Employment effects of unconventional monetary policy: Evidence from QE

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Employment effects of unconventional monetary policy: Evidence from QE Stephan Luck and Tom Zimmermann 2018-071 Please cite this paper as: Luck, Stephan, and Tom Zimmermann (2018). “Employment effects of unconventional monetary policy: Evidence from QE,” Finance and Economics Discussion Series 2018-071. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2018.071. 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 effects of unconventional monetary policy: ∗ Evidence from QE Stephan Luck†and Tom Zimmermann‡ Federal Reserve Board, University of Cologne October 9, 2018 Abstract This paper investigates the effect of the Federal Reserve’s unconventional monetary policy on employmentviaabanklendingchannel. Wefindthatbankswithhighermortgage-backedsecurities holdingsmaderelativelymoreloansafterthefirstandthirdroundsofquantitativeeasing(QE1and QE3). While additional volume is concentrated in refinanced mortgages after QE1, increases are driven by newly originated home purchase mortgages and additional commercial and industrial lendingafterQE3. Usingspatialvariation,weshowthatregionswithahighshareofaffectedbanks experiencedstrongeremploymentgrowthafterbothQE1andQE3. Whiletheabilityofhouseholds torefinancemortgagesafterQE1spurredlocaldemand,theresultingadditionalemploymentgrowth was relatively weak and confined to the non-tradable goods sector. In contrast, the increase in employment after QE3 is sizable and can be attributed to the supply of additional credit to firms. To support this finding, we use new confidential loan-level data to show that firms with stronger tiestoaffectedbanksincreasedemploymentandcapitalinvestmentmoreafterQE3. Altogether,our findingssuggestthatunconventionalmonetarypolicycan,similartoconventionalmonetarypolicy, affectrealeconomicoutcomes. ∗ThispaperpresentstheviewoftheauthorsandnotoftheFederalReserve. WewouldliketothankDavidArseneau,Jose Berrospide,NeilBhutta,DavidByrne,DanielCooper,OlivierDarmouni,CynthiaDoninger,BurcuDuygan-Bump,Miguel Faria-e-Castro,MichaelConnolly,GiovanniFavara,AndreasFuster,PaulGoldsmith-Pinkham,SimonJaeger,ElizabethKlee, Arvind Krishnamurthy, Anna Kovner, Adrien Matray, Michael McMahon, Ralf Meisenzahl, Atif Mian, John Mondragon, ChristopherPalmer,MatthewPlosser,FarzadSaidi,EricSwanson,SkandervandenHeuvel,JoeVavra,JamesVickery,Egon Zakrasjekaswellasseminarparticipantsatthe2017NBERMonetaryEconomicsProgrammeeting,FederalReserveBank ofNewYork,FederalReserveBoard,FederalReserveBankofSt. Louis,MFASanAntonio,EABCNBarcelona,Barcelona SummerForumonMonetaryPolicyandCentralBanking,ITAMFinance,UniversityCollegeDublin,CEBRAFrankfurtfor helpfulcommentsandsuggestions. WethankZachFernandesandTylerWakeforexcellentresearchassistance. †FederalReserveBoard,stephan.luck@frb.gov ‡UniversityofCologne,tom.zimmermann@uni-koeln.de 1

1 Introduction Does unconventional monetary policy affect the real economy? And if so, what are the channels? Over the course of the 07-08 Financial Crisis, monetary policy reached the zero lower bound in several countries. As a consequence, central banks have turned toward unconventional monetary policies such as forward guidance and large-scale asset purchases (LSAPs). Because the channels through which unconventional monetary policy affects the real economy are yet not fully understood and because conclusive empirical evidence is scarce, the efficacy and desirability of such policies has been controversial.1 Against this background, this study investigates the real effects of the Federal Reserve’s unconventional monetary policy. Our study is the first to document a link between the Federal Reserve’s quantitativeeasing(QE)andemploymentoutcomes. Thisisaparticularlyimportantfindingasprevious research has highlighted an effect of QE on bank lending (Darmouni and Rodnyansky, 2017), and an effect of QE on household net-worth and consumption via the mortgage market (Di Maggio et al., 2016; Beraja et al., 2018), but has not been able to speak to the crucial question of its effect on improving broader economic conditions such as employment outcomes. In particular, we show that while the first round of QE (QE1) led to increased refinancing activity in the mortgage market by commercial banks, this additional lending activity had only weak effects on employment growth, confined to the non-tradable goods sector. In contrast, the third round of QE (QE3) induced additional commercial and industrial (C&I) lending as well as increased origination of home purchase mortgages, which in turn translated into economically sizable growth in total employment. Our evidence suggests that LSAPs can, similar to reductions of interest rates in times of conventional monetary policy, affect real economic outcomes via a bank lending channel.2 The empirical assessment of macroeconomic policy is generally difficult, given the natural absence of a control group.3 In order to overcome this inherent challenge, we proceed in three steps. First, we 1Forinstance,aftertheimplementationofQE1anumberofprominenteconomistswroteanopenlettertoBenBernanke (publishedintheWallStreetJournalonNovember152010): “WebelievetheFederalReserve’slarge-scaleassetpurchaseplan (so-called ‘quantitative easing’) should be reconsidered and discontinued. We do not believe such a plan is necessary or advisableundercurrentcircumstances. Theplannedassetpurchasesriskcurrencydebasementandinflation,andwedonot thinktheywillachievetheFed’sobjectiveofpromotingemployment.” 2GiventhatwedonotidentifystrongemploymenteffectsofQE1andQE2doesnotimplythattheseprogramsdidn’thave considerableimpactontherealeconomyviadifferentchannelsnotconsideredinthispaper. Likewise,giventhecross-sectional natureofouranalysis,onecannoteasilydrawconclusionswithrespecttoaggregateemployment. 3TheproblemisbestsummarizedbyBenBernankeinhismemoir,TheCouragetoAct: “Wecan’tknowexactlyhowmuchof theU.S.recoverycanbeattributedtomonetarypolicy,sincewecanonlyconjecturewhatmighthavehappenediftheFedhad nottakenthestepsthatitdid.” 2

exploit cross-sectional variation in the exposure of commercial banks to QE. Second, we exploit spatial variation in the activity of banks that were more affected by QE, allowing us to trace the effect of QE on local credit and labor markets. Third, we use confidential loan-level and mortgage-level credit registry data that allow us to shed light on whether additional credit results from changes in credit supply or credit demand. In the first step, we exploit a link between QE and bank lending that has recently been established in several papers (see, e.g. Darmouni and Rodnyansky, 2017). This literature uses cross-sectional variation of banks’ mortgage-backed securities (MBS) holdings in order to identify an effect of QE on bank lending. Complementary to these findings, we show that more exposed banks experienced higher stock returns on QE announcement days, controlling for many potentially confounding bank characteristics. In addition, we find that, within the mortgage market, QE1 mainly affected the refinancing volume of existing mortgages (Di Maggio et al., 2016; Beraja et al., 2018), and QE3 affected the origination of home purchase mortgages at more exposed banks. Within the C&I loan market, only QE3 increased lending, in particular in to smaller firms. These findings turn out to be important for understanding how QE affected employment. Inthesecondandcentralpartofthepaper,westudythelinkbetweenbanklendingandemployment at the county level. We exploit that, on top of the cross-sectional variation of MBS holdings, there is geographical variation in banks’ activity. Importantly, a bank’s regions of activity are very stable over time and highly predictable. Measuring banks’ local activity by either the historical amount of small business lending, mortgage origination, or deposit volume in a given county, we construct an exposure measure at the county level: We treat counties that have historical activity from banks with more MBS holdings as exposed counties and those with less activity from such banks as non-exposed counties. AconcernwithourapproachisthatbankswithhigherMBSmightsystematicallyselectintomarkets thataremosteconomicallydeveloped,orthatareexpectedtodisplayhighfutureeconomicgrowth. Our identificationaddressesthisconcernintwoways. First,whilebanks’locationchoicesaresystematic,they arealsohighlypersistent. Itishenceunlikelythatbanksselectedlocationsinanticipationofquantitative easing, and it is such correlation with the timing of unconventional monetary policy that would be of most concern. Second, our difference-in-differences approach allows for systematic differences across markets as long as markets would have co-moved in absence of QE. We provide evidence in support of this assumption. In particular, employment growth in counties with high and low exposure to banks with large MBS 3

holdings evolved very similarly during the Financial Crisis, suggesting that these counties were on similar trajectories absent QE. To be precise, employment growth in exposed and non-exposed counties followed the same trend for more than 18 consecutive quarters before the implementation of QE3. After QE3, however, we find that exposed counties experience higher overall employment growth. The effects are economically sizable: Our estimates suggest that counties in the upper tercile of the MBS–bank distribution experienced 40 basis points higher employment growth than counties in the lower tercile of the distribution. In contrast, after QE1, no such differential effect is observed. While exposed counties do experience higher employment growth in the non-tradable goods sector during QE1, the effect is weak, and overall employment growth is statistically indistinguishable between affected and unaffected counties. Finally, in the third step, we analyze whether additional credit is driven by changes in credit supply as opposed to credit demand. In order to control for demand in the C&I loan market, we use newly available confidential loan–level data from the Y-14 data collection, which, since 2011, requires large banks with more than $50 billion in assets to report any C&I loan on their balance sheet with a commitment of 1 million USD or more. Using the loan-level database, we provide evidence that the increase in lending after QE3 is not driven by increases in demand for loans by firms, but rather by additional supply of bank loans. To account for loan demand by firms, we control for firm-specific demand in the spirit of Khwaja and Mian (2008) in bank–firm–level regressions. Moreover, using firm-level data on investment and employment, we provide evidence that, after QE3, firms that historically tended to borrow from affected banks increased capital investment and employment by more than firms that had been borrowing from unaffected banks. Notably, our results in this part contrast to results documented by Chakraborty et al. (2016), who find a negative effect of QE3 on C&I credit and investment by large firms. The difference in findings can largely be explained by the fact that our data also contains medium-sized firms that are not active in the syndicated loan market. In particular, we show that the increased volume of C&I lending is more pronounced for non-syndicated borrowing by medium sized firms than for syndicated borrowing from larger firms, consistent with larger firms being less bank dependent. Inordertocontrolforcreditdemandinthemortgagemarket,weuseconfidentialmortgage-leveldata. In particular, we exploit that multiple banks are active in the same county and account for local credit demand by households by controlling for county-specific demand in county–bank–level regressions. We showthatafterQE1wasimplemented,affectedbanksaremorelikelytorefinancemortgagesirrespective 4

of whether they are GSE-conforming or non-conforming. However, when controlling for local demand, the effect becomes weaker in its magnitude, in particular for GSE-conforming mortgages, but remains statistically significant. We interpret this as evidence that the increase in refinancing activity after QE1 is partially driven by increased credit supply, complementary to the findings by Di Maggio et al. (2016) and Beraja et al. (2018), who emphasize the demand of households to refinance mortgages after QE1. While our results indicate that the increase in lending stems from an increase in credit supply to households as well as firms, the employment effect could be the result of two separate channels: It could either result from an increase in local credit supply to firms, or from increased local credit supply to households, which in turn affects employment through increasing demand for local consumption. While generally both channels may be relevant, our data gives us some sense of which type of lending is more important for employment. In particular, we exploit that increases in demand are more likely to affect the non-tradable goods sector than the tradable goods sector. Hence, if increased demand was driving the increase in economic activity, additional employment would more likely be generated in the non-tradable goods sector (Mian and Sufi, 2014). Even though we do not find an effect on total employment during QE1, employment in the nontradable sector increases subsequent to the implementation of the program. Moreover, the increase in employment in the non-tradable sector coincides with an increase in auto sales, which are a proxy for overall household consumption. Our findings thus suggest that the ability of households to refinance mortgages after QE1 spurred local demand for consumption by increasing household net-worth, resulting in increased employment in the non-tradable goods sector. In contrast, we find that the overall change in employment after QE3 is driven by changes of employment in sectors other than the non-tradable goods sector. We argue that the significant increase in employment after QE3 is more likely to be a consequence of the increased supply of C&I loans rather than increased mortgage origination, consistent with increases in lending, investment and employment at firms with relationships to more exposed banks as discussed above. Our results survive a large number of additional tests. First, the second round of QE (QE2) consisted entirely of Treasury purchases, and hence acts as a natural placebo test: Indeed, we do not find a significant relation between MBS exposure and bank lending or employment following QE2. Second, if QE3 induced bank lending and employment, the tapering of QE3 should have effects in the opposite direction: We provide evidence that supports this conjecture. Finally, we also show that our main results are robust to minor variations in the empirical specification and sample restrictions. 5

Altogether, our findings suggest that unconventional monetary policy can, similar to conventional monetary policy, spur economic activity via a bank lending channel. In particular, LSAPs can induce commercial banks to expand lending which may translate into additional employment. However, differences in findings across the different rounds of QE suggest that the effects are time varying, mirroring findings on the effects of conventional monetary policy (Vavra, 2014). We proceed as follows. After a brief summary of the existing literature, Section 2 discusses data used in our study. Section 3 presents the empirical strategy and discusses results on bank lending and employment.Section 4 separates supply and demand effects in the C&I loan and mortgage markets using loan-level data. Section 5 discusses evidence from additional events (QE2 and tapering) and robustness of results before we conclude in Section 6. 1.1 Related literature Our study contributes to the literature on the bank lending channel of monetary policy. During times of conventional monetary policy, the conventional wisdom is that more accommodative monetary policy is associated with an increase in bank lending and a decrease in unemployment (see, e.g. Bernanke and Blinder, 1992; Kashyap and Stein, 1994, 2000; Drechsler et al., 2017). Previous studies on the effect of unconventional monetary policy in the United States have mostly focused on outcomes other than employment. For instance, recent studies have investigated the effects of QE on aggregate financing cost (Hancock and Passmore, 2011; Gilchrist et al., 2015) or on outcomes in the housing market and at the household level (Di Maggio et al., 2016; Beraja et al., 2018), at the bank level (Darmouni and Rodnyansky, 2017; Kurtzman et al., 2017), or at the firm level (Foley-Fisher et al., 2016; Chakraborty et al., 2016).4 Our analysis of bank lending is closely related to Darmouni and Rodnyansky (2017) and Di Maggio et al. (2016), who exploit cross-sectional variation in MBS holdings. Darmouni and Rodnyansky (2017) document a positive effect of QE1 and QE3 on bank lending using Call Report data. While the focus of our study is on employment outcomes, we also provide complementary bank-level evidence beyond their study by distinguishing more granularly between different types of lending, and by investigating stock market reactions to QE announcements. Di Maggio et al. (2016) and Beraja et al. (2018) study the effect of QE on the mortgage market in 4TherearealsoseveralpapersthatinvestigatetheeffectsofunconventionalmonetarypolicyinEurope(see,e.g.,Acharya etal.,2016;CarpinelliandCrosignani,2017;Crosignanietal.,2018;Cahnetal.,2018;Cumming,2018). 6

more detail. In particular, Di Maggio et al. (2016) show that the Federal Reserve’s purchases of MBS are associated with a refinancing boom after QE1, which in turn triggered significant equity extraction and an increase in consumption. Beraja et al. (2018) show that refinancing and the subsequent growth in consumption is more pronounced in regions in which households have lower leverage and are hence less constrained in their refinancing decision. Both is consistent with the employment increase in the non-tradables sector that we document. Importantly, however, this employment effect is modest and does not translate into an effect for overall employment. As we document, only the additional lending subsequent to QE3 had a significant effect on overall employment and is arguably driven by credit supply rather than credit demand. Foley-Fisher et al. (2016) and Chakraborty et al. (2016) study the effect of unconventional monetary policy on firm behavior. In particular, Foley-Fisher et al. (2016) investigate the effects of the monetary expansion program (MEP), on non-financial firms and they find that firms that historically relied more on long-term debt issued more long-term debt after the MEP and that those firms increased investment. Chakraborty et al. (2016) study the effect of QE on bank lending and firm investment. While we find that QE3 induced additional C&I lending, Chakraborty et al. (2016) find that QE crowded out C&I lendingandledtoareductionofinvestmentbylargefirms. Thedifferencesinfindingscanbeexplained by the difference in the empirical specification and by differences in data coverage. First, their findings suggest a negative relationship between mortgage lending and C&I lending over a longer period, in line with more general trends documented by Chakraborty et al. (2018). In contrast, our findings suggest that, specifically around QE3, banks conduct additional mortgage lending as well as additional C&I lending. Second, our confidential loan-level data are less restricted in two important aspects: Our C&I lending data are not restricted to syndicated loans alone, and our sample at the bank-firm level is substantially larger and likely more representative of the C&I lending landscape. We find that the effect of QE3 is stronger for smaller firms that are not active in the syndicated loan market and that are arguably more bank dependent. Second, our data on C&I lending as well as our data on mortgages is at the quarterly rather than annual frequency and therefore allows us to capture the timing of QE effects more precisely. More generally, our analysis of the employment effects of QE contributes to the literature on how financial conditions affect real economic outcomes (see, e.g., Bernanke, 1983; Peek and Rosengren, 2000; Driscoll, 2004; Khwaja and Mian, 2008). In particular, a series of recent papers show how the 2007-08 financial crisis affected bank lending (see, e.g., Ivashina and Scharfstein, 2010), and real economic 7

