Household Balance Sheets, Consumption, and the Economic Slump

Household Balance Sheets, Consumption, and the Economic Slump Atif Mian University of California, Berkeley and NBER Kamalesh Rao MasterCard Advisors Amir Sufi University of Chicago Booth School of Business and NBER June 2012 Abstract We provide evidence that high levels of household debt in combination with the collapse in house prices was a primary factor in the onset and severity of consumption collapse from 2006 to 2009. Using novel county-level retail sales data, we show that the decline in household spending was much stronger in high leverage counties with large house price declines, and this decline was not offset by higher spending in low leverage counties. High leverage counties de-leveraged at a faster pace, experienced a larger drop in credit limits, and had difficulty refinancing mortgages into lower rates. The effect of declining house prices on household spending was much stronger in zip codes with higher leverage or lower net worth. The evidence suggests that household debt should play an important role in models trying to understand the large decline in aggregate consumption during the Great Recession. * Lucy Hu, Ernest Liu, Christian Martinez, Yoshio Nozawa, and Calvin Zhang provided superb research assistance. We are grateful to the National Science Foundation, the Initiative on Global Markets at Chicago Booth, and the Fama-Miller Center at Chicago Booth for funding. Seminar participants at Chicago Booth, Columbia Business School, Harvard, MIT Sloan, MIT Economics, NYU Stern, U.C. Berkeley, and the NBER Monetary Economics meeting provided valuable feedback. The results or views expressed in this study are those of the authors and do not reflect those of the providers of the data used in this analysis. Corresponding authors: Mian: (510) 643 1425, atif@haas.berkeley.edu; Sufi: (773) 702 6148, amir.sufi@chicagobooth.edu

The recent economic slump in the U.S. has been characterized by a highly indebted household sector and a large drop in aggregate consumption. Figure 1 shows these facts for the Great Recession. As the left panel shows, the household debt to income ratio rose sharply to historically unprecedented levels before 2007. The right panel shows a sharp decline in retail sales during the recession. These patterns have spurred a series of theoretical papers in macroeconomics that emphasize the role of household leverage in reducing aggregate consumption, output, and employment (Eggertsson and Krugman (2012), Guerrieri and Lorenzoni (2011), Hall (2011), Midrigan and Philippon (2011)). While different in some respects, these models share three important ingredients. First, a relaxation in credit constraints or rise in collateral value creates a group of highly indebted households. Second, a shock to debtors balance sheets such as falling house prices or credit tightening forces them to cut back on consumption and to start paying down debt. Third, in the absence of an equivalent increase in consumption by unlevered households, there is a drop in aggregate consumption. The failure of unlevered households to increase consumption is typically related to nominal price rigidities including the zero lower bound on nominal interest rates. This household balance sheet explanation of the recession has distinct implications for our understanding of the economy and the appropriate policy response. However, empirically evaluating the household balance sheet channel is not possible without detailed microeconomic data on household spending. In particular, one needs to separately track the consumption response of levered and unlevered households when the economy is hit with a shock. In this study, we formally test the household balance sheet explanation for the decline in aggregate consumption by utilizing new data on household spending. The data cover a measure of all household purchases at the county level and new automobile purchases at the more 1

disaggregated zip code level. We find strong qualitative and quantitative support for the hypothesis that household debt accumulated from 2002 to 2006 in combination with the fall in house prices was responsible for the onset and severity of the subsequent consumption collapse. Our empirical strategy exploits the large amount of variation across U.S. counties in the accumulation of household debt prior to the slump. From 2002 to 2006, Mian and Sufi (2011) show that homeowners responded to higher house price growth in inelastic housing supply counties by aggressively borrowing against home equity. This aggressive borrowing increased leverage in inelastic counties substantially. As a result, in 2006, there were large differences across U.S. counties in accumulated debt, measured by the household debt to income ratio. Counties with high household debt as of 2006 experienced a severe shock to their balance sheets between 2006 and 2009. House prices in the highest decile of the household debt to income distribution declined by almost 40% from 2006 to 2009, whereas house prices in the lowest decile fell by less than 10%. This dramatic decline in house prices in the high leverage counties wiped out a large fraction of the net worth of households indeed many households ended up under water with negative home equity. The models described above hypothesize that the aggregate consumption decline is driven by the differential negative shock to highly levered households in combination with frictions that prevent full offset from unlevered households. We find strong support for this view. Using county-level data on auto purchases from R.L. Polk and expenditure by merchant type from MasterCard Advisors, we find that counties with high leverage experienced a much larger decline in household spending across all types of goods, including even groceries. For example, a one standard deviation increase in household leverage as of 2006 was associated with a ten percentage point additional drop in spending on durable goods and a five percentage point 2

