The Employment Eects of Credit Market Disruptions: Firm-level Evidence from the Financial Crisis

Transcription

1 The Employment Eects of Credit Market Disruptions: Firm-level Evidence from the Financial Crisis Gabriel Chodorow-Reich October 2012 Abstract This paper investigates the eect of bank lending frictions on employment outcomes. I construct a new dataset that combines information on banking relationships and employment at two thousand nonnancial rms during the crisis. The paper rst veries empirically the importance of banking relationships, which imply a cost to borrowers that switch lenders. I then use the dispersion in lender health following the Lehman crisis as a source of exogenous variation in the availability of credit to borrowers. I nd that credit matters. Firms that had pre-crisis relationships with less healthy lenders had a lower likelihood of obtaining a loan following the Lehman bankruptcy, paid a higher interest rate if they did borrow, and reduced employment by more compared to pre-crisis clients of healthier lenders. Consistent with frictions deriving from asymmetric information, the eects vary by rm type. Lender health has an economically and statistically signicant eect on employment at small and medium rms, but the data cannot reject the hypothesis of no eect at the largest or most transparent rms. I use a general equilibrium model to relate the cross-sectional empirical approach to the aggregate eect of the credit frictions on employment. I nd that the withdrawal of credit accounts for between one-third and one-half of the decline in employment at small and medium rms in the sample in the year following Lehman. UC Berkeley Department of Economics, 530 Evans Hall #3880, Berkeley, California (gabecr@econ.berkeley.edu). I thank Christina Romer and David Romer for extensive advice and support. Peter Ganong, Yuriy Gorodnichenko, Joshua Hausman, Ulrike Malmendier, Atif Mian, Michael Reich, Ricardo Reis, Jesse Rothstein, Johannes Wieland, Michael Woodford, seminar participants at the UC Berkeley GEMS, Columbia Summer Workshop, and the NBER Summer Institute Finance and Macroeconomics Workshop, and my discussant Robert Hall at the NBER Summer Institute all provided many valuable comments. The National Science Foundation and the National Science Foundation Graduate Research Fellowship provided nancial support. This research was conducted with restricted access to Bureau of Labor Statistics (BLS) data. The views expressed here do not necessarily reect the views of the BLS or the U.S. government. I am particularly grateful to Jessica Helfand and Michael LoBue of the BLS for their help with the Longitudinal Database. Any remaining errors are of course my own.

2 1. Introduction Does the health of banks on Wall Street aect economic outcomes on Main Street? In the wake of the nancial crisis, this question has generated substantial interest among politicians, the popular press, and the public at large. The renewed interest partly reects the deeply unpopular government support for nancial institutions, which policymakers defended by arguing that not providing such support would have dire implications for jobs in sectors far removed from banking (see e.g. Bernanke 2008). Notwithstanding the policy interventions, nancial sector lending to nonnancial rms in the United States contracted signicantly during the crisis, and the economy experienced the sharpest decline in employment since the Great Depression. This paper investigates the link between credit market frictions and employment using a newly constructed dataset. Previous studies of the eects of nancial constraints at the rm level have either not observed non-nancial outcomes, or limited attention to rms with information in databases that by construction contain only the most transparent, largest rms. The dataset I construct merges the Dealscan syndicated loan database, which contains the borrowing history of both public and private rms that have accessed the syndicated loan market, with condential employment data from the Bureau of Labor Statistics Longitudinal Database. The new dataset contains information on employment outcomes and banking relationships at two thousand nonnancial rms, ranging in size from fewer than fty employees to more than ten thousand. I then compare employment outcomes at rms that had borrowed before the crisis from relatively healthy nancial institutions with otherwise similar rms that had borrowed from lenders more adversely aected during the crisis. Having a relationship with a troubled bank had large real consequences. Comparing rms that had pre-crisis relationships with lenders at the 10th and 90th percentile of the distribution of lender health during the crisis, the client of the more constrained lender had a one-third lower likelihood of receiving a loan during the crisis, and reduced employment by an additional four percentage points in the year following the bankruptcy of Lehman Brothers. The importance of the credit supply channel also varied greatly by rm type. The credit channel had a signicant eect on employment at small and medium-sized rms (fewer than one thousand employees), and at rms without access to public debt markets. For these rms, the withdrawal of credit explains between one-third and one-half of the employment decline in the year following Lehman. In contrast, the data cannot reject that the relative availability of bank credit supply had no eect on relative employment outcomes at the largest or most transparent rms. The paper's methodological approach relies on two key facts established below. First, 1

