysis with stronger controls in Table 4. Economic magnitudes of results are similar for HPA shocks, and stronger for unemployment shocks, when non-parametrically allowing for flexible relationships to time period. 3.2 Survey Evidence Using the credit access survey component of the 2016 SCE, I find that voluntary credit closures by self-reported higher credit score borrowers (680 and above) are 9 times as prevalent as lender-driven closures % state voluntarily closing an account in the last 5 When answering this question, please consider all kinds of credit you and your spouse/partner have, including credit cards (including retail/store cards), mortgages, home-based loans (such as home equity 10

13 12 months, which is remarkably on top of my CCP-based estimate of 16.5% for borrowers 700 credit score and above. The SCE number is positively biased as it includes terminations of other, more long-lived, loans such as mortgages. While the closure question in the survey is backwards looking, and surveyed expectations are forward looking, I also find statistically significant positive correlations between the expected probability of a future need for an unexpected $2,000 expense and voluntarily closing an account. 6 A 1% more likely unexpected $2,000 expense is related to a 0.07% greater likelihood of closing an account. This seems to be, on the surface, a breakdown in the standard model or the joint hypothesis of rational expectations and income smoothing. It is consistent with some subjectively expected costs to the recourse nature of debt (default costs) and neglected default risk in normal times, that outweighs any of consumption smoothing benefits. There is a long list of empirical evidence suggesting that households make various avoidable mistakes in managing their finances, for example prepaying optimally, remembering to make monthly payments, or in renegotiation of defaulted loans (Campbell, 2006; Hundtofte, 2012, 2015). 57% of people are taught how to manage their finances from their parents (most common source of learning other than life experience ) What were you taught as a child about debts? A promise you should keep 65% A useful way to shift money from periods when you have it to periods when you do not 8.6% A useful way to split up payments on a high cost item that would be useful to you sooner rather than later 27% Something to be avoided 52% A necessary evil 29% Something to be used in moderation - considered carefully and paid off when possible 70% lines of credit), auto loans, student loans as well as all other personal loans. In the past 12 months, did any of the following happen?... Closed an account Y/N 6 What do you think is the percent changes that you could come up with $2,000 if an unexpected need arose within the next month? 11

14 3.3 Within-Person Analysis Holding the individual constant, the analysis of Table 5 confirms the earlier results that closures vary by economic conditions with similar magnitude to changes in economic conditions. Are individuals who previously close a card and subsequently miss a payment on a debt less likely to close a credit card in the future? Table 6 finds that individuals are less likely to close a card after previously closing a card and missing a payment on one of their other debts. I use a subsequent delinquency in the next year after paying a credit card off as a proxy for having experienced a binding liquidity constraint, which the option to borrow could have alleviated. Consistent with both individual level fixed effects and with there being some change in behaviour, the coefficient on the interaction of having previously closed a card and subsequently experiencing delinquency is negative on closing a card in the future, albeit very small (0.3% less likely) controlling for age and credit score. 3.4 Costs and Benefits Associated With Closure Table 7 examines the credit outcomes associated with card closure in the next quarter, essentially trying to compare individuals who are the same age and credit score in the same zipcode at the same time, by whether one closes a credit card after paying it down. Clearly these two borrowers can be different along unobservable dimensions, such as in their type (sophistication) or income. Mechanically, we know there must be costs imposed on credit score, and that tighter liquidity constraints can only lead to a weakly positive effect on new delinquencies. The CCP is available on a quarterly basis, and while other changes could occurring in the intervening time between closure and the end of a quarter. According to these unobservable differences, the point estimates are in the directions as if there were no meaningful unobservables, and if anything possibly muted in magnitude according to what might otherwise be expected based on closure alone. This could also arise from noise in the ap- 12

15 proximation of credit card closure. Closing a credit card is associated with -5 points of credit score (or 1/8 th a standard deviation), and a 12 bps greater chance of experiencing a new delinquency, compared to a base rate of 98bps in this sample. Table 8 examines credit outcomes 4 quarters later, limiting analysis to the first zerobalance event for each individual to aid clarity. Any estimated changes are net any subsequent credit choices and activity by the individual. On this smaller sample, the relationship to new delinquencies and non-current balances are statistically insignificant, while the average relationship to credit score is a small positive change of Conclusion In this paper I have documented the prevalence of credit aversion by U.S. households, and noted that the propensity to voluntarily close a credit card is greater in the presence of negative local economic shocks. Voluntary closures of costless unsecured lines of credit are anomalous to the standard model. That they appear so prevalent in the data suggests meaningful frictions or biases. Closing credit cards precisely when the marginal returns to credit should be at their greatest is difficult to reconcile with constant frictions. It could be that procyclical credit closures occur because of the institutions or types of contracts available to borrowers, even in a developed country such as the U.S. Credit cards might not be the appropriate means to access capital markets for consumption-smoothing purposes. But then we must ask ourselves why many households are happy to engage in such high interest rate borrowing, if not for utility gains from smoothing transitory shocks. More than half of individual-month level observations show positive credit card balances, at a non-negligible average greater than $6,000. An open question in the household finance literature is whether the general amount of credit available to borrow is too little, too much, or just right (Zinman, 2014). It appears that when the future looks bleak, borrowers often decide the answer is too much. This 13