outcomes via various lending channels (see, e.g., Chodorow-Reich, 2014; Duygan-Bump et al., 2015; Haltenhof et al., 2014; Bord et al., 2015; Benmelech et al., 2017; Greenstone et al., 2015). Our empirical strategy is closest to Benmelech et al. (2017), Bord et al. (2015), or Greenstone et al. (2015), who exploit spatial variation in the exposure to the financial crisis. Unlike our paper, which is concerned with the effect of unconventional monetary policy on employment, Benmelech et al. (2017) trace the effect of the run in the asset-backed commercial paper market on auto sales, and Bord et al. (2015) and Greenstone et al. (2015) analyze the employment effects of the financial crisis through reductions in small business lending. 2 Data Our study brings together data from different sources at the bank, county, and firm level. We take commercial bank balance sheet and income statement information from the Consolidated Reports of Condition and Income for commercial banks in the United States (Call Reports, based on Forms FFIEC 031 and FFIEC 041). Information on bank holding companies are obtained from the the Consolidated Financial Statements for Holdings Companies (FR Y-9C). All items are adjusted for bank mergers. We match publicly traded BHC’s with daily stock returns from the Center for Research in Security Prices (CRSP) using the publicly available crosswalk provided by the Federal Reserve Bank of New York. We make use of three further administrative sources that collect data on bank lending. The Home Mortgage Disclosure Act (HMDA) requires most banks to report on their mortgage lending activity. Data are reported at the mortgage-level and include information on the mortgage amount, whether the amount refers to the origination of a mortgage for a home purchase or the re-financing of an existing loan, and the geographic location of the property down to the census tract level. Virtually all banks in the Unites States that conduct any mortgage lending in an MSA report on their activity, so the HMDA data are considered a near census of mortgage lending. Importantly, the public version of the HMDA data only identifies the year of the observation. We use a confidential version of the data, maintained by the Board of Governors of the Federal Reserve, that reports exact application and action dates for each observation and therefore enables us to measure the timing of potential lending effects more precisely. For our analysis, we aggregate the HMDA data to the bank-quarter level and to the county-quarter level. Similar to the HMDA data, the Community Reinvestment Act (CRA) requires banks to report on 8

small business lending activity by geographic area. Each bank breaks down data by geographic region down to the census tract and by loan size bins. In addition, banks report total lending to businesses with revenues of less than $1m. The CRA data are fairly representative of the universe of small business lending by commercial banks: Greenstone et al. (2015) estimate that small business lending in the CRA data captures 86% of total small business lending in the U.S. In addition, we use data from the Summary of Deposits (SoD) collected by the Federal Deposit Insurance Corporation (FDIC). The SoD data consist of annual branch-level reports of total deposits by bank branch, and we aggregate the data to the bank–county level in our analysis. County-level data come from different sources. Employment data are from the Quarterly Census of Employment and Wages (QCEW) that is collected by the Bureau of Labor Statistics. The data provide quarterly counts of employment and wages by industry as reported by employers, and they cover more than 95% of U.S. jobs. We complement employment data with the County Business Patterns (CBP), an annual data collection by the U.S. Census that includes the number of establishments and number of employees by industry and county in the first week of March each year. The CBP data often report employmentonlyinbrackets. WeusethemethodinAutoretal.(2013)toestimateemploymentnumbers within brackets. As additional data on local economic conditions, we obtain regional house prices from Moody’s, median income from Haver Analytics, auto sales data from Polk, and population from the U.S. Census. To investigate directly the link between banks’ MBS holdings, corporate lending and real outcomes forasubsetofbanksandfirms,wemakeuseofnewlyavailableadministrativedataonbankloans. Since 2012, regulators have collected loan-level data on bank lending from any bank with total assets of more than 50 billion USD on a quarterly basis. The purpose of the data collection is to assess capital adequacy and to support stress test models, and the loan schedule (Y-14Q-H1) requires banks to report any loan on their balance sheet with a commitment of 1 million USD or more. Data include loan characteristics such as interest rates and the dates on which those rates can be re-set, collateral requirements, and the purpose of the loan. Moreover, the data allow to distinguish between term loans and credit lines, and whether a loan is syndicated or not; and include borrower characteristics, such as total assets, total debt, and capital investment. Firms can be followed across banks in the data set by their tax identification numbers. Importantly, relative to other data sets (such as DealScan or SNC), this data set includes syndicated loans but also many smaller and/or non-syndicated loans, and therefore has much broader coverage of firms than has been available in previous studies. 9

When we look at employment effects at the firm-level, we match the loan registry data with firms’ annual employment figures from Compustat based on tax identification numbers. Analysis of firm-level employment effects is therefore restricted to firms that can be matched in both datasets, and the overlap amounts to roughly 2700 companies. 3 Evidence on lending and employment 3.1 Empirical Strategy Weexploittwosourcesofvariationtostudytheeffectofunconventionalmonetarypolicyonemployment viabanklending: First, weusethefactthatbankswerearguablydifferentiallyexposedtotheFed’sMBS purchases during QE1 and QE3, which led to differential lending responses at the bank level. Second, banks’ business activities vary across different regions, allowing us to investigate how local changes in bank lending correlate with changes in employment. We start out by briefly5 reviewing the Federal Reserve’s unconventional monetary policy before laying out our empirical strategy in more detail. 3.1.1 The Federal Reserve’s LSAPs By the end of 2008, the federal funds target rate effectively hit the zero lower bound, and LSAPs became an important tool for the Federal Open Market Committee (FOMC) to conduct monetary policy. On November 25, 2008, the FOMC announced what came to be known as QE1: the Federal Reserve would buy up to $100 billion of direct debt obligations issued by Fannie Mae and Freddie Mac, and an additional $500 billion of agency MBS. The program was extended in the FOMC meeting on March 18, 2009, and, by the end of QE1 (March 2010), the Federal Reserve had bought $1.25 trillion in MBS, $175 billion in Federal Agency debt, and $300 billion in U.S. Treasuries. In March 2010, the Fed held about one quarter of all available MBS. Overthecourseof2010and2011theFederalReserveimplementedtwoadditionalprogramsfocused solely on the purchase of Treasuries. First, on August 10, 2010, the FOMC indicated the start of a second round of quantitative easing (QE2), which ultimately led to the total purchase of $778 billion in long-term Treasury securities. Second, on September 21, 2011, the Federal Reserve announced the 5Foramoredetailedchronologyofevents,seeTable1inGilchristetal.(2015). 10

maturity extension program (MEP), which involved the sale of short-term U.S. Treasuries and the purchase of long-term Treasuries. Givendisappointingeconomicactivityandstillrelativelyhighunemployment,theFOMCannounced athirdroundofquantitativeeasing(QE3)initsstatementonSeptember13,2012. Largelyunanticipated, QE3 initially dictated the purchase of $40 billion in agency MBS per month, and another $45 billion in U.S.Treasuries(addedtothepolicyonDecember12, 2012). Afterimprovementsoftheeconomybecame apparent, Chairman Ben Bernanke first indicated the Tapering of QE3 in his testimony to the Joint Economic Committee on May 22, 2013. The potential tapering was confirmed in the FOMC statement on June 19, 2013, and the Federal Reserve reduced the purchase amounts to $35 billion in agency MBS and $40 billion in U.S. Treasuries, respectively in December, 2013, and the program formally ended on October 29, 2014. 3.1.2 Variation at the bank level In this study, we focus on the effect of the Federal Reserve’s purchases of MBS during QE1 and QE3. We use cross-sectional variation of banks’ mortgage-backed securities (MBS) holdings in order to identify an effect of QE on bank lending. In particular, we argue that those banks that held more MBS prior to QE were relatively more affected by the MBS purchases. There are several reasons to believe that large-scale purchases of agency MBS affected banks with relatively larger MBS holdings more than banks with relatively lower MBS holdings. First, differences in MBS holdings may capture differences in banks’ business models. Some banks tend to be more active in the mortgage market and are thereby more exposed to housing in general. Hence, when the Federal Reserve purchases agency MBS and the prospects of the housing market arguably rise, banks that are more exposed to the housing market may benefit more. Second, large-scale purchases of MBS may directly raise the values and liquidity of banks’ MBS holdings;fortheoreticalmechanismssee,e.g.,GertlerandKaradi(2011)andBrunnermeierandSannikov (2014). Krishnamurthy and Vissing-Jorgensen (2013) show that the Fed’s actions had a considerable effect on MBS prices, especially during QE1. Moreover, they argue that QE operated through a “narrow channel” and MBS prices changed more than other asset prices. Thus, one may argue that banks with higherMBSsharesbenefitedrelativelymorefromtheFed’sactions. Inlinewiththisevidence,Darmouni and Rodnyansky (2017) find that additional lending after QE1 stems from an improved capital position of affected banks, and that additional lending after QE3 was driven by an improved liquidity position. 11

Third, the increase in lending may be the result of a more general portfolio re-balancing channel: given that low-yield assets, such as reserves, are not perfect substitutes for higher yielding assets, such as MBS, large increases in the supply of central bank money can induce banks to invest in more higher yielding assets. This can be achieved, for instance, by issuing new loans or mortgages. In a first step, we test the empirical connection between banks’ MBS holdings and MBS purchase announcements by the Federal Reserve. Mirroring the approach used by Foley-Fisher et al. (2016) to analyze the effect of the MEP on non-financial firms, we investigate stock returns of bank holding companies on the day of the announcement of a given round of QE using the following model: (cid:18) (cid:19)(j) MBS r = α+β +θX +(cid:101) (1) bt bt bt Total Assets b where r is the (risk-adjusted) stock return of bank b on day t. We proxy a bank’s exposure to QE by bt a bank’s share of agency MBS holdings relative to total assets, (cid:0) MBS (cid:1)(j) , averaged over the four TotalAssets b quarters prior to round j = 1,2,3 of QE. X is a vector of bank-level characteristics available at time t. bt On average, around 8% of all bank assets are MBS (see Section A in the appendix for details). In 2008, prior to QE1, more than a quarter of all commercial banks held no MBS at all, and the average MBS-to-assets ratio was around 12% in the upper quartile of the cross-sectional distribution across banks. Among the set of publicly traded bank holding companies, the average MBS-to-assets ratio is around 6.5 percent in the 25th percentile and around 17 percent in the 75th percentile. Moreover, bank with larger MBS holdings tend to be larger, more leveraged, and more exposed to the housing market in general – observable characteristics we control for in Equation (1). Table 1 shows estimates of equation (1) on the announcement days of QE1 and QE3, respectively. A higher MBS share is associated with higher raw and risk-adjusted returns: In particular, the stock return ofabankatthe75thpercentileofthecross-sectionalMBSsharedistributionexceededthestockreturnof a bank at the 25th percentile of the distribution by about 78 basis points when QE1 was announced and byabout25basispointswhenQE3wasannounced, controllingforotherobservablebankcharacteristics. These findings suggest that the market valued MBS exposure beyond e.g. bank size or leverage on QE announcement days. [TABLE 1 ABOUT HERE] Figure 1 plots coefficient estimates of β when we estimate equation (1) for five days before and 12

after the announcements of QE1 and QE3. In line with QE1 occurring in a period of higher volatility, coefficients in QE1 are less precisely estimated. Both panels suggest that banks with a higher MBS share outperformed banks with a lower MBS share on or shortly after the day of the announcement of QE1 and QE3, respectively. [FIGURE 1 ABOUT HERE] Further evidence for the empirical connection between banks’ MBS holdings and the MBS purchases of the Federal Reserve comes from the response of bank balance sheets after the implementation of the policy. In appendix A, we find that banks that held more MBS prior to QE expanded their lending more after the implementation of QE1 and QE3. Our analysis there, which builds on a previous study by Darmouni and Rodnyansky (2017), shows that overall bank lending volumes increased after both, QE1 andQE3,butC&IlendingincreasedonlyafterQE3andnotafterQE1. Withinthemortgagemarket,QE1 mainly affected the refinancing volume of existing mortgages, in line with the findings of Di Maggio et al. (2016), and QE3 affected the origination of home purchase mortgages at more exposed banks. 3.1.3 Variation at the regional level BuildingontheempiricallinkbetweenbanksandtheFederalReserve’sMBSpurchasesdescribedabove, we exploit that there is spatial variation in the local activity of banks with different MBS holdings. This allows us to test whether the activity of affected bank correlates with increased local lending subsequent to a round of QE, and whether an increase in lending correlates with employment and consumption growth. For our main specification, we observe counties at different points in time (quarterly or annually) and we calculate the exposure of each county c to a round QE j as (cid:18) (cid:19)(j) Exposure (j) = ∑ w (j) MBS . (2) c bc Total Assets b b (j) Here, w describes the market share of bank b in county c prior to QEj, where market share is bc computed as bank b’s deposits, its volume of small business lending or its volume of mortgage lending in county c prior to QEj relative to the total deposits (or total small business loans, or total mortgage lending) of all banks active in county c. In other words, a county’s exposure is a bank-activity weighted 13

average of banks’ exposure to QE where a bank’s exposure is measured by its MBS holdings. Note that since MBS are held by the respective commercial bank, the MBS share varies only at the bank-level and not at the bank-county level. In our main specifications, we use the exposure measure that is most closely related to the outcome ofinterest, e.g. whentheoutcomeismortgagegrowth,weusetheexposuremeasurebasedonmortgage volume. In general, results are robust to using different exposure measures. As one would expect, all three measures are highly correlated, with raw correlations being above .5 and rank correlations being above .8. Figure B.2 in the appendix shows scatter plots of the three exposure measures based on deposits, small business lending and mortgage lending, respectively. Figure 2 shows our measure based on small business lending across U.S. counties before QE1. The spatial distribution does not seem to follow a systematic pattern, except for a cluster of counties with higher values of the exposure variable in the north east. In robustness checks, we show that results do not depend on including the north east in the estimations. Moreover, note that since MBS shares are very persistent within banks over time (see, e.g., Kurtzman et al., 2017) and spatial variation in geographical activity of banks is very stable over time as well, the map looks very similar atother points in time. [FIGURE 2 ABOUT HERE] Notethatourexposuremeasureis,however,correlatedwithseveralobservablecountycharacteristics. Splitting the sampleinto the upper and lower tercile, Table 2shows thatthose counties that arearguably more affected tend to be more urban counties. Hence they have a higher population, a higher median income, higher housing price levels and higher recent population growth. However, reassuring for our identification strategy, both types of counties experienced comparable declines in house prices and income during the financial crisis. [TABLE 2 ABOUT HERE] In our main analysis we estimate the following difference-in-differences specifications around the 14

first and third round of quantitative easing: K K y = α+β×Exposure (j)×QE (j) +γ +τ + ∑ θ (0) X (k) + ∑ θ (1) X (k) QE (j) +(cid:101) (3) ct c t c t k ct k ct t ct k=1 k=1 K K y = α+β×Treat (j) ×QE (j) +γ +τ + ∑ θ (0) X (k) + ∑ θ (1) X (k) QE (j) +(cid:101) (4) ct c t c t k ct k ct t ct k=1 k=1 Here, y is an outcome in county c at time t, and γ and τ denote county and time fixed effects. ct c t (j) QE is a dummy variable equal to 1 in all time periods after QEj and 0 otherwise. Equation (3) uses t (j) the continuous exposure measure directly, while Equation (4) uses a binary version with Treat equal c to 1 if a county’s exposure is in the upper tercile of the exposure distribution over all counties, and 0 (k) if it is in the lower tercile of that distribution. X is a vector of county-level time-varying controls, ct including levels and annual growth of population and median income. We allow for changes in the relation between controls and outcome variables in response to QE by interacting control variables with QE event dummies. Also note that we restrict data to counties with a population no less than 2500 registered inhabitants, as data from such small counties are less reliable. We estimate our main equation for three different types of outcomes at the county level. First, we consider the effect on local lending, distinguishing between mortgage lending to households and lending to firms, the latter proxied by small business lending. Second, in our main analysis, we consider the effect on local employment. Finally, we investigate the effect on local employment by different industries and the effect on local consumption, proxied by auto sales. Since our outcomes variables may not immediately react to the implementation of a round of QE, we estimate each regression with using quarterly data from three quarters before to three quarters after each respective QE event. In annual regressions we use the data from two years, the year in which a specific round of QE was started as well as the subsequent year. Our main outcome variable of interest is a county’s employment growth. County-level employment is measured quarterly in the QCEW data and we compute the four-quarter harmonized growth rate as Emp −Emp ∆ Emp = ct c,t−4 . (5) ct 0.5(Emp +Emp ) t c,t−4 Wefurthercalculategrowthinsmallbusinessloanswithfacevaluebetween$250kand1millionreported in the CRA data, mortgage origination, and refinancing of existing mortgages reported in the HMDA ∆ ∆ data, and auto sales as reported in the Polk data. We denote them as C&ILending, Origination, 15