additional drop in non-durable consumption. Moreover, there were no large increases in consumption in the least levered counties to compensate for the drop in consumption in high leverage counties. We also show evidence of high leverage counties paying back debt at a faster pace from 2006 onwards. Using individual level data on consumer borrowing, we show that highly levered households in 2006 experienced a significantly larger decline in debt balances between 2007 and 2010, even when we condition on the household not defaulting. The faster pace of deleveraging in counties with higher debt is consistent with the explanation for the severity of the recent recession given in Eggertsson and Krugman (2012) and Guerrieri and Lorenzoni (2011). Along with deleveraging, homeowners in high leverage counties were unable to refinance their mortgages into historically low interest rates from 2008 to 2010. Further, home equity credit availability for homeowners in high leverage counties declined from 2007 to 2010. These findings are consistent with the argument in Midrigan and Philippon (2011) that a reduction in liquidity services from housing was a key reason for the decline in household spending. Since counties with high household debt also experienced a larger decline in house prices, one could argue that the drop in consumption in high leverage counties is entirely driven by a pure wealth effect of the decline in house prices. However, the magnitude of the elasticity of household spending with respect to house price declines is too large to be justified by a pure wealth effect argument; instead, they suggest an important role of leverage in amplifying the effect of falling house prices. Further, we exploit within-county zip code level variation in automobile purchases and household balance sheets to examine more carefully the interactive effect between leverage and house price declines. We find that house price declines lead to much larger cutbacks in 3

household spending among highly levered households. The estimated elasticity of auto purchases with respect to house price growth is three times larger for households in the 90th percentile of the leverage distribution relative to households in the 10th percentile. In other words, within a county, levered households cut back spending by much more for a given drop in house prices, consistent with the amplification role of debt. Could the large drop in consumption in high leverage counties with large house price drops have been driven by a local macro shock such as a shock to expectations of future productivity that drove both house prices and consumption down? If this were the case, then the correlation of household debt with spending declines may be spurious. However, we show that the timing of the effects was more consistent with the importance of housing and debt, and less consistent with alternative views based on expectation shocks that were unrelated to these factors. For example, the initial decline in house prices and the initial increase in default rates occurred in 2006 in high leverage counties--well before other aggregate or local economic variables moved. In fact, in high leverage areas, problems related to debt and housing emerged a full four quarters before employment responded. Similarly, as described above, within a given county--where expectations shocks should be similar--the spending decline was much larger among highly levered households. We view our central contribution as providing evidence in support of the idea that elevated levels of household debt in combination with a shock to asset values can generate a significant decline in aggregate consumption. While this idea goes back at least to Fisher (1933), it has not played a prominent role in macroeconomic models until recently. Other studies providing empirical support this view are Glick and Lansing (2009, 2010), King (1994), Mishkin (1978), and Olney (1999). Our analysis here is most closely related to Mian and Sufi (2010) and 4

Midrigan and Philippon (2011) who examine spending patterns across the U.S. during the Great Recession. Relative to these studies, we use a novel measure of geographically disaggregated consumption, which allows us to estimate consumption elasticities for a broader set of goods for almost the entire population. Further, the zip code level data on auto sales is new and allows for a finer test of the household balance sheet view of the consumption collapse. The rest of the analysis proceeds as follows. The next section presents the theoretical motivation. Section 2 discusses the data and summary statistics. Section 3 presents the household spending results, and Section 4 shows the importance of debt in explaining the decline in spending. Sections 5 and 6 discuss alternative hypotheses for the results. In Section 7, we conclude. 1. Theoretical Motivation Figure 1 summarizes the key motivating facts behind this paper. The years from 2002 to 2006 saw an unprecedented buildup in household debt relative to income followed by a sharp collapse in household spending. House prices also plummeted after 2006. Is there a connection between the increase in debt, the fall in house prices, and the collapse in aggregate consumption? In this section we discuss the theoretical work that links changes in aggregate consumption to household leverage and house price shocks. Eggertsson and Krugman (2012) build a model where an unexpected shock to the borrowing capacity of borrowers forces them to de-lever and cut back on current consumption. The authors think of the shock as a sudden realization that assets were overvalued... which corresponds to the house price declines observed in the data. Should lower consumption by borrowing households lead to a decline in aggregate consumption? In standard neo-classical 5