16 suggests that in addition to monitoring lending conditions we should also monitor households willingness to consume using borrowed money. Self-imposed financial constraints are an understudied phenomenon compared to lender-imposed constraints. The apparent unwillingness to even hold an option to borrow suggests additional targets for policy interventions in recessionary or crisis periods. 14

17 5 Tables and Figures Figure 1 Illustration of Credit Card Closures in the Time Series The blue line (left axis) represents the proportion of individuals closing at least one bank card out of those that either closed or were holding a zero balance at the end of each quarter. The red bars (right axis) indicate the total number of individuals who closed one (or more) bank cards each quarter. The sample is and restricted to borrowers with credit score of 700 or above in the period prior to adding a zero balance credit or closing a card, using the 0.1% CCP sample (results are scaled to present national estimates). Total Closures 0 10,000,000 20,000,000 30,000,000 Quarterly Credit Card Closures (High Credit-Scored Households) Year % Closed Total Closures % Closed 15

18 Table 1 Summary Statistics This table presents summary statistics of the CCP, using a sample size of 0.1% of U.S. households. Observations are at the individual-quarter level. The analysis sample is conditional on adding 1 or more zero balance credit cards or closed accounts. (1) All of sample (2) Analysis Sample Mean Std Dev Mean Std Dev Min Max Age Median Home Price ($) 207, , , ,485 15,595 2,350,000 Last 12m HPA Per-capita Income 30,705 9,347 32,088 9,933 10,156 88,640 Unemployment Rate Unemployment Rate Credit Score Total Debt ($) 60, , , , ,183,778 Non-performing Debt ($) 3,576 32, , ,061,887 Number of Cards Closed Card (0/1) Observations 17,298,335 1,707,184 16

19 Table 2 Propensity to Close a Credit Card This table presents Linear Probability Model estimates of credit card closures. All regressions are estimated with unweighted OLS using panel data. The unit of observation is a single individual s quarterly credit report from the FRBNY/Equifax CCP. The dependent variable is a binary indicator variable, which equals 1 if one or more bank cards were closed by that individual, and 0 otherwise. Analysis is conditional on adding one or more zero balance bank card tradelines that quarter. Age and Credit Attributes are at the individual level, all other demographics at the county level. Avg. Per-capita Income is BEA average for that county in Median Home Price is the Core Logic reported median for the last month in each quarter for that county. Unemployment Rate is the BLS reported statistic for that county-quarter. Time Controls are a second order polynomial for the sequential quarter of observation. (1) (2) (3) (4) Closed Credit Card Age ( ) ( ) ( ) ( ) Lag Credit Score ( ) ( ) ( ) ( ) Lag # Cards ( ) ( ) ( ) ( ) Lag CC Debt Current ($1000s) ( ) ( ) ( ) ( ) Lag CC Debt Non-current ($1000s) (0.0022) (0.0019) (0.0020) (0.0020) Lag Total Debt ($1000s) e-7 4.3e e-06 (2.5e-06) (3.0e-06) (2.9e-06) (2.8e-06) Lag Total Debt Non-current ($1000s) ( ) ( ) ( ) ( ) Median Home Price ($1000s) (7.0e-06) (5.6e-06) (4.7e-06) Avg. Per-capita Income ($1000s) ( ) ( ) ( ) Unemployment Rate (0.034) (0.042) (0.039) 1{Y ear [2008, 2010]} (0.0015) Time Controls (2nd order polynomial) Y Y Constant (0.013) (0.014) (0.013) (0.013) N 1,445,565 1,228,826 1,228,826 1,228,826 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p < 0.001

20 Table 3 Relationship of Closures to Changes in Household Beliefs This table presents Linear Probability Model estimates of the extensive margin (having a new zero balance or closed credit card) amongst all borrowers meeting minimum credit score criteria. The unit of observation is the individual-quarter of credit reporting from the FRBNY/Equifax CCP. The dependent variable is a binary indicator variable, which equals 1 if an individual closed or held a new zero balance credit card, 0 otherwise. Trailing 12m HPA is the House Price Appreciation in that county for the last year. Unemployment Rate is instrumented using Bartik-style county-industry share shocks. Constant omitted. Trailing 12m HPA (1) (2) (3) (4) Closed Credit Card (0.0052) (0.0056) UnemploymentRate (0.085) (0.13) 1{Y ear [2008, 2010]} (0.0019) (0.0036) Age and Lagged Credit Attributes Y Y Y Y Time period 2nd order poly. Y Y Y Y N 1,161,272 1,161,272 1,100,878 1,100,878 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p <

