Liquidity from Two Lending Facilities

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Liquidity from Two Lending Facilities Sriya Anbil and Angela Vossmeyer 2017-117 Please cite this paper as: Anbil, Sriya, and Angela Vossmeyer (2017). “Liquidity from Two Lending Facilities,” Finance and Economics Discussion Series 2017-117. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2017.117. 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.

Liquidity from Two Lending Facilities Sriya Anbil Angela Vossmeyer Federal Reserve Board(cid:3) Claremont McKenna Collegey September 8, 2017 Abstract During (cid:12)nancial crises, the lender of last resort (LOLR) uses lending facilities to inject critical funding into the banking sector. The facilities need to be designed in such a way that banks are not reluctant to seek assistance due to stigma and that banks with liquidity concerns are attracted rather than those prone to risk-taking and moral hazard incentives. We use an unexpected disclosure that introduced stigma at one of two similar LOLRs during the Great Depression to evaluate whether banks used LOLR assistance to improve their liquidity needs using a novel trivariate model with recursive endogeneity. We (cid:12)nd evidence that banks that approached the facility with stigma were less liquid and reduced their position of safe assets in comparison with banks that approached the facility with no stigma. Thus, stigma forced the pool of LOLR borrowers to separate into different groups of banks that ex-post revealed their liquidity preferences. This (cid:12)nding sheds light on why and when banks approach their LOLR. 1 Introduction During the recent (cid:12)nancial crisis, the Federal Reserve acted as the lender of last resort (LOLR) to inject critical liquidity into the banking sector through its main emergency lending facility, the discountwindow(DW)(Armantieretal.,2015). TheDWwasdesignedtoalleviatefundingstresses in the banking sector, thereby lessening a \credit crunch" to the real economy. However, banks were reluctant to borrow from the Federal Reserve’s DW because if it somehow became known, it would lead market participants to infer weakness { the so-called stigma problem (Bernanke, 2009; Ennis and Weinberg, 2013). The implications of this problem are that banks contract their lending (cid:3)Division of Monetary Affairs, Board of Governors of the Federal Reserve System; email: sriya.l.anbil@frb.gov yRobertDaySchoolofEconomicsandFinance, ClaremontMcKennaCollege; email: angela.vossmeyer@cmc.edu. We thank Ayushi Bajaj, Jim Barth, Mark Carlson, Jim Clouse, Matt Jaremski, Meredith Rector, Gary Richardson, JonathanRose,ZeynepSenyuz,HeatherTookes,KimRosasTyson,PatrickVanHorn,MarcWeidenmier,andseminar participants at the Federal Reserve Board, Western Economic Association, and Federal Reserve Bank of St. Louis for valuable suggestions. We also thank Grant Carney, Maher Latif, and Anna Balderston for excellent research assistance. All remaining errors are our own. The views expressed in this paper are solely the responsibility of the author and should not be interpreted as re(cid:13)ecting the views of the Federal Reserve Board, its staff, or any other person associated with the Federal Reserve System. 1

and experience deposit withdrawals after their LOLR loans are revealed to the public (Anbil, 2017; Vossmeyer, 2017). An important consideration when designing LOLR lending facilities is to lend to banks that are solvent but illiquid rather than to banks that increase their risk-taking because of the presence of anLOLR(Madigan,2009). TraditionalLOLRtheorysuggeststhatbanksshouldborrowfromtheir LOLR to stop runs, and that the monetary authority should lend unsparingly at a penalty rate (Bagehot, 1873). Then, the LOLR could ease funding constraints during a (cid:12)nancial crisis, limiting its transmission to the real economy. However, the very presence of the LOLR may create moral hazard incentives for banks where they increase their risk-taking (Carpinelli and Crosignani, 2017). In this scenario, LOLR lending facilities may increase overall systemic risk in the (cid:12)nancial system, rather than ease funding constraints. It is difficult to ex-ante determine each bank’s liquidity preferences when they pool together at one LOLR facility. In this paper, we examine which banks borrow from the LOLR and when they do so. We shed light on how to design lending facilities that achieve three objectives: (1) ease funding constraints, (2) are the least subject to a stigma problem, and (3) attract banks with liquidity concerns. However,becauseabank’sdecisiontoborrowfromtheLOLRisalsoafunctionofthesethreeobjectives, it is especially difficult to empirically disentangle each component of the bank’s decision. In typical identi(cid:12)cation settings, we cannot isolate a bank’s risk-taking from its funding needs because its decision to borrow from the LOLR is affected by both objectives. Motivated by this difficulty, we use an unexpected event from the Great Depression that allows us to cleanly analyze why banks approach their LOLR. The Great Depression was the worst (cid:12)nancial crisis in U.S. history during which LOLR lending was considerable (Bernanke, 1983). We use a unique setting with two LOLRs: the Reconstruction Finance Corporation (RFC) and the Federal Reserve’s Discount Window (DW). Beginning on February 2, 1932, each LOLR had a nearly identical lending facility that provided loans con(cid:12)dentially to banks. However, on August 22, 1932, the Clerk of the House of Representatives took it upon himself to publish several partial lists of banks that had secretly borrowed from the RFC, which unexpectedly introduced a stigma problem at the RFC (Anbil, 2017; Vossmeyer, 2017). Using a unique hand-collected data set of balance sheet, DW, and RFC loan information for banks in the Federal Reserve Sixth District from January 1931 to September 1933, we develop a novel 2