3 bank-borrower relationships are sticky. This means that rms that borrowed before the crisis from banks that did less lending during the crisis would have greater diculty obtaining bank nancing than rms that had borrowed from healthier lenders. Second, the origins of the 2008 crisis lay outside of the corporate loan sector, implying that the cross-sectional variation in banks' willingness to make corporate loans was plausibly orthogonal to the characteristics of each banks' pre-crisis borrowers. To further clarify the logic of the exercise, consider the examples of U.S. Bankcorp and Credit Suisse, both active lenders in the syndicated market in the United States. During the nancial crisis, Credit Suisse suered large losses from exposure to mortgage-backed securities, and its stock price declined by 60 percent during U.S. Bankcorp had relatively little exposure to mortgage-backed securities, experienced one of the smallest stock price declines among major banks during the crisis, and in July 2009 became among the rst banks to fully repay its TARP commitment to the U.S. Treasury. In the nine months between the bankruptcy of Lehman Brothers and the end of the recession, Credit Suisse reduced its lending in the syndicated market by about 79 percent relative to before the crisis, whereas U.S. Bankcorp reduced lending by only 14 percent. As a result, rms with pre-crisis lending syndicates where U.S. Bankcorp had a lead role were nearly four times as likely to receive a loan during the crisis as rms with syndicates where Credit Suisse had a lead role. The paper establishes the causal eect of credit supply in the general case in a series of steps. I rst document the importance of borrower-lender relationships using data from the syndicated loan market. In a syndicated loan, one or more lead arrangers set up the loan deal with the borrower and provide most of the nancing, and recruit other participant lenders to provide the remainder of the funds. Among borrowers that have previously accessed the syndicated market, about three-quarters return to the same lead arranger as they used for their prior loan. This empirical stickiness in borrower-lender relationships exceeds by a factor of seven what one would predict based only on lenders' market shares, indicating frictions to switching lenders. I next discuss measures of bank health in order to isolate the eect of credit supply. Unfortunately, banks' internal cost of funds are not directly observable. Instead, I measure the relative health of a rm's lenders using the amount of lending to other borrowers during the crisis by the rm's pre-crisis syndicate. The validity of this measure relies in part on systematic variation in bank-level total lending during the crisis coming from variation in bank health. Two pieces of evidence corroborate this assumption. First, banks that had larger reductions in the quantity of loans raised the prices of new loans by more. The negative comovement of quantity and price in the cross-section of lenders strongly suggests a shift in the loan supply curve. Second, new lending during the crisis correlates robustly with variables 2

4 such as the stock return, exposure to Lehman Brothers through the syndicated market, and various balance sheet and income statement items. The reasons to prefer the lending-based measure to these other measures of bank health include its consistent construction across banks, that it applies to the relevant level of bank ownership, and that it likely has a tighter relationship with the internal cost of funds. With this measure in hand, I turn to the consequences of the sharp contraction in credit supply following the collapse of Lehman Brothers. Pre-crisis clients of banks in worse nancial condition had a lower likelihood of receiving a loan from any lender in the nine months following the Lehman failure. Moreover, among banks that did obtain a loan, interest spreads increased more. These ndings resemble those of Ivashina and Scharfstein (2010) and Santos (2011) using similar data, but dier from those papers by focusing on borrower outcomes rather than bank outcomes, thereby establishing that borrowers of weaker banks could not simply switch to healthier banks during the crisis. I then present evidence that the eects in the loan market translated into eects on real outcomes for the borrowers. To do so, I link the Dealscan syndicated loan dataset with information on quarterly rm employment from the Bureau of Labor Statistics Longitudinal Database. In principle, rms that lose access to the syndicated loan market could compensate by switching to other forms of nancing, without having to reduce production inputs (Becker and Ivashina 2010; Adrian, Colla and Shin Forthcoming). However, the data indicate a signicant eect on employment at these rms. As noted above, employment at pre-crisis clients of lenders at the 10th percentile of bank health fell by an average of four percentage points more than at clients of lenders at the 90th percentile in the year following the Lehman Bankruptcy. The gap does not close at all over the following twelve months, suggesting that the nancing frictions have persistent eects. The employment shortfall at rms that had borrowed from less healthy lenders can inform an estimate of the aggregate eects of the nancial frictions under certain conditions. In general equilibrium, some demand shifts from the more credit-constrained to the less constrained rms, inducing an increase in labor demand at the less constrained rms. In the opposite direction, the reduction in aggregate expenditure caused by the nancial crisis lowers labor demand at the less constrained rms. I present a general equilibrium model that illustrates these channels and the link between the estimated employment regressions and the aggregate employment decline. Calibration of the model suggests that the general equilibrium eects either magnify the employment losses observed by comparing borrowers of dierent lenders or have at most a modest attenuating eect. I confront the key identication challenge the possibility that borrowers of banks in worse nancial condition share other unobservable characteristics that inuence loan demand 3