∆ ∆ Refinancing, and Auto, respectively. Note that the data on mortgage lending are available at a quarterly frequency, while the data on small business lending are only available at an annual frequency. Before turning towards results, note that the success of our empirical strategy depends on a number of assumptions. First, as in any difference-in-differences design, outcomes need to evolve similarly absent treatment, i.e., follow parallel trends. Below, we provide suggestive evidence that counties with high and low MBS exposure followed similar trends before quantitative easing episodes. Second, relevant in our specific setup, our measure of a county’s MBS exposure is a proxy for its actual exposure, and any definition that we use might introduce measurement error in the regressor, leading to attenuation bias in the estimated coefficients. While our results are largely unaffected by the exact definition, this concern is an additional reason to focus on Equation (4), our specification with a binary treatment: Even if there is measurement error in the continuous variable, that should not affect the ordinal ranking of counties much, especially if, as we do, one compares counties in the highest tercile of the MBS exposure distribution to the ones in the lowest tercile. Third, and related to the previous point, the success of our strategy also depends on the extent to which bank lending markets are local. Throughout the main body of our analysis, we define each county as a local market. If there were no frictions in lending across regions, an expansion of lending at the bank level should not extend to the regional level, as a bank with a high MBS share would be, conditional on local loan demand, equally likely to increase lending in any region. Hence, all of our regressions at the regional level are a joint test: we do not just test whether there is an effect of QE on bank lending and employment but also whether banking markets are sufficiently local. Note that existing evidence suggests that the markets for C&I and small business loans tend to be local (see, e.g., Brevoort et al., 2010; Greenstone et al., 2015; Nguyen, 2015). In the case of mortgage lending activity, market locality is more debatable and the existing evidence is ambiguous. Beraja et al. (2018) do not find regional frictions in mortgage lending, while Scharfstein and Sunderam (2014) find that mortgage lending is characterized by local markets. 3.2 Results 3.2.1 Local lending We start out by testing whether more affected counties experience stronger lending growth subsequent to QE1 and QE3. Table 3 displays estimates of Equation (3) and Equation (4) with mortgage lending 16

variables as outcomes, distinguishing between refinancing of existing mortgages (Panel A) and origination of home purchase mortgages (Panel B). In both panels, we calculate exposure to QE using weights given by banks’ historical mortgage lending activity in each county. We find that mortgage refinancing activity increases more in more exposed counties after QE1. This finding is in line with the findings in Di Maggio et al. (2016) and Beraja et al. (2018), who show that when long-term interest rates were reduced during QE1, a refinancing boom followed. Counties in the upper tercile of the exposure distribution experienced roughly 3 percentage points higher refinancing activity than counties in the lower tercile of the distribution after QE1. Moreover, while we do not find consistenteffectsonmortgageoriginationactivityduringQE1,thepatternreversesaroundQE3: Wefind that mortgage origination increased in more exposed counties but refinancing activity was unaffected. Counties in the upper tercile of the exposure distribution experienced roughly 2.6-2.7 percentage points higher mortgage origination growth than counties in the lower tercile of the distribution after QE3. [TABLE 3 ABOUT HERE] Table 4 estimates difference-in-difference regressions using small business lending growth as the outcome variable for both the lending-based (Panel A) and the deposit-based exposure measures (Panel B). Counties in the upper tercile of the exposure distribution experienced roughly 6-7 percentage points higherlendinggrowththancountiesinthelowertercileofthedistributionafterQE3,dependingonhow the exposure is calculated. Point estimates after QE1 are modestly lower and effects are less precisely estimated. This finding could reflect the fact that the time period around QE1 was a period of generally higher volatility, making it harder to detect lending effects. Thesefindingsatthecountylevelarealsolargelyinlinewithevidenceatthebanklevel. InSectionA in the appendix, we use Call Report, HMDA, and CRA data to show that additional overall lending after QE1 by banks with higher MBS shares is mostly driven by an increase in the refinancing of existing loans. In addition, we show that increases in lending after QE3 stem mostly from origination of home purchase mortgages and additional C&I lending. [TABLE 4 ABOUT HERE] While the results in Table 3 and Table 4 report average effects, before and after quantitative easing episodes, we can study the dynamics of the effect in more detail. We interact time effects with the exposure variable and estimate the following regression: 17

3 K K y = α+ ∑ β (j) Treat (j) τ + ∑ θ (0) X (k) + ∑ θ (1) X (k) QE (j) +γ +τ +(cid:101) . (6) ck k c k k ct k ct t c k ck k=−3 k=1 k=1 In Equation (6), time (k) is measured relative to the start of each QE episode, and the regression includes county controls and county and time fixed effects as before. We normalize coefficients to 0 in the period before the start of QE. Given that only mortgage lending data, but not C&I lending data, are available at the quarterly level, we estimate the model for mortgage lending only. [FIGURE 3 ABOUT HERE] Figure3showsestimatesformortgagerefinancingaroundQE1andformortgageoriginationaround QE3. First, note that, reassuringly, there are no differential trends prior to either QE1 or QE3. In line with the result on the average effects above, coefficient estimates are large and significant in the three quarters immediately after the start of QE1. Similarly, for QE3, we see an uptick of mortgage origination growth quickly after the start of QE3. In line with estimates of the average effect, the magnitude of the effect is slightly weaker than the magnitude of the effect for the refinancing boom after QE1. Takentogether,ourevidencesuggeststhatexposedregionsexperiencedhighergrowthinrefinancing after QE1, and stronger growth in C&I lending and mortgage origination after QE3. These differences in lending responses across the two rounds of QE turn out to be crucial when explaining the employment effects of the respective rounds of QE. 3.2.2 Employment Figure 4 shows the employment growth of counties with high and low MBS exposure, defined as being in, respectively, the upper or lower tercile of the cross-sectional exposure distribution, with the dashed vertical lines denoting the start of QE1 and QE3. The figure reveals, that both types of counties experienced very similar employment growth rates during the recession and throughout much of the recovery. This helps to validate our empirical design in Equation (3) and Equation (4) as the parallel trends assumption appears to hold. Employment growth rates diverge, however, two quarters after the start of QE3. At up to 40 basis points per quarter, the difference is sizable. [FIGURE 4 ABOUT HERE] 18

Panel A of Table 5 presents our main results and provides a more rigorous analysis of the visual patterns in Figure 4. We estimate Equation (3) and Equation (4) around QE1 and QE3, controlling for county- and time-fixed effects and county-level time-varying characteristics. In line with Figure 4, we do not find any effect around QE1; the coefficient estimates are small and insignificant, irrespective of how we calculate the exposure measure. Effects around QE3, though, are larger and significant. Estimates based on the binary treatment specification suggest that county’s in the upper tercile of the MBS distribution experienced 40 basis points higher employment growth than counties in the lower tercile of the distribution, after accounting for other time-varying and time-invariant county characteristics. Panel B and Panel C of Table 5 confirm the results when we use the deposits-based and mortgage-based exposure measures with similar magnitudes. [TABLE 5 ABOUT HERE] As above, we can again study the timing of the effects in more detail by estimating Equation (6), now using the change in county-level employment as dependent variable. Figure 5 shows estimates for QE1 and QE3. First, note that there is again no discernible pre-trend before either QE1 or QE3, again, providing support to the parallel trends assumption. In line with the result on the average effect for QE1, coefficient estimates are small and insignificant in each quarter following the start of QE1. For QE3, we see an uptick of employment growth that becomes statistically significant two quarters after QE3 started. [FIGURE 5 ABOUT HERE] Given that QE3 was implemented in mid September 2012, Figure 5 suggests that the employment effect therefore becomes apparent between 6 and 9 months after the implementation. The literature has found a delayed effect of conventional monetary policy on real outcomes. The classic study by Bernanke and Blinder (1992) finds that employment reacts around 8 to 12 month after the monetary policy shock. More generally, the literature typically estimates the delay to be between 6 and 24 months (see Christiano et al. (1999) for an overview, and Olivei and Tenreyro (2007) for a recent study). Our estimate for unconventional monetary policy is at the lower end of that range. The question of whether this is true for unconventional monetary policy more generally can only be answered from a larger set of episodes of unconventional monetary policy implementation. 19

Magnitude To summarize, we find that small business lending and mortgage origination increased in arguably more exposed counties after QE3, while only mortgage refinancing activity increased in those counties after QE1. Our estimates imply that, after QE3, employment growth was approximately 30-50 basis points higher in more exposed counties, while small business lending growth was about 4-6 percentage points higher and growth in mortgage origination was about 2.7 percentage points higher. To get a better sense of the economic significance, we put our estimates in the historical context. RecallthattheU.S.economywasinarecoveryphasefromthefinancialcrisisduring2012. Ourestimates suggest that employment growth was 50 basis points higher in counties that were more affected by QE. Total employment in the upper tercile of counties by MBS exposure before QE3 was roughly 37.6m in September of 2012 and 38.3m a year later, resulting in an employment growth rate of 1.7 percent, such that roughly a third of employment growth in more affected counties (or, equivalently, 200k new jobs) can be attributed to the additional lending induced by QE3. Notethat,duetothecross-sectionalnatureofouranalysis,wecannotconcludethatwearemeasuring an aggregate employment effect of QE. To illustrate why this is the case, consider two extreme cases. At one extreme, the effect could be merely redistributive: jobs that would have been created in unaffected counties, were, due to QE, created in affected regions instead. At another extreme, the aggregate effect may be larger than what we are observing as there may be complementarity between different regions: QEinducedjobcreationinsomeaffectedregionsthatcouldhavespurredadditionaleconomicactivityin unaffected regions. Note, however, that our results also hold at the MSA level (see Section 5): Assuming that labor market mobilityis relatively high within MSA’s but lowacross MSA’s, the documentedeffects are thus unlikely to be purely redistributive. 3.2.3 Non-tradable goods, auto sales, and industrial financial dependence In this section, we present additional evidence on the employment effect by industry. This allows us to derive two additional insights. First, it reveals that even though there is no effect of QE1 on overall employment, employment in the non-tradable goods sector increases. Second, it allows us to shed some light on whether the additional local economic activity results from additional demand by households or additional investment and employment by firms. We estimate the main specification, Equation (3) and Equation (4), for employment growth in the non-tradable goods sector as well as for employment growth in manufacturing and the service sector ∆ ∆ (denoted by EmpNonTrad and EmpTradOther) separately. We define sectors using the definitions 20

of Mian and Sufi (2014).6 Because the QCEW data report many missing values for employment by industry, we focus on the annual CBP data in this part of the analysis. [TABLE 6 ABOUT HERE] Results reported in Table 6, Panel A, reveal two important findings. First, even though there is no overall employment effect of QE1, there is an economically and statistically significant expansion of employment in the non-tradable goods sector, see columns (3) and (4). Counties in the upper tercile of the exposure distribution experienced roughly 1.6 percentage points higher employment growth in the non-tradable goods sector than counties in the lower tercile of the distribution after QE1. Second, the overall employment effect of QE3 documented above results from both, an expansion of employment in the manufacturing and services sector, see columns (5) and (6), as well as an expansion of employment in the non-tradable goods sector, see columns (7) and (8). According to the estimates in Table 6, counties in the upper tercile of the exposure distribution experienced roughly 1.4 percentage point higher employment growth in the service industry and tradable goods sector, and 0.5 percentage point higher employment growth in the non-tradable goods sector. Note, however, that around QE3 the estimates for the non-tradable goods sector are relatively imprecisely estimated and not statistically significant. Distinguishing the employment effect by industry also allows to shed light on whether additional local economic activity is driven by increases in local demand from households or by increased credit supply for firms. Assume for the moment that the increase in local mortgage origination and C&I lending can be attributed to an increase in credit supply. This will be discussed in more depth in the next two sections. While in theory both channels, lending to household and lending to firms, may be relevant for determining the level of employment at the same time, the former would work via changes inlocaldemandforconsumption, andthelatterwouldworkviaimprovedfinancingconditionsforlocal firms. In other words, QE may operate via spurring local demand if the improved access to mortgages for households increases housing net-worth (Mian and Sufi, 2014), or QE may operate via spurring labor supply if improved access to credit for local firms increases their labor (and capital) investment (Chodorow-Reich, 2014), or both. 6MianandSufi(2014)calculatethatin2007,around20%ofallemploymentisinthenon-tradablegoodssector,10%arein thetradablegoods/manfuactringsector,60%aredefinedas“’others”,whichmainlycontainstheservicesectorthatdoesnot offernon-tradablegoods,andanother10%areinconstruction. Inouranalysis,wedistinguishbetweennon-tradablegood industriesandindustriesthatproducetradablegoodsorservices. 21

During QE1, we find that there is little or no additional lending to firms. Hence, any effect on employment should be driven by increases in local demand. The fact that employment increases in the non-tradable sector is in line with the findings by Mian and Sufi (2014), who show that increases in local demand, e.g., changes in economic activity due to changes in household net-worth, are more likely to affect non-tradable goods sector employment than employment in other sectors. The underlying idea is that non-tradable employment relies heavily on local demand, whereas tradable goods related employment is related to aggregate demand. Building on this insight, our evidence suggests that the additional refinancing of existing mortgages during QE1 indeed positively affected household net-worth and therewith demand: Given that householdsfacelowerinterestpaymentsontheiroutstandingdebt,theyarewealthierandcansustainahigher level of consumption. The additional consumption is reflected in a higher demand for non-tradable goods, leading to an increase of employment in the non-tradable goods sector. However, even though the effect is statistically significant and economically sizable within the non-tradable goods sector, the effect does not translate into additional growth in overall employment. This can be attributed to the fact that only 20% of the work force are employed in the non-tradable good sector. TotesttheeffectofQE1onlocalconsumptionmoredirectly,weprovidecomplementaryevidenceon how QE1 affected auto sales. Auto sales represent a good proxy for consumption of durables and they are available at the county level. Table 7 shows that sales of new and used cars increased subsequent to QE1. Counties in the upper tercile of the exposure distribution experienced roughly 3.5 percentage points higher auto sales growth than counties in the lower tercile of the distribution after QE1. The evidence is in line with consumption increasing after a positive shock to household net worth after QE1, resulting in additional employment in the non-tradable goods sector. [TABLE 7 ABOUT HERE] This interpretation is further supported by studying the timing of the effect on auto sales in more detail. Figure 6 plots coefficient for estimating Equation (6) during QE1 and QE3, using the growth in auto sales as the dependent variable. Panel A, which shows results for the episode of QE1, suggesting an immediate response that fades over the subsequent quarters. Importantly, the timing of the growth in auto sales in exposed counties also closely follows the increase in local refinancing documented in Figure 3 over time. 22

[FIGURE 6 ABOUT HERE] Recall that in contrast to QE1, QE3 led to an expansion of bank lending not only to households but also to firms. It is a priori not clear and empirically challenging to analyze which one of the two types of lending is more important for generating additional employment. While both channels may be important, consider the fact that the employment effect of QE3 seems strongest outside the non-tradable goods sector. To see this, compare again columns (5) and (6) with (7) and (8) in Table 6. Hence, unlike during QE1, employment growth does not result only from growth in the those industries that are more likely to be affected by local demand, indicating that lending to households was less relevant relative to the lending to firms to generate employment. Likewise, the effect of QE on the growth of auto sales documented in Table 7 is substantially weaker after QE3 than after QE1. Correspondingly, the dynamics of the auto sales response in Panel B of Figure 6 are also dampened relative to the response after QE1. These findings suggest that the additional employment after QE3 likely results at least in part from a stimulus to credit supply. To further test this interpretation, we investigate whether the employment change is concentrated in industries that are more dependent on external financing. If so, that would also suggest that credit supply forces were more relevant than local demand forces to drive the effect. As a measure of external financial dependence, we use the definition of Almeida et al. (2010), which gives an ordering of financial dependence of all industries. We define an industry as dependent on external financing if it is in the upper tercile of the financial dependence distribution and non-dependent if it is in the lower tercile. Using the industry codes in the CBP data, we construct the change in employment by county and ∆ ∆ financial dependence, denoted EmpFin and EmpNonFin. Panel B of Table 6 reports results from estimating our main specification, Equation (3) and Equation (4), using these outcome variables. The results in column (5) and (6), even if imprecisely estimated, indicate that employment growth subsequent to QE3 comes from employment growth in financially dependent industries, as opposed to less financially dependent industries, see column (7) and (8). Again, this suggests that the main employment effect during QE3 is driven by improved credit supply. Moreover,inlinewiththeabsenceofanoverallemploymenteffectduringQE1,noeffectonemployment in either industry is detected during QE1. Altogether, our evidence suggests that the increase in employment in the non-tradable goods sector 23

after QE1 was due to the additional refinancing activity subsequent to QE1, with positive effects on household net-worth and local demand. In contrast, additional overall employment growth subsequent to QE3 was less likely to be driven by additional lending to households but rather related to additional credit provision to firms. 4 Supply and demand in the credit market Results in the previous section show that the expansion of bank credit to firms and households subsequent to QE3 is associated with an effect on overall employment. Moreover, the additional refinancing of existing mortgages after QE1 is associated with an increase of employment only in the non-tradable goods sector. Additional lending and employment subsequent to QE1 and QE3 could result from additional credit supply by banks, but also from additional credit demand by firms and households. This section provides additional evidence to disentangle supply and demand in the C&I loan market and in the mortgage market. We first use confidential loan-level data to analyze supply and demand in the C&I loan market, Section 4.1, before using confidential mortgage-level data to discuss the role of demand and supply in the mortgage market, Section 4.2. 4.1 C&I loan market: Evidence from confidential loan-level data Using confidential loan-level data that are part of the Y14 data collection allows us to provide direct evidence of a link from QE3 to bank lending, and from bank lending to firms. The advantage of using this data set is that it provides a direct link between banks and firms, and firms can be followed across banks via their tax identification numbers. In particular, the latter allows us to control for firm-specific credit demand by comparing loan amounts from different banks to the same firm in the same period. Moreover, the data include information on firms’ capital expenditures and therefore allow us to trace the impact of lending on real firm activity directly. There are, however, also two main drawbacks of the data. First, the data collection started only in 2012; too late to cover the episodes of QE1 and QE2. Second, it only includes corporate lending from the largest banks, and to the extent that firms borrow from other banks, these loans would not show up in our data. Note, however, that the banks covered by the Y-14 data collection conduct roughly 75% of total C&I lending by banks reported in the Call Reports. 24

