Employment in the Great Recession: How Important Were Household Credit Supply Shocks?

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Employment in the Great Recession: How Important Were Household Credit Supply Shocks? Daniel Garc´ıa 2018-074 Please cite this paper as: Garc´ıa, Daniel (2018). “Employment in the Great Recession: How Important Were Household Credit Supply Shocks?,” Finance and Economics Discussion Series 2018-074. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2018.074. NOTE: Staff working papers in the Finance and Economics Discussion Series (FEDS) are preliminary materials circulated to stimulate discussion and critical comment. The analysis and conclusions set forth are those of the authors and do not indicate concurrence by other members of the research staff or the Board of Governors. References in publications to the Finance and Economics Discussion Series (other than acknowledgement) should be cleared with the author(s) to protect the tentative character of these papers.

Employment in the Great Recession: How Important Were Household Credit Supply Shocks? Daniel Garc´ıa∗ Federal Reserve Board† November 9, 2018 Abstract I pool data from all large multimarket lenders in the U.S. to estimate how many of the over seven million jobs lost in the Great Recession can be explained by reductions in the supply of mortgage credit. I construct a mortgage credit supply instrument at the county level, the weighted average (by prerecession mortgage market shares) of liquidity-driven lender shocks during the recession. The reduction in mortgage supply explains about 15 percent of the employment decline. The job losses are concentrated in construction and finance. JEL codes: E44, G21, R31 ∗This paper presents the view of the author and not of the Federal Reserve. I would like to thank Chris Carroll, Jon Faust, and Jonathan Wright for their helpful comments, encouragement, and guidance, as well as Olivier Jeanne, Laurence Ball, Robert Barbera, Greg Duffee, Pedro Gete, Vadim Elenev, John Driscoll, Wayne Passmore, Shane Sherlund, Steven Laufer, Raven Molloy, Neil Bhutta, Laura Feiveson, Andrew Paciorek, Karen Pence, Alejandro Justiniano, Gadi Barlevy, Gene Amromin, Marco Bassetto and seminar participants at The Johns Hopkins University and the Federal Reserve Board and Bank of Chicago. †Federal Reserve Board, daniel.i.garcia@frb.gov. 1

1 Introduction Employment fell by over 7 million in the Great Recession. Possible explanations include declines in credit supply (Eggertsson and Krugman 2012; Guerrieri and Lorenzoni 2017), household net worth (Mian and Sufi 2014; Giroud and Mueller 2015), and increases in uncertainty (Baker et al. 2016; Bloom 2014). The goal of this paper is to isolate and empirically assess the credit supply hypothesis: to what extent did reductions in credit supply play a causal and independent role in explaining the job losses that occurred in the period 2007- 2010? To do so, I measure plausibly exogenous variation in credit supply (specifically for mortgages) at the county level, based on the interaction of prerecession county-lender market shares, which measure the importance of each lender to each locality immediately prior to the recession, and heterogeneous aggregate lender shocks during the recession. The county level estimates show that the reduction in mortgage supply negatively affected the health of residential markets, leading to declines in home buying, home prices, and employment in the construction sector; in other industries less directly linked to real estate, the job losses were much more muted and close to zero. A partial equilibrium aggregation exercise suggests that the decline in mortgage supply could explain close to 13 percent of the total job losses during the recession, or about 1 million of the total jobs lost. The starting point of this paper is the observation that there is a strong OLS association between declines in local employment and mortgage credit issuance during the recession. This suggests that reductions in mortgage supply could have played an important role in driving the job losses. On the other hand, the OLS association could be entirely driven by reverse causality – declines in local employment and economic activity could have led to the decline in mortgage issuance. To isolate the effects of reductions in mortgage supply on economic activity, I construct a mortgage credit supply instrument at the county level. The identification strategy exploits the well-known fact in the mortgage literature (discussedandfurtherdocumentedinthepaper)thatcreditrelationshipsinthemortgagemarket – as in the corporate market – are persistent and not easily substitutable.1 1Market shares at the county-lender level are highly persistent year on year; for example, 2005 countylender shares explain 2007 shares almost 1-for-1. In the recession, there were few cases of lender entry into new localities: of 2008-2010 county-lender pairs, less than 8 percent were new. Even in ‘normal’ times there is limited shopping; mortgage borrowers tend to shop too little despite significant price dispersion for 2

The instrument measures the average supply response of a county’s traditional lenders during the recession for ‘nonlocal reasons’ – reasons unrelated to the condition of local economies. The instrument is based on two sources of variation: (i) the heterogenous aggregate supply response of lenders during the recession, and (ii) variation in the reliance of localities to different lenders prior to the recession (measured with 2005-2007 market shares). To measure (i) aggregate differences in lender supply, I estimate lender fixed effects explaining variation in credit changes at the county-lender level during the recession, while holding constant local economic conditions (via county fixed effects). The lender fixed effects estimates are highly robust to alternative specifications, such as controlling for census tract fixedeffectsor loancharacteristicsvarying atthecounty-lenderlevel.2 County-lender market shares (ii) come straight from the main data source, the Home Mortgage Disclosure Act. The county level credit supply instrument is the weighted average (by 2005-2007 market shares) of the lender fixed effects. This paper is the first to construct a Bartik-style instrument based on the interaction of heterogeneous aggregate lender shocks and local market shares in the mortgage market during the recession. Working in parallel, Mondragon (2018) also studies the employment effects of household credit shocks during the recession, though his main instrument is based on county exposure to a single troubled lender during the recession (discussed shortly). My approach instead pools data from essentially all large multimarket lenders in the U.S., and follows in the tradition of recent related work studying the employment effects of reductions in corporate credit supply via Bartik-style instruments, such as Chodorow-Reich (2014), Greenstone et al. (2015), and Amiti and Weinstein (2018).3 The main contribution of this paperisthefocusonmortgagesupplyduringthecrisis, whichmaybeparticularlyimportant, since mortgages are the largest category of private credit, and funding markets for mortgages were severely disrupted during the crisis. Notably, the private secondary market for mortgages fully collapsed at the onset of the crisis and remained inactive throughout. In line with borrowers of similar characteristics (Alexandrov and Koulayev 2017; Woodward and Hall 2012; Lacko and Pappalardo 2010). 2Specifically, I regress credit changes at the county-lender level over 2007-2010 on county fixed effects and lender fixed effects in the baseline specification. The lender fixed effects are highly correlated (close to 1) when controlling instead for census tract fixed effects, using only high-income or low-income loans, and controlling for variation in precrisis county-lender loan characteristics. 3See Goldsmith-Pinkham et al. (2018) for a discussion of Bartik instruments. 3

previous research documenting that low liquidity contributed to lower credit issuance during the crisis (Ivashina and Scharfstein 2010; Cornett et al. 2011; Irani and Meisenzahl 2017), I document that mortgage lenders were more likely to cut supply if they relied on funding sources that proved fragile in the crisis. In particular, bank reliance on wholesale debt, loan sales in the secondary market, and loan sales to private buyers, explains 74 percent of the variation in lender supply during the recession (the lender fixed effects). The 2SLS results controlling for a detailed set of county observables and region fixed effects are as follows. Declines in mortgage supply negatively affected the health of residential markets. For example, a supply-driven plausibly exogenous 10 percent decline in local mortgage issuance led to an 8 percent decline in new residential permits and a 7 percent decline in home prices. Areas with larger declines in mortgage supply also experienced higher default and foreclosure rates. The next question is whether the negative shock to real estate spilled over into local labor markets, both in directly related industries such as construction and in other industries. The employment effects are largely direct and concentrated in construction and finance, a category of employment where about a third of workers are real estate intermediaries.4 The main mechanism is that declines in mortgage supply reduce demand for housing, which contributes to job losses in industries reliant on housing demand. As evidence, I find that, for a given decline in mortgage credit, job losses in construction are larger in counties where housing supply is more elastic – areas where construction is more responsive to changes in housing demand. The estimated effects on other categories of employment – total private employment excluding construction and finance, and nontradable employment – are close to zero and not significant. The 2SLS estimates on these broad employment categories contrast with their OLS counterparts, which are 2-4 times larger and highly significant, suggesting the OLS estimates are biased upward due to reverse causality. Overall, a 10 percent exogenous decline in local mortgage credit leads to a 1 percent decline in total private employment. To gain a sense of the aggregate implications of the county-level estimates, I perform a partial equilibrium aggregation exercise that exploits 4Construction and finance accounted for close to 35 percent of the job losses in the recession. Typically, their share in total employment is between 10-15 percent. 4

the in-sample distribution of the credit supply instrument, similar to the approaches in Chodorow-Reich (2014) and Mian and Sufi (2014). The exercise suggests reductions in supply could explain close to 13 percent of the total jobs lost, or about 1 million jobs. The bottomlineisthatthereductioninmortgagesupplylikelyaggravatedthefallinemployment to a meaningful, but moderate, extent. The main concern regarding instrument validity is that lender location might be correlated with unobserved local characteristics associated with job losses during the recession. Identification requires that below-average suppliers were not systematically sorted into localities experiencing below-average (or above-average) employment shocks. To the extent that counties and lenders are matched along observables, controlling for those characteristics isolates the remaining ‘as good as random’ variation in lender location. To that end, I control for a highly detailed set of local characteristics that explain about 60 percent of the variation in mortgage credit issuance across localities during the recession, including the share of subprime borrowers, the run-up in home prices during the boom, and various other demographic, housing, and industry characteristics. It is not possible to control for everything that may be relevant, however, and so I also rule out specific hypotheses about non-random lender location. For example, risky lenders may have moved to risky localities during the boom years. However, measuring the exposure of lenders to localities using 2000-2002 (instead of 2005-2007) market shares yields very similar results – the first-stage is weaker due to the loss in precision, but 2SLS point estimates are not statistically different. I also show that results are very similar when using region, division, or state fixed effects – this rules out hypotheses such as the possibility that weak suppliers in the recession were more heavily concentrated in the Sand States.5 This paper is part of the literature exploring the extent to which credit shocks explain the fall in employment in the Great Recession. Most empirical work focuses on the employment effects of corporate credit shocks.6 Chodorow-Reich (2014) estimates that credit shocks explain between about one-fifth and one-third of the aggregate employment decline in the 5Theresultsarerobusttoanumberofchecksincluding‘placebo’tests;controllingforthedeclineinsmall business lending in the recession; and the inclusion of large failed lenders (e.g., IndyMac) in the analysis. 6A related empirical literature studies the international transmission of the financial crisis through the banking sector (Cetorelli and Goldberg (2011); Haas and Lelyveld (2014); Schnabl (2012)). 5

year following the Lehman bankruptcy. Greenstone et al. (2015) find that credit shocks to small business loans help explain declines in borrowing, but produce only small employment effects. The credit supply variation I measure is specific to mortgages because county-lender market shares for mortgages and small business loans are largely uncorrelated – in other words, the mortgage lenders to one locality are often not the same as the small business loan lenders. The results in the paper are robust to controlling for declines in small business loans. The most closely related paper is Mondragon (2018), whose credit supply instrument is exposure to Wachovia Bank, a troubled lender in the recession. He estimates that a 10 percent decline in instrumented mortgage credit leads to a 3 percent decline in employment, an elasticity two times as large as the OLS counterpart, and three times as large as my own estimate. One concern is potential ‘bad’ bank in a ‘bad’ region matching – Wachovia had a larger presence in states in the South Atlantic (e.g., FL, SC, NC) where job losses were among the worst in the country. His estimates would have an upward bias if employment shocks and Wachovia location have correlated spatial fixed effects. In fact, when using division or state fixed effects, the Wachovia instrument significantly weakens. In contrast, my paper pools information from all large lenders located across the U.S. and employs a richer set of county controls, and so is more robust to potential concerns about non-random county-lender matching. For example, the 2SLS point estimates in this paper are essentially the same (not statistically different) when using region, division, or state fixed effects. The paper is also related to the work of Mian and Sufi (2014), Mian et al. (2013), Kaplan et al. (2017), and others, on the household net worth channel, which assesses the hypothesis that declines in household net worth led to declines in aggregate demand and employment. The credit supply and household net worth channels are related – for example, the bursting of the housing boom helped precipitate the financial crisis. However, the run-up in house prices does not explain all of the ensuing economic decline. To isolate the credit supply channel, this paper asks: holding house prices constant, what were the employment effects of reductions in mortgage supply during the recession? I therefore condition on house price changes during the boom years, as well as various prerecession characteristics of localities 6

associated with the housing boom and bust. More broadly, this paper is part of the literature studying the effects of changes in mortgage supply on housing and labor markets. Most empirical work has focused on the former. Related work includes Favara and Imbs (2015); Mian and Sufi (2011); Adelino et al. (2012); Berrospide et al. (2016); Glancy (2015); Anenberg et al. (2016); Vojtech et al. (2016); Gropp et al. (2014); Gete and Reher (2016); Haltenhof et al. (2014); Chen et al. (2017). Only a handful of papers focus on the employment effects of reductions in mortgage supply. In the boom years, DiMaggio and Kermani (2016) use a federal preemption of national banks from local anti-predatory lending laws in 2004 to estimate the elasticity of nontradable employment with respect to mortgage supply. In the bust, Passmore and Sherlund (2016) find that counties more reliant on GSEs for mortgage credit experienced healthier labor markets in the Great Recession. I contribute to this literature by highlighting the heterogeneous industry effects of mortgage supply shocks on construction and financial employment. 2 Data Sources I assemble a detailed county-level dataset including home prices, home sales, employment, mortgagecredit, creditscores, demographics, borrowercharacteristics, industrycomposition, and various other local characteristics. The main source for mortgage data is the Home Mortgage Disclosure Act (HMDA). Mortgage lenders with offices in metropolitan areas are required to publicly disclose detailed information each year, including the dollar amount and number of mortgages issued, as well as the location (census tract, county) of the property securing the loan. Throughout the mid to late 2000s, HMDA covered over 90% of residential mortgage lending by dollar amount (Dell0Ariccia et al. 2012). I use mortgages for home purchase and improvement as the main measure (loan purpose 1 and 2 in HMDA). Figure 1 plots aggregate trends in mortgage originations, total private employment, and the S&P Case-Schiller U.S. National Home Price Index, with the series indexed to their 2006 value. Data on home prices, permits, and delinquency and foreclosure rates are obtained from CoreLogic, the Census, and Black Knight McDash, respectively. For employment, I rely on 7