models, the answer is no because a reduction in spending acts as a positive savings shock which puts downward pressure on prices and interest rates. As a result, declines in prices and interest rates induce spending by the unlevered households that compensates for the fall in consumption by borrowers. However, Eggertsson and Krugman (2012) point out that in the presence of nominal price rigidity and a zero lower bound for nominal interest rates, it may not be possible to compensate for the fall in consumption by borrowers. If the initial level of household leverage and the decline in borrowing capacity are large enough, the economy will be stuck in a liquidity trap, characterized by reduced household spending and zero nominal interest rates. 1 Guerrieri and Lorenzoni (2011) present a model in which borrowing by debtors is driven by a desire to smooth idiosyncratic income shocks. Agents with negative prior idiosyncratic shocks have higher debt levels given a desire to smooth consumption over these shocks. An unexpected shock to the borrowing limit forces debtors to reduce consumption due to deleveraging and a need for higher precautionary savings. 2 Midrigan and Philippon (2011) present a model where housing serves the dual purpose of a consumption good and a transactions role given a cash in advance constraint. A shock to the ability of households to extract equity from their homes forces levered households to cut back on consumption. They explain how frictions related to labor mobility across sectors and nominal price adjustment can translate the decline in spending into a severe recession. 1 See also Hall (2011) for a similar argument where zero lower bound on nominal interest rates prevents real interest rates from turning sufficiently negative to clear the goods market. 2 Guerrieri and Lorenzoni (2011) also allow levered households to increase their labor supply in response to the shock; they show that under reasonable parameters, levered households will choose to reduce consumption instead of increasing labor supply. For more on the employment implications of our findings here, see Mian and Sufi (2012). 6

While the zero lower bound on nominal interest rates plays a prominent role in some general equilibrium models of aggregate demand, others assign an important role for frictions in labor force adjustment, labor mobility, and nominal wage rigidities. We do not take a stand on which of these frictions is more relevant. Instead, our focus is on the common prediction of the models mentioned above: in response to an unexpected shock to housing collateral, the decline in spending will be stronger among levered households, and the fall in consumption is not compensated for by an equivalent increase by non-borrowing households. It is important to emphasize the underlying economic mechanism that leads to reduced spending in these models. In all three models, the decline in house prices represents a shock that disproportionately affects levered households by introducing a constraint in their optimal intertemporal Euler equation. This tightened constraint generates a sharp pull-back in spending today. An alternative view is that house price declines could affect household spending through what we call a "levered wealth effect". In this alternative view, households are not pushed off their Euler equation by the shock. Instead, lower values of home equity represent lower resources for consumption in the future, and as a result households respond to the lower value of home equity by reducing spending in all periods going forward. 3 This effect is amplified by higher leverage given that highly levered households start with lower home equity. We will show evidence below that a pure housing wealth effect without leverage is refuted by the data, but we do not attempt to distinguish between the constraints view and the levered wealth effect view given that both have very similar cross-sectional implications. 4 3 One obvious problem with the housing wealth effect argument is that housing is a consumption good that must be purchased going forward. In the aggregate, a decline in house prices should not be viewed as a net wealth effect given that many households that are short housing benefit from lower prices (Campbell and Cocco (2007), Sinai and Souleles (2005)). To the degree that a levered wealth effect exists, it must be because of the distribution of leverage across households. 4 Separating these views would likely require household level data on balance sheet strength and household spending, which we do not have for our analysis. It is therefore material for future research. 7

2. Data and Summary Statistics A. Data Our goal is to test the empirical validity of the models described above. The key data requirement is the ability to follow consumption patterns separately for households with varying degrees of leverage. Historically, consumption data has only been available either at the aggregate level, or at more micro level based on survey responses. While survey data are useful, they are typically based on very small samples, and there have been concerns raised about the accuracy of survey data. 5 This study introduces new sources of consumption data based on actual household expenditure (as opposed to survey responses). The first is zip code level auto sales data from R.L. Polk covering 1998 to 2011. These data are collected from new automobile registrations and provide information on the total number of new automobiles purchased in a given zip code and year. The address is derived from registrations, so the zip code represents the zip code of the person that purchased the auto, not the dealership. The second source of consumption data is at the county level from 2005 to 2009 from MasterCard Advisors. These data provide us with total consumer purchases in a county that use either a credit card or debit card for which MasterCard is the processor. The data are based on a 5% sample of the universe of all transactions from merchants in a county. An important advantage of the MasterCard data is that they break down total consumer expenditure by the NAICS code attached to the merchant providing the data. There are ten categories for merchants we use: furniture, appliances, home centers (i.e., home improvement), groceries, health-related 5 See for example Attanasio, Battistin, and Ichimura (2007) and Cantor, Schneider, and Edwards (2011) for criticism of the Survey of Consumer Expenditure in particular. 8