As above, we measure a bank’s exposure to QE by its MBS holdings relative to its total assets. Table C.2 in the appendix gives a sense of the variation of Y14-banks’ MBS holdings: While the average BHC covered in the Y14 data holds around 9.5% of its assets as MBS, the average share is 7% below the median and 15% above the median. Moreover, Table C.1 shows that banks with higher MBS shares tend to be very similar to banks with relatively low MBS shares across a number of observables. Nonetheless, regressions below include a set of time-varying controls (X ) in all specifications to alleviate concerns bt that results could be driven by differences in observables. We start with quarterly bank-firm level regressions that allow for bank-firm and firm-time fixed effectsinthespiritofKhwajaandMian(2008). Giventhecredit-registrylikenatureofourdata,ourdata includes several thousand firms with a relationship with more than one banks, an order of magnitude larger than in previous applications that relied on publicly available data of syndicated loans in the United States which could not observe loans by medium-sized firms, see for instance Chodorow-Reich (2014) or Chakraborty et al. (2016). However, firms that are active in the syndicated loan market are typically larger and hence are more likely to have the ability to substitute bank financing with other types of financing such as the bond market or public equity issuance. Figure B.4 in the appendix shows that not only large firms but firms of all sizes in the data set tend to have relationships with multiple banks. Exploiting the fact that banks borrow from multiple banks in the same time period makes it feasible to control for firm-specific credit demand. To this end, we estimate the following two equations: y = α+β (cid:18) MBS (cid:19)(3) QE (3) +γ +δ + ∑ K θ (0) X (k) + ∑ K θ (1) X (k) QE (3) +(cid:101) (7) bit Total Assets t it ib k bt k bt t bit b k=1 k=1 K K y = α+βTreat (3) QE (3) +γ +δ + ∑ θ (0) X (k) + ∑ θ (1) X (k) QE (3) +(cid:101) (8) bit b t it ib k bt k bt t bit k=1 k=1 where y is the quarterly change in the amount firm i borrows from bank b between time t−1 and bit t. In Equation (7) we measure bank b’s exposure as a continuous variable given by the average MBS holding prior to the implementation of QE3, (cid:0) MBS (cid:1)(3) . In Equation (8) we use a binary dummy TotalAssets b variable that takes the value one if a bank has MBS holdings above the median MBS holdings among all (3) banks that report in the Y14 data, Treat . b Further, γ is a firm-time fixed effect that absorbs firm-specific demand, and δ is a bank-firm it ib fixed effect that controls for relationship-specific unobservables between bank b and firm i. Due to the 25

inclusion of γ , estimation requires multiple observations of the same firm within one time period and it hence the sample is restricted to firms that borrow from multiple banks. We also include K bank-level (k) controls X as individual regressors and interacted with the QE3 time dummy. The bank-level control bt variables are listed in Table C.2 in the appendix. Table 8 shows results. Panel A uses the continuous treatment variable and Panel B uses the binary treatment variable. The specifications in column (2), (4), and (6) include firm-time fixed effects and suggestthatevenwhencontrollingforcreditdemand,theoveralleffectonlendinggrowthisstillsizable and statistically significant. In particular, as shown in column (2) in Panel B, lending by banks with an above median MBS share grew about 3 percentage points more than lending by banks with a below median MBS share after QE3. The remaining columns split total lending into firms that report and do not report in Compustat. While the coefficient for bank lending to Compustat firms is positive and statistically significant (see column (6)), it is lower than for firms that do not report in Compustat: lending by banks with an above median MBS exposure grows at an approximately 1.7 percentage points higher rate after the implementation of QE3. In contrast, credit to firms that do not report in Compustat grew around 3 percentage points faster for banks that have an above median MBS exposure (see column (4)). [TABLE 8 ABOUT HERE] The difference in credit outcomes between the different types of firms reflects the relationship between firm size and bank dependence. The effect on firms that do not report in Compustat is about twice the effect of lending to Compustat firms. This can be rationalized by the difference in firm size and bank dependence: Compustat firms are typically larger, tend to borrow via the syndicated loan market and have access to the corporate bond market, and are arguably less bank dependent than smaller firms that are not public, and do not have the same access to sources of outside financing. In sum, the analysis reveals that bank lending to firms increased subsequent to QE3, even when controllingfordemand. ThisevidencesuggeststhatadditionalC&IlendingsubsequenttoQE3stemmed from additional credit supply. Table C.11 in the appendix shows that effects are somewhat larger in magnitude in non-syndicated lending than in syndicated lending, the latter of which is also considered in Chakraborty et al. (2016). 26

Firm-level evidence on investment and employment As indicated above, the Y-14 data has an additional significant advantage: it allows us to link banks’ credit supply directly to firm investment and employment decisions. Using additional information on capital expenditures from the Y14 C&I schedule, which is reported for a subset of around one third of the firms documented in the data, we can further analyze how QE3 affected firms’ investment decisions. Moreover, for a subset of those firms that we match with Compustat, we do not only have information on capital expenditures, but we also obtain information on firm-level employment decisions. We first calculate a firm’s exposure to QE3 as (cid:18) (cid:19) ∑ MBS Exposure = w , (9) i bi Total Assets b b (cid:0) (cid:1) where MBS is the average MBS share of bank b over the four quarters prior to QE3, and w TotalAssets b bi is the average total lending volume from bank b to firm i over the same period. Note that results are robust to alternative exposure measures, e.g., calculating the exposure as the MBS share of only the most important bank/lead bank for firm i over the last four quarters. To investigate the effect of QE3 on bank credit, investment, and employment, we then run the following cross-sectional regression: y = α+βExposure +θX +(cid:101) (10) i i i i y = α+βTreat +θX +(cid:101) (11) i i i i where y is either the one-year growth in bank credit or investment at the firm level from 2012Q4 to i 2013Q4, or the one- or two-year growth rate of employment at the firm level. As above, we estimate the model both with a continuous exposure measure, Equation (10), as well as a binary treatment that takes the value of 1 if firms i’s exposure is in the upper tercile of the exposure distribution and 0 if in the lower tercile, Equation (11). Note that while we calculate the change in the stock variables (credit and employment) as growth rates, we calculate the change of investment as the increase in capital expenditures normalized by the total assets of the firm, i.e. CapEx −CapEx ∆ Investment = 2013Q4 2012Q4 . i Total Assets 2012Q4 27

[TABLE 9 ABOUT HERE] Panels A, B and C of Table 9 show results for credit, investment, and employment, respectively. Starting with Panel A, firms that were more exposed to banks with a higher MBS share, increased bank borrowing more after QE3 for both the subsample of public firms from Compustat and for the larger sample of private firms in the Y14 data. In particular, moving from the lower tercile of the exposure distribution at the firm-level to the upper tercile implies a 1.3 percentage point higher growth in bank credit for firm reporting in Compustat and a 1.7 percentage point higher growth in bank borrowing for firms not in Compustat. This is consistent with the fact that firms that report in Compustat tend to be larger and have access to relatively more sources of financing than firms that do not report in Compustat. Panel B of Table 9 shows that the growth in bank credit is associated with an increase in investment: For firms reporting in Compustat, Columns (2) and (4) report relatively imprecise estimates suggesting that moving from the lower to the upper tercile of the exposure distribution has a positive effect on growth of capital expenditures over assets. For firms that do not report in Compustat, growth in capital expenditures is also larger, and estimated to be around 21 basis points higher when controlling for other firm characteristics. Finally, Panel C shows that more exposed firms were also relatively more active in hiring new employees after QE3. As noted above, because employment is not available in the Y14 data, results are based on those firms that can be matched to Compustat. Effects in the first four columns show employment growth over the four quarters after QE3, and the remaining columns show employment growth over eight quarters. The employment effect is already detectable in the four quarters after QE3: Employment growth is about 1 percentage point higher for a firm in the upper tercile of the exposure distribution compared to a firm in the lower tercile. Given that it takes time to hire, estimates are somewhat higher for employment growth over two years, but the estimates show that relative employment growth is slowing down over time. To get a sense of the relative magnitudes of all three effects, consider the median firm reporting in Compustat that we are able to match with the Y-14 data. The median firm in the Compustat subsample reports around $2 billion in total assets (see Table C.3 in the appendix). Moreover, firms in our Compustat sample finance on average 20 percent of their assets with bank loans. Our estimate in 28

panel A implies that a firm in the upper tercile of the exposure distribution borrows an additional $5.2 million relative to a firm in the lower tercile of the distribution (20% of $2b times .013). Our estimates in panels B and C suggest that the more exposed firm increases capital expenditure by 7-14 basis points relative to total assets, i.e. by about $1.4m - $2.8m, and that it hires an additional 34 workers (with median employment at 3,400 workers and a 1 percentage point higher growth rate). Of course, these estimates come with relatively large standard errors. In general, though, the three estimates appear to have the same order of magnitude. While the estimated number of additional employees relative to additional bank credit might seem high, its important to keep in mind that the firms considered have several other ways of raising additional funds. 4.2 Mortgage market We now turn to analyzing demand and supply in the mortgage market. Recall that the additional employmentinthenon-tradablegoodssectorsubsequenttoQE1iscorrelatedwithadditionalborrowing from households. As with C&I lending, we cannot easily conclude that the additional lending to household results from additional supply of credit by banks. Therefore, we provide additional evidence to test whether the additional refinancing after QE1 and the additional origination of mortgages subsequent to QE3 can at least partially be attributed to additional supply by banks. Observe that separating supply and demand empirically in the mortgage market is considerably more difficult than in the C&I loans market. This is due to the fact that a household, unlike firms, typically does not have a (mortgage borrowing) relationship with more than one bank. Hence, we can only control for local demand for mortgage financing by exploiting that multiple bank are active in the same county. I.e., we can control for county-specific credit demand by estimating the following county-bank level regressions: y = α+β× (cid:18) MBS (cid:19)(j) ×QE (j) +γ +δ + ∑ K θ (0) X (k) + ∑ K θ (1) X (k) QE (j) +(cid:101) , (12) bct Total Assets t ct bc k bt k bt t bct b k=1 k=1 where y is the harmonized annualized growth of refinancing or mortgage origination of bank bct b in county c from quarter t−4 to t, (cid:0) MBS (cid:1)(j) is bank b’s MBS exposure prior to round j of QE, TotalAssets b and δ is a bank-county fixed effect, andγ is a county-time fixed effect that absorbs all local economic bc ct conditions, including county-specific credit demand. X contains all time-varying bank-level controls, bt also interacted with the QE event dummy. We distinguish between GSE-conforming mortgages and 29

non-conforming mortgages.7 [TABLE 10 ABOUT HERE] Table 10 shows results. In line with findings in Section 3.2.1, we find that additional lending to households after QE1 results from mortgage refinancing. Interestingly, additional refinancing takes place in the conforming as well as non-conforming segment of the market (Panel A), even though the non-conforming part of the market is relatively small by historical standards (also documented in Di Maggio et al. (2016)). Also in line with finding in Section 3.2.1, we find that the origination of mortgages for home purchases (Panel B) increased after QE3 at more affected banks. This is, however, entirely driven by an expansion of lending in the non-conforming segment of the market, which recovered relative to the period around QE1. Most important for our purposes, Table 10 suggests that local demand is not the sole driver of our results. In particular, when county-time fixed effects are included in columns (2), (4), (6), and (8), the overall effects remain statistically and economically significant. How credible is the finding that the refinancing subsequent to QE1 is at least partially driven by additional credit supply? At first glance, the structure of the US mortgage market implies that credit demand rather than credit supply should be driving observed increases in mortgage lending, in particular during QE1. Mortgages are typically highly standardized products that can be securitized and sold in secondary markets. Moreover, when interest rates were lowered during QE1, the incentives to refinance mortgages were strongest for households. In line with that logic, existing evidence by Di Maggio et al. (2016) and Beraja et al. (2018) document a refinancing boom subsequent to QE1. Moreover, Di Maggio et al. (2017) show that refinancing is more likely to be demand-driven as those household with the strongest incentive to refinance also have the highest propensity to refinance. However, credit supply may nonetheless be important. First, note that originating GSE-conforming mortgages, even if the ultimate default risk does not lie with the originating bank, may yet be associated withsignificantrisks. Forinstance,GSE-conformingmortgagesaretypicallyassociatedwitharepurchase riskthatrealizeswhenamortgageisreturnedtothebankasnon-eligible. Thisriskhadbeenparticularly high in the period after the Financial Crisis. Moreover, non-conforming mortgages are often kept on the banks balance sheets and hence the default risk is likely to remain with the issuing bank. 7Wedefineamortgageasconformingiftheamountoflendingliesbelowtheconformingamountspecifiedonfhfa.gov (oneunitlimit). Observethataconformingmortgagealsoneedstofulfillotherrequirements,suchasminimumLTVratios andFICOscores,unobservedintheHMDAdata. Wehencefalselydefinesomenon-conformingmortgagesasconforming. 30

In line with this logic, observe that around QE1, the potential bias from local demand effects should be relatively lower for non-conforming mortgages than for conforming ones: Comparing the coefficient for conforming mortgages in the specification with and in the specification without county-time fixed effects, columns (1) and (2) in Panel A, we observe that the coefficient is reduced significantly when controlling for demand. In contrast, the change in coefficients for non-conforming mortgages across specifications, columns (5) and (6) in Panel A and columns (7) and (8) in Panel B, is relatively small. 5 Robustness This section provides additional evidence to corroborate our main findings. First, we consider effects of QE2 and the Tapering of QE3 on bank lending and employment. Second, we show robustness of our main results to different specifications and exposure measures. 5.1 QE2 and the Tapering of QE3 QE2wasannouncedattheFOMCmeetingofAugust10,2010. Giventhattheprogramconsistedentirely of Treasury purchases, it acts as a natural placebo test: If the effects of the Federal Reserve’s LSAPs were a result of MBS purchases rather than Treasury purchases, one would expect to find no significant relation between MBS exposure and bank lending or employment following QE2.8 The Tapering of QE3 was first indicated by chairman Bernanke’s opening remarks in the Federal Reserve’s semiannual report to Congress (“Humphrey Hawkins hearing”) on May 22, 2013 and formally announcedattheFOMCmeetingofJune19,2013. Likewise,ifadditionalbanklendingandemployment was a result of MBS purchases, the Tapering should have a negative effect on lending and employment, thereby providing additional support for our QE1 and QE3 findings from a different event. To analyze both programs in more detail, we estimate our main specification, Equation (4), for the three quarters before and after the announcement of QE2 and of the Tapering of QE3, using mortgage lending or employment as outcome variables. In addition, we estimate Equation (7) at the bank-firmquarter level to assess effects on C&I lending for the three quarters around the start of the Tapering of QE3. Table 11 shows results for mortgage origination and mortgage refinancing. Columns (1) - (4) in both panels indicate that there is little difference in mortgage related lending across counties with different 8AsQE2focusedontreasurypurchases,were-ranallQE2specificationsreportedbelowwithtreasuryexposureinsteadof MBSexposure. Resultsareverysimilartotheonesdiscussedbelowandareavailableonrequest. 31

exposure to QE2. These findings are consistent with bank-level results in Section A in the appendix that show there is no differential mortgage lending response after QE2 across banks with different MBS shares. [TABLE 11 ABOUT HERE] Incontrast,columns(5)-(8)inpanelAshowasignificantnegativeresponseofmortgagerefinancing growth in more exposed counties after the Tapering of QE3. In addition, columns (5) - (8) in Panel B suggest lower growth in the origination of home purchase mortgages after the Tapering of QE3 in more exposed counties, although the effects are imprecisely estimated. Overall, the Tapering of QE3 is associated with a reduction in mortgage-related bank lending. Table 12 shows results for employment growth. Consistent with the muted lending response after QE2, it is not surprising that there is no considerable employment effect of QE2, regardless of which exposure measure is used (columns (1) - (4) in all panels). In contrast, after the Tapering of QE3, coefficient estimates are consistently negative although not precisely estimated across all exposure measures (columns (5) - (8) in all panels). The negative point estimates are consistent with the negative effects of Tapering on bank lending. [TABLE 12 ABOUT HERE] Table13reportsestimatesoftheeffectoftaperingonC&Ilendingusingbank-firmleveldataandthe specification in Equation (7). Recall that, as above, this specification is particularly attractive, because we can include firm-time fixed effects that absorb changes in credit outcomes resulting from changes in credit demand rather than credit supply. However, data availability restricts the analysis to events after late 2011 and hence to the Tapering of QE3 only. Across all specifications and subsamples in panels A and B of Table 13, our estimates suggest a significant negative effect of credit supply to firms. In particular, column (2) of Panel B shows that, in the three quarters after the Tapering of QE3, bank-firm credit growth is on average 1.3 percentage points lower at banks with above median MBS exposure. Columns (4) and (6) show that this effect is somewhat larger for non-public firms and somewhat weaker for public firms. [TABLE 13 ABOUT HERE] 32

Altogether, the absence of an effect of QE2 on bank lending or employment, and the negative effects of the Tapering of QE3 on bank lending and employment reinforce our main finding that the Federal Reserve’s MBS purchases (QE1 and QE3) were associated with additional bank lending that affected the real economy. 5.2 Specifications and sample restrictions Table C.4 in the Appendix shows that the employment effect of QE3 is robust to a number of sample restrictions and to different definitions of the exposure measure. Columns (1) and (2) report results for our main regression (equation (4)) when we restrict the sample to counties that are relatively small as our banking market definition is arguably more likely to hold for smaller counties. The visual evidence in Figure 2 suggests that the concentration of banks with high MBS is particularly high in the Northeast corridor of the United States. Columns (3) and (4) show that results are unchanged when this region is excluded from the estimation. In columns (5) and (6), we calculate the exposure measure by using the MBS holdings of banks in the four quarters prior to QE1 instead of those holdings prior to QE3. Results are robust to this change, which is unsurprising as MBS holdings are highly autocorrelated within bank over time. Finally, columns (7) and (8) report results for calculating the exposure measure as MBS over total securities instead of MBS over total assets. A related concern is that quantitative easing led not only to a decline of MBS yields but also to a decline of long-term treasury yields (see, e.g., Krishnamurthy and Vissing-Jorgensen, 2011). As such, a bank’s exposure to quantitative easing might be better captured by the sum of MBS and long-term treasury holdings. While we cannot measure long-term treasury holdings in the call reports separately, we re-estimated our main specifications using the sum of MBS and all treasury holdings as exposure measure. Given that only a small share of banks’ assets is invested in treasuries (the average share during our sample period is less than 1%), the alternative exposure measure is very similar to our original one and, as a result, estimations with the alternative measure had little effect on conclusions (results available upon request). Our main results are also robust to defining markets at the MSA level instead of the county level. Firms may borrow and employees may work across county-lines, and therefore a county might not comprise a local credit or labor market. Table C.7 in the appendix shows that employment results are robust to conducting the analysis at the MSA level. 33