twosources,bothofwhichareestablishment-basedandprovidenearlyfullcoverageofprivate employment: the Quarterly Census of Employment and Wages (QCEW), and the County Business Patterns (CBP). I use the CBP to measure tradable and nontradable employment using the definitions in Mian and Sufi (2014), and the QCEW for the other employment data. Table 1 shows summary statistics for over 1,000 of the largest counties in the U.S. Each of these localities had over 15,000 households in the 2000 Decennial Census and account for about 85% of aggregate employment. Table 2 provides definitions and sources for the data used throughout the paper. While mortgage credit declined over 2007-2010 in virtually all counties, there is significant cross-sectional variation in the decline, with credit falling by more than 53% in ten percent of the counties in the sample and falling by less than 25% in the top decile. Figure 2 shows there is a strong positive OLS association between declines in mortgage credit issuance and declines in both home prices and employment. This suggests that declines in mortgage issuance could have driven employment losses. On the other hand, the relationship might be entirely explained by reverse causality – declines in local economic activity could have driven the decline in employment and credit issuance. I obtain lender-level data from HMDA, which provides loans by lender subsidiaries (respondents) and locality. I match subsidiaries belonging to the same parent company using the crosswalk maintained by Robert Avery, and aggregate to the level of the parent company (bank holding company, for banking institutions).7 To calculate changes in lending at the lender level without including changes due to acquisitions, I use the standard approach (Bernanke and Lown 1991; Greenstone et al. 2015) of treating the acquired and acquiring institutions as part of the same entity throughout the sample period, which in this paper is over 2000-2010. I also conservatively drop failed institutions for most of the paper, because the extent to which their credit decline was supply- or demand-driven cannot be credibly estimated. Dropping these institutions is a conservative choice: it reduces the potential for biased estimates at the expense of statistical power. I show, however, that including the failed lenders, by assuming all of their credit decline was supply-driven and nonlocal, increases the explanatory power of the credit supply instrument, while leaving coefficient 7Available upon request at Robert.Avery@fhfa.gov 8

estimates in the second-stage essentially unchanged. In measuring the exposure of counties to lenders, I focus on large multimarket lenders operating in multiple counties who did not file for bankruptcy during the crisis. Specifically, I include lenders operating in at least 100 counties in 2007, and who issued over $1 billion in mortgage originations in the same year. Table 3 gives a summary of lender-level statistics. The 57 lenders account for 75 percent of mortgage lending over 2005-2007, so they cover the majority of lending by market share, even though there were over 6,000 mortgage lenders in that period. I roll up the remaining small institutions into a single entity. 3 Differences in Lender Supply Thereweresubstantialdifferencesinsupplyacrosslendersduringtherecession. Somelenders almost fully halted originations, while a few even expanded. For example, mortgage originations fell by 69 percent at Citibank but increased by 17 percent at US Bank (Table 3). The empirical challenge, a variant of the reflection problem in Manski (1993), is that those differences could reflect borrower characteristics rather than differences in lender supply. For example, it is possible that US Bank’s typical customers experienced above-average credit demand during the recession. The main empirical strategy is to estimate lender fixed effects explaining variation in credit changes during the recession, while holding various characteristics of loans constant including the location of the property via locality fixed effects; other work employing similar methods includes Khwaja and Mian (2008), Greenstone et al. (2015), and Amiti and Weinstein (2018). This strategy exploits the richness in the HMDA data which provides originations at the locality-lender level and includes various loan characteristics. The lender fixed effects reveal substantial differences in aggregate supply across lenders. They are largely driven by differences in lenders’ funding strategy: reliance on funding sources that proved fragile in the crisis, such as wholesale debt and private loan sales in the secondary market, explain close to 75 percent of the variation in lender supply. In contrast, credit growth in the boom years (2003-2006) does not help explain either differences in lender supply or credit growth over 2007-2010, as shown in Figure 3. Therefore, I interpret 9

thesupplydifferencesaslargelyreflectingexposuretounexpectedfundingcostshocksduring the recession. Specifically, I estimate versions of the following linear model that specifies credit changes during the recession as a function of lender fixed effects, locality fixed effects, and lenderlocality interaction effects:8 ∆L = α +φ +γD +υ (1) i,b i b i,b i,b where ∆L are percent changes in mortgage credit originations at the county-lender level i,b over 2007-2010; α are locality fixed effects (county or census-tract); φ are lender fixed i b effects; and D are prerecession county-lender characteristics. The parameters of interest i,b are those associated with the vector of lender fixed effects φ , which capture the idiosyncratic b lender factor common across localities explaining variation in credit changes, net of locality fixed effects and prerecession county-lender characteristics. The model captures many of the reasons for variation in credit changes at the lenderlocality level during the Great Recession. For example, if originations to a locality declined sharply because of deteriorating local economic conditions – declines in local productivity, house prices, or credit scores – that will be captured by the locality fixed effects α . Similarly, i if originations decline because it is difficult for lenders to fund new mortgages, that would be captured in the lender fixed effects φ . It is also possible that the variation is driven by b interaction effects D – for instance, Citibank’s traditional borrowers could have tended to i,b experience below-average credit demand shocks, even within localities. In the baseline specification, I control only for county fixed effects. In this case the identifying assumption is that within-county credit demand shocks are uncorrelated with lender shocks. For example, supply contractions for Citibank would be overestimated if their borrowerstendtobelow-income,andlow-incomeborrowersexperiencedworsecreditdemand shocks than average, even within-counties. To address this possibility, I estimate equation 1 using only high income loans, but estimates are very similar. Specifically, I estimate equation 8The lender fixed effects are estimated using 30,161 county-lender observations, for the 57 lenders in the sample, and for county-lender pairs where the dollar value of originations is larger than $1 million. The lender fixed effects explain about a fifth of the variation in within-county lending changes over 2007-2010. 10

1 using only loans to borrowers with income over $70,000 the median income of borrowers in 2007. The correlation coefficient between the lender fixed effects estimates in the baseline and the specification with only high income loans is 0.96; see Figure 4. When using only low-income loans (borrower income below $70,000), the correlation coefficient is also high, 0.94. I also estimate equation 1 using census tract fixed effects rather than county fixed effects. Census tracts are statistical subdivisions of counties, each generally having a population size between 1,200 and 8,000 people. Census tracts are smaller and are more homogeneous than counties.9 The lender fixed effects estimates when using census tract fixed effects are also highly correlated with the baseline (0.91). This shows that using a more detailed local control for changes in credit demand has very little bearing on the lender fixed effects estimates. Another alternative is to directly control for differences in the prerecession profile of borrowers and lenders via county-lender characteristics D . The county-lender characterisi,b tics observed in HMDA are borrower income, fraction of loans classified as being high-risk, race, type of loan (owner-occupier), and credit growth in the peak boom years 2003-2006 by county-lender. When including D in equation 1, the lender fixed effects estimates are i,b again highly correlated. Table 4 shows sample statistics for the 25 largest lenders in the sample. Column 2 provides percent changes in national mortgage originations over 2007-2010, and Column 3 ranks them by percent changes in mortgage originations. Column 4 ranks lenders by the lender fixed effects estimates; above-average lender fixed effects indicate above-average supply. Changes in the ranking (going from Column 3 to 4) indicate differences in the degree to which national changes in mortgage originations were driven by geographic variation in exposure to credit demand shocks. For example, the drop in Flagstar’s ranking from 8th to 14th (from Column 1 to Column 2) indicates that lending changes for this bank remained relatively robust in the recession partly because of its exposure to above-average geographies (in this case the Midwest). Conversely, the improvement in the ranking of JPMorgan Chase 9I rank census tracts within a county by borrower income, and divide the census tracts into four equalsized groups by income, i.e. the top quartile consists of the high-income census tracts in the county. Census tract-income groups are more homogeneous than the county – in 2007, the median within-group standard deviation of HMDA borrowers in the census tract-income groups was $92 thousand, 27% lower than in counties. 11

from 52th to 35th indicates that part of its national decline in originations was driven by exposure to underperforming areas. While there are some changes, overall the rankings are highly correlated (correlation coefficient = 0.89), indicating that the lender fixed effects are only weakly correlated with locality fixed effects. 3.1 Funding Fragility and Differences in Supply What explains the dispersion in aggregate supply across lenders, the variation in the lender fixed effects φ ? In line with previous research documenting that low liquidity contributed b to lower credit issuance during the crisis (Ivashina and Scharfstein 2010; Cornett et al. 2011; Brunnermeier 2008; Gorton and Metrick 2012; Kacperczyk and Schnabl 2010; Ramcharan et al. 2016), this section shows mortgage lenders were more likely to cut supply during the recession if they relied on funding sources that proved fragile in the crisis. As discussed in Passmoreetal.(2005),mortgageloanscanusuallybefundedinoneofthreeways: (i)vialoan sales in the secondary market, or through balance sheet retention; (ii) if kept in the balance sheet, through wholesale debt or deposit-like liabilities; (iii) if sold in the secondary market, through loan sales to the GSEs (e.g. Fannie Mae, Freddie Mac, Ginnie Mae), or through sales to private buyers. I measure each of these three funding strategies by combining lender data from HMDA and the Federal Reserve’s FRY-9C. Table 5 reports results from regressions of differences in lender supply (φ ) against difb ferences in funding strategy over 2006-2007 (see also Figure 5) for the banks in the sample. Specifically, Iregressthelenderfixedeffectsontheprecrisisratiosofwholesaledebttoassets, loan sales to total originations, and private loan sales to total loan sales. These measure the reliance prior to the crisis of banks on wholesale debt and the originate-to-distribute lending model. Column 1 shows variation in these three funding strategies explain 74 percent of the variation in supply differences φ .10 Column 2 shows that lower prerecession capital ratios b are also associated with declines in credit supply, though this factor is relatively minor, judging by its 3 percentage point contribution to the R-squared (Column 2). Column 3 shows 10In complementary work, Dagher and Kazimov (2012) find that mortgage lenders more reliant on wholesalefundingweremorelikelytorejectmortgageapplicationsduringtherecession,aftercontrollingforvarious borrower characteristics. 12

that, in contrast, prerecession credit growth (over 2003-2006) is not helpful in explaining variation in differences in supply during the Great Recession. Observations are weighted by the dollar amount of mortgage originations in 2007, although the weighting is not critical, as shown in Column 4. I measure bank-level exposure to wholesale funding as the ratio of non-core funding (sum of large time deposits, foreign deposits, repo sold, other borrowed money, subordinated debt, and federal funds purchased) to total assets, from the Federal Reserve’s FRY-9C form, a standard definition in the literature (Irani and Meisenzahl 2017). To measure lender exposure to the secondary market, I use data from HMDA, which provides loan sales in the secondary market by year and type of buyer. Reliance on loan sales is measured as the share of loans originated and sold to total originations over 2005-2007. Exposure to private securitization is measured as the ratio of private investor loan sales to total loan sales over 2005-2007.11 Measuring reliance on loan sales to private buyers is important since private-label residential mortgage securitization, which funded about 30% of mortgages over 2005-2007, went to essentially zero in 2008-2010 (Frame et al. 2015); see also Avery et al. (2011), and Nadauld and Sherlund (2009)). Because private investors stopped purchasing nongovernment-insured mortgages, lenders reliant on those sales likely cut supply during 2008-2010. For example, Calem et al. (2013) find that banks who were pre-recession more dependent on loan sales experienced more severe declines in jumbo lending, which are loans too large to be purchased by GSEs, and thus can only be sold to private investors, during the recession. Loan sales to GSEs also became more expensive. G-fees, the monthly insurance fee GSEs charge as a fixed fraction of the loan balance, increased from about 20 basis points in 2005- 2007 to 30 basis points in 2008-2010 (Fuster et al. 2013). Putback risk also increased in 2008. Lenders are required to repurchase loans sold to GSEs if it is found that those loans fail to satisfy original underwriting standards. While putbacks were rare, they rose during the recession, with Fannie Mae estimating that 3.7 percent of single-family loans purchased over 2005-2008 were putback to lenders, whereas the figure in other periods tended to be 11Private loan sales are defined as loan sales to any buyers excluding FNMA, FAMC, GNMA, FHLMC, and lender affiliates. 13