21 Table 4 Robustness: More Flexible Controls for Time Period This table repeats the analysis of Table 1, estimating Linear Probability Models of credit card closures. The unit of observation is a single individual s quarterly credit report from the FRBNY/Equifax CCP. The dependent variable is a binary indicator variable, equalling 1 if one or more bank cards were closed by that individual, 0 otherwise. Analysis is conditional on adding one or more zero balance bank card tradelines that quarter. Column (1) uses binary indicators to flexibly control for time-varying relationships. Column (2) ignores the crisis years of , parametrically controlling for time period as before. Trailing 12m HPA is the House Price Appreciation in that county for the last year. Unemployment Rate is instrumented using Bartik-style industry county-share shocks. Age and Credit Score are measured at the individual level, all other demographics at the county level. Trailing 12m HPA (1) (2) (3) (4) Closed Credit Card (0.0064) (0.0064) UnemploymentRate (1.31) (0.99) Age and Lagged Credit Attributes Y Y Y Y Year FEs Y Y Y Y County-level controls Y Y Zipcode FEs Y Y N 1,161,272 1,163,320 1,100,878 1,227,421 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p <

22 Table 5 Within-person Analysis This table estimates Linear Probability Models of credit card closures, with individual fixed effects. The unit of observation is a single individual s quarterly credit report from the FRBNY/Equifax CCP. The dependent variable is a binary indicator variable, equalling 1 if one or more bank cards were closed by that individual, 0 otherwise. Analysis is conditional on adding one or more zero balance bank card tradelines that quarter. Trailing 12m HPA is the House Price Appreciation in that county for the last year. Age and Credit Score are measured at the individual level, all other demographics at the county level. (1) (2) Closed Credit Card Trailing 12m HPA (0.0057) (0.0074) Individual FEs Y Y Age and Lagged Credit Attributes Y Y Time controls: 2nd order poly. Y Time controls: Year FEs Y N 1,161,272 1,161,272 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p <

23 Table 6 Learning Out of It? Correlations with Previous Closure Experiences Event study around zero balance credit card events (either new zero balance or credit card closed within quarter). Subsequent Delinquency Previously Closed Zero Balance Previously Closed * Delinquency Age and Lagged Credit Attributes Zip-Year FEs (1) Closed Card (0.0055) ( ) ( ) Y Y N 1,196,033 R Standard errors clusted by county in parentheses p < 0.05, p < 0.01, p <

24 Table 7 Immediate Changes in Credit Outcomes This table estimates Linear Probability Models of credit card closures, with individual fixed effects. The unit of observation is a single individual s quarterly credit report from the FRBNY/Equifax CCP. The dependent variables are the change in credit score (in points) and the change in delinquent accounts. Closed Credit Card is a binary indicator variable, equalling 1 if one or more bank cards were closed by that individual that quarter, 0 otherwise. Analysis is conditional on adding one or more zero balance bank card tradelines that quarter. Trailing 12m HPA is the House Price Appreciation in that county for the last year. Age and Credit Score are measured at the individual level, all other demographics at the county level. (1) (2) Credit Score Delinquency Closed Credit Card (0.064) ( ) Age and Lagged Credit Attributes Y Y Zip-Year FEs Y Y N 1,445,452 1,445,565 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p <

25 Table 8 Long-run Changes in Credit Outcomes This table estimates Linear Probability Models of credit card closures, with individual fixed effects. The unit of observation is a single individual s credit report from the FRBNY/Equifax CCP four quarters after the first zero balance or credit card closure event. The criteria for selecting borrowers only above 700 credit score is applied prior to the event only. The dependent variables are the presence of a delinquency (non-current) account, the change in credit score (in points), and change in credit card borrowing. Closed Credit Card is a binary indicator variable, equalling 1 if one or more bank cards were closed by that individual that quarter, 0 otherwise. Credit Attributes consist of credit score, number of credit cards, total credit card debt (current, non-current), and total debts (current, non-current). (1) (2) (3) Outcome 1 year later New Delinquency Non-Current Balance ($) Credit Score Closed Credit Card (0.0020) (219.4) (0.52) Lagged Credit Attributes & Age Y Y Y Zip-year FEs Y Y Y N 130, , ,609 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p <