6 Conclusion How can monetary policy affect real economic outcomes at the zero-lower bound? Almost a decade after the start of quantitative easing in the United States, the channels by which such monetary policies can affect employment are not yet fully understood and controversial. While existing evidence shows how LSAPs can affect bank lending as well as activity in the mortgage market, the effect of QE on economic activity and employment remains unclear. However, such evidence is particularly important, given that the Federal Reserve’s statutory objective is concerned with not only price stability, but also employment.9 Against this background, our study brings to bear cross-sectional variation across banks and regions to shed light on the effect of unconventional monetary policy on employment. We show that banks with a higher share of assets in MBS increased lending more after QE1 and QE3, and we exploit spatial variation in banks’ activity to trace the effect of lending on employment. While QE1 and QE3 were both successful in spurring bank lending in general, we show that there is considerable variation in the type of lending that expanded. The differential lending response is crucial to understanding the ultimate effect on the real economy. In particular, we show that only QE3, after which banks extended C&I lending, led to additional overall employment. In contrast, QE1, after which mortgage refinancing activity picked up, likely increased household net-worth and spurred local demand. Thiseffect, however, didnottranslateintooverallemploymentgrowthbutonlyintoadditional employment in the non-tradable goods sector. Altogether, our results indicate that LSAPs as conducted by the Federal Reserve can generally affect employment via a bank lending channel. Our evidence suggests that the Federal Reserve’s action, in particular QE3, were helpful for the economic recovery, in line with the stated goals of the program.10 However,theabsenceofastrongemploymenteffectafterQE1alsoindicatesthattheultimaterealeffects of LSAPs depend on other economic variables that are not under the control of the central bank.Hence, while our results indicate that transmission from unconventional monetary policy to the real economy is not entirely impaired at the zero-lower bound, its efficacy – analogous to the efficacy of conventional monetary policy (Vavra, 2014) – varies over time. 9NotethattheFederalReserveActalsoexplicitlymentionsmoderatelong-terminterestratesasastatutoryobjective. 10See for instance Ben Bernanke in The Courage to Act: “[QE3] was risky. Either we reached our goal of substantial labormarketimprovementorwewouldhavetodeclaretheprogramafailureandstopthepurchases,astepsuretorattle confidence”. 34

References Acharya, V. V., T. Eisert, C. Eufinger, and C. W. Hirsch (2016). Whatever it takes: The real effects of unconventional monetary policy. Working paper. Almeida, H., M. Campello, et al. (2010). Financing frictions and the substitution between internal and external funds. Journal of Financial and Quantitative Analysis 45(03), 589–622. Autor, D. H., D. Dorn, and G. H. Hanson (2013). The china syndrome: Local labor market effects of import competition in the united states. American Economic Review 103(6), 2121–68. Benmelech, E., R. R. Meisenzahl, and R. Ramcharan (2017). The real effects of liquidity during the financial crisis: Evidence from automobiles. Quarterly Journal of Economics 132(1), 317–365. Beraja, M., A. Fuster, E. Hurst, and J. Vavra (2018). Regional heterogeneity and monetary policy. Quarterly Journal of Economics, forthcoming. Bernanke, B. S. (1983). Nonmonetary effects of the financial crisis in the propagation of the great depression. American Economic Review 73(3), 257–276. Bernanke, B. S. and A. S. Blinder (1992). The federal funds rate and the channels of monetary transmission. American Economic Review 82(4), 901–921. Bord, V., V. Ivashina, and R. Taliaferro (2015). Large banks and the transmission of financial shocks. Working paper. Brevoort, K. P., J. D. Wolken, and J. A. Holmes (2010). Distance still matters: the information revolution in small business lending and the persistent role of location, 1993-2003. Working paper. Brunnermeier, M. K. and Y. Sannikov (2014). A macroeconomic model with a financial sector. American Economic Review 104(2), 379–421. Cahn, C., A. Duquerroy, and W. Mullins (2018). Unconventional monetary policy and bank lending relationships. Working paper. 35

Carpinelli, L. and M. Crosignani (2017). The effect of central bank liquidity injections on bank credit supply. Working paper. Chakraborty, I., I. Goldstein, and A. MacKinlay (2016). Monetary stimulus and bank lending. Working paper. Chakraborty, I., I. Goldstein, and A. MacKinlay (2018). Housing price booms and crowding-out effects in bank lending. Review of Financial Studies. Chodorow-Reich, G. (2014). The employment effects of credit market disruptions: Firm-level evidence from the 2008–9 financial crisis. Quarterly Journal of Economics 129(1), 1–59. Christiano, L. J., M. Eichenbaum, and C. L. Evans (1999). Chapter 2 monetary policy shocks: What have we learned and to what end? Handbook of Macroeconomics 1, 65 – 148. Correia, S. (2017). Linear models with high-dimensional fixed effects: An efficient and feasible estimator. Working paper. Crosignani, M., M. Faria-e Castro, and L. Fonseca (2018). The (unintended?) consequences of the largest liquidity injection ever. Working Paper. Cumming, F. (2018). Monetary policy and its effect on regional employment. Working paper. Darmouni, O. and A. Rodnyansky (2017). The effects of quantitative easing on bank lending behavior. Review of Financial Studies 30(11), 3858. Di Maggio, M., A. Kermani, B. Keys, T. Piskorski, R. Ramcharan, A. Seru, , and V. Yao (2017). Monetary policy pass-through: Mortgage rates, household consumption and voluntary deleveraging. American Economic Review, forthcoming 107(11), 3550–88. Di Maggio, M., A. Kermani, and C. Palmer (2016). How quantitative easing works: Evidence on the refinancing channel. Working paper. Drechsler, I., A. Savov, and P. Schnabl (2017). The deposits channel of monetary policy. Quarterly Journal of Economics 132(4), 1819–1876. Driscoll, J. C. (2004). Does bank lending affect output? evidence from the us states. Journal of Monetary Economics 51(3), 451–471. 36

Duygan-Bump,B.,A.Levkov,andJ.Montoriol-Garriga(2015). Financingconstraintsandunemployment: evidence from the great recession. Journal of Monetary Economics 75, 89–105. Federal Financial Institutions Examination Council (U.S.). Home mortgage disclosure act (confidential data). Foley-Fisher, N., R. Ramcharan, and E. Yu (2016). The impact of unconventional monetary policy on firm financing constraints: Evidence from the maturity extension program. Journal of Financial Economics 122(2), 409 – 429. Gertler, M. and P. Karadi (2011). A model of unconventional monetary policy. Journal of Monetary Economics 58(1), 17 – 34. Gilchrist, S., D. López-Salido, and E. Zakrajšek (2015). Monetary policy and real borrowing costs at the zero lower bound. American Economic Journal: Macroeconomics 7(1), 77–109. Greenstone, M., A. Mas, and H.-L. Nguyen (2015). Do credit market shocks affect the real economy? quasi-experimental evidence from the great recession and ‘normal’ economic times. Working paper. Haltenhof, S., S. J. Lee, and V. Stebunovs (2014). The credit crunch and fall in employment during the great recession. Journal of Economic Dynamics and Control 43, 31 – 57. Hancock, D. and W. Passmore (2011). Did the federal reserve’s mbs purchase program lower mortgage rates? Journal of Monetary Economics 58(5), 498 – 514. Ivashina,V.andD.Scharfstein(2010). Banklendingduringthefinancialcrisisof2008. JournalofFinancial Economics 97(3), 319 – 338. Kashyap,A.K.andJ.C.Stein(1994). Monetarypolicyandbanklending. InMonetarypolicy,pp.221–261. The University of Chicago Press. Kashyap,A.K.andJ.C.Stein(2000). Whatdoamillionobservationsonbankssayaboutthetransmission of monetary policy? American Economic Review 90(3), 407–428. Khwaja, A. I. and A. Mian (2008). Tracing the impact of bank liquidity shocks: Evidence from an emerging market. American Economic Review 98(4), 1413–42. Krishnamurthy, A. and A. Vissing-Jorgensen (2011). The effects of quantitative easing on interest rates: Channels and implications for policy. Brookings Papers on Economic Activity 42(2), 215–287. 37

Krishnamurthy, A. and A. Vissing-Jorgensen (2013). The ins and outs of lsaps. Kansas City Federal Reserve Symposium on Global Dimensions of Unconventional Monetary Policy. Kurtzman, R., S.Luck, andT.Zimmermann(2017). Didqeleadtolaxbanklendingstandards? evidence from the federal reserve’s lsaps. Journal of Banking & Finance, forthcoming. Mian, A. and A. Sufi (2014). What explains the 2007–2009 drop in employment? Econometrica 82(6), 2197–2223. Moody’s Analytics, Inc. U.S. National & Regional Historical Database (Data Buffet). http://www. economy.com/home/products/databases.asp?src=DataSolutions. Nguyen, H.-L. Q. (2015). Do bank branches still matter? the effect of closings on local economic outcomes. Working paper. Olivei, G. and S. Tenreyro (2007). The timing of monetary policy shocks. AmericanEconomicReview 97(3), 636–663. Peek, J. and E. S. Rosengren (2000). Collateral damage: Effects of the japanese bank crisis on real activity in the united states. American Economic Review 90(1), 30–45. R.L. Polk & Co. New Vehicle Registration Data. https://www.ihs.com/btp/polk.html. Scharfstein, D. and A. Sunderam (2014). Market power in mortgage lending and the transmission of monetary policy. Working paper. Vavra, J. (2014). Inflation dynamics and time-varying volatility: New evidence and an ss interpretation. Quarterly Journal of Economics 129(1), 215–258. 38

Figures 39

(a)QE1: November25,2008 (b)QE3: September13,2012 Figure1: Coefficientestimatesaroundeventdates. ThisfigureplotsMBSexposurepointestimatesand95%confidencebands obtainedfromaseriesofdailyregressionsaroundQEannouncementdates(November25,2008,forQE1andSeptember13, 2012,forQE3),indicatedbyverticallinesat0. Resultsarebasedonaregressionofabnormalstockreturnsforagivendayon abank’sMBSexposureandotherbankcharacteristics,seeeq.(1). 40

)1(LBS .seitnuocrevo1EQotroirp, erusopxE,erusopxeSBM :2erugiF c 41

htworgnoitanigiroegagtroM :3EQ)b( htworggnicnanfieregagtroM :1EQ)a( neewtebhtworggnidnelegagtromniecnereffidehtrofslavretniecnedfinocdnastneicfifeocnoissergerstolperugfisihT .setadtnevednuorasetamitsetneicfifeoC :3erugiF )j( ehtni0otdezilamronerastneicfifeocehT .)6(.qemorf β .e.i,retrauqhcaeninoitubirtsiderusopxeSBMytnuoc-ssorcehtfoelicretrewolehtdnareppuehtniseitnuoc k ecnedfinoc%59etonedsenillacitreV .htworgnoitanigiroegagtromrohtworggnicnanfieregagtromrehtiesielbairavemoctuoehT .tneveEQevitcepserehterofebretrauq .retrauqdnaytnuocybderetsulcsrorredradnatsnodesabslavretni 42

Figure4: Employmentgrowthbetween2008and2015. Annualemploymentgrowthrate,∆Emp ,averagedseparatelyover ct countiesintheupperandinthelowertercilesofthecross-countyMBSexposuredistributionpriortoQE3. Thetwovertical linesindicatethequartersoftheimplementationofQE1andQE3,respectively. 43

3EQ)b( 1EQ)a( neewtebhtworgtnemyolpmeniecnereffidehtrofslavretniecnedfinocdnastneicfifeocnoissergerstolperugfisihT .setadtnevednuorasetamitsetneicfifeoC :5erugiF )j( ehtni0otdezilamronerastneicfifeocehT .)6(.qemorf β .e.i,retrauqhcaeninoitubirtsiderusopxeSBMytnuoc-ssorcehtfoelicretrewolehtdnareppuehtniseitnuoc k .retrauqdnaytnuocybderetsulcsrorredradnatsnodesabslavretniecnedfinoc%59etonedsenillacitreV .tneveEQevitcepserehterofebretrauq 44

3EQ)b( 1EQ)a( neewtebhtworgnoitartsigerracniecnereffidehtrofslavretniecnedfinocdnastneicfifeocnoissergerstolperugfisihT .setadtnevednuorasetamitsetneicfifeoC :6erugiF )j( ehtni0otdezilamronerastneicfifeocehT .)6(.qemorf β .e.i,retrauqhcaeninoitubirtsiderusopxeSBMytnuoc-ssorcehtfoelicretrewolehtdnareppuehtniseitnuoc k .retrauqdnaytnuocybderetsulcsrorredradnatsnodesabslavretniecnedfinoc%59etonedsenillacitreV .tneveEQevitcepserehterofebretrauq 45

Tables Table1: TheannouncementeffectofQE1(25November2008)andQE3(13September2012) Dependent variable: Daily return Daily risk-adjusted Return (1) (2) (3) (4) (5) (6) (7) (8) (cid:0) MBS (cid:1)(1) 0.069** 0.081** 0.055* 0.078* TotAssets b (0.033) (0.041) (0.033) (0.042) (cid:0) MBS (cid:1)(3) 0.019** 0.032*** 0.014 0.025** TotAssets b (0.009) (0.012) (0.008) (0.011) log(Assets) 0.003 0.003*** 0.002 0.002** (0.003) (0.001) (0.003) (0.001) Equity/Assets 0.164 0.081*** 0.080 0.076** (0.139) (0.030) (0.139) (0.030) Profitability 0.071 -0.127 0.111 -0.105 (0.125) (0.082) (0.130) (0.085) Real Estate Ratio -0.017 -0.007 -0.003 -0.007 (0.038) (0.009) (0.038) (0.008) R2 0.012 0.049 0.012 0.115 0.008 0.035 0.007 0.066 a N 325 325 315 315 325 325 315 315 Controls No Yes No Yes No Yes No Yes ThistablestudiestheimpactofMBSexposureonstockreturnsonQEannouncementdays. Thetable reportscoefficientsfromacross-sectionalregressionofbanks’dailystockreturnsonbankcharacteristics oneitherNovember25,2008(QE1)oronSeptember13,2012(QE3). Theoutcomevariableiseitherthe bankholdingcompany-levelrawdailyreturnortherisk-adjustedreturnthatcontrolsforthemarket return using a one-factor model, see eq. (1) and section 3.1.2 for details. Standard errors allow for heteroskedasticity. Starsindicatesignificanceatthe10%,5%and1%levels,respectively. 46

Table2: CountieswithhighandlowMBSexposure Low MBS exposure High MBS exposure Difference Mean Std Mean Std Diff t-stat Population (in thousand) 40.611 186.277 134.527 364.124 135.311 8.364 Median Income 41035.527 9181.616 44032.547 10519.178 4228.495 9.000 HousingMoodys 97.755 46.440 124.612 64.390 34.980 12.513 ∆ Population 0.001 0.010 0.005 0.010 0.004 8.058 2007Q4:2008Q4 ∆ Housing -0.042 0.049 -0.041 0.047 0.000 0.125 2007Q4:2008Q4 ∆ Income -0.027 0.053 -0.030 0.051 -0.001 -0.450 2007Q4:2008Q4 ThistablecomparescountycharacteristicspriortoQE1forcountiesintheupperandinthelowertercileofthe cross-countryMBSexposuredistribution. MBSexposureismeasuredastheweightedaverageofbank-specificMBS ratiosaveragedoverthefourquarterspriortoQE1,withweightsgivenbybanks’depositsharespercountry,see eq.(2). 47

Table3: QEandcounty-levelmortgagelending Panel A: Mortgage refinancing ∆ Dependent variable: % Refinancing (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)HMDA 0.337** 0.327** t c (0.144) (0.144) QE (1) × Treat (1)HMDA 0.031*** 0.030*** t c (0.011) (0.011) QE (3) × Exposure (3)HMDA 0.048 0.060 t c (0.147) (0.147) QE (3) × Treat (3)HMDA -0.010 -0.009 t c (0.010) (0.010) R2 0.593 0.593 0.563 0.563 0.189 0.189 0.174 0.174 No. Counties 2854 2854 1899 1899 2855 2855 1922 1922 N 19065 19065 12637 12637 19505 19505 13112 13112 Panel B: Mortgage origination ∆ Dependent variable % Origination (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)HMDA -0.149 -0.159 t c (0.156) (0.157) QE (1) × Treat (1)HMDA -0.019 -0.020 t c (0.012) (0.012) QE (3) × Exposure (3)HMDA 0.219 0.228 t c (0.160) (0.161) QE (3) × Treat (3)HMDA 0.026** 0.027** t c (0.012) (0.012) R2 0.239 0.239 0.220 0.220 0.213 0.213 0.192 0.192 No. Counties 2804 2804 1865 1865 2807 2807 1896 1896 N 19047 19047 12597 12597 19109 19109 12882 12882 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes This table reports estimates of the effect of MBS exposure on mortgage lending. The outcome variable is county-levelquarterlygrowthofmortgagerefinancingvolumeinpanelA,andcounty-levelquarterlygrowth inmortgageoriginationvolumeinpanelB.MBSexposureismeasuredasbanks’MBS-to-assetratios,weighted by banks’ average mortgage origination volume in a county in the four quarters prior to each QE event. All specificationsincludecountyandtimefixedeffectsandcounty-levelcontrols,seeEquation(3)andEquation(4) fordetails. Column(1)-(2)and(5)-(6)reportcoefficientsforthecontinuoustreatmentvariableandcolumn(3)-(4) and(7)-(8)reportcoefficientsfortheindicatortreatmentvariable. Standarderrorsinparenthesesareclusteredby countyandtime. Starsindicatesignificanceatthe10%,5%and1%levels,respectively. 48