such as pharmacies and drug stores, gasoline, clothing, sports and hobby, department stores, and restaurants. 6 We group the MasterCard purchases into three categories: durable goods (furniture, appliances, home centers), groceries, and other non-durable goods (all remaining categories). In the appendix, we provide further detail on the MasterCard data and how it compares to the aggregate retail sales information from the Census. We also address concerns that consumption patterns using credit card and debit card purchases may affect inference on the consumption declines in high versus low debt counties. In this regard, it is useful to keep in mind that our auto sales data from R.L. Polk represent the universe of all auto purchases and can therefore be used as a cross-check on the results using MasterCard data. Further, as we show in the appendix, we find quantitatively similar results if we use state-level sales tax revenue data from the Census as our measure of household spending. As we explain in the appendix, the bottom line is that we believe that results using the MasterCard measures of retail sales are not systematically biased relative to the results we would obtain if we had the geographic micro data underlying the Census retail sales aggregate data. We augment our county and zip code level data on consumption with equally granular data on household debt, house prices, and other demographics. Given the large number of data sources that we have put together for our analysis, we have placed much of the data discussion in the appendix. The appendix includes information on the data sets that are publicly available and the contacts for buying data that we cannot share by license agreement. Measures of household balance sheets, including debt, defaults, and home equity limits, come from Equifax Predictive Services. These data are available at the zip code level and are 6 These correspond to 3-digit NAICS codes of 442, 443, 444, 445, 446, 447, 448, 451, 452, and 722, respectively. For more information on the exact types of stores included in each NAICS, see http://www.naics.com/free-codesearch/sixdigitnaics.html?code=4445. These categories are identical to those used by the Census measures of retail sales. 9

fully described in Mian and Sufi (2009). We also have a second data set available from Equifax that includes anonymous individual level data based on a random sample of 266,005 individuals taken in 1998 and followed through 2010. These data are fully described in Mian and Sufi (2011). The individual level data allow us to split individuals by default status. In the appendix, we discuss an issue related to the total debt numbers in Equifax. In the data we have been provided, debt is sometimes "double-counted" given joint accounts. As Lee and van der Klaauw (2010) show, the Federal Reserve Bank of New York Consumer Credit Panel is also based on the Equifax data, but this problem is corrected in their data. We have used the FRBNY data to correct for the double-counting problem in our data, a process we describe in the appendix. We have also made publicly available the county-level debt to income ratio using the FRBNY data to measure debt. This is also described in the appendix. Zip code- and county-level data on house prices are from CoreLogic. Zip code-level data on the amount of mortgages refinanced are from HMDA. Information at the state level on the number of mortgages underwater is from CoreLogic. County-level information on new residential investment is from the Census, and county-level information on employment in the construction sector is from the Census County Business Patterns. B. Summary statistics We combine all of the data described above into a county-year level data set. Table 1 presents summary statistics. The household debt to income ratio as of 2006 was on average 1.6 across counties in our sample. House prices declined on average by 16% from 2006 to 2009. The sharp economic downturn from 2006 to 2009 is reflected in spending patterns. When we weight by county population, auto sales on average dropped by 48%, other durables purchases dropped 10

by 17%, groceries grew by 12%, and other non-durables grew by 4%. Default rates jumped sharply from 2006 to 2009, and home equity limits declined by 3.1%. Finally, Table 1 shows that there is a substantial amount of variation across the sample counties in terms of population; there are only five thousand people at the 10th percentile and 191 thousand at the 90th percentile. Given a number of very small counties, we weight all of our regressions by the total population as of 2006. The last two columns give the weighted mean and weighted standard deviation for all variables. 3. Household Leverage, the House Price Shock, and the Consumption Decline The household balance sheet models of the economic slump described in Section 1 start with heterogeneity in debt levels in the household sector. The subsequent collapse in consumption is driven by a negative balance sheet shock that disproportionately affects levered relative to unlevered households. The key testable prediction of the models is that high leverage households cut back on spending the most when the balance sheet shock materializes, and there is no equivalent increase in consumption by unlevered households. In this section, we empirically evaluate this prediction using county-level data on spending through the recession. However, before conducting this test, we provide a brief discussion of two important issues. First, how did some counties end up with higher leverage as of 2006? And second, what was the precise shock that hit levered households? A. What explains the variation in leverage across counties? We use geographic variation across counties in accumulated leverage measured by the debt to income ratio as of 2006 to test whether more levered households cut back relatively more on consumption during the recession. The debt to income ratio is a widely used measure of 11

household leverage in both models and practice. Indeed, this is precisely the measure used in the Midrigan and Philippon (2011) model to measure the degree to which households were affected by the decline in house prices. Nonetheless, the use of income in the denominator is not critical; as we show in the appendix, all the results are similar if we scale debt by total population of the county instead of income. Table 1 shows that there is a large amount of variation in the 2006 debt to income ratio across U.S. counties: the ratio is less than 1 at the 10 th percentile and 2.3 at the 90th percentile. What is the source of variation in county-level debt to income ratio in 2006? In other words, why did households in certain counties borrow a lot more than households living in other counties? Mian and Sufi (2009) show that the increase in household debt in the U.S. was driven by an expansion in the supply of credit (e.g., relaxation in lending standards) and not an improvement in credit demand conditions (e.g., higher productivity). Mian and Sufi (2011) exploit geographic variation in housing supply elasticity across CBSAs to explain cross-sectional variation in household debt. Taken together, these two studies suggest that the expansion in the availability of credit led to higher house price growth during the 2002 to 2006 period in counties with more inelastic supply of housing due to geographical factors. The sharp increase in house prices in more inelastic counties led to large scale home-equity based borrowing that dramatically increased homeowners debt to income ratio in these counties. Mian and Sufi (2011) provide an extensive analysis of this channel, and also provide evidence for the validity of housing supply elasticity as an instrument for house price growth and home-equity based debt growth from 2002 to 2006. Columns 1 and 2 of Table 2 summarize these findings by regressing house price growth from 2002 to 2006 and the debt to income ratio in 2006 on the housing supply elasticity measure 12