less than 0.5 percent12 4 The Nonlocal Lending Shock Differences in lender supply affected counties differently, because of variation in the intensity of preexisting county-lender relationships, as measured by market shares prior to the recession. The credit supply instrument – the nonlocal lending shock – is the weighted average, for county i, of lender supply shocks in the recession φ (from equation 1). The weights are b county-lender 2005-2007 mortgage origination market shares. The sum is taken over all large multimarket lenders in the sample B, as discussed in section 2: X Nonlocal Lending Shock = Share φ (2) i i,b b B Counties had below-average access to mortgage credit, all else equal, if they had existing relationships (as measured by 2005-2007 market shares) with lenders with below-average supply in the recession. New lender entry would work towards offsetting the decline in credit supply by the locality’s traditional lenders. In the extreme case of perfectly substitutability, lender entry would fully offset the reduction in supply by the locality’s traditional lenders. The instrument, however, is not weak with the first-stage Kleibergen-Paap F statistic in the baseline over 40. I provide evidence of both highly persistent county-lender relations prior to the crisis, and of limited new lender entry during the recession. First, county-lender market shares are highly persistent year-on-year. Table 6 shows results from regressing 2007 county-lender market shares on 2005 shares. Column 1 shows that 2005 shares explain 91 percent of the variation in 2007 shares, with the coefficient on the 2005 shares equal to 0.92. The left panel of Figure 6 plots 2007 shares against 2005 shares. Moreover, the relationship between 2005 and 2007 shares is highly stable across localities. The correlation coefficient and R-squared are very similar when focusing only on high credit score counties or only low credit score counties (Columns 2 and 3), or when using county fixed effects (Column 4). The persistence in credit relationship extends to at least the early 2000s. The right panel 12source: Fannie Mae 10-K 2013, p. 143 14

in Figure 6 plots 2000 market shares against 2007 market shares; there is a strong positive association, with 2000 shares explaining 71 percent of the variation in 2007 shares. As for limited entry, I find few cases of lenders entering new counties in the recession: of all county-lender pairs in 2008-2010, only 7.85% were new matchings. The lack of entry suggests substantial switching costs across lenders during the recession. Part of the reason for low new entry may be that only a handful of lenders were expanding during the recession. Because most lenders were contracting, they may not have been looking to expand into new localities.13 The contraction in lending by many mortgage lenders, particularly the larger ones, is also documented in Gete and Reher (2016) and Chen et al. (2017). The findings in this paper on persistent credit relationships and limited entry during the recession are in line with the literature documenting stickiness in mortgage credit relationships and limited shopping in the mortgage market in spite of significant price dispersion. In a survey of recent mortgage borrowers, Alexandrov and Koulayev (2017) report that close to half of the borrowers did not do any shopping. Woodward and Hall (2012) also find that borrowers engage in too little shopping, and “sacrifice at least $1,000 by shopping from too few brokers.” Lacko and Pappalardo (2010) shows that mortgage borrowers are often severely uninformed about key costs associated with getting a mortgage, with half of respondents having problems identifying the loan amount, and two-thirds being unaware of prepayment penalties, for example. Moreover, Mondragon (2018) and Nguyen (2014) find evidence for stickiness in the mortgage market, in line with the large literature showing substantial switching costs for firms, as recently discussed in Chodorow-Reich (2014). The main concern with instrument validity is that the credit supply instrument, county exposure to lender supply shocks, may be correlated with unobserved characteristics of counties affecting employment. It would be sufficient (but not necessary) if lender location is randomly distributed across counties. That is unlikely to be the case, however. Belowaverage suppliers in the recession may have been more likely to locate in subprime counties (for example) prior to the crisis. To the extent I can observe and control for the fraction of subprime borrowers (and other relevant local characteristics), I can isolate the ‘as good as random’ variation in lender location. To that end, I employ a detailed set of prerecession 13These statistics are based on the 57 lenders in the sample as described in section 2. 15

county characteristics, including the subprime share, that explains close to 60 percent of the cross-sectional variation in mortgage credit changes over 2007-2010, described in Table 2. The controls include: 2006 household debt to income used in Mian and Sufi (2009); the run-up in house prices over 2003-2006; the fraction of subprime borrowers in 2006; industry composition such as the construction share of employment in 2006; loan characteristics such as local incidence of FHA or investor loan over 2003-2006; demographics; and measures of local lending competitiveness.14 Figure 7 is a map of the nonlocal lending shock, after controlling for a detailed set of county covariates. The map appears balanced with no apparent trends by region. Conditional on the detailed set of county observables used in the paper, I find evidence consistent with ‘as good as random’ county-lender matching both in the boom and before. Theresultsinthepaperarerobusttomeasuringcountyexposuretolendershocksusing2000- 2002 shares (instead of 2005-2007 shares). This addresses the concern that risky lenders may havelocatedinriskycountiesduringthehousingboom. Asforpotentialnon-randomcountylender matching before the 2000s, I estimate the models using different regional fixed effects, including state, division, and region fixed effects; substantially different estimates would be evidence of correlated fixed effects at regional levels for employment outcomes and lender location i.e., regional county-lender matching. However, estimates are consistent across specifications. As discussed shortly, I perform various other checks that find support for the exclusion restriction, including ‘placebo’ tests; controlling for declines in small business lending in the recession; and the inclusion of large, failed lenders (e.g. IndyMac) in the analysis. 14Previousliteraturehasestablishedthatdifferenthouseholdcharacteristicsareassociatedwiththeseverity ofthehousingboomandbust. Fortheincidenceinsubprimelending,see: Keysetal.(2010),Demyanykand Hemert (2011), Dell0Ariccia et al. (2012), Gerardi et al. (2008), and Mian and Sufi (2009). For the growth inhouseholddebttoincome, see: MianandSufi(2014), andCarrollandKimball(1996). Fordemographics: Elsby et al. (2010). For loan characteristics: Haughwout et al. 2011; Chinco and Mayer 2016; Bhutta 2015; Bhutta and Ringo 2014 16

5 Empirical Framework and Results I now discuss results based on the following 2SLS specification: ∆Outcomej =θX +β∆ (cid:92) Credit +f +(cid:15) (3) i i i s i ∆Credit =δX +ρNonlocal Lending Shock +f +v (4) i i i s i where observations are at the county i level; changes are over 2007-2010 for different outcome variables j (house sales, house prices, employment) each estimated separately; and f are s fixed effects that could be at the region, division, or state level – I report results for each. Table 2 defines the set of prerecession county controls X as well as the outcome variables. i The nonlocal lending shock is the credit supply instrument defined in equation 2. All of the outcome variables are expressed as percent changes over 2007-2010. For employment categories and the home price index, changes are taken between 2007Q4 and 2010Q4. For mortgage credit (a flow) changes are taken between the average dollar flow over 2008-2010 with respect to the value in 2007.15 Mortgage flows are deflated using the GDP deflator.16 I use data on approximately the largest 1,000 counties in the U.S. (those having over 15,000 households in the 2000 Decennial Census), which account for 85% of aggregate employment. I drop states having 3 or fewer counties, to have at least a few observations per state for the specifications that use state fixed effects. Observations are weighted by the number of households in the 2000 Decennial Census, though results are very similar without weighting.17 Extreme observations (1% from each tail) are dropped from each dependent variable.18 Standard errors are clustered at the division level to allow for correlated shocks within broad geographic regions due to, for example, state or division-specific institutional 15Using 2005-2007 as the base period produces nearly identical results, for example, the correlation coefficient between ∆Credit using 2007 as the base period and using 2005-2007 as the base is ρ=.87. Table 10 i in the Online Appendix reports the main estimation results using 2005-2007 as the base. 16Alternatively, ∆Credit could be defined as the percent change in the number of mortgage originations, i with very similar results; Table 11 in the Online Appendix shows the main 2SLS results when doing so. 17Table 12 in the Online Appendix reports unweighted results for counties with over 40,000 households in the 2000 Decennial Census – these close to 500 counties account for 76% of total employment. 18For example, I drop house price growth outliers from the house price regression, but I don’t drop those counties from the private employment growth regression (unless they are also outliers in that variable). The only exception is growth in house permits for which I winsorize 5% of observations. 17

arrangements and spatial correlation.19 Estimates are robust to alternatives, such as clustering at the commuting zone level (Table 13 in the Online Appendix). 5.1 First Stage Results The nonlocal lending shock has significant independent explanatory power over local changes inmortgagecreditintheGreatRecession,consistentwithhighswitchingcostsacrosslenders. Table 7 reports first-stage regression results; all the controls listed in Table 2 are included (e.g. the share of subprime borrowers, measures of the severity of the housing boom, and various demographic, industry, and loan characteristics) though only the nonlocal lending shock coefficient estimates are reported, to economize on space. Columns 1-4 include varying degrees of spatial fixed effects, ranging from no spatial fixed effects (Column 1) to region, division, and state fixed effects specifications (Columns 2-4 respectively). The R-squared is reasonably high in all specifications (55 percent or higher), indicating that the regression controlsarehelpfulinexplainingvariationinmortgagecreditissuance. Acrossspecifications, the coefficient estimate on the instrument is positive and strongly significant. For example, in the specification without spatial fixed effects (Column 1), a 10 percent reduction in the nonlocal lending shock is associated with a 4.79 percent decline in mortgage credit issuance; the first-stage Kleibergen-Paap F statistic is slightly above 20. The instrument has considerably explanatory power in all models with different spatial fixed effects. The F statistic, with degrees of freedom adjusted for division level clustering, is over 10 in all the specifications, a rule of thumb commonly used to indicate weak instrument problems (Stock and Yogo 2002). The F statistic is lowest in the specifications with state fixed effects (13.95), since this specification uses less information (only within-state variation in the instrument). In some models with state fixed effects, such as in the residential permits model (Table 1, Column 4 of the Online Appendix), the first stage F statistic is just above 10. I report in the Online Appendix p-values for LM tests of underidentification based on the Kleibergen-Paap rk statistic; the null of underidentification can be rejected at the 5% significance level across all models including those with state fixed effects. 19The Census divides the US into 9 divisions – New England, Middle Atlantic, East North Central, West North Central, South Atlantic, East South Central, West South Central, Mountain, and Pacific. 18

5.2 Effects of Supply Reductions on Residential Markets Supply-driven,exogenousdeclinesinmortgagecreditarestatisticallyassociatedwithdeclines in home sales, home prices, and increases in delinquency rates as well as foreclosure rates. This is evidence of the negative effects of declines in mortgage supply on the health of local housing markets. The mechanism is that reductions in mortgage supply reduce the ability of households to buy homes and to refinance. Table 8 reports two stage least squares results for different housing market outcomes in the models with region fixed effects and standard errors clustered at the division level. I use region fixed effects in the baseline, though I provide results with other spatial fixed effects in the Online Appendix and in some cases throughout the paper. Declines in credit supply are associated with declines in home permit issuance. Column 1 shows that a 10 percent reduction in mortgage credit (when instrumented using the nonlocal lending shock) is associated with a 8.24 percent decline in the issuance of new residential permits – close to a one-to-one effect. This is evidence that households were unable to offset the reduction in credit availability originating from nonlocal sources by borrowing from private sources or from lenders other than their traditional, prerecession lenders. The effect operates through the extensive margin – fewer loans were taken out, which led to lower housing demand and caused declines in new permits. The effect is very similar (8.55 percent decline) when measuring changes in mortgage credit using declines in the number of loans, rather than in the real dollar value, as reported in Table 11 of the Online Appendix. Declines in mortgage credit are also associated with declines in home prices. A plausibly exogenous 10 percent decline in mortgage credit is associated with a 7.35 percent decline in home prices.20 Delinquency rates and foreclosure rates also increased more in counties with belowaverage supply. Table 8 shows that a 10 percent decline in mortgage credit is associated with 1.42 and 0.87 percentage point increases in delinquency and foreclosure rates. This is evidence of the contractionary effects of reductions in mortgage supply on the health of local 20This is consistent with other articles finding that supply-driven changes in credit have real effects on home prices, such as Favara and Imbs (2015), Mian and Sufi (2011), Adelino et al. (2012), Favara and Imbs (2015),DiMaggioandKermani(2016),Anenbergetal.(2016),Vojtechetal.(2016),PassmoreandSherlund (2016), and Kung (2015). 19

housing markets. The fall in home prices induced by the credit shock would make it more likely for households to go underwater. In the Online Appendix I present results for the each dependent variable with no fixed effects, region, division, or state fixed effects. The main conclusions are essentially the same. The point estimates are very similar. For example, a 10 percent reduction in mortgage credit is associated with a 6.73, 7.57, 8.28, and 7.98 percent decline in home prices in the models without spatial fixed effects, and with region, division, and state fixed effects respectively (Table 2 of the Online Appendix). 5.3 Effects of Supply Reductions on Employment Declinesinmortgagesupplycontributedtothejoblossesintherecession, thoughtoamoderate extent. The job losses explained by the mortgage shock are concentrated in construction and financial services, a category of employment where over a third of workers are real estate intermediaries. The likely mechanism is that reductions in supply caused declines in housing demand, which negatively affected employment in industries reliant on housing demand. As evidence for this, I find that the construction losses are stronger in areas where housing supply is more elastic, that is, in areas where construction responds more to changes in housing demand. Overall, a supply-driven plausibly exogenous decline in mortgage credit issuance is associated with about a 1 percent decline in total private employment. Using the in-sample variation of the nonlocal lending shock, I estimate that about 15 percent of the aggregate employment losses in the Great Recession can be explained by declines in mortgage supply. Weak mortgage supply contributed to job losses in the construction sector. Table 9 shows that a 10 percent decline in mortgage credit originating from nonlocal sources is associated with a 3.81 percent decline in construction employment for the model with region fixed effects, with point estimates similar for the other specifications. The mechanism is that declines in mortgage supply reduce housing demand, which is associated with lower employment in construction. The employment losses in construction were, for a given decline in instrumented credit, more severe in areas where housing supply is more elastic.21 That is, in areas where con- 21I add the interaction of credit changes and the housing supply elasticity to the regression model with 20