Table4: QEandcounty-levelsmallbusinesslending Panel A: Exposure measured with bank-county small business lending activity ∆ Dependent variable % C&ILending (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)SBL 0.663 0.693 t c (0.740) (0.743) QE (1) × Treat (1)SBL 0.071 0.061 t c (0.059) (0.059) QE (3) × Exposure (3)SBL 0.923*** 0.923*** t c (0.325) (0.325) QE (3) × Treat (3)SBL 0.071** 0.070** t c (0.030) (0.030) R2 0.292 0.293 0.278 0.280 0.326 0.326 0.329 0.330 No. Counties 2794 2794 1829 1829 2729 2729 1764 1764 N 5588 5588 3658 3658 5458 5458 3528 3528 Panel B: Exposure measured with bank-county deposit activity (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)DEP 0.282 0.242 t c (0.452) (0.454) QE (1) × Treat (1)DEP 0.058 0.051 t c (0.052) (0.053) QE (3) × Exposure (3)DEP 0.258 0.256 t c (0.251) (0.252) QE (3) × Treat (3)DEP 0.064** 0.066** t c (0.031) (0.031) R2 0.298 0.300 0.312 0.313 0.325 0.325 0.317 0.317 No. Counties 2788 2788 1828 1828 2735 2735 1796 1796 N 5576 5576 3656 3656 5470 5470 3592 3592 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes ThistablereportsestimatesoftheeffectofMBSexposureonsmallbusinesslendingusingtheannualCRA data. The outcome variable is county-level annual growth of small business lending. MBS exposure is measuredasbanks’MBS-to-assetratios,weightedbybanks’averagesmallbusinesslendingvolumeina county(panelA)andweightedbybanks’averagedepositvolumeinacounty(panelB)priortoeachQE event. Allspecificationsincludecountyandtimefixedeffectsandcounty-levelcontrols,seeEquation(3) andEquation(4)fordetails. Dataisrestrictedtotwoannualdatapoints. Column(1)-(2)and(5)-(6)report coefficientsforthecontinuoustreatmentvariableandcolumn(3)-(4)and(7)-(8)reportcoefficientsforthe indicatortreatmentvariable. Standarderrorsinparenthesesareclusteredbycountyandtime. Starsindicate significanceatthe10%,5%and1%levels,respectively. 49

Table5: QEandcounty-levelemployment Panel A: Exposure measured with bank-county small business lending activity ∆ Dependent variable Emp (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)SBL 0.017 0.026 t c (0.023) (0.024) QE (1) × Treat (1)SBL 0.001 0.002 t c (0.002) (0.002) QE (3) × Exposure (3)SBL 0.051** 0.050** t c (0.020) (0.020) QE (3) × Treat (3)SBL 0.004** 0.004** t c (0.002) (0.002) R2 0.624 0.626 0.613 0.615 0.483 0.483 0.479 0.479 No. Counties 2947 2947 1964 1964 2947 2947 1964 1964 N 20205 20205 13436 13436 20293 20293 13478 13478 Panel B: Exposure measured with bank-county deposit activity (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)DEP -0.036 -0.031 t c (0.023) (0.023) QE (1) × Treat (1)DEP -0.003* -0.002 t c (0.001) (0.001) QE (3) × Exposure (3)DEP 0.014 0.013 t c (0.018) (0.018) QE (3) × Treat (3)DEP 0.003*** 0.003*** t c (0.001) (0.001) R2 0.599 0.600 0.588 0.589 0.501 0.501 0.507 0.507 No. Counties 2946 2946 1965 1965 2952 2952 1968 1968 N 20597 20597 13730 13730 20636 20636 13741 13741 Continued on next page 50

Table5: (continued) Panel C: Exposure measured with bank-county mortgage lending activity (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)HMDA 0.012 0.025 t c (0.031) (0.031) QE (1) × Treat (1)HMDA 0.002 0.003 t c (0.002) (0.002) QE (3) × Exposure (3)HMDA 0.059** 0.060** t c (0.026) (0.026) QE (3) × Treat (3)HMDA 0.005*** 0.005*** t c (0.002) (0.002) R2 0.605 0.607 0.590 0.592 0.485 0.485 0.488 0.488 No. Counties 2954 2954 1972 1972 2945 2945 1966 1966 N 20204 20204 13457 13457 20284 20284 13505 13505 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes ThistablereportsestimatesoftheeffectofMBSexposureonemployment. Theoutcomevariableiscountylevelquarterlygrowthofemployment. MBSexposureismeasuredasbanks’MBS-to-assetratios,weightedby banks’averagesmallbusinesslendingvolumeinacounty(panelA),weightedbybanks’averagedeposit volumeinacounty(panelB)andweightedbybanks’averagemortgageoriginationvolume(panelC)prior to each QE event. All specifications include county and time fixed effects and county-level controls, see Equation(3)andEquation(4)fordetails. Column(1)-(2)and(5)-(6)reportcoefficientsforthecontinuous treatment variable and column (3)-(4) and (7)-(8) report coefficients for the indicator treatment variable. Standarderrorsinparenthesesareclusteredbycountyandquarter. Starsindicatesignificanceatthe10%,5% and1%levels,respectively. 51

Table6: QEandcounty-levelemployment: Industryeffects Panel A: Employment in tradable/other and nontradable goods sector ∆ ∆ ∆ ∆ Dependent variable EmpTradOther EmpNonTrad EmpTradOther EmpNonTrad (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)SBL -0.026 0.218** t c (0.072) (0.089) QE (1) × Treat (1)SBL -0.005 0.016** t c (0.006) (0.006) QE (3) × Exposure (3)SBL 0.095* -0.001 t c (0.056) (0.043) QE (3) × Treat (3)SBL 0.014** 0.005 t c (0.006) (0.005) R2 0.476 0.484 0.440 0.427 0.475 0.490 0.425 0.425 No. Counties 2959 1970 2959 1970 2951 1964 2951 1964 N 5918 3940 5918 3940 5902 3928 5902 3928 Panel B: Employment by financial dependence of industry ∆ ∆ ∆ ∆ Dependent variable EmpFin EmpNonFin EmpFin EmpNonFin (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)SBL 0.015 -0.098 t c (0.125) (0.106) QE (1) × Treat (3)SBL 0.003 -0.011 t c (0.010) (0.008) QE (3) × Exposure (3)SBL 0.116 0.066 t c (0.089) (0.076) QE (3) × Treat (3)SBL 0.013 0.006 t c (0.008) (0.008) R2 0.435 0.430 0.478 0.473 0.454 0.463 0.424 0.426 No. Counties 2959 1970 2959 1970 2960 1973 2960 1973 N 5918 3940 5918 3940 5920 3946 5920 3946 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes ThistablereportsestimatesoftheeffectofMBSexposureonemploymentbyindustrysplitsusingtheannual QCEW data. In panel A, the outcome variable is county-level annual employment growth by tradable or non-tradable industry. We use the definition of non-tradable industries of Mian and Sufi (2014). In panel B,theoutcomevariableiscounty-levelannualemploymentgrowthbyindustry’sfinancialdependence. We measurefinancialdependenceasinAlmeidaetal.(2010). MBSexposureismeasuredasbanks’MBS-to-asset ratios, weighted by banks’ average small business lending volume in a county prior to each QE event. All specificationsincludecountyandtimefixedeffectsandcounty-levelcontrols,interactedwithQEeventdummies, seeEquation(3)andEquation(4)fordetails. Dataisrestrictedtotwoannualdatapoints. Column(1)-(2)and (5)-(6)reportcoefficientsforthecontinuoustreatmentvariableandcolumn(3)-(4)and(7)-(8)reportcoefficients fortheindicatortreatmentvariable. Standarderrorsinparenthesesareclusteredbycountyandtime. Stars indicatesignificanceatthe10%,5%and1%levels,respectively. 52

Table7: QEandcounty-levelautopurchases Exposure measured with bank-county mortgage lending activity ∆ Dependent variable Auto (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)HMDA 0.505*** 0.471*** t c (0.105) (0.107) QE (1) × Treat (1)HMDA 0.036*** 0.035*** t c (0.009) (0.009) QE (3) × Exposure (3)HMDA 0.054 0.057 t c (0.087) (0.088) QE (3) × Treat (3)HMDA 0.014* 0.013* t c (0.007) (0.007) R2 0.551 0.551 0.529 0.530 0.206 0.206 0.200 0.200 No. Counties 2683 2683 1794 1794 2685 2685 1795 1795 N 18775 18775 12544 12544 18790 18790 12560 12560 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes ThistablereportsestimatesoftheeffectofMBSexposureonautopurchases. Theoutcomevariableiscounty-level quarterlygrowthofcarregistrations. MBSexposureismeasuredasbanks’MBS-to-assetratios,weightedbybanks’ averagemortgageoriginationvolumepriortoeachQEevent. Allspecificationsincludecountyandtimefixed effectsandcounty-levelcontrols,seeEquation(3)andEquation(4)fordetails. Column(1)-(2)and(5)-(6)report coefficientsforthecontinuoustreatmentvariableandcolumn(3)-(4)and(7)-(8)reportcoefficientsfortheindicator treatmentvariable. Standarderrorsinparenthesesareclusteredbycountyandtime. Starsindicatesignificanceat the10%,5%and1%levels,respectively. 53

Table8: QEandbank-firm-levellending Panel A: Continuous treatment ∆ Dependent variable C&I lending Sample Entire sample Not in Compustat Compustat (1) (2) (3) (4) (5) (6) QE(3)× (cid:0) MBS (cid:1)(3) 0.319*** 0.354*** 0.326** 0.424*** 0.151*** 0.130** TotAssets b (0.021) (0.052) (0.119) (0.074) (0.053) (0.054) R2 0.207 0.528 0.213 0.559 0.174 0.486 No Banks 25 25 25 25 25 25 No Firms 127305 9768 124729 8005 2638 1779 No Bank-Firm-Relationships 152803 33642 141636 23349 11231 10318 No obs 641048 145669 587074 96160 53696 49288 Panel B: Binary treatment ∆ Dependent variable C&I lending Sample Entire sample Not in Compustat Compustat (1) (2) (3) (4) (5) (6) QE(3)× Treat (3) 0.025*** 0.030*** 0.024*** 0.031*** 0.021*** 0.017*** b (0.002) (0.004) (0.007) (0.006) (0.005) (0.005) R2 0.207 0.528 0.213 0.558 0.175 0.486 No Banks 25 25 25 25 25 25 No Firms 127305 9768 124729 8005 2638 1779 No Bank-Firm-Relationships 152803 33642 141636 23349 11231 10318 No obs 641048 145669 587074 96160 53696 49288 Bank-Firm FE Yes Yes Yes Yes Yes Yes Time FE Yes No Yes No Yes No Bank Controls Yes Yes Yes Yes Yes Yes Interacted Bank Controls Yes Yes Yes Yes Yes Yes Firm-Time FE No Yes No Yes No Yes Thistablereportsestimatesoftheeffectofbanks’MBSexposureonbank-firm-levellending. Theoutcome variableisgrowthoftotallendingfromeachbanktoeachfirm. MBSexposureisabank’saverageMBS-toassetsratiointhefourquarterspriortoQE3. Allspecificationsincludebank-firmfixedeffectsandbank-level controls. Specificationsincludeeithertimefixedeffectsorfirm-timefixedeffects,seeeq.(7)fordetails. We usethecomputationallyefficientestimatoroflinearmodelswithmultiplehigh-dimensionalfixedeffects proposed by Correia (2017). Exposure is continuous in panel A and binary in panel B, i.e. in panel B, treatment is equal to 1 if a bank’s MBS exposure is above the median of the cross-bank MBS exposure distribution. Incolumns(1)and(2),thedataconsistofallbank-firmrelationshipsidentifiedintheY14data. Columns(3)and(4)reportestimatesforfirmswithintheY14datathatcannotbematchedtoCompustat, andcolumns(5)and(6)reportestimatesthatcanbematchedtoCompustat. Standarderrorsinparentheses areclusteredatthefirm-timelevelandstarsindicatesignificanceatthe10%,5%,and1%level,respectively. 54

Table9: QEandfirm-levellending,investmentandemployment Panel A: Firm-level borrowing ∆ Dependent variable Credit Sample Compustat Not in Compustat (1) (2) (3) (4) (5) (6) (7) (8) (3) Exposure 0.255 0.334* 0.190*** 0.203*** i (0.158) (0.201) (0.036) (0.040) (3) Treat 0.012* 0.013* 0.017*** 0.017*** i (0.006) (0.008) (0.002) (0.003) R2 .0014 .0028 .0045 .0097 .00048 .0018 .00067 .0021 No Firms 1852 1171 1493 919 40319 27576 33623 23067 Panel B: Firm-level investment ∆ Dependent variable Investment (in ppt) Sample Compustat Not in Compustat (1) (2) (3) (4) (5) (6) (7) (8) (3) Exposure 3.628** 1.641 13.461*** 12.599*** i (1.703) (1.910) (0.846) (0.832) (3) Treat 0.141** 0.075 0.431*** 0.217*** i (0.071) (0.082) (0.046) (0.047) R2 .0021 .0033 .014 .015 .0068 .003 .0096 .047 No Firms 1852 1171 1690 1050 40319 27576 33623 23067 Panel C: Firm-level employment ∆ ∆ Dependent variable Emp Emp 2012Q4:2013Q4 2012Q4:2014Q4 Sample Compustat (1) (2) (3) (4) (5) (6) (7) (8) (3) Exposure 0.458** 0.431** 0.640** 0.665** i (0.180) (0.200) (0.273) (0.290) (3) Treat 0.010* 0.011* 0.011 0.016* i (0.006) (0.006) (0.009) (0.010) R2 .014 .01 .02 .018 .029 .026 .064 .067 No Firms 1894 1256 1578 1045 1775 1173 1466 970 Firm Controls No No Yes Yes No No Yes Yes This table reports estimates of the effect of firms’ MBS exposure (via relationships to exposed banks) on lending, investmentandemployment. Theoutcomevariableisafirm’slendinggrowthbetween2012Q4and2013Q4inpanel A,thedifferenceincapitalexpenditurebetween2012Q4to2013Q4overassetsin2012Q4inpanelB,andemployment growthfrom2012Qto2013Q4orfromfrom2012Qto2014Q4inpanelC.Exposureismeasuredasbanks’MBS-to-asset ratiosaveragedoverfourquarterspriortoQE3,weightedbylendingvolumeatthebank-firmlevelaveragedoverfour quarters prior to QE3, see eq. (9) for details. Panels A and B report estimates for firms in the Y14 data that can be matchedtoCompustatincolumns(1)to(4),andestimatesforfirmsthatcannotbematchedtoCompustatincolumns(5) to(8). PanelCusesonlyfirmsthatcanbematchedtoCompustatbecauseemploymentisnotavailableintheY14data. Firmcontrolsincludethefirms’sleverage,z-Score,currentincome,andsizemeasuredaslogoftotalassets,seeeq.(11) fordetails. Robuststandarderrorsinparentheses. Starsindicatesignificanceatthe10%,5%,and1%level,respectively. 55

Table10: QEandcounty-banklevelmortgagelending Panel A: Mortgage refinancing ∆ ∆ Dependent variable Refinancing (conforming) Refinancing (non-conforming) (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × (cid:0) MBS (cid:1)(1) 0.740*** 0.182** 0.863*** 0.659** t Assets b (0.060) (0.071) (0.243) (0.304) QE (3) × (cid:0) MBS (cid:1)(3) 0.091 0.087 0.099 0.141 t Assets b (0.087) (0.110) (0.186) (0.199) R2 .14 .26 .037 .14 .05 .28 .016 .18 N 173910 172138 217239 212450 25565 21781 46110 42604 No Banks 3078 3075 2757 2738 1923 1638 1776 1595 No Counties 3078 2910 3070 2910 1852 1009 1932 1172 Panel B: Mortgage origination ∆ ∆ Dependent variable Origination (conforming) Origination (non-conforming) (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × (cid:0) MBS (cid:1)(1) 0.108 0.104 0.561 0.212 t Assets b (0.126) (0.140) (0.485) (0.553) QE (3) × (cid:0) MBS (cid:1)(3) 0.100 0.113 0.486*** 0.627*** t Assets b (0.068) (0.077) (0.169) (0.199) R2 .055 .18 .034 .16 .18 .36 .042 .23 N 190282 187973 224625 216700 26843 23828 38568 35318 No Banks 3330 3330 3004 3000 1698 1502 1504 1315 No Counties 3083 2909 3070 2867 1501 881 1515 872 Bank-County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes No Yes No Yes No Yes No Bank Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Bank Controls Yes Yes Yes Yes Yes Yes Yes Yes County-Time FE No Yes No Yes No Yes No Yes This table reports estimates of the effect of MBS exposure on county-bank level mortgage lending. The outcome variable is county-bank level mortgage refinancing growth for conforming and non-conforming mortgages in panel A,andcounty-banklevelmortgageoriginationgrowthforconformingandnon-conformingmortgagesinpanelB.All specificationsincludebank-countyfixedeffectsandbank-levelcontrols. Specificationsincludeeithertimefixedeffectsor county-timefixedeffects,seeeq.(12)fordetails. Weusethecomputationallyefficientestimatoroflinearmodelswith multiplehigh-dimensionalfixedeffectsproposedbyCorreia(2017). Standarderrorsaretwo-wayclusteredatthecounty andtimelevels. Starsindicatesignificanceatthe10%,5%and1%levels,respectively. 56