5 and employment outcomes along several dimensions. First, I make use of a rich set of observed borrower characteristics, including state, industry, and a measure of ex-ante riskiness, to constrict the space of unobservable characteristics that could contaminate the results. Second, restricting the variation in bank health to coming from particular sources yields results quantitatively similar to the baseline estimate. Third, I do two specication checks with a subsample constructed to eliminate the possibility of unobserved loan demand driving the loan outcome results. The rst of these shows that, among borrowers that actually obtain a loan during the crisis, the likelihood they used their pre-crisis lead lender increases with the overall amount of lending by the lead lender. The latter implements the Khwaja and Mian (2008) within-rm estimator, which uses borrower xed eects to absorb all possible variation in unobserved borrower characteristics. Finally, I present placebo regressions corresponding to employment changes near the end of the last business cycle expansion and during the 2001 recession. Employment at rms that borrowed from less healthy lenders and had worse outcomes during the period did not dier systematically from employment at other rms during the placebo periods. The paper relates to a number of strands of literature in both macroeconomics and nance. Bernanke (1983) popularized the concept of a credit channel that translates shocks to lending institutions into outcomes in the real economy through the destruction of bankspecic intermediary capital. Most subsequent papers that have attempted to demonstrate the credit channel empirically fall into one of two categories. Small versus large papers use the insight that due to greater asymmetric information or smaller buer savings, smaller, less transparent borrowers should exhibit greater sensitivity to credit supply constraints. In an important contribution, Gertler and Gilchrist (1994) found that sales, inventory, and debt at smaller manufacturing rms respond disproportionately to contractionary monetary policy shocks. Duygan-Bump, Levkov and Montoriol-Garriga (2011) show that job losses in the United States during the recession occurred disproportionately in small rms, but only in industries with high external nancial dependence. A second category of empirical papers uses natural experiments that induce variation in the cross-section of credit availability. In an early example of this approach, Slovin, Sushka and Polonchek (1993) studied the share price response of borrowers of Continental Illinois Bank around its failure and rescue in Subsequent papers have used variation in lending stemming from the bursting of the Japanese real estate bubble (Peek and Rosengren 1997; Peek and Rosengren 2000; Gan 2007; Amiti and Weinstein 2011); the response to the 1998 Pakistani nuclear tests (Khwaja and Mian 2008); the 1998 Russian crisis (Chava and Purnanandam 2011); and the 2002 WorldCom bankruptcy (Lin and Paravisini 2011). Papers that use variation generated by the crisis include Albertazzi and Marchetti (2011) 4

6 and Aiyar (2012) on lending by exposed banks in Italy and the United Kingdom, respectively; De Haas and Van Horen (2012) on cross-border lending; Almeida et al. (2012) on corporate investment at rms with dierent term-structures of debt; Ivashina and Scharfstein (2010) on lending by banks with exposure to Lehman Brothers through cosyndication; and Santos (2011) on loan pricing by banks with varying levels of mortgage-backed securities holdings. The natural experiment papers robustly nd contractions in lending at aected banks. However, rms facing a withdrawal of credit from one nancing source may be able to substitute with nancing from an alternative source, and data limitations make it dicult to link the shocks to real borrower economic outcomes. Two exceptions, Gan (2007) and Almeida et al. (2012), nd contractions in investment at aected borrowers, but, in violation of the logic of the small versus large literature, necessarily restrict attention to rms that have regulatory lings with borrower-level information. 1 Conversely, Peek and Rosengren (2000) show that U.S. states that had higher penetration by Japanese banks had less construction activity following the bursting of the Japanese credit bubble, but do not have evidence of rm-level eects. 2 A key innovation of the current paper is to build a dataset that contains nancial information and employment outcomes at a wide range of U.S. rms. The nding of heterogeneous treatment eects then provides a bridge between the small versus large and natural experiment literatures. 3 To my knowledge this is also the rst natural experiment paper to examine eects on employment, a variable of key economic and popular interest. The outline of the paper is as follows. Section 2 reviews the theoretical reasons for the formation of banking relationships, explains the key features of the syndicated loan market, and presents empirical evidence for the presence of frictions to switching lenders. Section 3 provides a brief overview of the nancial crisis. Section 4 describes the data sources and the construction of the merged loan-employment dataset. Section 5 details the identication strategy. Sections 6 and 7 provide the paper's key empirical results, demonstrating the eect of lending supply on loan market outcomes and employment outcomes, respectively. Section 8 constructs an estimate of the decline in employment in the sample due to nancial frictions. Section 9 presents the general equilibrium model linking the results in section 8 to aggregate employment uctuations. Section 10 concludes. 1 Slovin, Sushka and Polonchek (1993) and Lin and Paravisini (2011) link lending supply shocks to borrower stock price and nancial outcomes. They face the same data restrictions that exclude non-public rms or rms not in the Compustat database. Amiti and Weinstein (2011) examine the eect on exports at Japanese rms, but only include rms listed on a stock exchange. A parallel corporate nance literature that studies the investment sensitivity of rm cash ows has had the same rm-size limitations (Fazzari, Hubbard and Petersen 1988; Kaplan and Zingales 1997). 2 Benmelech, Bergman and Seru (2011) also use the Peek and Rosengren experiment but analyze the eect on MSA-level unemployment rates. 3 In a very dierent institutional setting, Khwaja and Mian (2008) nd a greater ability for larger borrowers to substitute nancing following the Pakistani nuclear shock. 5