trivariate model with recursive endogeneity to estimate the effect of the clerk’s publication on banks’ choice of LOLR and their subsequent liquidity seeking behavior. We (cid:12)nd that the pool of LOLR borrowers ex-post separated into speci(cid:12)c groups after the publication of the list: banks that continued borrowing from the RFC (\RFC banks"), banks that switched away from the RFC (\switched banks"), and banks that only borrowed from the DW (\DW banks"). This separation of banks revealed information about their liquidity preferences to market participants after the publication of the list. Prior to the publication of the list, this information was unavailable since all LOLR borrowers pooled together. To the best of our knowledge, the combination of DW and RFC loan information makes our paper the (cid:12)rst to study the entirety of LOLR lending to (cid:12)nancial intermediaries during a crisis. We use a trivariate model to jointly model a bank’s choice of LOLR and its performance. The publication of the list adds a useful dimension to our model by allowing us to endogenously accommodate the timing of when banks approached a facility. After the revelation, banks’ choice of LOLR revealed information about their liquidity preferences to market participants. Furthermore, the joint model allows us to estimate the covariance (and implied correlation) between the determinants of each LOLR choice and bank liquidity, giving insights as to how the unobservables are related. This is important because we also examine the performance of the three groups that banks separated into based on their LOLR choice using a simple reduced-form approach. From our trivariate model, we have three signi(cid:12)cant (cid:12)ndings. First, RFC banks decreased their position of safe assets (as a ratio of total assets) by 9 percentage points, while switched and DW banks maintained their position of government securities. The unexpected introduction of stigma attheRFCex-postrevealedthatthosebanksthatcontinuedtoborrowfromtheRFC(RFCbanks) were less concerned with liquid assets on their balance sheets than banks that switched or stayed at the DW. Second, our model allows us to take advantage of a fourth category of banks { banks that did not apply to an LOLR (\non-applicant banks") { because comparing banks that approached an LOLR to banks that did not approach an LOLR introduces a sample selection bias into a reduced form approach (Vossmeyer, 2016). Further, with the non-applicant category, we are able to capture the entire banking population eligible for support from these LOLRs. We (cid:12)nd that non-applicant banks increased their position of safe assets (as a ratio of total assets) by 7 percentage points. 3

Because we expect non-applicant banks to be the most-well-capitalized in this economy, as these banks did not seek LOLR assistance, switched and DW banks have the most similar balance sheets to non-applicant banks. The DW attracted banks that sought more liquid assets than the RFC after stigma was unexpectedly introduced at the RFC. These banks valued the anonymity of the DW. However, banks with higher funding stress and plausibly less liquid securities on their balance sheets approached the RFC after the publication of the list. Third, banks that borrowed from the DW in 1931 were likely to approach the RFC in 1932. Speci(cid:12)cally, borrowing from the DW prior to the RFC’s establishment, increased the probability that a bank approached the RFC by 14.6 percentage points. This result suggests that the choice of approaching either the DW or RFC was similar, reinforcing the validity of our reduced form approach when we compare the balance sheets of banks that chose one facility over the other. Next, forclarity, westudyhowthepublicationof thelistaffected thebalancesheet composition ofbanksbasedontheirchoiceofLOLRusingalinearpaneldatamodel. Ouridenti(cid:12)cationstrategy, basedontheunexpectedintroductionofstigmaattheRFC,allowsustoanalyzethebalancesheets ofbanksthatapproachedtheLOLRwithstigma(RFC)incomparisonwithbanksthatapproached the LOLR with no obvious stigma (DW). Fromour linearpaneldata model, wehavethreesigni(cid:12)cant(cid:12)ndings. First, banksthat switched awayfromtheRFC(switchedbanks)maintainedsimilarlevelsoftheirloans-and-discountportfolios toDWbanks. ThisresultsuggeststhatswitchedandDWbankswerenotforcedtoeitherwritedown their loans-and-discounts portfolios, or contract their lending any differently. However, switched banks did experience a 4.1 percentage point drop in their bonds-and-securities portfolios, but the magnitude of the drop was quite small, suggesting only slightly higher funding stress to that of DW banks. The identities of both switched and DW banks remained con(cid:12)dential throughout this period. Second, banks that were unexpectedly revealed on a list (\revealed banks") were 52 percent more likely to borrow from the RFC after the publication of the list despite the expectation that subsequentlistsofbankidentitieswouldbepublished. Revealedbanksexperienceda9.8percentage pointdropintheirbonds-and-securitiesportfoliosrelativetoDWbanks. Inaddition,revealedbanks reduced their loans-and-discounts portfolios by 15.3 percentage points. These results suggest that revealed banks either wrote down their loan-and-discounts portfolios, or contracted their lending in 4