struction responds more strongly to changes in housing demand, the employment effects of a given credit decline were stronger. To see this, I focus on the sample of counties for which the Saiz (2010) measure of the elasticity of housing supply is available.22 Table 10 reports results for changes in home permits for new construction and construction employment for the model with region fixed effects. The coefficient estimate is positive for the interaction of credit changes and housing supply elasticity and significant at the 1% level for construction employment (Column 2). For the permits model (Column 4), the interaction is also positive and significant at the 10% level. That is, the same relative decrease (increase) in credit is associated with lower (higher) permit issuance and construction employment in areas with higher housing supply elasticities. This is evidence for the mechanism that reductions in mortgage supply reduced housing demand and contributed to employment losses in construction. Declines in mortgage supply also caused job losses in finance. Table 9 shows that a 10 percent reduction in mortgage credit is associated with a 4.40 percent decline in employment in financial services in the model with region fixed effects. The likely mechanism again is that reductions in supply negatively affected housing demand, and therefore demand for housing intermediaries. Via the effects on construction and financial employment, declines in mortgage credit led to declines in total private employment. Table 9 shows two stage least squares results for different employment categories, including total private employment. Column 3 shows that a 10 percent reduction in mortgage credit originating from nonlocal sources is associated with a significant 1.14 percent decline in total private employment. The models with other types of spatial fixed effects have similar point estimates, as reported in Table 7 of the Online Appendix, though confidence intervals are wider especially when state fixed effects are used, sincethesemodelsuselessinformation(onlywithin-statevariation). Inthespecificationwith state fixed effects, for example, a 10 percent reduction in instrumented mortgage credit is associated with a (not significant) 8.3 percent decline in employment. The Online Appendix region fixed effects. For the two endogenous regressors (credit changes and the interaction of credit changes and the housing supply elasticitiy), I use two instruments – the nonlocal lending shock, and the interaction of the nonlocal lending shock with the housing supply elasticity. 22Saiz (2010) estimates housing supply elasticity as a nonlinear combination of data on physical and regulatory building constraints and population levels in 2000 at the metro area level. 21

reports estimates for all of the dependent variables discussed in the paper for specifications with no spatial fixed effects, region, division, and state fixed effects. Declines in mortgage supply are only weakly associated with declines in employment in other, broader employment categories – ‘other employment’ (total private excluding construction and finance) and nontradable employment, which mostly consists of local retail and food. These are shown in Table 9, Columns 4 and 5 respectively. The coefficient estimates are close to zero and not significant. That is, the negative shock on local real estate markets did not appear to significantly spillover to broader local employment categories. One possibility is that the real estate shock did have large spillover effects, but that those effects were nonlocal, and were instead dispersed through localities through the tradable sector. However, there is little evidence that the local real estate shock had large spillover effects on the local nontradable sector (Column 5). In Boldrin et al. (2012) the spillover between a housing shock to the rest of the economy depends on the elasticity of substitution between consumption and housing. The results in this paper suggest the (local) elasticity is relatively low. The elasticity estimates of other and nontradable employment also contrast with their OLS counterparts, which are about twice as large and strongly significant, with t-statistics ranging from 3 to 8 across specifications, as shown in Table 11. That the OLS coefficients are larger suggests that they are biased upward, due to reverse causality – employment losses may lead to declines in mortgage issuance. The credit supply instrument is strong, and helps predict declines in real estate activity, such as declines in home permits, home prices, and construction employment. But it does not help explain substantial job losses in industries less directly related to real estate. This ameliorates concerns about reverse causality – if local employment shocks were correlated with the instrument, then the 2SLS estimates for broad employment categories would likely be large and significant. In parallel work Mondragon (2018) also estimates the county level elasticity of employment with respect to mortgage supply during the recession. We both find that reductions in mortgage supply mattered for employment in the recession, though the estimated effects in Mondragon (2018) are substantially higher. He estimates that a 10 percent decline in instrumented mortgage credit is associated with a 3 percent decline in employment, an elas- 22

ticity about two times as large as the OLS counterpart, and three times as large as my own estimate.23 The main difference between the papers is the credit supply instrument; his instrument is prerecession exposure to Wachovia Bank, a troubled lender acquired by Wells Fargo in late 2008.24 One reason his estimates are likely larger is ‘bad’ bank in a ‘bad’ region matching – Wachovia had a larger presence in states in the South Atlantic such as Florida, South Carolina, and North Carolina where job losses were among the worst in the country. The Wachovia instrument significantly weakens when controlling for characteristics of localitiescorrelatedwithbothWachovialocationandemploymentlossesduringtherecession. For example, using only division or state fixed effects greatly diminishes the statistical power of the Wachovia instrument. To see this, I obtain Wachovia 2005-2006 purchase shares from HMDA and restrict the sample to counties in the South and East. The first-stage F statistic associated with the Wachovia instrument is 14.47, absent other controls including regional fixed effects. When including division (state) fixed effects, the F statistic drops to 4.33 (0.93).25 In contrast, the results in my paper are very similar when using no fixed effects, or region, division, orstatefixedeffects. Moreover,itisnotthecasethattheresultsaredifferent because Wachovia was a particularly troubled lender. In fact, Wachovia was acquired by Wells Fargo, the strongest lender of the top 4. As discussed shortly, the results in this paper are very similar even when including large, failed lenders in the analysis such as IndyMac, which was not rescued by another institution. 5.4 Aggregate Implications Overall, I find that reductions in mortgage supply could explain close to close to 13 percent of the employment losses in the U.S. over 2007-2010, or about 1 million of the jobs lost. This is evidence that reductions in mortgage supply mattered for employment. The imputation is based on a partial equilibrium aggregation exercise that answers the counterfactual ques- 23TheseestimatesreplicatetheearlierMondragon(2014). Inmorerecentversions,changesinMondragon’s specifications such as variable standardization and sample restrictions make replicating Mondragon (2018) less straightforward. 24Mondragon (2018) continues to use Wachovia as the key source of identification as in earlier versions (Mondragon 2014), though the more recent version uses a few other regional lenders as a robustness check. 25Observations are weighted by population in 2007, and standard errors are clustered by state. These results are available upon request. 23

tion: what if counties, all else equal, had experienced the best credit shock in the sample – specifically, the credit shock of the counties in the top 5 percent of the distribution? The improvement in supply generates employment gains via the estimated elasticity of employment with respect to mortgage supply. This approach addresses the challenge that the level effect of supply reductions cannot be recovered from the cross-section by assuming that the top percentile of counties by the credit supply instrument represent a ‘no credit shock’ scenario. This is a standard aggregation exercise in this literature, with similar approaches in Chodorow-Reich (2014) and Mian and Sufi (2014). The estimate would be biased downwards if the top percentile counties also experienced a reduction in supply. The severe disruptions in mortgage supply in the recession affecting wholesale funding markets and loan sales in the secondary market suggest the assumption is conservative. First,definethecounterfactualemploymentchangeincountyi,∆Empcf,asthepredicted i employment if county i had experienced the nonlocal lending shock of county zero (NLS ), 0 rather than its own (NLS ), after conditioning on all other observables X : i i ∆Empcf =E[∆Emp |NLS = NLS ,X ] i i i 0 i (cid:91) (cid:92) (cid:92) =∆Emp +β(∆Credit (NLS )−∆Credit (NLS )) i i 0 i i (cid:92) =∆Emp +βρ(NLS −NLS ) i 0 i (cid:92) where ∆Emp denotes the fitted value from the private employment regression model with i region fixed effects, β is the estimated elasticity of employment with respect to mortgage supply, and ρ is the coefficient on the nonlocal lending shock in the first-stage regression. I then recover the end-period levels of employment corresponding to both the counterfactual and fitted changes in employment, using the initial-period employment level: Empcf = i,2010Q4 Emp (1+∆Empcf) and E (cid:91) mp = Emp (1+∆E (cid:91) mp ). Then, the total job i,2007Q4 i i,2010Q4 i,2007Q4 i loss explained by variation in the nonlocal lending shock is given by: Total jobs lost explained by lending shock = X [Empcf −E (cid:91) mp ] (5) i,2010Q4 i,2010Q4 i:NLSi<NLS0 24

The fraction of jobs lost that is explained by the lending shock is given by: P [Empcf −E (cid:91) mp ] i:NLSi<NLS0 i,2010Q4 i,2010Q4 (6) P [Emp −Emp ] i:NLSi<NLS0 i,2010Q4 i,2007Q4 The exercise indicates that the decline in mortgage supply can explain about 13 percent of the employment losses in the Great Recession, when defining county zero as the 95th percentile county by the credit supply instrument, and using the coefficient point estimate β = 0.114 from the region fixed effects model. There is uncertainty around β, however. For example, β = 0.083 in the model with state fixed effects. Using the latter, the aggregation exercise suggests the mortgage credit supply shock explains about 9 percent of the job losses during the recession. Alternatively, the 95 percent confidence interval for β in the region fixedeffectsspecificationrangesfrom0.048to0.179; usingthisrange, thedeclineinmortgage supply explains between 4 and 22 percent of the job losses during the recession. Another important parameter choice is which counties are used as the ‘no credit shock’ reference. The baseline uses the 95th percentile as the baseline. If localities in the top 5 percent of the credit supply distribution also experienced a reduction in credit supply, the aggregation exercise will deliver an under-estimate. When using the top 1 percent as a reference instead, the aggregation exercise suggests declines in mortgage supply can explain 19 percent of the job losses in the recession. The bottom line of these aggregation exercises is that the reduction in mortgage supply likely aggravated the job loss during the recession, though moderately so. 13 percent of the totaljoblossesissizable–about1millionjobslostishardlysmall–butitisfarfromthebulk of the job losses, as argued by Mondragon (2018) and particularly Mondragon (2014) which attributed about 60 percent of the total job losses (at a minimum) to household credit supply shocks. In sum, the evidence in this paper adds nuance to the debate of “what explains the job losses during the recession?” The answer provided by this paper is that mortgage supply shocks mattered, though moderately. This suggests that other factors, such as the decline in household net worth (Mian and Sufi 2014) or increase in uncertainty (Baker et al. 2016) may explain the bulk of the job losses in the recession. Credit supply shocks to firms may have also been an important contributor, though evi- 25

dence is mixed, with Chodorow-Reich (2014) finding that credit shocks explain between onefifth and one-third of aggregate employment losses in the year after Lehman’s bankruptcy. However, for the smaller corporates taking out small business loans Greenstone et al. (2015) find that credit shocks help explain declines in borrowing, but led to only small overall employment effects. Duygan-Bump et al. (2015) also find that employment fell more in small firms in industries with high external financial dependence. They conclude that the small firm-high external financial dependence channel may explain about 8% of the rise in the aggregate unemployment rate, so overall the channel they identify may have had moderate aggregate implications as well. 5.5 Robustness I test for the validity and interpretation of the main results of the paper along several dimensions. As discussed, a concern is that lenders with below-average supply systematically located in counties with below-average employment shocks during the boom – perhaps risky lenders moved to risky counties during the boom years. I measure the credit supply instrument as in equation 2, with the same lender shocks during the recession φ , but this time b using 2000-2002 market shares (instead of 2005-2007 as in the baseline). Nonlocal Lending Shock2000−2002shares = X Share2000−2002φ (7) i i,b b B Figure 8 plots the baseline credit supply instrument measure against the instrument measured with 2000-2002 shares; the R-squared is close to 64 percent. Table 12 reports 2SLS results based on county exposure to lender shocks, with the exposure measured in 2000- 2002. For identification, the important aspect is the point estimates are very similar, which is evidence that βj are estimated consistently for different models j. The point estimates are indeed similar, though standard errors tend to be higher. For example, β is .087 in the total employment model (Column 3) while it is 0.114 in the baseline reported in Table 9, well within one standard error. The estimates are noisier – in the baseline, the first-stage F statistic was 43.83 whereas in this specification it is 20.10 – as is expected, due to the noise in measuring lender location in the early 2000s rather than immediately prior to the crisis. 26