7 2. Relationship lending The econometric approach in this paper requires that borrowers and lenders form relationships. Otherwise, pre-crisis clients of banks that restricted lending during the crisis could costlessly switch to borrowing from less constrained banks, with no reason to expect dierential outcomes at pre-crisis borrowers of dierent banks. This section reviews the theoretical reasons for banking relationships, provides a brief overview of the syndicated loan market, and presents new evidence of the presence of frictions to switching lenders Theory Economic theory explains banking relationships as the result of asymmetric information. In the Sharpe (1990) adverse selection model, banks learn about borrower quality through lending. In equilibrium, borrowers that try to switch lenders face a lemons problem, since intuitively prospective new lenders will wonder why the more informed prior lender has not oered a new loan to the borrower. Su (2007) adapts the Holmstrom and Tirole (1997) moral hazard model to the syndicated loan market. Here, a lead lender must retain enough skin in the game to persuade other potential syndicate members that it has done its costly due diligence. If lenders have nite resources, then retaining a larger portion of the loan raises costs that get passed on to the borrower. Lending to a prior borrower can serve as a signal to other potential lenders, in which case both the borrower and lead lender gain from maintaining a relationship. Finally, ex-ante or ex-post monitoring costs that fall for repeat borrowers generate persistent relationships by creating positive match capital. 4 In these models, a wedge develops between the interest rate available from a prior lender and a prospective new lender. A bank that has a higher internal cost of funds will move its interest rate oer to existing borrowers closer to the open market level, shrinking the wedge and raising costs to the borrower. At some point, the borrower may decide either to take the open market rate from a new lender or not to borrow at all. Moreover, both the Holmstrom and Tirole model and the costly state verication model generate a tradeo that results in a maximum interest rate that a lender will oer. 5 In that case, even borrowers 4 See Williamson (1987) for an exposition of ex-post monitoring costs in the costly state verication model, and Montoriol-Garriga and Wang (2011) for a recent empirical application. 5 In the Holmstrom and Tirole (1997) model, borrowers without enough skin in the game can choose a technology with lower probability of success in exchange for capturing private benet. The borrower's incentive compatibility constraint puts a limit on the fraction of output that can be claimed by lenders, yielding a maximum incentive compatible interest rate. In the costly state verication model, lenders trade o the extra income from charging a higher interest rate with the increased probability that the borrower will be unable to repay all of the principal and interest and will declare bankruptcy. Deadweight loss in bankruptcy implies a maximum interest rate the lender is willing to charge. 6

8 with positive loan demand at an interest rate above their lender's internal cost of funds may nd themselves unable to borrow, and the likelihood of getting rationed out of the market rises for clients of banks with larger increases in their internal cost of funds. Summing up, the asymmetric information models predict that borrowers of banks with larger increases in their internal cost of funds should have a lower likelihood of obtaining a loan, and face higher borrowing costs conditional on obtaining a loan at all. The link between banking relationships and asymmetric information also generates testable hypotheses for the heterogeneity of treatment eects. First, the benet to using a previous lender, or conversely the lemons cost to switching lenders, declines with the transparency of the borrower. Second, if the per-dollar monitoring cost falls with the size of the loan, then the cost of asymmetric information falls with borrower size. Thus, banking relationships should matter more to less transparent, smaller borrowers The syndicated loan market In a typical syndicated loan deal, the lead arranger comes to preliminary agreement on the terms of the loan with the borrower, performs the due diligence, and recruits other participant lenders to provide some of the nancing. The modal deal contains one lead arranger and one participant; however, about one out of three deals contain at least ve participant lenders, and one-quarter of loans contain multiple lead arrangers. Besides handling the borrower relationship, the lead arranger also retains a larger share of the loan than participants, with the lead share usually fty to one hundred percent larger than each participant's share. The lead arranger typically cannot sell its portion of the loan to a third party. Both traditional deposit-taking banks and investment banks act as lead arrangers. Participants also include hedge funds and pension funds. The syndicated loan market has become the dominant source of bank nance for midsized and large corporations. The market serves both publicly-traded and private rms, with about half of the rms private. The median borrower in my sample (described in further detail in the next section) had sales of about $500 million (in 2005 dollars), and had 620 employees in However, the 10th percentiles of sales and employment were $60 million and 77 employees. For comparison, 71 percent of private sector employees in the United States work at rms with at least 50 employees. Previous work has documented a number of features of the market that indicate the importance of asymmetric information. For example, Dennis and Mullineaux (2000) and Su (2007) show that the number of participants in the syndicate increases with the transparency of the borrower. They interpret this nding as indicating that lead arrangers retain a larger portion of loans to less transparent borrowers as a signal of the quality of their private 7