comparison with DW banks. Furthermore, their lower bonds-and-securities portfolios also suggests that revealed banks were experiencing higher funding stresses perhaps leading to the sale of assets in comparison with DW banks. Third, RFC banks also experienced drops of 5.2 and 5.5 percentage points in their bondsand-securities and loans-and-discounts portfolios, respectively, in comparison with DW banks, but the magnitudes of these drops were far smaller than those of revealed banks. RFC banks also wrote down their loans-and-discounts portfolios and reduced their bonds-and-securities portfolios in comparison with DW banks, but by not as much as revealed banks. Anbil (2017) and Vossmeyer (2017) (cid:12)nd that the publication of the lists caused deposit withdrawals at the revealed banks that forced them to contract their lending and sell assets off their balance sheet, which is likely driving the larger drops. Overall, the results suggest that although not all RFC banks were revealed to the public, there was enough desperation for funds for banks to risk their identities being revealed. This desperation is not evident for switched or DW banks. Altogether, our results imply that the presence of two lending facilities where one guarantees anonymity while the other does not might separate banks in a way that reveals their liquidity preferences to market participants. We (cid:12)nd that the DW attracted banks that purchased more government securities onto their balance sheets. Moreover, these banks did not contract their lendingincomparisonwithRFCbanks. Becauseacrucialconcernwhendesigningalendingfacility is to attract solvent yet illiquid banks that would continue lending to the real economy during a (cid:12)nancialcrisis, thepresenceoftwolendingfacilitiesthatforcesbankstoseparateaccordingtotheir liquidity preferences may achieve these goals. The facility with no stigma would reduce the ex-ante concern that LOLR assistance goes to less liquid banks. While a setting with two lending facilities where one is stigmatized may seem unlikely, in addition to the 1930s, the setting existed during the recent (cid:12)nancial crisis where the Term Auction Facility (TAF) was not stigmatized but the DW was. The concerns presented in this paper remain an active topic of interest central banking. The empirical literature on why banks approach their LOLR is limited most likely due to researchers facing major challenges in identi(cid:12)cation. Drechsler et al. (2016) show that weakly capitalized banks took out more LOLR loans and used riskier collateral than strongly capitalized banks. Weareabletoshedlightonthelendingfacilitythatwouldattractmorestronglycapitalized banksduringa(cid:12)nancialcrisis. Ourresultsalignwiththeirsinthatwe(cid:12)ndweakbanksborrowedto 5