Table11: Robustness: QE2,Taperingandmortgagelending Panel A: Mortgage refinancing ∆ Dependent variable: Refinancing (1) (2) (3) (4) (5) (6) (7) (8) QE (2) × Exposure (2)HMDA 0.045 0.082 t c (0.156) (0.155) QE (2) × Treat (2)HMDA -0.006 -0.000 t c (0.011) (0.011) Taper × Exposure (3)HMDA -0.228 -0.229 t c (0.174) (0.175) Taper × Treat (3)HMDA -0.029*** -0.029*** t c (0.011) (0.011) R2 0.382 0.383 0.348 0.349 0.219 0.219 0.198 0.198 No. Counties 2869 2869 1907 1907 2860 2860 1906 1906 N 19762 19762 13081 13081 19679 19679 13054 13054 Panel B: Mortgage origination ∆ Dependent variable Origination (1) (2) (3) (4) (5) (6) (7) (8) QE (2) × Exposure (2)HMDA 0.068 0.045 t c (0.174) (0.184) QE (2) × Treat (2)HMDA -0.014 -0.018 t c (0.013) (0.014) Taper × Exposure (3)HMDA -0.243 -0.317 t c (0.209) (0.211) Taper × Treat (3)HMDA -0.020 -0.030* t c (0.016) (0.017) R2 0.257 0.257 0.246 0.246 0.333 0.334 0.300 0.301 No. Counties 2847 2847 1893 1893 2856 2856 1927 1927 N 19311 19311 12794 12794 19502 19502 13138 13138 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes ThistablereportsestimatesofMBSexposureonmortgagelending. Theoutcomevariableiscounty-levelquarterly growthofmortgagerefinancingvolumeinpanelA,andcounty-levelquarterlygrowthinmortgageorigination volumeinpanelB.MBSexposureismeasuredasbanks’MBS-to-assetratios,weightedbybanks’averagemortgage originationvolumeinacountyinthefourquarterspriortoeitherQE2orTapering. Allspecificationsinclude countyandtimefixedeffectsandcounty-levelcontrols,seeEquation(3)andEquation(4)fordetails. Column (1)-(2)and(5)-(6)reportcoefficientsforthecontinuoustreatmentvariableandcolumn(3)-(4)and(7)-(8)report coefficientsfortheindicatortreatmentvariable. Standarderrorsinparenthesesareclusteredbycountyandtime. Starsindicatesignificanceatthe10%,5%and1%levels,respectively. 57

Table12: Robustness: QE2,Taperingandemployment Panel A: Exposure measured with bank-county small business lending activity ∆ Dependent variable Emp (1) (2) (3) (4) (5) (6) (7) (8) QE (2) × Exposure (2)SBL 0.006 0.000 t c (0.012) (0.012) QE (2) × Treat (2)SBL 0.000 -0.000 t c (0.001) (0.001) Taper × Exposure (3)SBL -0.028** -0.023* t c (0.013) (0.013) Taper × Treat (3)SBL -0.002** -0.002* t c (0.001) (0.001) R2 0.547 0.549 0.527 0.529 0.527 0.527 0.527 0.527 No. Counties 2947 2947 1967 1967 2950 2950 1967 1967 N 20278 20278 13494 13494 20335 20335 13549 13549 Panel B: Exposure measured with bank-county deposit activity (1) (2) (3) (4) (5) (6) (7) (8) QE (2) × Exposure (2)DEP 0.012 0.013 t c (0.009) (0.009) QE (2) × Treat (2)DEP 0.001 0.001 t c (0.001) (0.001) Taper × Exposure (3)DEP -0.031 -0.032 t c (0.026) (0.026) Taper × Treat (3)DEP -0.005 -0.005 t c (0.004) (0.004) R2 0.521 0.521 0.512 0.512 0.370 0.370 0.345 0.345 No. Counties 2929 2929 1949 1949 2946 2946 1966 1966 N 20225 20225 13422 13422 20608 20608 13743 13743 Continued on next page 58

Table12: (continued) Panel C: Exposure measured with bank-county mortgage lending activity (1) (2) (3) (4) (5) (6) (7) (8) QE (2) × Exposure (2)HMDA 0.043 0.010 t c (0.035) (0.035) QE (2) × Treat (2)HMDA 0.004* 0.002 t c (0.002) (0.003) Taper × Exposure (3)HMDA -0.023 -0.020 t c (0.018) (0.018) Taper × Treat (3)HMDA -0.003 -0.003 t c (0.002) (0.003) R2 0.539 0.540 0.529 0.530 0.516 0.517 0.478 0.478 No. Counties 2900 2900 1974 1974 2945 2945 1965 1965 N 20279 20279 13797 13797 20493 20493 13661 13661 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes ThistablereportsestimatesofMBSexposureonemployment. Theoutcomevariableiscounty-levelquarterly growthofemployment. MBSexposureismeasuredasbanks’MBS-to-assetratios,weightedbybanks’average smallbusinesslendingvolumeinacounty(panelA),weightedbybanks’averagedepositvolumeinacounty (panelB)andweightedbybanks’averagemortgageoriginationvolume(panelC)priortoeachQEevent. All specificationsincludecountyandtimefixedeffectsandcounty-levelcontrols,seeEquation(3)andEquation(4) fordetails. Column(1)-(2)and(5)-(6)reportcoefficientsforthecontinuoustreatmentvariableandcolumn (3)-(4)and(7)-(8)reportcoefficientsfortheindicatortreatmentvariable. Standarderrorsinparenthesesare clusteredbycountyandtime. Starsindicatesignificanceatthe10%,5%and1%levels,respectively. 59

Table13: Robustness: Taperingandfirm-levellending Panel A: Continous treatment ∆ Dependent variable C&I lending Sample Entire sample Not in Compustat Compustat (1) (2) (3) (4) (5) (6) Taper × (cid:0) MBS (cid:1)(3) -0.051*** -0.120*** -0.050 -0.141*** -0.071*** -0.070*** TotAssets b (0.009) (0.025) (0.039) (0.036) (0.025) (0.026) R2 0.148 0.473 0.153 0.512 0.122 0.419 a No Banks 26 26 26 26 26 26 No Firms 145122 11396 142480 9532 2718 1890 No Bank-Firm-Relationships 175129 39483 163053 28231 12188 11318 No obs 997172 226112 921785 156069 75102 69785 Panel B: Binary treatment ∆ Dependent variable C&I lending Sample Entire sample Not in Compustat Compustat (1) (2) (3) (4) (5) (6) Taper × Treat (3) -0.007*** -0.013*** -0.008** -0.016*** -0.011*** -0.008*** b (0.001) (0.002) (0.004) (0.003) (0.003) (0.003) R2 0.147 0.473 0.153 0.512 0.127 0.419 a No Banks 25 25 25 25 25 25 No Firms 144847 11360 142220 9506 2704 1880 No Bank-Firm-Relationships 174682 39280 162683 28099 12112 11247 No obs 996278 225698 921045 155800 74950 69640 Bank-Firm FE Yes Yes Yes Yes Yes Yes Time FE Yes No Yes No Yes No Bank Controls Yes Yes Yes Yes Yes Yes Interacted Bank Controls Yes Yes Yes Yes Yes Yes Firm-Time FE No Yes No Yes No Yes Thistablereportsestimatesofbanks’MBSexposureonbank-firm-levellending. Theoutcomevariableis growthoftotallendingfromeachbanktoeachfirm. MBSexposureisabank’saverageMBS-to-assetsratio inthefourquarterspriortoQE3. Allspecificationsincludebank-firmfixedeffectsandbank-levelcontrols. Specificationsincludeeithertimefixedeffectsorfirm-timefixedeffects,seeeq.(7)fordetails. Weusethe computationallyefficientestimatoroflinearmodelswithmultiplehigh-dimensionalfixedeffectsproposed byCorreia(2017). ExposureiscontinuousinpanelAandbinaryinpanelB,i.e. inpanelB,treatmentis equalto1ifabank’sMBSexposureisabovethemedianofthecross-bankMBSexposuredistribution. In columns(1)and(2),thedataconsistofallbank-firmrelationshipsidentifiedintheY14data. Columns (3) and (4) report estimates for firms within the Y14 data that cannot be matched to Compustat, and columns(5)and(6)reportestimatesthatcanbematchedtoCompustat. Standarderrorsinparenthesesare clusteredatthefirm-timelevelandstarsindicatesignificanceatthe10%,5%,and1%level,respectively. 60

Description of appendices • Appendix A: Bank-level analysis of QE and mortgage and small business lending • Appendix B: Supplementary figures • Appendix C: Supplementary tables • Appendix D: Variable construction 61

Appendix A Bank-level analysis A.1 Empirical Strategy This section provides detailed bank-level results complementary to those of previous studies. We construct our bank-level exposure measure as a bank’s MBS scaled by total assets. Figure B.1 shows the distribution of the exposure measure averaged over the four quarters prior to QE1. More than a quarter of all commercial banks held no MBS at all, and the average MBS-to-asset ratio was around 12% in the upper quartile of the cross-sectional distribution across banks. [FIGURE B.1 ABOUT HERE] While banks with higher MBS shares tend to be larger and tend to operate with higher leverage than banks with relatively lower MBS shares, banks sorted by MBS shares are otherwise very similar in other observable characteristics. Table C.5 gives a sense of those differences, splitting the sample of all commercial banks in the United States by the median of the average MBS share in the 4 quarters prior to QE1. [TABLES C.5 and C.6 ABOUT HERE] In order to assess the effect of the Fed’s actions at the bank level, we employ a difference-indifferences(DiD)designwithacontinuoustreatmentvariable. Theunitofobservationisthecommercial bank and the main specification is given by: (cid:18) (cid:19)(j) ln(y ) = α+β MBS QE (j) +θX +γ +τ +(cid:101) (13) bt Total Assets t bt b t bt b Here, y is the amount of lending of bank b at time t. We use different categories of lending in our b,t regressions, including total lending, mortgage lending, and C&I lending. Moreover, we distinguish betweennewlyoriginatedandrefinancedmortgagesaswellasbetweensmallbusinessloansofdifferent sizes. We estimate the regression for each episode of quantitative easing, j = 1,2,3, with a time window of four quarters before and after the introduction of the respective program.11 QE (j) is an indicator t variable equal to 1 after the introduction of the j-th round of quantitative easing.12 We measure bank b’s 11Allresultsarerobusttochangingthetimewindowaswellastopoolingeventsinasingleregression. 12Giventhatthedataonthecommercialbanklevelisquarterly,wechoose2009Q1astheeventdateforQE1,and2010Q4 62

MBS-to-assets ratio, (cid:0) MBS (cid:1)(j) , as the average ratio over the 4 quarters prior to the j-th round of QE. TotalAssets b Additionally, our regression includes bank-fixed effects, γ , and time-fixed effects, τ, to control b t for fixed differences between banks and for differences over time that affect all banks. We include bank-specific time-varying controls X to control for remaining differences between banks. Controls bt included in the regression are listed in Table C.6. A.2 Results Table C.8 estimates Equation (13) for three different dependent variables: the natural logarithm of the overall amount of lending, of the amount of mortgage lending, and of the amount of C&I lending. [TABLE C.8 ABOUT HERE] We make three observations about the results depicted in Table C.8. First, for total lending volume, the interaction term between QE and a bank’s MBS-to-asset ratio prior to QE is positive and statistically significant for QE1 and QE3. This suggests that banks with large MBS holdings issued relatively more loans after QE1 and QE3. Second, focusing on real estate lending and C&I lending, we find that real estate related lending increased more at affected banks after QE1 and QE3, while C&I lending increased only after QE3. Third, there is no effect on lending during QE2. 13 Finally, the results are economically sizable. For instance, moving a bank from the 25th percentile of the MBS distribution to the 75th percentile implies an increase in C&I lending of about 1.3% after QE3. The overall pattern of our results confirms the findings of Darmouni and Rodnyansky (2017) of a stimulating effects of QE1 and QE3 on bank lending. However, our analysis goes further. Using data collected under the HMDA and aggregated to the bank-level, we provide further evidence on the effect on lending for residential housing. Importantly, the (confidential) version of HMDA data that we use, allows us to estimate regressions at the quarterly level (rather than the annual level which is the frequency in public HMDA data) to get a more precise and2012Q4astheeventdatesforQE2andQE3,respectively. 13ThisisconsistentwiththefocusofthedifferentprogramsonMBSpurchases(QE1,QE3)andU.S.Treasurypurchases (QE2). WeshouldnotexpectQE2toaffectbankswithhighandlowMBSsharesdifferentlyviaanarrowchannel. Notethat banksgenerallyholdlessTreasuriesthanMBS,andthepriceeffectonTreasuriesisgenerallyweaker. Notealsothatlowering Treasuryyields,asdonebyQE2,canyetaffectbanklending: mortgageratesmovewithTreasuryrates,andhenceQE2could spurrefinancinginthemortgagemarket. Indeed,thepositivecoefficientforoveralllendingandmortgagelending,Column (2)andColumn(5),pointtoaweakeffectonmortgage-relatedlending. However,theabsenceofasignificanteffectcanbe rationalizedbythehighrefinancingactivityafterQE1andtherelativelyweakereffectonlong-termratesofQE2comparedto QE1(seeGilchristetal.(2015)). 63

sense of the timing of effects. Table C.9 shows the estimates from Equation (13), but uses refinanced mortgages as well as newly issued mortgages as dependent variables. Columns (1) and (2) confirm the results for total real estate lending from the Call Report data in Table C.8. The additional specifications distinguish between new origination of mortgages and refinancing of existing mortgages. Even though aggregate lending related to housing increased during QE1 and QE3, the underlying type of lending is different. In particular, results in columns (3) and (4) reveal that the effect during QE1 is driven by increased refinancing activity of affected banks, consistent with the findings by Di Maggio et al. (2016), who show that QE1 spurred refinancing activity in the mortgage market. Columns (5) and (6) show that the effect of QE3 is driven by origination of mortgages for new home purchases and by refinancing of existing mortgages. [TABLE C.9 ABOUT HERE] Using additional data collected under the Community Reinvestment Act (CRA) on small business lending, we estimate the main specification using four related types of small business lending as dependent variables: loans to small business with face value of $ 0 to 100k, $ 100k to 250k, and $ 250k to 1m, as well as loans to businesses with an annual revenue of less than $ 1 million. Table C.10 shows results. ConsistentwiththeresultsinTableC.8,smallbusinesslendingdoesnotrespondinanycategory after QE1 and estimated coefficients fluctuate widely (columns (1), (3), (5), and (7) of Table C.10). For QE3, however, we find consistent effects across all categories (columns (2), (4), (6), and (8)): Coefficients are only significant in two of the categories but are positive and of similar magnitude across all of three of them. In particular, note that the coefficient for loans with a face value between 250k to 1m, the category with the highest aggregate volume, is significant. [TABLE C.10 ABOUT HERE] Abankthatholds12%ofitsassetsinMBSinsteadofhavingnoMBSholdings,increasedtheissuance of small business loans with a face value between 250k and 1m by about 4% after QE3. The magnitude is comparable to the magnitude of the effect on total C&I lending in Table C.8. 64

Appendix B Supplementary Figures FigureB.1: Distributionofbanks’averageMBSsharespriortoQE1 65

45.0=ρ,stisopeddnagnidnelssenisubllamS)c( 65.0=ρ,stisopeddnagnidnelegagtroM)b( 94.0=ρ,gnidnelssenisubllamsdnaegagtroM)a( egareva’sknabybdethgiew,soitartessa-ot-SBM’sknabsaderusaemsierusopxeSBMs’ytnuocA .3EQotroirpserusaemerusopxetnereffidfotolprettacS :2.BerugiF emulovnoitanigiroegagtromegareva’sknabybdethgiewdna)Blenap(ytnuocaniemulovtisopedegareva’sknabybdethgiew,ytnuocaniemulovgnidnelssenisubllams .tneveEQhcaeotroirp)Clenap( 66

FigureB.3: Employmentgrowthbetween2008and2014intheannualCBPdata. Annualemploymentgrowthrate,∆Emp , ct averagedseparatelyovercountiesintheupperandinthelowertercilesofthecross-countyMBSexposuredistributionprior toQE3. FigureB.4: Distributionoffirms’averagetotallendingin2012intheY-14databythenumberofbank-firmrelationships. Afirmandabankhavearelationshipifthebanklendstothefirminatleastthreequartersbetween2011Q3and2016Q2. 67

Appendix C Supplementary Tables TableC.1: BankswithhighandlowMBSshares: BankinY14data Low MBS Share High MBS Share Difference Mean Std Mean Std Diff t-stat log(assets) 19.428 1.418 19.042 0.948 -0.339 -0.696 Equity/TotalAssets 0.106 0.029 0.116 0.015 0.009 1.125 Tier 1 Ratio 0.129 0.024 0.123 0.021 -0.005 -0.694 RoA 0.009 0.012 0.005 0.009 -0.003 -1.165 Real Estate Ratio 0.175 0.182 0.279 0.145 0.096 1.452 Loans to Assets 0.469 0.262 0.560 0.214 0.083 0.862 C&I Loan Ratio 0.317 0.321 0.231 0.091 -0.086 -0.929 Non-performing Loans Ratio 0.010 0.010 0.012 0.009 0.001 0.505 ThistablereportsbankcharacteristicspriortoQE3averagedseparatelyforbanksaboveandbelowthe medianofthecross-bankMBS-to-assetsdistribution,forbanksintheY14dataonly. MBS-to-assetsare averagedoverthefourquarterspriortoQE3foreachbank. TableC.2: Descriptivestatisticsforbank-levelcontrolvariables: BanksinY14data Mean Std 10th Perc 25th Perc Median 75th Perc 90th Perc N MBS/TotalAssets 0.095 0.058 0.009 0.031 0.103 0.154 0.162 502 MBS/TotalSecurities 0.553 0.293 0.195 0.236 0.578 0.806 0.907 502 log(assets) 19.237 1.223 17.826 18.208 18.961 20.434 21.365 502 Equity/TotalAssets 0.111 0.024 0.080 0.096 0.111 0.123 0.135 502 Tier 1 Ratio 0.126 0.023 0.101 0.111 0.122 0.137 0.158 502 RoA 0.007 0.012 -0.001 0.003 0.006 0.011 0.017 502 Real Estate Ratio 0.227 0.173 0.002 0.029 0.225 0.369 0.468 502 Loans to Assets 0.514 0.243 0.075 0.340 0.630 0.699 0.746 502 C&I Loan Ratio 0.282 0.279 0.053 0.148 0.221 0.303 0.603 502 Non-performing Loans Ratio 0.011 0.010 0.001 0.004 0.009 0.016 0.024 502 This table reports means, standard deviations and various percentiles of variables used as controls in bank-level regressions. Thedatasetrunsfrom2011Q3to2016Q3andincludes25banks. 68