introduced by Saiz (2010). There are two important takeaways. First, one could potentially use housing supply elasticity as an instrument for the county level debt to income ratio when testing whether more levered households cut back more on spending. The advantage of doing so is that one knows the precise channel that is generating the variation in household leverage. We report the instrumental variables version of all our tests that use debt to income ratio in the appendix. We also discuss the relative merits of the instrumental variables approach there. 7 Second, cross-sectional variation in house price growth and leverage growth over the 2002 to 2006 period was driven by a common factor. There is therefore a high degree of correlation between house price growth from 2002 to 2006 and leverage accumulated by 2006 (correlation of 0.51). Since counties with the largest house price increase from 2002 to 2006 also suffered the largest decline in house prices afterwards, there is also a strong correlation between household leverage in 2006 and house price decline from 2006 to 2009 (correlation of -0.64). This implies that more levered households suffered a larger decline in house prices on average during the recession. Therefore, it is the combination of high leverage and large house price declines that generates the cross-county variation in the household balance sheet shock that we explain below. B. The household balance sheet shock In the models discussed in Section 1, the decline in aggregate household spending is driven by a shock that disproportionately affects the most highly levered households in the economy. In Eggertsson and Krugman (2012) and Guerrieri and Lorenzoni (2011), the shock is a 7 The key reason we do not report the instrumental variables estimation in the main results is that the instrument is correlated with the increase in house prices and the increase in spending from 2002 to 2006, as shown in Mian and Sufi (2011). Housing supply elasticity affected all three outcomes: house price growth, spending growth, and debt growth from 2002 to 2006. As a result, it is not an "exogenous" determinant of household leverage as of 2006, and so the advantage of the instrumental variables estimation over the OLS estimates is not large. It does help to rule out some alternative explanations, as we show in Section 6. 13

tightened borrowing constraint that can be interpreted as a collapse in asset values. In Midrigan and Philippon (2011), the shock is explicitly a decline in house prices. The empirical analog to the shock described in the models is the collapse in house prices that began in 2007. This collapse was concentrated in high leverage counties. This implies that high leverage counties experienced a double shock: they already had higher leverage before the housing decline, and the housing decline was disproportionately felt in the same areas. The left panel of Figure 2 reports house price growth relative to 2006 for high and low leverage counties, where high and low are defined to be the top and bottom decile of the household debt to income distribution as of 2006. The deciles are weighted by population so that both represent 10% of households. The decline in house prices in high household leverage counties was 40% from 2006 to 2009, whereas it was only 10% in low leverage counties. High leverage counties started with higher debt levels in 2006 and then experienced a larger decline in house prices. The right panel of Figure 2 shows an important implication of these two facts: states with higher debt to income ratios as of 2006 ended up with a much larger fraction of homeowners underwater on their mortgages. Columns 3 and 4 of Table 2 present coefficients from the following specification: Δ,, Γ (1) where Y is county-level house price from CoreLogic. These are cross-sectional first difference regression specifications that relate house price growth from 2006 to 2009 to the 2006 debt to income ratio. Equation (1) represents the general specification that we use throughout the analysis for a variety of outcomes including household spending growth. In columns 5 and 6, we use the fraction of homeowners that were underwater on their mortgage in 2009. The latter two 14

specifications are conducted at the state level, given that we only have the underwater information at that level. As Table 2 shows, the negative correlation between house price growth and the debt to income ratio is large and statistically significant. A one standard deviation increase in the 2006 debt to income ratio (0.6) is associated with an additional 12% decline in house prices. It is also associated with an 11 percentage point increase in the fraction of homeowners underwater (from column 6). Inclusion of a host of control variables does not affect the coefficients on the debt to income ratio as of 2006. C. The effect of the balance sheet shock on household spending More highly leveraged counties experienced a much larger shock to their balance sheets in the form of house price declines. How did this shock affect household spending? Figure 3 splits counties into high and low leverage deciles based on the 2006 debt to income ratio, and shows that across all types of goods, household spending declined disproportionately in high leverage counties. The magnitudes are large. For auto purchases, sales declined by 20 percentage points more in high relative to low leverage counties. The analogous numbers for other durables, groceries, and other non-durables are 25, 15, and 15 percentage points, respectively. Table 3 presents county-level regression coefficients from the estimation of equation (1) using household spending growth from 2006 to 2009 as the left hand side variable. Across all four types of goods, household spending declined much more in high leverage counties from 2006 to 2009. In terms of magnitudes, a one standard deviation increase in the 2006 debt to income ratio for a county is associated with a 10% additional decline in durable goods purchases and a 5% additional decline in non-durable goods purchases. 15