I also run ‘placebo’ tests on the first and second stage equations. First, I regress yearly changes between mortgage credit (2000-2013) at the county-level on the 2007-2010 nonlocal lending shock and all the county controls used in the baseline specification. Figure 9 plots the coefficient estimates and associated 95 percent confidence intervals on a year-by-year basis. The mortgage credit shock helps explain credit changes over 2007-2008 and 2008- 2009 only, and not during any of the prerecession years.26 Second, I repeat the main 2SLS elasticity estimates, holding the right-hand side constant, but instead measuring left-hand side variables (e.g. employment changes) over the last two recessions: i) 1990-1992, during which the unemployment rate increased from 5.6 to 7.5 percent; and ii) 2000-2003, during which the unemployment rate increased from 4 percent (lowest since 1970) to 6 percent in 2003. Table 13 reports elasticity estimates for construction and total private employment. If coefficient estimates are positive and significant, that would indicate counties with belowaverage supply during the Great Recession tend to experience below-average employment shocks during other recessions, possibly for other unobserved characteristics of localities. However, the estimates are not significant, except for changes in total private employment over 1990-1992, though in this case the coefficient estimate has the opposite sign (negative rather than positive). Inthebaselineresultsofthepaper, Ididnotincludeinstitutionsthatfiledforbankruptcy (and were not acquired by another lender), because the portion of lending changes that is nonlocal cannot be plausibly isolated for these lenders, since lending for these institutions fell by 100% everywhere (there is no variation across localities). This is a conservative choice. The inclusion of these lenders might lead to biased elasticity estimates. On the other hand, their exclusion likely decreases the statistical power of the estimation approach. I add to the sample the ten largest multimarket lenders who failed over 2005-2010.27 Table 14 reports two stage least squares estimates when the credit supply instrument includes these large failed lenders. Their addition leads to a small increase in the first-stage F statistic. Moreover, the second stage estimates are very similar to the baseline. Some are a bit higher and some a 26Making a similar coefficient plot using a different dependent variable (e.g. total private employment) yields coefficient estimates which are not significant. The reason is there are efficiency gains with lumping the recession years into a single cross-section. Results available upon request. 27American Home Mortgage, New Century Financial, IndyMac, Fremont Investment, WMC Mortgage, Lehman, Ameriquest, Option One, First Magnus, and Taylor, Bean, & Whitaker Mortgage. 27

bit smaller, though all within one standard error of the baseline estimates. Ialsocheckwhethercoefficientestimatesarestatisticallydifferentwhenaddingadditional controls. In particular, I add squared and cubed terms of some of the most important drivers of the housing boom and bust identified in the literature: the runup in home prices over 2003-2006, 2006 debt-to-income, and the fraction of borrowers in a county with Equifax Risk Score 3.0 less than 620. Table 15 reports the main regressions of the paper (with region fixed effects), this time including as additional explanatory variables the squared and cubed prerecession terms of these three variables. The results are essentially identical, ameliorating concern about omitted variable bias. The total private employment coefficient estimate is 0.091, compared with 0.114 in the baseline. Finally, I show that the results in the paper are robust to controlling for realized declines in small business lending over 2007-2010, which I obtain from the Community Reinvestment Act dataset. I average the flow of new business originations over 2008-2010, and compute percent changes with respect to 2007. Table 16 shows that controlling for the change in small business lending does not affect the main results of the paper. This is evidence that the mortgage credit shock discussed in this paper is carefully identified, and pertains specifically to changes in the availability of mortgage credit. The total private employment coefficient estimate is 0.111, compared with 0.114 in the baseline. Part of the reason why the two channels are distinct is that the exposure of localities to small business and mortgage lenders is only weakly correlated. In other words, the small business lenders to a locality are often not the same as the mortgage lenders. Figure 10 plots HMDA shares against CRA shares for the top 4 banks; they are only weakly correlated. 6 Conclusion One of the leading narratives of the Great Recession is the credit crunch view – disruptions in financial markets limited the supply of new credit, which reduced the spending capacity of households and firms and lowered aggregate demand and employment, as discussed in prominent models of the Great Recession (Eggertsson and Krugman 2012; Guerrieri and Lorenzoni 2017; Midrigan and Philippon 2016). This paper contributes to this literature 28

by empirically quantifying the employment effects of changes in mortgage credit supply. The emphasis on mortgages complements existing research the majority of which focuses on corporatecreditsupplyshocks(Chodorow-Reich2014;Greenstoneetal.2015;Duygan-Bump et al. 2015). To do so, I construct a county level mortgage credit supply instrument, which exploits two sources of heterogeneity: differences in the extent to which lenders cut supply in the Great Recession for nonlocal reasons, and variation in the intensity of county-lender relations coming into the recession. I then estimate the effect of changes in mortgage supply on employment, net of other possibly confounding factors affecting spending during the recession. By quantifying the effects of mortgage supply reductions, this paper adds nuance to the debate on the drivers of the job losses during the Great Recession. Overall, the bottom line is that mortgage supply shocks mattered for employment, though only moderately so. Declines in mortgage supply caused declines in local real estate activity – in residential permits, house prices, and construction employment, for example – but the evidence does not suggest there were large spillover effects in other, broader employment categories. A partial equilibrium aggregation exercise, based on the estimated local elasticity of total private employment with respect to mortgage supply, indicates that the reduction in mortgage supply could explain about 13 percent of the employment losses in the Great Recession, or close to 1 million of the jobs lost. In other words, the reduction in mortgage supply likely aggravated the job losses to a meaningful extent. But, other factors – the decline in household net worth, increase in uncertainty, or credit supply reductions to firms – together likely explain the bulk of the job losses in the recession, particularly in sectors less directly linked to real estate. 29

References Adelino, M., Schoar, A., and Severino, F. (2012). Credit supply and house prices: Evidence from mortgage market segmentation. NBER Working Papers 17832, National Bureau of Economic Research, Inc. Alexandrov, A. and Koulayev, S. (2017). No Shopping in the U.S. Mortgage Market: Direct andStrategicEffectsofProvidingInformation. Unpublishedworkingpaper2017-01,CFPB Working Paper. Amiti, M. and Weinstein, D. E. (2018). How much do idiosyncratic bank shocks affect investment? evidence from matched bank-firm loan data. Journal of Political Economy, 126(2):525–587. Anenberg, E., Hizmo, A., Kung, E., and Molloy, R. (2016). The Effect of Mortgage Credit Availability on House Prices and Construction: Evidence from a Frontier Estimation Approach. Unpublished working paper. Avery, R., Brevoort, K., and Canner, G. (2011). The Mortgage Market in 2010: Highlights from the Data Reported under the Home Mortgage Disclosure Act. Federal Reserve Bulletin. Baker, S. R., Bloom, N., and Davis, S. J. (2016). Measuring Economic Policy Uncertainty. The Quarterly Journal of Economics, 131(4):1593–1636. Bernanke, B. S. and Lown, C. S. (1991). The Credit Crunch. Brookings Papers on Economic Activity, 22(2):205–248. Berrospide, J. M., Black, L. K., and Keeton, W. R. (2016). The cross-market spillover of economic shocks through multimarket banks. Journal of Money, Credit and Banking, 48(5):957–988. Bhutta, N. (2015). The ins and outs of mortgage debt during the housing boom and bust. Journal of Monetary Economics, 76(C):284–298. 30

Bhutta, N. and Ringo, D. R. (2014). The 2013 Home Mortgage Disclosure Act Data. Federal Reserve Bulletin, (Nov):A107–A146. Bloom,N.(2014). FluctuationsinUncertainty. Journal of Economic Perspectives,28(2):153– 76. Boldrin, M., Garriga, C., Peralta-Alva, A., and Sa´nchez, J. M. (2012). Reconstructing the great recession. Working Papers 2013-006, Federal Reserve Bank of St. Louis. Brunnermeier, M. K. (2008). Deciphering the Liquidity and Credit Crunch 2007-08. NBER Working Papers 14612, National Bureau of Economic Research, Inc. Calem, P., Covas, F., and Wu, J. (2013). The Impact of the 2007 Liquidity Shock on Bank Jumbo Mortgage Lending. Journal of Money, Credit and Banking, 45:59–91. Carroll, C. D. and Kimball, M. S. (1996). On the Concavity of the Consumption Function. Econometrica, 64(4):981–92. Cetorelli,N.andGoldberg,L.S.(2011). GlobalBanksandInternationalShockTransmission: Evidence from the Crisis. IMF Economic Review, 59(1):41–76. Chen, B. S., Hanson, S. G., and Stein, J. C. (2017). The Decline of Big-Bank Lending to Small Business: Dynamic Impacts on Local Credit and Labor Markets. NBER Working Papers 23843, National Bureau of Economic Research, Inc. Chinco, A. and Mayer, C. (2016). Misinformed speculators and mispricing in the housing market. The Review of Financial Studies, 29(2):486–522. Chodorow-Reich, G. (2014). The Employment Effects of Credit Market Disruptions: Firmlevel Evidence from the 2008-9 Financial Crisis. The Quarterly Journal of Economics, 129(1):1–59. Cornett, M. M., McNutt, J. J., Strahan, P. E., and Tehranian, H. (2011). Liquidity risk management and credit supply in the financial crisis. Journal of Financial Economics, 101(2):297–312. 31

Dagher,J.andKazimov,K.(2012). Banks’LiabilityStructureandMortgageLendingDuring the Financial Crisis. IMF Working Papers 12/155, International Monetary Fund. Dell0Ariccia, G., Igan, D., and Laeven, L. (2012). Credit Booms and Lending Standards: Evidence from the Subprime Mortgage Market. Journal of Money, Credit and Banking, 44:367–384. Demyanyk, Y. and Hemert, O. V. (2011). Understanding the Subprime Mortgage Crisis. Review of Financial Studies, 24(6):1848–1880. DiMaggio, M. and Kermani, A. (2016). Credit-Induced Boom and Bust. Working paper, Unpublished working paper. Duygan-Bump, B., Levkov, A., and Montoriol-Garriga, J. (2015). Financing constraints and unemployment: Evidence from the Great Recession. Journal of Monetary Economics, 75(C):89–105. Eggertsson, G. B. and Krugman, P. (2012). Debt, Deleveraging, and the Liquidity Trap: A Fisher-Minsky-Koo Approach. The Quarterly Journal of Economics, 127(3):1469–1513. Elsby, M.W.L., Hobijn, B., andSahin, A.(2010). TheLaborMarketintheGreatRecession. Brookings Papers on Economic Activity, 41(1 (Spring):1–69. Favara, G. and Imbs, J. (2015). Credit Supply and the Price of Housing. American Economic Review, 105(3):958–92. Frame, W. S., Fuster, A., Tracy, J., and Vickery, J. (2015). The rescue of Fannie Mae and Freddie Mac. Staff Reports 719, Federal Reserve Bank of New York. Fuster, A., Goodman, L., Lucca, D. O., Madar, L., Molloy, L., and Willen, P. S. (2013). The rising gap between primary and secondary mortgage rates. Economic Policy Review, (Dec):17–39. Gerardi, K. S., Shapiro, A. H., and Willen, P. S. (2008). Subprime outcomes: risky mortgages, homeownership experiences, and foreclosures. Working Papers 07-15, Federal Reserve Bank of Boston. 32

Gete, P. and Reher, M. (2016). Systemic Banks, Mortgage Supply and Housing Rents. Unpublished working paper. Giroud, X. and Mueller, H. M. (2015). Firm Leverage and Unemployment during the Great Recession. NBER Working Papers 21076, National Bureau of Economic Research, Inc. Glancy, D.(2015). HousingBust, BankLending&Employment: EvidencefromMultimarket Banks. Unpublished working paper. Goldsmith-Pinkham, P., Sorkin, I., and Swift, H. (2018). Bartik Instruments: What, When, Why, and How. NBER Working Papers 24408, National Bureau of Economic Research, Inc. Gorton, G. and Metrick, A. (2012). Securitized banking and the run on repo. Journal of Financial Economics, 104(3):425–451. Greenstone, M., Mas, A., and Nguyen, H.-L. (2015). Do Credit Market Shocks affect the Real Economy? Quasi-Experimental Evidence from the Great Recession and ‘Normal’ Economic Times. NBER Working Papers 20704, National Bureau of Economic Research, Inc. Gropp, R., Krainer, J., and Laderman, E. (2014). Did consumers want less debt? consumer credit demand versus supply in the wake of the 2008-2009 financial crisis. Working Paper Series 2014-8, Federal Reserve Bank of San Francisco. Guerrieri,V.andLorenzoni,G.(2017). Creditcrises,precautionarysavings,andtheliquidity trap*. The Quarterly Journal of Economics, 132(3):1427–1467. Haas, R. and Lelyveld, I. (2014). Multinational Banks and the Global Financial Crisis: Weathering the Perfect Storm? Journal of Money, Credit and Banking, 46(s1):333–364. Haltenhof, S., Jung Lee, S., and Stebunovs, V. (2014). The credit crunch and fall in employment during the Great Recession. Journal of Economic Dynamics and Control, 43(C):31– 57. 33

Haughwout, A. F., Lee, D., Tracy, J., and van der Klaauw, W. (2011). Real estate investors, the leverage cycle, and the housing market crisis. Staff Reports 514, Federal Reserve Bank of New York. Irani, R.M.andMeisenzahl, R.R.(2017). LoanSalesandBankLiquidityRiskManagement: Evidence from a U.S. Credit Register. Review of Financial Studies. Ivashina, V. and Scharfstein, D. (2010). Bank lending during the financial crisis of 2008. Journal of Financial Economics, 97(3):319–338. Kacperczyk, M. and Schnabl (2010). When Safe Proved Risky: Commercial Paper during the Financial Crisis of 2007-2009. Journal of Economic Perspectives, 24(1):29–50. Kaplan, G., Mitman, K., and Violante, G. L. (2017). The Housing Boom and Bust: Model Meets Evidence. NBER Working Papers 23694, National Bureau of Economic Research, Inc. Keys, B. J., Mukherjee, T., Seru, A., and Vig, V. (2010). Did Securitization Lead to Lax Screening? Evidence from Subprime Loans. The Quarterly Journal of Economics, 125(1):307–362. Khwaja, A. I. and Mian, A. (2008). Tracing the Impact of Bank Liquidity Shocks: Evidence from an Emerging Market. American Economic Review, 98(4):1413–42. Kung, E. (2015). The Effect of Credit Availability on House Prices: Evidence from the Economic Stimulus Act of 2008. Unpublished working paper. Lacko, J. M. and Pappalardo, J. K. (2010). The Failure and Promise of Mandated Consumer Mortgage Disclosures: Evidence from Qualitative Interviews and a Controlled Experiment with Mortgage Borrowers. American Economic Review, 100(2):516–521. Manski, C. F. (1993). Identification of Endogenous Social Effects: The Reflection Problem. Review of Economic Studies, 60(3):531–542. Mian, A., Rao, K., and Sufi, A. (2013). Household Balance Sheets, Consumption, and the Economic Slump. The Quarterly Journal of Economics, 128(4):1687–1726. 34