TableC.3: Descriptivestatisticsforfirm-levelvariables Mean Std 10th Perc 25th Perc Median 75th Perc 90th Perc N Sample Compustat Total Assets (in million) 15005.26 108343.67 192.69 611.90 2019.86 6728.73 21728.00 1831 Total Bank Credit 3823.87 7029.06 162.74 500.00 1819.39 4747.69 9231.01 1831 Capital Expenditures 414.85 1449.08 1.13 10.18 51.92 234.46 904.52 1831 No of Employees 16927.64 67383.27 238.00 999.00 3400.00 11800.00 35283.00 1831 Sample Not in Compustat Total Assets (in million) 1381.05 27442.46 1.57 4.94 16.05 71.55 421.42 40319 Total Bank Credit 211.18 765.16 12.07 18.59 40.20 129.14 402.39 40319 Capital Expenditures 30.29 1013.57 -0.17 0.00 0.01 1.17 11.94 40319 This table reports means, standard deviations and various percentiles of firm characteristics for firms in the Y14 data collection. PanelAincludesfirmsthatcanbematchedtoCompustat,andalsoincludesemploymentwhichisnotavailable intheY14data. PanelBincludesfirmsthatcannotbematchedtoCompustat. Dataareasreportedin2012Q4. 69

TableC.4: Robustness: Mainemploymenteffect ∆ Dependent variable Emp Population Exclude Northeast QE1 exposure MBS Exposure (1) (2) (3) (4) (5) (6) (7) (8) QE (3) × Exposure (3) 0.044** 0.045** t c (0.022) (0.022) QE (3) × Treat (3) 0.004** 0.004** t c (0.002) (0.002) QE (3) × Exposure (1) 0.037* t c (0.022) QE (3) × Treat (1) 0.004* t c (0.002) QE (3) × Exposure (3)MBS 0.012** t c (0.005) QE (3) × Treat (3)MBS 0.004* t c (0.002) R2 0.457 0.452 0.436 0.433 0.433 0.428 0.435 0.420 a No. Counties 1750 1161 2590 1734 2888 1922 2784 1862 N 12069 8079 17628 11756 19718 13076 18978 12675 County FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls Yes Yes Yes Yes Yes Yes Yes Yes ThistablereportsestimatesoftheeffectofMBSexposureonemploymentinanumberofrobustnesschecks. The outcomevariableiscounty-levelquarterlygrowthofemployment. MBSexposureismeasuredasbanks’MBS-to-asset ratios,weightedbybanks’averagedepositvolumeinacountypriortoQE3. Allspecificationsincludecountyand timefixedeffectsandcounty-levelcontrols,seeEquation(3)andEquation(4)fordetails. Incolumns(1)and(2),the sampleisrestrictedtocountieswithpopulationofmorethan15000andnomorethan250000. Incolumns(3)and(4), thesampleexcludesthefollowingstates: Connecticut,Delaware,Massachusetts,Maine,NewYork,NewJersey,New Hampshire,Pennsylvania,RhodeIsland,Vermont. Incolumns(5)and(6),exposureisbasedontheMBSholdingsof banksinthe4quarterspriortoQE1. Incolumns(7)and(8),exposureisbasedonMBSovertotalsecuritiesinsteadof MBSovertotalassets. Standarderrorsinparenthesesareclusteredbycountyandquarter. Starsindicatesignificance atthe10%,5%and1%levels,respectively. 70

TableC.5: BankswithhighandlowMBSshares Low MBS Share High MBS Share Difference Mean Std Mean Std Diff t-stat log(assets) 11.622 1.262 12.171 1.396 0.549 17.660 Equity/TotalAssets 0.133 0.119 0.115 0.075 -0.018 -8.263 Deposit Ratio 0.799 0.152 0.798 0.101 -0.001 -0.365 Trading Book Ratio 0.001 0.013 0.001 0.006 -0.000 -0.245 Profitability 0.006 0.183 0.003 0.015 -0.003 -1.598 Overhead Ratio 0.820 1.225 0.807 1.129 -0.012 -0.612 Net interest margin 0.020 0.685 0.023 0.011 0.003 0.451 Non-performing Loans Ratio 0.010 0.018 0.010 0.017 -0.001 -1.480 Delinquency Ratio 1.950 1.104 1.916 1.066 -0.034 -1.476 Real Estate Ratio 0.676 0.180 0.693 0.163 0.016 4.080 C&I Loan Ratio 0.154 0.114 0.155 0.099 0.002 0.698 Tier 1 Ratio 0.233 1.116 0.173 0.350 -0.061 -3.360 ThistablereportsbankcharacteristicspriortoQE1averagedseparatelyforbanksaboveandbelowthe medianofthecross-bankMBS-to-assetsdistribution. MBS-to-assetsareaveragedoverthefourquarters priortoQE1foreachbank. TableC.6: Descriptivestatisticsforbank-levelcontrolvariables Mean Std 10th Perc 25th Perc Median 75th Perc 90th Perc N Agency MBS/TotalAssets 0.08 0.10 0.00 0.00 0.05 0.12 0.21 166563 Agency MBS/TotalSecurities 0.36 0.33 0.00 0.02 0.31 0.60 0.84 166563 Treasuries/TotalAssets 0.01 0.03 0.00 0.00 0.00 0.00 0.01 166563 Treasuries/TotalSecurities 0.03 0.13 0.00 0.00 0.00 0.00 0.04 166563 log(Assets) 12.07 1.31 10.62 11.23 11.93 12.71 13.61 166563 Equity/TotalAssets 0.11 0.04 0.08 0.09 0.10 0.12 0.15 166563 Deposit Ratio 0.84 0.07 0.75 0.81 0.85 0.88 0.90 166563 Trading Book Ratio 0.00 0.01 0.00 0.00 0.00 0.00 0.00 166563 Profitability 0.00 0.01 -0.00 0.00 0.00 0.01 0.01 166563 Overhead Ratio 0.77 2.59 0.53 0.61 0.71 0.82 0.97 166560 Net interest margin 0.02 0.01 0.01 0.01 0.02 0.03 0.04 166563 Non-performing Loans Ratio 0.01 0.02 0.00 0.00 0.01 0.02 0.03 166563 Delinquency Ratio 2.11 1.73 0.31 0.85 1.71 2.93 4.42 166563 Real Estate Ratio 0.69 0.19 0.43 0.59 0.73 0.83 0.89 166563 C&I Loan Ratio 0.15 0.10 0.05 0.08 0.12 0.19 0.26 166563 Tier 1 Ratio 0.16 0.24 0.10 0.12 0.14 0.18 0.24 166563 This table reports means, standard deviations and various percentiles of variables used as controls in bank-level regressions. Thedatasetrunsfrom2007Q1to2015Q2,andincludesuptoaround7000banks. 71

TableC.7: QEandMSA-levelemployment ∆ Dependent variable % Emp (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × Exposure (1)SBL -0.003 0.001 t c (0.014) (0.014) QE (1) × Treat (1)SBL -0.000 0.000 t c (0.001) (0.001) QE (3) × Exposure (3)SBL 0.044*** 0.048*** t c (0.013) (0.013) QE (3) × Treat (3)SBL 0.004*** 0.004*** t c (0.001) (0.001) R2 0.723 0.724 0.704 0.704 0.541 0.541 0.560 0.560 No. Counties 950 950 633 633 928 928 617 617 N 6383 6383 4236 4236 6419 6419 4266 4266 MSA FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes MSA Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls No Yes No Yes No Yes No Yes ThistablereportsestimatesoftheeffectofMBSexposureonemployment. TheoutcomevariableisMSA-levelquarterly growthofemployment. MBSexposureismeasuredasbanks’MBS-to-assetratios,weightedbybanks’averagesmall businesslendingvolumeinaMSApriortoeachQEevent. AllspecificationsincludeMSAandtimefixedeffectsand MSA-levelcontrols,seeEquation(3)andEquation(4)fordetails. Column(1)-(2)and(5)-(6)reportcoefficientsforthe continuous treatment variable and column (3)-(4) and (7)-(8) report coefficients for the indicator treatment variable. StandarderrorsinparenthesesareclusteredbyMSAandquarter. Starsindicatesignificanceatthe10%,5%and1% levels,respectively. 72

gnidnellevel-knabdnaEQ :8.CelbaT )I&C(gol )gnidnel ER(gol )gnidneL(gol elbairav tnednepeD )9( )8( )7( )6( )5( )4( )3( )2( )1( 500.0 ***201.0 ***370.0 )1((cid:1) SBM (cid:0) × )1( EQ b stessAtoT t )530.0( )330.0( )620.0( 620.0 510.0 531.0 )2((cid:1) SBM (cid:0) × )2( EQ b stessAtoT t )601.0( )650.0( )711.0( **870.0 ***760.0 ***370.0 )3((cid:1) SBM (cid:0) × )3( EQ b stessAtoT t )530.0( )620.0( )420.0( 800.0 900.0 640.0 130.0 440.0 471.0 370.0 350.0 891.0 2R a 4895 4146 0196 5995 3246 9196 4106 8446 7496 sknaB .oN 78725 66365 02506 30925 97465 66606 18035 81765 13906 N seY seY seY seY seY seY seY seY seY EF knaB seY seY seY seY seY seY seY seY seY EF emiT seY seY seY seY seY seY seY seY seY slortnoC seY seY seY seY seY seY seY seY seY slortnoC detcaretnI sielbairavemoctuoehT .atadtroperllaCgnisugnidnelnosoitartessa-ot-SBM’sknabfotceffeehtfosetamitsestroperelbatsihT ,)6(ot)4(snmulocnignidneletatselaerylretrauqfogoleht,)3(ot)1(snmulocnignidnellatotylretrauqs’knabafogoleht llA .tneveEQhcaeotroirpsretrauqruofehtrevodegarevasistessa-ot-SBM .)9(ot)7(snmulocnignidnelI&Cfogolehtdna sesehtnerapnisrorredradnatS .sliatedrof)31(.qeees,slortnoclevel-knabdnastceffedexfiemitdnaknabedulcnisnoitacfiiceps .ylevitcepser,slevel%1dna%5,%01ehttaecnacfiingisetacidnisratS .slevelretrauqdnaknabehttaderetsulcyaw-owtera 73

TableC.9: QEandbank-levelmortgagelending Dependent variable log(Total) log(Refinance) log(Origination) (1) (2) (3) (4) (5) (6) QE (1) × (cid:0) MBS (cid:1)(1) 0.452*** 0.712*** 0.212 t TotAssets b (0.161) (0.206) (0.148) QE (3) × (cid:0) MBS (cid:1)(3) 0.229** 0.220* 0.214* t TotAssets b (0.113) (0.129) (0.127) R2 0.057 0.032 0.119 0.064 0.065 0.108 a No. Banks 4178 3581 4039 3495 4102 3493 N 24958 25449 22811 23206 23381 23780 Bank FE Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Interacted Controls Yes Yes Yes Yes Yes Yes Thistablereportsestimatesoftheeffectofbanks’MBS-to-assetratiosonmortgagelending usingHomeMortgageDisclosureAct(HMDA)data. Theoutcomevariableisthelogofa bank’squarterlytotalmortgagelendingincolumns(1)and(2),thelogofmortgagerefinancing volumeincolumns(3)and(4),andthelogofmortgageoriginationvolumeincolumns(5)to (6). MBS-to-assetsisaveragedoverthefourquarterspriortoeachQEevent. Allspecifications includebankandtimefixedeffectsandbank-levelcontrols,seeeq.(13)fordetails. Standard errors in parentheses aretwo-wayclustered at the bank andquarter levels. Stars indicate significanceatthe10%,5%and1%levels,respectively. 74

TableC.10: QEandbank-levelsmallbusinesslending Dependent variable log(C&I lending) Loan size [0, 100k] [100k, 250k] [250k, 1m] Rev<1mil (1) (2) (3) (4) (5) (6) (7) (8) QE (1) × (cid:0) MBS (cid:1)(1) -0.126 0.012 0.107 0.248* t TotAssets b (0.124) (0.120) (0.122) (0.150) QE (3) × (cid:0) MBS (cid:1)(3) 0.137 0.183** 0.395*** 0.299** t TotAssets b (0.114) (0.089) (0.125) (0.134) R2 0.316 0.072 0.322 0.136 0.319 0.150 0.350 0.119 a No. Banks 743 652 742 653 744 654 743 650 N 1826 1674 1821 1674 1818 1681 1826 1665 Bank FE Yes Yes Yes Yes Yes Yes Yes Yes Time FE Yes Yes Yes Yes Yes Yes Yes Yes Controls Yes Yes Yes Yes Yes Yes Yes Yes Interacted Controls Yes Yes Yes Yes Yes Yes Yes Yes Thistablereportsestimatesoftheeffectofbanks’MBS-to-assetratiosonmortgagelendingusingCommunity Reinvestment Act (CRA) data. The outcome variable is the log of a bank’s total C&I loans with loan volume lessthan$100kincolumns(1)and(2),thelogoftotalC&Iloanswithloanvolumebetween$100kand$250kin columns(3)and(4),thelogoftotalC&Iloanswithloanvolumebetween$250kand$1mincolumns(5)and(6), andthelogoftotalC&Iloanstobusinesseswithrevenuelessthan$1mincolumns(7)and(8). MBS-to-assetsis averagedoverthefourquarterspriortoeachQEevent. Allspecificationsincludebankandtimefixedeffectsand bank-levelcontrols,seeeq.(13)fordetails. Standarderrorsinparenthesesaretwo-wayclusteredatthebankand quarterlevels. Starsindicatesignificanceatthe10%,5%and1%levels,respectively. 75

TableC.11: Firm-levelregression: InvestmentandemploymenteffectsforfirmsinY14andforfirms matchedtoCompustat Panel A: Continous treatment ∆ Dependent variable C&I lending Sample Entire sample Not in SNC SNC sample (1) (2) (3) (4) (5) (6) QE(3)× (cid:0) MBS (cid:1)(3) 0.344*** 0.402*** 0.301** 0.501*** 0.206*** 0.226*** TotAssets b (0.023) (0.054) (0.120) (0.145) (0.041) (0.046) R2 0.206 0.524 0.217 0.560 0.189 0.529 No Banks 25 25 25 25 25 25 No Firms 127305 9768 117713 4101 9592 5667 No Bank-Firm-Relationships 152803 33642 124696 9726 28107 23916 No obs 641048 145669 511503 37675 129545 107994 Panel B: Binary treatment ∆ Dependent variable C&I lending Sample Entire sample Not in SNC SNC sample (1) (2) (3) (4) (5) (6) QE(3)× Treat (3) 0.022*** 0.026*** 0.021*** 0.030*** 0.024*** 0.022*** b (0.001) (0.004) (0.007) (0.011) (0.004) (0.004) R2 0.208 0.533 0.217 0.560 0.189 0.529 No Banks 25 25 25 25 25 25 No Firms 127305 9768 117713 4101 9592 5667 No Bank-Firm-Relationships 152803 33642 124696 9726 28107 23916 No obs 641048 145669 511503 37675 129545 107994 Bank-firmlevelanalysiswithfirm-timefixedeffect. Observationarerestrictedto7quarters,3quarters beforeand3quartersafterQE3. Standarderrorsinparenthesesareclusteredatthefirm-timelevel andstarsindicatesignificanceatthe10%,5%,and1%level,respectively. Weusethecomputationally efficientestimatoroflinearmodelswithmultiplelevelsoffixedeffectsproposedbyCorreia(2017). 76

Appendix D Variable construction Bank variables are from the merger-adjusted Consolidated Reports of Condition and Income (FFIEC031 and FFIEC041). Banks are indexed by b, time is indexed by t (quarters). • Total securities (available for sale, fair value) : RCFD1773 b,t • Total securities (held to maturity, amortized cost) : RCFD1754 b,t • MBS (held to maturity, amortized cost) : The sum of all item in Schedule RC-B, item 4, column A, b,t excluding items a.(3), b.(3), and c(1b) and c(2b). • MBS (available for sale, fair value) : The sum of all item in Schedule RC-B, item 4, column C, b,t excluding items a.(3), b.(3), and c(1b) and c(2b). • Bank size : the log of total assets: Log(RCFD2170) b,t • ReturnonAssets : Income(loss)beforediscontinuedoperationsoverassets: RIAD4300/RCFD2170 b,t • Overhead ratio : The ratio of Noninterest expense (RIAD4093) divided by revenue. Revenue is b,t the sum of net interest income (RIAD4074) and noninterest income (RIAD4079)). • Net-interest margin : The ratio of Annualized net interest income (RIAD4074) divided by (30-day b,t average) interest-earning assets (RCFD3381+ RCFDB558 + RCFDB559 + RCFDB560 + RCFD3365 + RCFD3360 + RCFD3484 + RCFD3401) • (Delinquencies/Loan Loss Reserves) : The ratio of Delinquencies on all loans and leases (RC-N) b,t divided by reserves for loan losses (RCFD3123) • Ratio of non-performing loans : The sum of all loans that are past due 90 days or more and still b,t accruing (Schedule RC-N, Items 1 – 9 Column B) divided by total loans (RCFD2112) • Equity ratio : Total equity capital over assets: RCFDG105/RCFD2170 b,t • Realestateloanratio : Loanssecuredbyrealestateovertotalloansandleasesheldforinvestment b,t and held for sale: RCFD1410/RCFD2122 • Deposit ratio : Deposits in foreign and domestic offices over assets: b,t (RCON2200 + RCFN2200)/RCFD2170 77

• Loans Loans : Total loans and leases held for investment and held for sale over assets: b,t RCFD2122/RCFD2170 • C&I Lending /Loans : Commercial and industrial loans over total loans: b,t b,t (RCFD1763+RCFD1764)/RCFD2122 • Tier 1 capital ratio : RCFA7206 b,t 78