9 information about the borrower. 6 Su (2007) also shows that lead arrangers tend to recruit participants that have formed part of a previous syndicate for the borrower, again consistent with the existence of private information about borrowers that banks can learn over time New evidence A direct test of whether banking relationships form is whether borrowers use their previous lead arranger when accessing the market for a new loan. Let m = [m 1, m 2,..., m N ] denote the vector of market shares of N lead banks. As a benchmark, if borrowers always randomly select a lender based only on the lender's overall market share, the percent of loans in which the borrower uses the same lead arranger as in its previous loan would equal m m, the Herndahl index. Column 1 of the top panel of table 1 shows that this index ranges from a low of 9.1 percent for private rms to a high of 13.5 percent for rms with a credit rating, with a mean over all rms of 9.7 percent. The next columns in the top panel give the actual repeat-borrowing propensity for the whole post-2000 period and for two subperiods corresponding to before and during the nancial crisis. 7 Over the full period, borrowers use their previous lead arranger in just under three-quarters of loan pairs, more than seven times the frequency which would occur if borrowers selected lenders randomly. Comparing the pre-crisis and crisis periods, the repeatborrowing propensity rises by about nine percentage points in the crisis for private borrowers. It is essentially unchanged for public borrowers without a credit rating, and falls by about eight percentage points for the most transparent segment of publicly-traded borrowers with a credit rating. The higher repeat-borrowing propensity among private borrowers accords with the stickiness deriving from asymmetric information. It also suggests that pre-crisis lender health would have greater eect on the outcomes of less transparent rms, since these rms faced particularly high barriers to switching lenders during the crisis. The lower panel of table 1 explores whether borrowers also develop a relationship with syndicate participants. In particular, it asks how often participants assume the role of lead lender in loans where the borrower does not use a previous lead lender. This metric has particular relevance in a crisis where some borrowers lose access to their previous lead lender. Again, the rst column provides a benchmark of the likelihood of such a role switch 6 A small literature has developed to determine the optimal size and composition of the lending syndicate. On the one hand, larger syndicates provide greater diversication to the borrower in maintaining lending relationships (Detragiache, Garella and Guiso 2000) and oer opportunities for specialization in monitoring tasks for lenders (Francois and Missonier-Piera 2007). However, larger syndicates also complicate the cost of restructuring in the event of default (Bolton and Scharfstein 1996; Bris and Welch 2005). 7 To remove any possible inuence of loans misclassied as separate packages instead of separate facilities of the same package, the sample includes only loan pairs where the loans start at least 31 days apart. Results are similar using no minimum interval restriction or restricting the loan pairs to at least 365 days apart. 8

10 Table 1: Banking relationships to Jun-07 Oct-08 to Jun-09 p-value At random Actual Actual Actual of benchmark share Obs. share Obs. share Obs. dierence P ercent of loans in which borrower uses previous lead lender Private , , Public (Unrated) , , Public (Rated) , , All , , P ercent of remaining loans in which previous participant replaces lead lender Private Public (Unrated) Public (Rated) All , , Notes: In the top panel, the sample includes loans to non-fire borrowers that have previously accessed the syndicated market, and the rst column shows the fraction of loans in which a borrower would use a previous lead lender if borrowers always selected their lead lender randomly using only lenders' market shares. In the bottom panel, the sample includes loans to non-fire borrowers that have previously accessed the syndicated market and do not use a previous lead lender, and the rst column shows the fraction of loans in which the borrower would use a previous participant lender as its lead lender if borrowers always selected their lead lender randomly using only lenders' market shares. In both panels, the sample includes only loans made for working capital or corporate purposes, where the second loan has at most two lead lenders, and the loans start more than 30 days apart. The last column of the table reports the p-value from a t-test of equality bewteen the period before Jun-07 and from Oct-08 to Jun-09. if borrowers always selected their lead lender randomly. Over the full period, borrowers switched to a previous participant in just over one-quarter of loans where the borrower did not use a previous lead lender, roughly double what would obtain if borrowers selected their new lead lender at random. During the crisis, the fraction rises to one-third of all such loans. These results suggest an informational advantage for syndicate participants as well, consistent with the result in Su (2007) that lead lenders are also more likely to recruit participants that have previously formed part of a syndicate with the borrower. 3. Background on the crisis The rst signs of the possibility of problems at major nancial institutions came in June 2007, with the rescue by the investment bank Bear Stearns of a subsidiary hedge fund that had invested heavily in subprime mortgages. A month and a half later, the French bank BNP Paribas announced the freezing of three investment funds based on an inability to value the 9

11 funds' subprime assets. The announcement sparked a rise in the interest rate at which banks lend to each other in the interbank market (see Figure 1). 8 Concerns mounted as a wave of bank writedowns on their subprime portfolios ensued. The panic reached a brief crescendo in March 2008, when the withdrawal of short-term nancing to Bear Stearns forced its sale to J.P.Morgan. Financial conditions stabilized somewhat over the summer, but then deteriorated sharply in September On September 10 Lehman Brothers reported a $3.9 billion loss for the third quarter of its scal year. Five days later, unable to nd a buyer and unable to obtain short-term nancing, Lehman Brothers led for bankruptcy. The cost of interbank lending spiked immediately. A cascade of market and policy events followed, including an $85 billion loan from the New York Federal Reserve Bank to the insurer AIG; the announcement by the money market fund Reserve Management Corporation that its net asset value had fallen below par, prompting widespread withdrawals from other money market funds; the forced sales of the investment bank Merrill Lynch and the commercial bank Wachovia; and direct capital injections by the federal government into major nancial institutions through the TARP, to name a few. The stress in the interbank lending market began to ameliorate during the fall and winter of 2008, but remained elevated until the summer of Figure 1: 600 Stress in the interbank lending market (3 month Treasury-Eurodollar spread, basis points) First TARP commitments; FDIC lifts deposit insurance cap, guarantees new bank debt 400 Writedowns begin Bear Stearns purchased BNP Paribas freezes funds SCAP (stress test) results released 200 Lehman Brothers bankruptcy 0 Jun-07 Dec-07 Jun-08 Dec-08 Jun-09 Dec-09 Source: Federal Reserve Board of Governors (H.15 Release). The quantity of new bank lending followed a similar pattern. Figure 2 shows the dollar 8 The Treasury-Eurodollar (TED) spread became a widely watched indicator of nancial distress during the crisis. Gertler and Kiyotaki (2011) provide theoretical justication for why stress in the interbank market matters in an economy with lending relationships of the type described in section 2. 10