Mian, A. and Sufi, A. (2009). The consequences of mortgage credit expansion: Evidence from the u.s. mortgage default crisis*. The Quarterly Journal of Economics, 124(4):1449. Mian, A. and Sufi, A. (2011). House Prices, Home Equity-Based Borrowing, and the US Household Leverage Crisis. American Economic Review, 101(5):2132–56. Mian, A. and Sufi, A. (2014). What Explains the 2007–2009 Drop in Employment? Econometrica, 82:2197–2223. Midrigan, V. and Philippon, T. (2016). Household Leverage and the Recession. CEPR Discussion Papers 11407, C.E.P.R. Discussion Papers. Mondragon, J. (2014). Household Credit and Employment in the Great Recession. Unpublished working paper, UC Berkeley. Mondragon, J. (2018). Household Credit and Employment in the Great Recession. Unpublished working paper, Kellogg School of Management. Nadauld, T. D. and Sherlund, S. M. (2009). The Role of the Securitization Process in the Expansion of Subprime Credit. Working Paper Series 2009-9, Ohio State University, Charles A. Dice Center for Research in Financial Economics. Nguyen, H.-L. (2014). Do Bank Branches Still Matter? The Effect of Closings on Local Economic Outcomes. Unpublished working paper, MIT. Passmore, W. and Sherlund, S. M. (2016). Government-Backed Mortgage Insurance, Financial Crisis, and the Recovery from the Great Recession. Technical report. Passmore, W., Sherlund, S. M., and Burgess, G. (2005). The Effect of Housing Government- Sponsored Enterprises on Mortgage Rates. Real Estate Economics, 33(3):427–463. Ramcharan, R., Verani, S., and Heuvel, S. J. V. D. (2016). From Wall Street to Main Street: The Impact of the Financial Crisis on Consumer Credit Supply. Journal of Finance, 71(3):1323–1356. Saiz, A. (2010). The Geographic Determinants of Housing Supply. The Quarterly Journal of Economics, 125(3):1253–1296. 35

Schnabl, P. (2012). The international transmission of bank liquidity shocks: Evidence from an emerging market. The Journal of Finance, 67(3):897–932. Stock, J. H. and Yogo, M. (2002). Testing for Weak Instruments in Linear IV Regression. NBER Technical Working Papers 0284, National Bureau of Economic Research, Inc. Vojtech, C., Kay, B., and Driscoll, J. (2016). The Real Consequences of Bank Mortgage Lending Standards. Unpublished working paper, Office of Financial Research. Woodward, S. E. and Hall, R. E. (2012). Diagnosing consumer confusion and sub-optimal shopping effort: Theory and mortgage-market evidence. American Economic Review, 102(7):3249–76. 36

Figure 1: National Trends in Employment, House Prices, and Mortgage Originations 1382 1272 1154 972 912 826 800 745 680 655 624 605 547 521 500 segagtroM ,pmE ,PH dexednI 6002 2.1 1 8. 6. 4. 2000 2005 2010 2015 year Mortgage Originations Private Emp Case Schiller HPI Mortgage originations are defined as the dollar value (in trillions) of originations for 1-4 residential loans for home purchase and improvement. Source: HMDA. Figure 2: County Level Changes in Employment against Changes in Mortgage Credit Issuance, 2007-2010 Coef = 0.16 Std Err = 0.03 01-7002 pmE latoT D 1. 0 1.- 2.- -.6 -.4 -.2 0 D Mortgage Credit 2007-10 The figure plots changes in total private employment (y-axis) against changes in mortgage credit issuance (x-axis) over 2007-2010 at the county level for locations with over 15,000 housing units in the 2000 Census. The figure shows the linear coefficient estimate when regressing changes in employment on changes in mortgage credit issuance. Observations weighted by housing units in 2000 Census. Standard errors clustered at the division level.

Figure 3: Credit Changes 2007-2010 Versus Changes in 2003-2006 Plainscapital Corp Franklin Ame. US Bank Sierra Plaza Bancwest Ci PA t H i SP zR s K W G u F y e sH B R n e N n J n R lP o uP Z e T a C Ut y B o P s y M V c a i u i eu b g D n Fo b v F lC W i B A M r lR agl s A a s F oW g a n aat i u B f & e M e t ni t tr C l i t t s e s non N e s a e C N l r M h o t nd e y T E o k s l knt C r r D d Y l E n y i ta i i s t f & T t v c sr F S i H v q a B B T A e h B h CI F i y u n u l H I r B a e ei a a o r b n B i F n n r t d S V n Sr m a t y t a g c n C r k B A a O n Q u n m o . Un C k s c P n u t t . e a r ic o n k v d e id e n e r L n o t a I n N s G Ba M n & k T PNC Home U L S o A a A n Center PenFed Freedom M CaMpoitfaOl mOanhea 0102-7002 tiderC egagtroM D 5.1 1 5. 0 5.- 1- 0 2 4 6 D Mortgage Credit 2003-2006 Thefigureplotschangesinmortgagecreditover2007-2010versuschangesinmortgagecreditover2003-2006 for the large multimarket lenders in the sample. Figure 4: Lender Fixed Effects Estimates in Baseline vs Only High Income Loan Specification Plaza NVBRB&FTifth Third RS TU By DS nC A oSv AieursBraBVA ZioAnssPsoPFcHilaa Mtgesodtr atBranc Pulaski UAWMeCPlClrsoi tFviziadeBrengansonFtcwest M DH & Hu IWn R I E tP y iev n l u a ie gc l nr tth M b eo d ea n& rnt T k KRJeePygMGbio Q a P CuFnn e a u SNrsk en r iN c c Sa SaeF h kY vnua dee wy tnn o C d y n I att Fm N ra o Bb L uC Gn m a o sMU d tn am Bek n ra s nk CenBtr aolf BAanc. PFiNrCsCt Ai tTilly e n F n in . HSBC EquityOne CHaopmitael OLonaen Center Webster MMoofrOgmana hSatanley enilesab , f stceffe dexif redneL b 0 5.- 1- -1 -.5 0 Lender fixed effects f , high income loans only b Thefigureplotslenderfixedeffects(equation1)inthebaseline(y-axis)againstaspecificationthatusesonly high-income loans to estimate equation 1. Observations are weighted by the 2007 dollar value of mortgage originations. Outliers (5% upper tail) removed.

Figure 5: Funding Fragility and Lender Supply Bancw BB Ces V itti A zensF TZ B Di B o & n R s T A B s C so S c y Wi n a o tee v l F dlP u s i f s B u t F h l aaa n T rsgc h koi ird M&I HunMtin&gTton KeySbcSahnawknatabJnRPdMeegrCionsSuntrust Equ C Hi i t t Sy i OBnC B e No C CfY a eA p n C i t t o r a a m l l O m BP n aNF e n iCr c s . t Tenn. Webster b fstceffe dexif redneL 0 2.- 4.- 6.- 8.- 1- AssociateCdit F iBz i W f ea th nnBe B scal T lF B sn h & c i F r w dT aergsot SynovusRBC BB Z V ioT A nDs Pulaski HuntingtonM&T M&I SanStaunndterurst RegionsJPMKSCechybwaanbk NY Comm FirCsti tTien B n o . f A PNCCentral Banc. HSBCEquiCtyaOpniteal One Webster 0 .2 .4 .6 .8 1 Loans Orig. and Sold/Loans Originated, 2006-07 2.- 4.- 6.- 8.- 1- 0 .2 .4 .6 .8 1 Loans Sold to Private/Loans Sold, 2006-07 TDZion W s F e i l f l t s h BR S F T P BB a y h uC V A n r Ci lgB r i B a oA st d oisas v B zk o u ne & i c s cn T iwsaFetesdt Banc M&THuMn&tinIgton Schwab RegKioeSnyusbnatJnrPuksMtC Santander PNCCe C n a t p ra it Cl a Bi l B t i O a on n fc e A F . E ir N q sH Y t u TSi t C y eBO o nC m n n . e m Webster b f stceffe dexif redneL 0 2.- 4.- 6.- 8.- 1- Bancwest TD Ci B B tPiWz B B Z ueA & i V el o na s T l A n sls ssF s k o Fi ca R F iar i B g f t t e C o h d T B h a ir n S d c ynovus M&THuntingtoMn&I SJaPnMtaRSCnedugenirtorunssStchKweaybbank Citi B of A FPirNs N tC Y Te C n o n m . m Central Banc. HSBC EquityOneCapital One Webster .1 .2 .3 .4 .5 .6 .7 .8 Wholesale Funding/Assets, 2006-07 2.- 4.- 6.- 8.- 1- .02 .04 .06 .08 .1 Tier1 Capital/Assets, 2006-07 The variable on the y-axis measures differences in lender supply over 2007-2010, φ from equation 1. Varib ables on the x-axis are different measures of funding fragility over 2006-2007: ratio of mortgages originated and sold to total mortgages originated (top left); loans sold to private investors to total sales (top right); wholesalefundingtoassets(bottomleft);andTier1capitaltoassets(bottomright). Observationsweighted by mortgage originations in 2007. The banks in the sample are large multimarket lenders located in at least 100 counties and with originations in excess of $1 billion in 2007. Outliers (5% upper tail) removed. Figure 6: Persistent Market Shares The left panel plots county-lender HMDA market shares in 2007 (y-axis) against market shares in 2005 (xaxis). The right panel plots county-lender HMDA market shares in 2007 (y-axis) against market shares in 2000(x-axis). Lendersinthesamplewerelocatedinatleast100counties, issuedover$1billioninmortgage originationsin2007, anddidnotfailduringthecrisis. Countiesinthesamplehadover15,000housingunits in the 2000 Census.

Figure 7: Nonlocal Lending Shock The map plots the residual variation in the credit supply instrument (the nonlocal lending shock) after regressing the credit supply instrument on the county controls used throughout the paper and defined in Table 2. The instrument is defined in equation 2. The map sorts the nonlocal lending shock into quartiles for counties in the sample. Darker tones indicate relatively stronger supply. Missing observations left blank (in white). Figure 8: Nonlocal Lending Shock using 2000-2002 Market Shares R-sq = 0.64 VI enilesaB 0 2.- 4.- 6.- -.6 -.4 -.2 0 IV based on 2000-2002 lender location This figure plots the baseline credit supply instrument (nonlocal lending shock) on the y-axis, against the credit supply instrument which measures lender location over 2000-2002 in the x-axis. The baseline instrument measures lender location using 2005-2007 county-lender market shares as defined in equation 2.

Figure 9: Regressing Yearly Mortgage Credit Changes on Nonlocal Lending Shock 59IC/etamitsE ffeoC VI 5. 0 5.- 200120022003200420052006200720082009201020112012201320142015 Figure shows coefficient estimates (ρ ) and 95 percent confidence intervals when regressing yearly mortgage t credit changes at the county-level on the nonlocal lending shock NLS and the other controls used in the i baseline specification: ∆Credit =ρ NLS +γX +υ for t=2001,2002,...2014 i,t t i i i Figure 10: County-Lender Market Shares in HMDA and CRA R-sq=0.08 7002 erahs ADMH 8. 6. 4. 2. 0 Bank of America R-sq=0.31 0 .1 .2 .3 .4 .5 CRA Share 2007 1 8. 6. 4. 2. 0 Wells Fargo 0 .2 .4 .6 .8 CRA share 2007 R-sq=0.06 7002 erahs ADMH 8. 6. 4. 2. 0 JPMorgan Chase R-sq=0.05 0 .1 .2 .3 .4 .5 CRA share 2007 5. 4. 3. 2. 1. 0 Citibank 0 .1 .2 .3 .4 CRA share 2007 Thefigureplotsmortgage2007marketsharesfromHMDA(y-axis)against2007smallbusinessloanmarket shares from the CRA (x-axis) for each of the big-4 lenders and for localities with over 15,000 housing units in the 2000 Decennial Census.