Table 3 includes a host of control variables that may potentially be responsible for the correlation between spending declines and leverage. These include controls for the prior construction boom, consumption boom, and any wage differentials. These control variables add significant explanatory power in most of the specifications, yet they have almost no effect on the coefficient estimate on the 2006 debt to income ratio. Highly levered counties experienced a much sharper pull back in consumption, and this effect is not mechanically driven by other covariates. Any alternative hypothesis for the decline in spending in high leverage counties must explain why control variables do not change the coefficients on debt to income as of 2006. Finally, one concern is that the results in Table 3 are driven by a differential decline in population in high leverage counties. Actually, the opposite is true. From 2006 to 2009, the population experienced a relative increase in high leverage counties. If we scale our spending variables by total population, we would see a slightly larger relative decline in household spending in high leverage counties. 4. The Role of Debt in Explaining the Decline in Household Spending The evidence in the previous section is consistent with implications of the models discussed in Section 1. Highly levered households, measured using county-level household debt to income ratios, experienced a large relative shock to their balance sheets from the collapse in house prices. These same areas experienced a sharp relative decline in household spending. As a further test of the underlying models discussed in Section 1, we examine the precise mechanism through which leverage affected household spending behavior. In Eggertsson and Krugman (2012) and Guerrieri and Lorenzoni (2011), the decline in household spending for levered households is driven by a tightened borrowing constraint that leads to deleveraging. In 16

Midrigan and Philippon (2011), the decline in household spending is driven by a sudden collapse in the housing-related liquidity services--i.e., home equity-based borrowing. All of these models predict a relative decline in the equilibrium debt amounts for highly levered households. Columns 1 and 2 of Table 4 show evidence consistent with this prediction. The coefficient on the 2006 debt to income ratio in column 2 suggests that a one standard deviation increase in the county-level debt to income ratio is associated with 3.3% additional reduction in debt, which is a 1/4 standard deviation of the left hand side variable. What are the potential channels that explain the reduction in debt in high leverage counties? The first is the delinquency channel: households can simply default on their loan obligations to reduce their debt burden. Second, there is the debt payback channel: households can start paying back accumulated principal at a faster pace. Finally, there is the liquidity services channel: the reduction in debt reflects the reduced ability of households to use their homes to directly finance consumption, as in Midrigan and Philippon (2011). As we show below, there is evidence to support all three channels. A. The delinquency channel Column 3 shows that high debt to income counties default at a significantly higher rate. 8 A one standard deviation increase in the debt to income ratio as of 2006 is associated with a 4.2 percentage point relative increase in household default rates, which is 2/3 a standard deviation of the left hand side variable. The theoretical models discussed in section 1 do not explicitly allow for default. Indeed deleveraging through default may be one reason not to cut back on consumption since the defaulter does not have to pay back the principal on debt. 8 The coefficient on debt to income ratio is similar without controls in all specifications. We do not report the specifications without controls in the interest of space. 17

However, default in the context of mortgage loans implies that the borrower has lost all equity in the house and is therefore likely to have lost a significant fraction of his savings. This is confirmed by column 4 that shows that high leverage counties also experienced a larger increase in foreclosures. Having lost a major fraction of their savings, defaulting homeowners are likely to cut back on consumption in an effort to rebuild their lost savings. Another reason for a defaulting homeowner to cut back on consumption is the additional cost associated with default and the negative spillover effects of foreclosures. In particular, default sours credit scores, which can have a negative effect on consumption (see for example Demyanyk, Koijen, and Van Hemert (2010)). Similarly, using state laws as an instrument for foreclosures, Mian, Sufi, and Trebbi (2012) show that higher foreclosures further depress local housing prices as well as local consumption since even non-defaulting homeowners feel poorer due to the further fall in house prices. B. The debt payback channel While default is an important channel for deleveraging, millions of non-defaulting homeowners also reduced their level of indebtedness via paying back principal. The county level data on debt do not allow us to separately observe the debt payback profile of non-defaulting borrowers. We therefore use the individual level data from Equifax described in Section 2 to separate debt dynamics for defaulters and non-defaulters. An individual is classified as a defaulter if at any time during 2008 to 2010 he is 120 days or more late on any of his debt payments, declares bankruptcy, or goes through a foreclosure. The left panel in Figure 4 separates the debt dynamics for the default channel and the debt payback channel. The solid blue line in Figure 4 plots total debt normalized to be 1 in 2007. Total household debt declined by about 13% from 2007 to 2010. 18