12 value of new lending by the forty-ve most active banks in the syndicated market. The market expanded rapidly during the mid-2000s, but began to contract during the fourth quarter of The initial fall in lending does not seem to reect expectations of a large decline in real activity; for example, in June 2008 members of the Federal Reserve Open Market Committee forecast that the unemployment rate would remain roughly unchanged (at around 5.5 percent) over the coming year. New lending troughed in the fourth quarter of 2008, coincident with the peak of stress in the interbank market. Lending started its rebound somewhat after the interbank market stabilized, with a slow recovery in volume beginning at the end of Figure 2: Aggregate new lending from top 45 lenders (Billions of 2005 dollars, seasonally-adjusted at quarterly rate) Term loans Credit lines 2001:1 02:1 03:1 04:1 05:1 06:1 07:1 08:1 09:1 10:1 Source: Dealscan. Notes: The figure shows the face value of new loans to non-fire borrowers for working capital or general corporate purposes in which one of the forty-five most active lenders had a lead role. Values seasonally-adjusted by author using Census-X12. The narrative timeline, the timepath of interbank lending spreads, and the timing of the trough in new bank lending all point to a division in the crisis between the periods before and after the Lehman bankruptcy. The acuity of both nancial distress and employment losses 9 Much discussion has centered on the build-up of liquid assets at nonnancial corporations following the crisis, with the implication that borrowing constraints could not matter in an environment where corporations had large amounts of cash or cash equivalents at their disposal. Leaving aside issues of aggregation - the measure of cash typically cited comes from the Flow of Funds and represents the entire nonnancial sector, and may therefore have less relevance for the small and medium rms identied as most aected in the empirical results below - the timing of the cash build-up merits mention. In nominal terms, liquid assets (Flow of Funds code FL Q) actually declined during the recession, and only in 2009:3 surpassed their pre-recession level. Liquid assets did rise substantially beginning in the latter part of 2009, but this build-up partly reects a return to the pre-crisis trend, as well as rms' increasing their optimal buer stock of savings in response to the credit crunch. The time-series of liquid asset holdings does not necessarily indicate that borrowing constraints did not bind in equilibrium. 11

13 following the Lehman bankruptcy suggest that nancial frictions may have had especially great inuence on employment outcomes during that period. My main analysis will therefore focus on the eects of the decline in lending during the post-lehman period, dened as October 2008 to June The choice of June 2009 as a terminal month reects both the timing of the U.S. recession, which ended that month, and the timing of the return to normalcy in the interbank lending market. 4. Data A principal innovation of this paper is to link a loan and employment dataset in order to observe employment outcomes of borrowers of dierent banks. The loan market data come from the Thomson Reuters Dealscan database. Dealscan collects loan-level information on syndicated loans from SEC lings, company statements, and media reports, and attempts to process the universe of such loans. 11 The data include the identities of the borrower and lenders present at origination, the terms of the loan, and the purpose of the loan (working capital, leverage buyout, etc.). The sample I use begins from all loans made to non-fire U.S. businesses with the primary purpose of the loan listed as working capital or corporate purposes. To restrict the sample to rms likely to have an active relationship with a lender during the crisis, I keep only borrowers with either at least one loan signed in or after 2004, or with a loan open in October 2007 or later. 12 To obtain lender nancial information, I merge the Dealscan lenders at the holding company level with data from the Federal Reserve FR Y-9C Consolidated Financial Statements for Bank Holding Companies (for lenders where the highest level parent is either a domes- 10 I set the post-lehman period to begin in October 2008 because the loan date reported in Dealscan corresponds to the start of the loan facility, which may lag the signing of the loan agreement and will certainly lag the beginning of the loan processing. 11 Public companies must report any new bank loan to the SEC by ling a form 8-K or as an attachment to their quarterly or annual ling. Thomson Reuters publishes a quarterly set of League Tables using the Dealscan data, which rank lenders according to their level of activity in the syndicated market over the prior period. The public ranking of lenders gives banks an incentive to report to Dealscan loans that Dealscan might otherwise miss. For about ten percent of loans Dealscan reports a single lead lender and zero participants. In some of these cases, Dealscan does not observe which lenders serve as syndicate participants. In others, the loan is an add-on to a previous loan facility. Loans with a single lead arranger and zero participants are if anything slightly larger than other loans in the dataset. 12 Dealscan contains 11,740 unique U.S. borrowers that either obtained a syndicated loan between 2004 and August 2008 or obtained a loan prior to 2004 that matured after October Of these, about twothirds (7,885) report the primary purpose of the loan as working capital or corporate purposes. (The next most common purpose is for a corporate takeover.) Eliminating borrowers in nance or real estate (FIRE), dened as SIC codes , further reduces the sample to 6,569. Finally, I remove borrowers with missing industry, state, or public/private status, leaving a sample of 4,932 borrowers. 12