Table 1: County Summary Statistics Dependent Variables, 2007-2010 percent changes Mean SD p10 Median p90 N ∆ Private Emp -0.065 0.047 -0.125 -0.063 -0.009 1013 ∆ Construction Emp -0.240 0.142 -0.413 -0.247 -0.059 987 ∆ Finance Emp -0.081 0.088 -0.189 -0.084 0.023 1012 ∆ Other Emp -0.048 0.049 -0.109 -0.047 0.011 1012 ∆ Nontradable Emp -0.045 0.063 -0.117 -0.051 0.034 1010 ∆ Home Prices -0.141 0.105 -0.290 -0.128 -0.013 1008 ∆ Home Permits -0.465 0.142 -0.645 -0.477 -0.267 932 ∆ Delinquency +90 Days 0.040 0.025 0.018 0.032 0.071 1012 ∆ Foreclosures 0.015 0.013 0.006 0.012 0.027 1013 ∆ Mortgage Credit -0.395 0.111 -0.530 -0.404 -0.248 1009 Prerecession Characteristics # Housing Units (thousands), 2000 94.599 179.490 17.803 41.248 216.948 1030 % White Pop, 2000 0.864 0.124 0.698 0.909 0.977 1030 % Educ ≥ College, 2000 0.216 0.089 0.121 0.196 0.341 1030 $HH Median Income (thousands), 2000 41.528 9.662 31.258 39.557 55.389 1030 $Home Value (thousands), 2000 108.300 47.862 66.600 95.250 161.750 1030 ∆ 2003-2006 Home Prices 0.250 0.181 0.063 0.203 0.516 1030 % Owner-Occupied Loans, 2003-2006 0.849 0.095 0.736 0.878 0.928 1030 % GSE-securitized Loans, 2003-2006 0.665 0.133 0.498 0.699 0.786 1030 % Nonconventional Loans, 2003-2006 0.184 0.111 0.050 0.172 0.324 1030 ∆ 2003-2006 #Lenders 0.419 0.296 0.111 0.364 0.797 1030 % Risk Score ≤ 620, 2006 0.270 0.082 0.174 0.257 0.392 1030 Median Risk Score, 2006 709.466 32.626 659.250 717.000 746.000 1030 HH Debt to Income, 2006 1.782 0.596 1.171 1.638 2.603 1030 Herfindahl Index, 2006 0.062 0.026 0.038 0.056 0.095 1030 Construction Share of Emp, 2006 0.122 0.045 0.075 0.113 0.184 1030 Tradable Share of Emp, 2006 0.135 0.082 0.044 0.120 0.247 1030 Unemployment Rate, 2007 4.749 1.410 3.300 4.500 6.400 1030 $Mortgage Credit (millions), 2007 905.114 2219.941 59.139 284.334 2185.790 1030 # Employed (thousands), 2007 97.738 214.502 11.192 35.106 229.155 1030 The table provides summary statistics for localities with over 15,000 households in the 2000 Decennial Census. The change in delinquency and foreclosure rates is in percentage points. For stocks (e.g. home prices), changes are taken between 2010Q4 and 2007Q4. For flows (e.g. mortgage originations), changes are taken between the average flow over 2008-2010 and the value in 2007.

Table 2: Data Definitions Variable Definition Source Dependent Variables, 2007-2010 percent changes Mortgage Credit By county-year, the dollar amount of HMDA originations for 1-4 residential loans for home purchase and improvement. ∆ Credit Percent change in average mortgage credit HMDA over 2008-2010 with respect to 2007 ∆ Residential Permits Percent change in average permits over Census 2008-2010 with respect to 2007 ∆ House Prices Percent change in house prices from 2007Q4 CoreLogic HPI to 2010Q4. ∆ Empj Percent change in employment category j QCEW from 2007Q4 to 2010Q4 ∆ Delinquency Rates Percentage point change in fraction of 90+ Black Knight McDash delinquent properties from 2007 to 2010 ∆ Foreclosures Rates Percentage point change in fraction of Black Knight McDash foreclosed properties from 2007 to 2010 Prerecession Characteristics Level Home Prices Log level median house price 2000 Census Household Income Median 2000 Census White population Fraction of population identified as white 2000 Census College population Fraction of population with a college degree 2000 Census or more Risk Score 3.0 Median 2006 FRBNY Consumer Credit Panel/Equifax Subprime Fraction of households in a county with Risk 2006 FRBNY Score less than 620 Consumer Credit Panel/Equifax Household Debt-to-Income Median household debt-to-income 2006 FRBNY Consumer Credit Panel/Equifax Nonconventional Loans One minus the fraction of loans issued over Black Knight McDash 2003-2006 identified as conventional loans GSE-securitized Loans Fraction of loans issued over 2003-2006 Black Knight McDash insured by GNMA, FNMA, or FHLMC Owner-Occupied Loans Fraction of mortgages over 2003-2006 HMDA identified as owner-occupied ∆ # Lenders Growth in the number of lenders per county HMDA over 2003-2006 ∆ House Prices Growth in house prices over 2003Q4-2006Q4 CoreLogic HPI Tradable Tradable share of employment, as defined in 2006 CBP Mian and Sufi (2014) Construction Construction share of employment, as 2006 CBP defined in Mian and Sufi (2014) Herfindahl Index Sum of squared market shares across lenders 2006 HMDA in county Unemp Rate Unemployment Rate 2007 BLS LAU Level Employment Log level of employed workers 2007 QCEW Level Mortgage Credit Log level of mortgage originations 2007 HMDA This table provides definitions and sources for the data used throughout the paper. HMDA: Home MortgageDisclosureAct;CBP:CountyBusinessPatterns;QCEW:QuarterlyCensusofEmploymentandWages; BLS: Bureau of Labor Statistics Local Area Unemployment Statistics.

Table 3: Lender Summary Statistics Mean SD p10 Median p90 N ∆ Mortgage Credit 2007-2010 -0.35 0.38 -0.76 -0.37 0.01 57 #Counties 2007 484.02 473.21 122.00 274.00 1117.00 57 Mortgage Credit 2007 (billions) 11.97 30.45 1.08 2.55 24.11 57 Sales/Originations 2006-2007 0.68 0.25 0.36 0.71 0.99 57 Private Loan Sales/Sales 2006-2007 0.66 0.34 0.18 0.73 1.00 57 Wholesale Funding/Assets 2006-2007 0.44 0.11 0.34 0.44 0.61 33 Tier 1 Capital/Assets 2006-2007 0.07 0.01 0.06 0.07 0.09 33 This table provides summary statistics for the lenders in the sample, which are large multimarket lenders located in at least 100 counties and with originations in excess of $1 billion in 2007. Table 4: Lender Rankings by Percent Changes in Mortgage Originations and Lender Fixed Effects Estimates Lender ∆Originations, Ranking by Ranking by Originations, 2007-2010 ∆Originations Baseline Lender $billions Fixed Effects 2007 Estimates US Bank .17 3 3 7.45 Flagstar -.13 8 14 10.37 USAA -.14 9 12 8.86 BB&T -.15 12 7 6.84 Provident -.18 14 19 5.64 Fifth Third -.21 18 5 6.41 PPH Mort -.25 19 16 12.11 Wells Fargo -.37 27 20 129.73 UAMC -.38 30 22 4.52 Navy FCU -.43 31 39 3.3 Pulte -.45 35 24 4.05 DHI -.48 39 27 5.09 Regions -.49 40 40 6.3 Suntrust -.53 41 36 27.86 NY Comm -.54 42 44 12.01 ING Bank -.58 44 32 5.34 EquityOne -.61 45 52 3.18 HSBC -.65 46 51 10.89 Bank of America -.65 47 45 182.57 Citibank -.69 48 49 29.27 Ally Fin -.71 49 47 16.63 PNC -.76 51 48 24.11 JPMC -.76 52 35 77.69 First Tennessee -.79 54 50 17.05 Capital One -.86 55 53 8.82 The table shows summary statistics for the 25 largest lenders in the sample ( large nonfailed multimarket lenders,seeSection2). Column2rankslendersbydeclineinnewmortgagelending. Column3rankslenders by baseline lender fixed effects estimates φ from equation 1. b

Table 5: Funding Fragility and Differences in Lender Supply Dependent variable: φb (1) (2) (3) (4) Coef./SE Coef./SE Coef./SE Coef./SE Wholesale Debt/Assets 2006-2007 -0.494*** -0.406*** -0.431*** -0.430*** (0.093) (0.104) (0.114) (0.141) Loan Sales/Originations 2006-2007 -0.584*** -0.629*** -0.572*** -0.542*** (0.130) (0.129) (0.164) (0.147) Private Loan Sales/Originations 2006-2007 -0.311*** -0.302*** -0.285*** -0.540*** (0.097) (0.094) (0.100) (0.143) Tier1 Capital 2006-2007 0.216 0.208 0.219 (0.128) (0.130) (0.143) ∆ Mortgage Credit 2003-2006 -0.065 0.048 (0.114) (0.141) Weighted Yes Yes Yes No N 32 32 32 32 R-squared 0.74 0.77 0.77 0.59 Adj R-squared 0.71 0.73 0.72 0.51 The dependent variable measures differences in lender supply over 2007-2010, φ from equation 1. The exb planatory variables measure the extent to which banks relied on fragile funding sources over 2005-2007, and credit growth over 2003-2006. Standard errors are in parentheses. Banks in the sample are large multimarket lenders as described in Section 2. All variables are standardized. 1% upper tail of lender fixed effects winsorized. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 6: Mortgage Market Shares are Highly Persistent Year-on-Year Dependent variable: County-Lender Market Shares 2007 Bottom credit risk quartile Top credit risk quartile County FE Coef./SE Coef./SE Coef./SE Coef./SE 2005 Market Shares 0.918*** 0.922*** 0.920*** 0.916*** (0.00) (0.01) (0.01) (0.00) County FE No No No Yes R-squared 0.92 0.89 0.93 0.92 Observations 33658 8169 8071 33658 This table show results from regressing 2007 county-lender market shares on 2005 county-lender market shares. Column2restrictsthesampletothelowEquifaxRiskScore3.0quartile, Column3tothehighRisk Score quartile, and Column 4 includes county fixed effects. The lenders in the sample are large multimarket lenders located in at least 100 counties and with originations in excess of $1 billion in 2007. Counties in the sample had over 15,000 households in the 2000 Decennial Census. Standard errors clustered at the county level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively.

Table 7: First Stage Results Dependent variable: ∆ Mortgage Credit 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE Nonlocal Lending Shock 0.479*** 0.526*** 0.430*** 0.247*** (0.11) (0.08) (0.09) (0.07) All other controls Yes Yes Yes Yes R-squared 0.56 0.62 0.67 0.76 Adj. R-squared 0.56 0.62 0.66 0.75 Kleibergen-Paap F stat 20.02 42.98 24.41 13.95 Observations 1009 1009 1009 1009 This table shows first-stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 2. Observationsweightedbythenumberofhousingunitsinthe2000DecennialCensus. Dependentvariable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 8: Housing Elasticities with respect to Mortgage Supply Dependent variables 2007-2010: ∆ Permits ∆ Home Price ∆ Delinq. Rate ∆ Foreclosure Rate Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.824*** 0.757*** -0.143*** -0.091*** (0.07) (0.15) (0.05) (0.04) All other controls Yes Yes Yes Yes Region fixed effects Yes Yes Yes Yes R-squared 0.44 0.75 0.77 0.56 Kleibergen-Paap F stat 45.54 40.06 50.06 52.36 Observations 919 991 997 998 This table shows the effects of changes in mortgage credit, when instrumented using the nonlocal lending shock, on changes in local outcomes for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation 2). All equations include region fixed effects and all characteristics of localities used throughout the paper defined inTable2. Observationsweightedbythenumberofhousingunitsinthe2000DecennialCensus. Dependent variable outliers (1 percent of each tail) are dropped. Standard errors are clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively.

Table 9: Employment Elasticities with respect to Mortgage Supply Dependent variables 2007-2010: ∆ Constr. Emp ∆ Fin Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.381*** 0.440*** 0.114*** 0.041 0.066 (0.08) (0.07) (0.04) (0.05) (0.10) All other controls Yes Yes Yes Yes Yes Region fixed effects Yes Yes Yes Yes Yes R-squared 0.63 0.16 0.53 0.41 0.41 Kleibergen-Paap F stat 43.32 44.29 43.83 43.50 45.42 Observations 967 991 992 991 989 This table shows the effects of changes in mortgage credit, when instrumented using the nonlocal lending shock, on changes in local outcomes for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation 2). All equations include region fixed effects and all characteristics of localities used throughout the paper defined inTable2. Observationsweightedbythenumberofhousingunitsinthe2000DecennialCensus. Dependent variable outliers (1 percent of each tail) are dropped. Standard errors are clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 10: Elasticity of Construction Employment with Housing Supply Interaction Dependent variables 2007-2010: ∆ Constr. Emp ∆ Constr. Emp ∆ Permits ∆ Permits Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.332*** 0.358*** 0.882*** 0.962*** (0.09) (0.10) (0.13) (0.15) ∆ Mortgage Credit 2007-10 × Elasticity 0.128*** 0.205* (0.05) (0.12) All other controls Yes Yes Yes Yes Kleibergen-Paap F stat 32.88 15.72 30.39 13.71 Observations 538 538 511 511 This table shows the effects of changes in mortgage credit over 2007-2010, interacted with the housing supply elasticity of Saiz (2010), on changes in construction employment and permit issuance during the recession. All regressions include region fixed effects and all other observed characteristics of localities used in the other tables in the paper (Table 2). The nonlocal lending shock measures the exposure of counties to lender shocks as defined in equation 2. Observations weighted by the number of housing units in the 2000 Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively.

Table 11: OLS Estimation Results Dependent variables: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.706*** 0.364*** 0.160*** 0.141*** 0.131*** (0.09) (0.08) (0.01) (0.02) (0.02) All other controls Yes Yes Yes Yes Yes R-squared 0.44 0.63 0.54 0.44 0.42 Adj. R-squared 0.43 0.63 0.52 0.43 0.40 Observations 919 967 992 991 989 This table shows that the OLS coefficients when regressing changes in outcome variables (e.g. home permits) on changes in mortgage credit at the county-level over 2007-2010 while controlling for all prerecession county characteristics listed in Table 2. Observations weighted by the number of housing units in the 2000 Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 12: Elasticities With IV Constructed Using 2000-2002 Shares Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.616*** 0.165 0.087 0.017 -0.046 (0.12) (0.18) (0.06) (0.07) (0.14) All other controls Yes Yes Yes Yes Yes R-squared 0.44 0.62 0.52 0.40 0.37 Kleibergen-Paap F stat 20.69 19.92 20.10 20.96 21.69 Observations 919 967 992 991 989 This table shows the effects of changes in mortgage credit over 2007-2010, when instrumented using the nonlocal lending shock based on 2000-2002 market shares (as opposed to the baseline measure which uses 2005-2007 shares); see equation 7. All regressions include region fixed effects and all other observed characteristics of localities used in the other tables in the paper (Table 2). The nonlocal lending shock measures the exposure of counties to lender shocks as defined in equation 2. Observations weighted by the number of housing units in the 2000 Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively.