Households that default at some point during 2008 to 2010 are represented by the thin solid line. The figure shows that these are precisely the households that increased their debt levels most dramatically prior to 2007, and then also saw the largest decline in outstanding debt as they start defaulting post 2007. Debt obligations for defaulting borrowers declined by 40% from 2007 to 2010. While it is natural to expect debt burdens to go down for defaulting borrowers, the thin dashed line in Figure 4 shows that debt burdens declined even for borrowers that never default during the 2008 to 2010 period. The decline was more modest in percentage terms, but it represents a much larger segment of the population. Moreover, the decline in nominal debt balances represents a sharp reversal for these borrowers, who increased debt levels substantially prior to 2007. The right panel of Figure 4 summarizes in dollars the relative contribution of the default and debt payback channels in reducing overall household debt. It normalizes the total debt reduction in individual level data from 2007 to 2010 to the actual aggregate reduction in debt (i.e. $790 billion), and shows that defaulters were responsible for $610 billion dollars reduction in household debt between 2007 and 2010, while non-defaulters reduced debt by $180 billion during the same period. Figure 5 separates debt growth for non-defaulting borrowers by high and low leverage counties as before. Household debt declined at a faster pace after 2007 in high leverage counties. Moreover, the decline in debt was negative in absolute nominal dollars highlighting the extent of deleveraging. In column 5 of Table 4, we use individual level data to construct debt growth for nondefaulting borrowers from 2006 to 2010 at the county level. We regress debt growth for non- 19

defaulting borrowers on the 2006 county debt to income ratio. The results confirm the findings of Figure 5: non-defaulting borrowers pay back debt at a significantly faster pace in counties with higher levels of household leverage in 2006. The number of counties is smaller in this sample because individual level data is not pulled from every county as explained in Section 2. 9 C. The liquidity services channel In Midrigan and Philippon (2011), household spending for levered households responds to house price declines because of a decrease in the liquidity services of housing through a collateral cash in advance constraint. Consistent with this mechanism, columns 6 and 7 of Table 4 show a sharp relative decline in home equity limits and refinancing volume in high leverage counties. The evidence on refinancing is especially striking. The regression estimates without control variables (coefficient of -0.55 and constant of 1.09) implies that households in the 90th percentile of the household debt to income ratio as of 2006 experienced a decline in refinancing of 17% (=1.093-0.55*2.3). In contrast, households in the 10th percentile saw an increase in refinancing of 56%, which is consistent with the lower interest rates available to borrowers in 2009. In other words, high household debt counties were unable to access historically low interest rates during the heart of the recession. 10 5. The House Price and Leverage Interactive Effect on Household Spending A. Could our results be driven by a pure housing wealth effect? The theoretical models and empirical analysis discussed above highlight that the decline in household spending among levered households is due to a combination of ex ante high 9 We use the 2006 to 2010 period instead of 2006 to 2009 because the paybacks by non-defaulting households lag the decline in household spending. Or in other words, households cut back on spending and then subsequently pay back principal on their loans. The full time series is in Figure 5. 10 The inability of underwater homeowners to refinance into lower rates has been an important factor in the recession. See Boyce, Hubbard, Mayer, and Witkin (2012). 20

leverage ratios and a relative decline in house prices faced by levered households. While there is an existing literature on the pure housing wealth effect of changes in house prices on consumption, there is only limited evidence on the interactive effect of house price changes and leverage. 11 Leverage amplifies the impact of house price changes on borrower net worth, which would affect consumption patterns directly. This is true even in the absence of collateral constraints. For example, assuming all assets are held in the form of housing, a 20% drop in house prices results in an equivalent 20% drop in net worth for a homeowner with 0% loan-tovalue ratio but a 100% drop in net worth for a homeowner with 80% loan-to-value ratio. We should therefore expect the elasticity of consumption with respect to house prices to be stronger for more levered households, even in a world without collateral constraints. Of course, if higher leverage reflects tighter collateral constraints, then a decline in house prices for a levered household will lead to an even larger reduction in spending relative to the unlevered household. The first point of this section is that the magnitude of the consumption house price growth elasticity implied by our estimates is much larger than the pure wealth effect of house price changes documented in the literature. Typical studies on the housing wealth effect rely on household level survey data and examine the correlation between household spending and house price growth. Most of the literature suggests an elasticity in the 0.05 to 0.10 range. With the exception of Disney, Gatherhood, and Henley (2010), the housing wealth effect literature does not examine how this elasticity varies by household leverage. A simple comparison of the 11 The literature on the housing wealth effect is too large to be completely summarized here--it includes Muellbauer and Murphy (1990, 1997), Attanasio and Weber (1994), Lehnert (2004), Case, Quigley, and Shiller (2005), Haurin and Rosenthal (2006), Campbell and Cocco (2007), Greenspan and Kennedy (2007), and Bostic, Gabriel, and Painter (2009), and Carroll, Otsuka, and Slacalek (2011)). 21