14 tic nancial holding company or a domestic bank holding company), and with data from Bankscope (for foreign holding companies and investment banks). I merge the lenders with data from CRSP to obtain stock price information. For reasons detailed in the next section, I limit the Dealscan sample to loans where at least one of the lead lenders was among the forty-ve most active lead lenders during the sample period. 13 the sample size by about ve percent. This last restriction reduces Employment data come from the condential Bureau of Labor Statistics Longitudinal Database (LDB). The LDB builds from unemployment insurance records at state workforce agencies. The database follows establishments longitudinally and contains monthly employment and quarterly payroll (wages and salaries) at every private sector establishment in the United States. An establishment is a single physical place of work. Within each quarter, the LDB reports the employer identication number (EIN) of the tax-ling unit to which the establishment belongs. I refer to a group of establishments reporting under the same EIN as a rm. 14 Employment corresponds to the total number of employees on payroll in the pay period containing the 12th day of the third month of the quarter, consistent with the standard U.S. statistical agency denition. The LDB does not share a common identier with the Dealscan data. For about onequarter of the sample, I use a linking table between Dealscan and Compustat as described in Sudheer and Roberts (2008), and then use the EIN reported in Compustat to link to the LDB. 15 Another twelve percent of the rms in the Dealscan sample have exact matches in the LDB along the dimensions of rm name, city, and zip code. In the remaining cases, I perform a fuzzy merge using geographic and industry identiers along with a bigram string comparator score of the rm name as reported in each dataset. 16 The nal merged sample 13 For purposes of constructing their league tables, Thomson Reuters identies one or more lead arrangers for each loan based on the descriptive role of each lender, and I adopt their classication. 14 This denition of a rm derives from tax purposes and reporting conventions, and does not always correspond to the economic denition of control, nor to the scope of activities controlled by the loan recipient in the Dealscan data. If, however, shared tax liability maps into shared internal capital markets, then EIN is the correct ownership level for matching rms to their borrowing history in Dealscan. Otherwise, failure to identify all EINs with common ownership would lead to measurement error in the growth rates derived from the LDB, raising the standard errors in the regressions reported in section 7. More problematic, some states allow establishments that use professional payroll rms to report the EIN of the payroll rm rather than the establishment's owner, and I have to drop these rms from the sample. I hand check any rms with a symmetric growth rate (dened below) greater in absolute value than 0.9. In many cases the extreme growth rates result from the use of multiple EINs by a single controlling economic unit. I either combine the EINs into a single new rm or drop the rm from the sample if I cannot identify all of the relevant EINs. 15 The matching le is available on request from Michael Roberts. I conducted an assessment of each match using the information on rm sales reported in Dealscan as well as industry and geographic identiers, resulting in a slightly dierent set of matches than in the Roberts le. 16 A bigram string comparator computes the fraction of consecutive character matches between two strings. I implement the fuzzy merge using the stata ado le reclink written by Michael Blasnik. I also did a manual review of all of the matches to ensure accuracy. 13

15 contains just over 2000 rms, or roughly half of the original Dealscan dataset. 17 Table 2 gives summary statistics for the full sample, the sample limited to loans with at least one lead lender among the forty-ve most active, and the merged Dealscan-LDB sample. Each borrower appears in the sample exactly once, and the summary statistics correspond to the borrower's last pre-crisis loan. Both the sample limited to loans with the most active lenders and the merged Dealscan-LDB sample look quite similar to the full sample along the observable dimensions of loan size and borrowers' sales. 18 The variable ˆL i, dened in section 5.1 below, measures the relative availability of credit to each borrower, and here as well the merged sample looks nearly identical. The sample (unweighted) average of the symmetric growth rate (dened in section 7) of employment from 2008:3-2009:3 is -9 percent. Aggregate employment in the sample declined by 5.8 percent (not shown), almost exactly equal to the 5.7 percent employment decline in the entire U.S. private sector. The merged Dealscan-LDB sample contains roughly twice as many rms as would merging Dealscan with Compustat after applying my sample lters. Crucially, while the Dealscan- LDB sample has about the same number of large (1000 or more employee) rms as a Dealscan-Compustat sample, it has more than ve times as many small and medium rms Identication To determine the eect of credit availability on employment, one cannot simply regress the change in employment on whether the rm received a loan. Firms that seek and obtain loans may have dierent characteristics than rms that do not want to borrow or cannot nd a willing lender. Instead, one needs to isolate a measure of loan supply. Formally, write the growth rate g y i,s of employment y at rm i that had a pre-crisis relationship with lending syndicate s as a function of a vector of (omitting time subscripts) loan characteristics L i,s ; observable characteristics of the rm X i ; unobservable characteristics U i ; 17 The unmatched rms fall into several categories. The merged sample does not include any rms with headquarters in eight states that did not grant access to their microdata. Some rms closed or merged prior to the nancial crisis, and do not appear in the LDB with the same ownership structure as in Dealscan. A few rms that moved their headquarters across states or changed their name in the interval between their last Dealscan loan and the nancial crisis may also be missing. As discussed in footnote 14, the merged sample also omits rms with establishments that use a professional employer organization to handle their payroll. Finally, rms that operate under multiple names may have generated too low a bigram string score to qualify as a match. The summary statistics presented in the next paragraph suggest that on balance the match attrition caused by these factors had limited eect on the composition of the sample. 18 The dierence between the samples described as All lenders and Top 45 lenders also reects the removal of about ve percent of rms that are excluded from the probit regressions reported in section 6. These rms are in industries in which no rm in the sample obtained a loan during the crisis. 19 See footnote 40 for a description of the Dealscan-Compustat sample. 14