Table 13: ‘Placebo’ Regressions ∆ Total 90-92 ∆ Total 00-03 ∆ Constr 90-92 ∆ Constr 00-03 Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 -0.107*** 0.054 -0.043 0.162 (0.03) (0.08) (0.12) (0.15) All other controls Yes Yes Yes Yes Observations 1009 1009 968 968 This table reports results from ‘placebo’ regressions over the previous two recessions. The dependent variables are in percent change over 1990-1992 and 2000-2003. All regressions include region fixed effects and all other observed characteristics of localities used in the other tables in the paper (Table 2). Observations weighted by 1990 payrolls. Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 14: Elasticity Estimates Including Failed Lenders Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.931*** 0.372*** 0.096** 0.028 0.065 (0.10) (0.10) (0.04) (0.06) (0.11) All other controls Yes Yes Yes Yes Yes R-squared 0.43 0.63 0.52 0.40 0.41 Kleibergen-Paap F stat 51.43 44.39 44.90 44.43 46.51 Observations 919 967 992 991 989 Thistableshowstheeffectsofchangesinmortgagecreditover2007-2010,wheninstrumentedusingthenonlocal lending shock, including large institutions who filed for bankruptcy over 2005-2010: American Home Mortgage, New Century Financial, IndyMac, Fremont Investment, WMC Mortgage, Lehman, Ameriquest, Option One, First Magnus, and Taylor, Bean, & Whitaker Mortgage. All equations include region fixed effects and all other observed characteristics of localities used in the other tables in the paper (Table 2). The nonlocal lending shock measures the exposure of counties to lender shocks as defined in equation 2. Observations weighted by the number of housing units in the 2000 Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively.

Table 15: Elasticity Estimates with Additional Controls Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.859*** 0.400*** 0.091** 0.010 0.051 (0.13) (0.09) (0.04) (0.06) (0.10) All other controls Yes Yes Yes Yes Yes R-squared 0.49 0.64 0.53 0.40 0.43 Kleibergen-Paap F stat 29.57 32.40 32.76 32.04 33.46 Observations 919 967 992 991 989 Thistableshowstheeffectsofchangesinmortgagecreditover2007-2010,wheninstrumentedusingthenonlocal lending shock, on local outcomes. These regressions include squared and cubed terms for household debt-to-income, the local fraction of subprime borrowers, and the runup in home prices over 2003-2006. All equations include region fixed effects and all other observed characteristics of localities used in the other tables in the paper (Table 2). The nonlocal lending shock measures the exposure of counties to lender shocks as defined in equation 2. Observations weighted by the number of housing units in the 2000 Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 16: Elasticity Estimates Including Changes in Small Business Lending (CRA) Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.821*** 0.376*** 0.111*** 0.037 0.063 (0.08) (0.08) (0.04) (0.05) (0.10) All other controls Yes Yes Yes Yes Yes R-squared 0.44 0.64 0.53 0.42 0.41 Kleibergen-Paap F stat 52.40 47.94 49.89 49.50 51.50 Observations 919 967 992 991 989 Thistableshowstheeffectsofchangesinmortgagecreditover2007-2010,wheninstrumentedusingthenonlocal lending shock, on local outcomes. I include changes in small business lending over 2007-2010 obtained from the Community Reinvestment Act. All equations include region fixed effects and all other observed characteristics of localities used in the other tables in the paper (Table 2). The nonlocal lending shock measures the exposure of counties to lender shocks as defined in equation 2. Observations weighted by the number of housing units in the 2000 Decennial Census. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively.

Employment in the Great Recession: How Important Were Household Credit Supply Shocks? Online Appendix The appendix consists of tables complementing the main results in the paper. Tables 1- 9 report second stage results for each of the main dependent variables in the paper, for specifications with no spatial fixed effects, region, division, and state fixed effects (Columns 1-4, respectively). In the paper, changes in mortgage credit issuance during the recession ∆Credit are defined as the percent change in the average real dollar value of mortgage originations over 2008- 2010 with respect to their value in 2007. Estimates are robust to alternative definitions, such as using the average real dollar flow of originations over 2005-2007 as the base (Table 10) or using the percent change in the number (rather than the real dollar value) of mortgage originations between 2007-2010 (Table 11). In the baseline regressions, observations are weighted by population. Table 12 shows results from unweighted regressions, for close to the largest 500 counties in the sample, where the number of households (from the 2000 Decennial Census) exceeds 40,000. Finally, Table 13 shows estimates when clustering standard errors at the commuting zone level. List of Tables 1 Elasticity of Residential Permits with respect to Mortgage Supply . . . . . . 3 2 Elasticity of Home Prices with respect to Mortgage Supply . . . . . . . . . . 3 3 Elasticity of Delinquency Rates with respect to Mortgage Supply . . . . . . 4 4 Elasticity of Foreclosure Rates with respect to Mortgage Supply . . . . . . . 4 5 Elasticity of Construction Employment with respect to Mortgage Supply . . 5 6 Elasticity of Financial Employment with respect to Mortgage Supply . . . . 5 7 Elasticity of Private Employment with respect to Mortgage Supply . . . . . 6 8 Elasticity of Other Employment with respect to Mortgage Supply . . . . . . 6 9 Elasticity of Nontradable Employment with respect to Mortgage Supply . . . 7 10 Estimates When Defining ∆Credit Using 2005-2007 as the Base Period . . . 7 11 Estimates When Defining ∆Credit Using Changes in the Number of Loans . 8 1

12 Estimates With No Population Weighting . . . . . . . . . . . . . . . . . . . 8 13 Estimates with Standard Errors Clustered at Commuting Zone . . . . . . . . 9 2

Table 1: Elasticity of Residential Permits with respect to Mortgage Supply Dependent variable: ∆ Permits 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.561*** 0.824*** 0.845*** 1.098*** (0.11) (0.07) (0.09) (0.20) All other controls Yes Yes Yes Yes K-P F stat 21.18 45.54 26.17 10.46 p-value K-P LM test 0.01 0.01 0.01 0.03 Observations 919 919 919 919 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 2: Elasticity of Home Prices with respect to Mortgage Supply Dependent variable: ∆ Home Price 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.673*** 0.757*** 0.828*** 0.789*** (0.16) (0.15) (0.17) (0.16) All other controls Yes Yes Yes Yes K-P F stat 18.42 40.06 21.77 13.10 p-value K-P LM test 0.02 0.01 0.02 0.02 Observations 991 991 991 991 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 3

Table 3: Elasticity of Delinquency Rates with respect to Mortgage Supply Dependent variable: ∆ Delinq. Rate 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 -0.187*** -0.143*** -0.141** -0.168*** (0.05) (0.05) (0.06) (0.06) All other controls Yes Yes Yes Yes K-P F stat 22.99 50.06 25.99 14.66 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 997 997 997 997 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 4: Elasticity of Foreclosure Rates with respect to Mortgage Supply Dependent variable: ∆ Foreclosure Rate 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 -0.118*** -0.091*** -0.084* -0.068** (0.04) (0.04) (0.05) (0.03) All other controls Yes Yes Yes Yes K-P F stat 23.52 52.36 29.50 15.50 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 998 998 998 998 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 4

Table 5: Elasticity of Construction Employment with respect to Mortgage Supply Dependent variable: ∆ Constr. Emp 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.209 0.381*** 0.217 0.672*** (0.15) (0.08) (0.15) (0.22) All other controls Yes Yes Yes Yes K-P F stat 20.33 43.32 26.29 16.35 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 967 967 967 967 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 6: Elasticity of Financial Employment with respect to Mortgage Supply Dependent variable: ∆ Fin Emp 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.435*** 0.440*** 0.478*** 0.893*** (0.08) (0.07) (0.09) (0.19) All other controls Yes Yes Yes Yes K-P F stat 20.45 44.29 24.79 13.47 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 991 991 991 991 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 5

Table 7: Elasticity of Private Employment with respect to Mortgage Supply Dependent variable: ∆ Total Emp 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.081 0.114*** 0.068 0.083 (0.05) (0.04) (0.06) (0.10) All other controls Yes Yes Yes Yes K-P F stat 19.92 43.83 24.56 14.95 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 992 992 992 992 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 8: Elasticity of Other Employment with respect to Mortgage Supply Dependent variable: ∆ Other Emp 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.021 0.041 0.009 -0.048 (0.06) (0.05) (0.07) (0.13) All other controls Yes Yes Yes Yes K-P F stat 19.91 43.50 24.72 14.17 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 991 991 991 991 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 6

Table 9: Elasticity of Nontradable Employment with respect to Mortgage Supply Dependent variable: ∆ Nontr. Emp 2007-2010 No FE Region FE Division FE State FE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 -0.010 0.066 0.053 0.183 (0.15) (0.10) (0.13) (0.11) All other controls Yes Yes Yes Yes K-P F stat 20.49 45.42 25.52 14.28 p-value K-P LM test 0.02 0.01 0.01 0.02 Observations 989 989 989 989 This table shows first stage results from regressing changes in mortgage credit issuance over 2007-2010 on the credit supply instrument (the nonlocal lending shock) for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation2). AllequationsincludeallcharacteristicsoflocalitiesusedthroughoutthepaperdefinedinTable 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Dpendent variable outliers (1 percent of each tail) are dropped. Standard errors clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 10: Estimates When Defining ∆Credit Using 2005-2007 as the Base Period Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.809*** 0.406*** 0.123*** 0.050 0.073 (0.11) (0.09) (0.04) (0.05) (0.10) All other controls Yes Yes Yes Yes Yes K-P F stat 21.11 18.79 19.99 19.92 19.55 p-value K-P LM test 0.01 0.02 0.02 0.02 0.02 Observations 917 964 989 988 986 This table shows the effects of changes in mortgage credit, when instrumented using the nonlocal lending shock,onchangesinlocaloutcomesforcountieswithover15,000housingunitsinthe2000Census. ∆Credit measured as the change in the real dollar value of morgage originations over 2008-2010 relative to the value in 2005-2007. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation 2). All equations include region fixed effects and all characteristics of localities used throughout the paper defined in Table 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Outliers (1 percent of each tail) are dropped. Standard errors are clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 7

Table 11: Estimates When Defining ∆Credit Using Changes in the Number of Loans Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.855*** 0.411*** 0.123*** 0.046 0.064 (0.10) (0.09) (0.04) (0.05) (0.11) All other controls Yes Yes Yes Yes Yes K-P F stat 66.29 62.70 69.67 68.60 72.04 p-value K-P LM test 0.01 0.01 0.01 0.01 0.01 Observations 918 967 993 991 990 This table shows the effects of changes in mortgage credit, when instrumented using the nonlocal lending shock,onchangesinlocaloutcomesforcountieswithover15,000housingunitsinthe2000Census. ∆Credit measured as the change in the average number of originations over 2008-2010 relative to the 2007 number. Thenonlocallendingshockmeasurestheexposureofcountiestononlocallendershocks(seeequation2). All equations include region fixed effects and all characteristics of localities used throughout the paper defined inTable3. Observationsweightedbythenumberofhousingunitsinthe2000DecennialCensus. Outliers(1 percent of each tail) are dropped. Standard errors are clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. Table 12: Estimates With No Population Weighting Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.808*** 0.366*** 0.110*** 0.039 0.034 (0.07) (0.09) (0.04) (0.05) (0.11) All other controls Yes Yes Yes Yes Yes K-P F stat 38.46 50.21 50.94 51.08 50.96 p-value K-P LM test 0.01 0.01 0.01 0.01 0.01 Observations 483 505 515 516 514 This table shows the effects of changes in mortgage credit, when instrumented using the nonlocal lending shock, on changes in local outcomes for counties with over 40,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation 2). All equations include region fixed effects and all characteristics of localities used throughout the paper defined in Table 3. Outliers (1 percent of each tail) are dropped. Standard errors are clustered at the division level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 8

Table 13: Estimates with Standard Errors Clustered at Commuting Zone Dependent variables 2007-2010: ∆ Permits ∆ Constr. Emp ∆ Total Emp ∆ Other Emp ∆ Nontr. Emp Coef./SE Coef./SE Coef./SE Coef./SE Coef./SE ∆ Mortgage Credit 2007-2010 0.824*** 0.381*** 0.114** 0.041 0.066 (0.21) (0.11) (0.04) (0.05) (0.08) All other controls Yes Yes Yes Yes Yes K-P F stat 55.03 59.19 62.21 62.34 58.30 p-value K-P LM test 0.00 0.00 0.00 0.00 0.00 Observations 919 967 992 991 989 This table shows the effects of changes in mortgage credit, when instrumented using the nonlocal lending shock, on changes in local outcomes for counties with over 15,000 housing units in the 2000 Census. The nonlocal lending shock measures the exposure of counties to nonlocal lender shocks (see equation 2). All equations include region fixed effects and all characteristics of localities used throughout the paper defined in Table 3. Observations weighted by the number of housing units in the 2000 Decennial Census. Outliers (1 percent of each tail) are dropped. Standard errors are clustered at the commuting zone level. *, **, and *** indicate significance at the 0.1, 0.05, and 0.01 levels, respectively. 9