existing estimates of housing wealth effect and our own estimate in this paper suggests that a second factor leverage is amplifying the traditional housing wealth effect. We conduct this exercise in Table 5, where we calculate an elasticity of household spending with respect to house prices that is implied by our estimates in Tables 2 and 3. We pick two points in the 2006 county-level debt to income ratio distribution--the 10th and 90th percentile--and use our estimated coefficients to project house price growth and household spending growth from 2006 to 2009 for these two points in the distribution. We can then calculate the differences for the outcomes, and scale the difference in spending with the difference in house price growth. This provides us an estimate of the elasticity of spending with respect to house prices implied by our results. As Table 5 shows, the estimated elasticities are far too large to be explained by a pure housing wealth effect. For non-durable goods, the estimated elasticities are in the 0.3 range, and for durable goods, they are in the 0.7 to 0.9 range. These estimates suggest that leverage significantly amplifies the effect of house price declines on household spending. Another interpretation of Table 5 is based on a two stage least squares estimation. The exercise we've reported in Table 5 is the same as a two stage least squares estimation in which household spending growth is regressed on the predicted value of house price growth instrumented with the 2006 debt to income ratio. Under this interpretation, it is obvious why the estimates are so large: leverage may have an independent effect on household spending beyond its effect on house price growth. Viewed in this manner, it is easy to see that the results in Table 5 are combining three effects: the pure housing wealth effect, the pure leverage effect, and the interaction of the two. In the next subsection, we attempt to separately estimate the interaction effect, which is at the heart of the models discussed in Section 1. 22

B. The interaction between leverage and house prices The preceding discussion suggests that the large decline in consumption in high leverage counties was unlikely to be driven by a pure housing wealth effect. Instead, leverage amplified house price declines in a powerful way to affect household spending. Our goal in this subsection is to separately estimate the direct effects of the house price decline and the initial leverage on consumption, and then also estimate the interactive effect of leverage and house price declines on consumption. The question we seek to answer is the following: for the same decline in house prices, how much more of a decline in spending do we predict for a levered versus unlevered household? However, this exercise requires meaningful orthogonal variation between house price growth and household leverage. For example, if house price declines from 2006 and 2009 and household leverage as of 2006 were perfectly collinear, it would be impossible to estimate their separate effects, or to estimate the interactive effect of house price declines and leverage on consumption. The statistical problem is that for reasons highlighted in Section 3, house price growth and leverage are naturally tied to each other at the county level via the housing supply elasticity channel. The high correlation between the two variables at the county level can be observed above in column 1 of Table 2, which shows an R 2 of 0.43 when regressing house price growth from 2006 to 2009 on county-level leverage. Given the high level of statistical correlation between household leverage and house price growth at the county level, it is not possible to separately estimate the effects of leverage and housing wealth. Instead, we exploit zip code level data on auto sales to conduct a within-county analysis of this question. Within counties, there is significant variation in house price declines that is less 23

correlated with ex ante leverage. The goal is to see whether the spending patterns of highly levered zip codes within the same county are more sensitive to the same drop in house prices. For this exercise, we construct the zip code leverage ratio as total debt in the zip code scaled by the value of housing assets and the value of financial assets in the zip code as of 2006. We describe in detail in the appendix how we obtain zip code level measures of housing and financial asset value as of 2006. Briefly, we use information on asset-generated income from the zip code level IRS Statistics of Income. We then assign a proportion of total financial wealth from the Federal Reserve Flow of Funds to a given zip code based on the zip code's proportion of asset-generated income. For housing wealth, we use the 2000 Decennial Census data to estimate total home value as of 2000 in a zip code as the product of the number of owners and the median home value. We then project forward this total home value into 2006 using the CoreLogic zip code level house price indices and an aggregate estimate of change in homeownership and population growth. 12 Once we have constructed the zip code level leverage ratio, we can then estimate coefficients for the following specification: Δ,,,, 2006,06 09 (2) The coefficient of interest from equation (2) is, which represents the differential impact of house price growth on auto sales growth from 2006 to 2009 by the 2006 leverage ratio of the zip code in question. 13 To be clear, measures whether spending in high and low levered zip codes respond differentially to the same decline in house prices. The specification in equation (2) 12 This procedure produces a median leverage ratio across zip codes of 0.24 and a housing wealth to (housing wealth+financial wealth) ratio of 0.30. From the flow of funds, the aggregate measures are 0.18 and 0.33, respectively. CoreLogic house price data is available for only 6,400 zip codes, which explains the smaller sample. 13 Auto purchases is the only spending information we have at the zip code level. 24