26 I. Appendix Table I.I Extensive Margin of Zero Balance and Closed Credit Cards This table presents Linear Probability Model estimates of credit card closures. All regressions are estimated with unweighted OLS using panel data. The unit of observation is the individual-quarter of credit reporting from the FRBNY/Equifax CCP. The dependent variable is a binary indicator variable, which equals 1 if one or more bank cards were closed by that individual, 0 otherwise. Analysis is conditional on adding one or more zero balance bank card tradelines that quarter. Trailing 12m HPA is the House Price Appreciation in that county for the last year. Unemployment Rate is instrumented using Bartik-style county-industry share shocks. Age and Credit Score are measured at the individual level, all other demographics at the county level. Trailing 12m HPA UnemploymentRate (1) (2) Zero Balance (Including Closed) Credit Card (0.0031) 1.17 (0.51) Age and Lagged Credit Attributes Y Y Year FEs Y Y County FEs Y Y N 5,124,154 5,367,618 R Standard errors clustered by county in parentheses p < 0.05, p < 0.01, p <

27 References Angeletos, G.-M., D. Laibson, A. Repetto, J. Tobacman, and S. Weinberg (2001). The hyperbolic consumption model: Calibration, simulation, and empirical evaluation. Journal of Economic Perspectives, Armantier, O., W. Bruine de Bruin, G. Topa, W. Klaauw, and B. Zafar (2015). Inflation expectations and behavior: Do survey respondents act on their beliefs? Economic Review 56 (2), International Armantier, O., G. Topa, W. van der Klaauw, and B. Zafar (2013). Introducing the frbny survey of consumer expectations: Survey goals, design and content. ht tp://libertystre eteconomics.newyorkfed.org/2013/12/introducing-the-frbny-survey-ofconsumer-expectations-survey-goals-design-and-content.html. Bertaut, C. C., M. Haliassos, and M. Reiter (2009). Credit card debt puzzles and debt revolvers for self control. Review of Finance 13 (4), Cadena, B. C. and B. J. Keys (2013). Can self-control explain avoiding free money? evidence from interest-free student loans. Review of Economics and Statistics 95 (4), Callender, C. and J. Jackson (2005). Does the fear of debt deter students from higher education? Journal of social policy 34 (4), Campbell, G., A. Haughwout, D. Lee, J. Scally, and W. van der Klauuw (2016). Just released: Recent developments in consumer credit card borrowing. Federal Reserve Bank of New York Liberty Street Economics (blog) libertystreeteconomics.newyorkfed.org/ 2016/ 08/ just-released-recent-developments-in-consumer-credit-card-borrowing.html. Campbell, J. Y. (2006). Household finance. The Journal of Finance 61 (4),

28 CFPB (2013). Card act report. Washington: CFPB (October). Available at: files. consumerfinance.gov/ f/ cfpb card-act-report.pdf. Chakrabarti, R., D. Lee, W. Van der Klaauw, and B. Zafar (2013). Household debt and saving during the 2007 recession. In Measuring wealth and financial intermediation and their links to the real economy, pp University of Chicago Press. Deaton, A. S. (1991). Saving and liquidity constraints. Ecomometrica 59 (5), Hundtofte, S. (2012). Working Paper. Promoting automation improves borrower repayment outcomes. Hundtofte, S. (2015). No such thing as a free option? mortgage modification offers under imprecise borrower beliefs. Working Paper. Hundtofte, S. and M. Pagel (2017). Credit smoothing. Ivashina, V. and D. Scharfstein (2010). Bank lending during the financial crisis of Journal of Financial economics 97 (3), Keynes, J. M. (1936). The general theory of employment, investment, and money. London and New York, quoted from: google. com/site/biblioeconomicus/keynesjohnmaynard-thegeneraltheoryof EmploymentInterestAndMoney. pdf. Kiyotaki, N. and J. Moore (1995). Credit cycles. Report, National Bureau of Economic Research. Kuchler, T. and M. Pagel (2017). Sticking to your plan: The role of present bias for credit card paydown. Lee, D. and W. Van der Klaauw (2010). An introduction to the frbny consumer credit panel. 26

29 Meier, S. and C. Sprenger (2010). Present-biased preferences and credit card borrowing. American Economic Journal: Applied Economics 2 (1), Minsky, H. P. (1977). The financial instability hypothesis: An interpretation of keynes and an alternative to standard theory. Challenge 20 (1), Minsky, H. P. (1992). The financial instability hypothesis. Prelec, D. and G. Loewenstein (1998). The red and the black: Mental accounting of savings and debt. Marketing science 17 (1), Prelec, D. and D. Simester (2001). Always leave home without it: A further investigation of the credit-card effect on willingness to pay. Marketing letters 12 (1), Schiller, R. J. (2000). Irrational exuberance. Princeton, New Jersey, Princeon University. Thaler, R. H. and H. M. Shefrin (1981). An economic theory of self-control. Journal of political Economy 89 (2), Zeldes, S. P. (1989). Consumption and liquidity constraints: an empirical investigation. Journal of political economy 97 (2), Zinman, J. (2014). Consumer credit: Too much or too little (or just right)? The Journal of Legal Studies 43 (S2), S209 S