16 Table 2: Sample summary statistics N Mean SD p10 p50 p90 Loan size (millions) All lenders 4, Top 45 lenders 4, Merged Dealscan-LDB 2, Sales at close (millions of 2005 dollars) All lenders 4, 091 1,808 4, ,584 Top 45 lenders 3, 619 1,921 4, ,834 Merged Dealscan-LDB 1, 731 2,014 4, ,795 L i Top 45 lenders 4, Merged Dealscan-LDB 2, Symmetric growth rate of employment 2008:3-2009:3 2, Memo: 2008 employment level 2, 050 2,978 9, ,148 Notes: The sample includes non-fire U.S. borrowers that obtained a loan for working capital or corporate purposes and with valid state, industry, and public/private status identiers, excluding the top 1% by sales. Statistics for loan size and borrower sales correspond to information as of the last loan obtained by each borrower prior to September The variable L i equals the weighted change in the number of loans made between October and June, to , by the members of the borrower's last pre-crisis loan syndicate, with the weights given according to each member's role as described in section 5.1 of the text. Rows indicated by Top 45 lenders include only borrowers whose last pre-crisis loan included one of the most active 45 lenders as a lead arranger. Rows indicated by Merged Dealscan-LDB and the employment statistics include only borrowers matched to the LDB. and an unobserved idiosyncratic component uncorrelated with X i or U i : g y i,s = f (L i,s, X i, U i, ɛ i ). (1) For the moment, suppose L i,s consists only of an indicator variable for whether the rm receives a loan, and that this depends on rm characteristics, the internal cost of funds at the pre-crisis lending syndicate R s, and an idiosyncratic disturbance η i uncorrelated with ɛ i and U i : L i,s = h (R s, X i, U i, η i ). (2) If U i R s (where denotes statistical independence) and assuming at a minimum separability of f ( ) between its rst two and second two arguments, then equations (1) and (2) could be estimated using the generalized method of moments, with the moment condition E [{ g y i,s f (L i,s, X i, 0, 0) } ] R s = 0. Two complications emerge. First, Rs is not directly observed. Instead, one needs an observable measure of loan supply M s such that 15

17 Corr (M s, R s ) 0 and U i M s. Second, employment may depend not only on success in obtaining a loan, but also on the interest spread, size, length, covenants, and other characteristics of the loan obtained, as well as on expectations of future credit availability if rms face costs to adjusting their labor input. In other words, h ( ) is a vector-valued function, and the system (1)-(2) is under-identied for determining the eect of any particular component of L i,s on employment. These considerations lead to a focus below on the reduced-form impact of lender health on employment. Substituting the arguments of (2) into (1), and replacing R s with the observed measure M s : Consistent estimation of 5.1. Measuring loan supply g y i,s = g (M s, X i, U i, ɛ i, η i ). (3) g M s follows from the orthogonality condition U i M s. I use the quantity of lending at each bank to proxy for the shadow price. Specically, to measure credit availability to borrower i during the crisis, I use lending during the crisis by i's pre-crisis syndicate to all other borrowers. If a bank's total lending reects its internal cost of funds, then this measure will satisfy the condition Corr (M s, R s ) 0, making it relevant for loan outcomes of borrower i. It will satisfy the exclusion restriction Corr (M s, U i ) = 0 as long as the unobserved characteristics of pre-crisis borrowers of syndicate s that inuence loan outcomes are uncorrelated. The following two assumptions formalize sucient conditions for the measure's validity: A1 (as good as random): along any dimension that aects whether a borrower obtained a loan during the crisis from its pre-crisis lenders, and conditional on observable borrower characteristics, the matching of borrowers and banks precrisis was as good as random. A2 (small): borrowers are small relative to the bank portfolio, so that the marginal change in the bank's internal cost of funds from making a loan to any single borrower is zero. I defer discussion of whether assumption A1 holds empirically to the robustness sections below. Assumption A2 states that the decision to lend to borrower i does not aect the decision to lend to other borrowers. This explains the sample restriction of including only the more active lenders. I implement the measure as follows. The matching of borrowers and lenders comes from the last loan obtained by each borrower with a start date before September For each 16