buy less-safe assets. Carpinelli and Crosignani (2017) (cid:12)nd that banks that experienced a wholesale fundingdry-upbeforetheEuropeanCentralBank’s(ECB)longtermre(cid:12)nancingoperation(LTRO) used their funding to restore credit supply, while banks that did not receive as much funding used it to increase their holdings of high-yield government bonds. We shed light on the type of lending facility where a bank \reaching for yield" may not occur. Finally, Acharya et al. (2016) (cid:12)nd that the ECB temporarily reduced funding pressure for banks but did not address solvency concerns via LTROs, suggesting it was difficult for the ECB to separate solvent but illiquid banks from those prone to risk-taking. In addition, our paper directly relates to a growing macroeconomic theory literature on how adverse selection affects markets. Bajaj (2017) studies the transition of a no-information revelation regime (pooling equilibrium) to information revelation regime (separating equilibrium). She shows that a negative shock to the quality of the regime implies a switch from no-information revelation regime to an information revelation regime. Our paper presents a no-information revelation regime where there are two similar LOLRs and market participants cannot determine information about the quality of banks that approach either facility. However, the revelation of banks that borrowed fromtheRFCin thepublicationcaused anegativeshocktothedesign ofthe RFC,whichled banks to separate into distinct groups that ex-post revealed information about their liquidity preferences to market participants. To the best of our knowledge, our paper is one of the (cid:12)rst to provide purely empirical evidence of these macroeconomic theories. Finally,ourpaperisalsorelatedtotheliteratureofhowbanksusetheirLOLRloans. Benmelech etal.(2017)(cid:12)ndthathadLOLRinterventionsbeeneffectiveinpreventingthecollapseoftheassetbacked commercial paper market, then the interventions might have contained the real effects of the crisis. We (cid:12)nd that DW and switched banks maintained their loans-and-discounts portfolios, suggestingtherewasnotasmuchofacontractionoflendingatthosebanks,whichcanbeinterpreted as a success of the DW. Sumit et al. (2015) (cid:12)nd that banks are less likely to lend to borrowers that most need funding during a (cid:12)nancial crisis, which may limit the effectiveness of LOLR lending facilities. However, Alves et al. (2016) (cid:12)nd that when Portuguese banks were prevented from going to repo markets during the European sovereign debt crisis, it was the virtually unlimited access to central bank funding that helped banks continue to provide funding to the real economy. Our papersuggeststhatananonymouslendingfacility(withnostigmaproblem)willattractbanksthat 6

will maintain their lending and that are concerned with the liquidity of their balance sheets. The remainder of the paper is organized as follows. Section 2 describes the RFC and DW as LOLRs to the U.S. banking system during the Great Depression, and details the publication of lists beginning on August 22, 1932. Section 3 describes the data, the development of our trivariate modelwithrecursiveendogeneity, andourlinearpaneldatamodel. Section4presentstheresultsof the trivariate model and our reduced form approach. Finally, Section 5 discusses the implications for future LOLR facilities and concludes. 2 Historical Background 2.1 The Reconstruction Finance Corporation and the Discount Window In response to an acceleration of bank suspensions after Britain left the gold standard in 1931, President Hoover created the RFC (Olson, 1977). The RFC began privately authorizing loans on February 2, 1932 to several types of institutions including commercial banks, insurance companies, and building and loan associations.1 We assume that the RFC and the DW served as simultaneous LOLRs to the U.S. banking system during the Great Depression. At the end of 1931, only 39 percent of banks were eligible to borrow from the DW at the Fed (henceforth referred to as member banks). There were 18,734 banks operating in the United States as of June 30, 1932. Of these banks, 7,246 were Federal Reserve member banks (FRB, 1959, 1932). Mitchener and Richardson (2016) show that the withdrawal pressures of nonmember banks on member banks magni(cid:12)ed liquidity risk during the Great Depression. If all banks had been member banks, systemic withdrawal pressures would have been substantially lower (Calomiris and Mason, 2003; FRB, 1932).2 As a result, President Hoover argued that another LOLR was needed to provide emergency liquidity assistance to the remaining nonmember banks (Olson, 1977). The RFC Act was submitted to Congress on December 7, 1931, and it was passed into law on January 22, 1932. Forty-four percent of all banks received loans from the RFC by June 30, 1933.3 1OfthetotalamountofbankloansrequestedfromtheRFC,80percentweregranted. AccordingtoRFC(1932), theamountofbankloanapplicationsreceivedin1932was$1,188,957,193. Theamountauthorizedwas$949,858,000. In contrast, member banks borrowed $518 million, on average, from the DW every month in 1932 (FRB, 1932). 2NationalbankswereFederalReservemembers,aswellassomestatebanks. SeeCalomirisetal.(2015)formore discussion on the decision to become a member bank. 3In fact, Rose (2016) shows that the RFC lent to a systemically important insurance company in 1933 acting as an LOLR. 7

We acknowledge that considering the RFC as an LOLR may be controversial. However, we merely use the terminology because during the period of invetigation we (cid:12)nd that the RFC’s operations were similar to those of the DW, and it was acting in a manner that aligns with the role of an LOLR. Furthermore, anecdotal evidence from DW and RFC loan applications suggests that many banks simultaneously applied to both the RFC and the DW, and offered similar reasoning as to why they needed assistance. Additionally, RFC loan applications cite DW examiner notes before authorizing a loan and vice versa, which suggests that RFC and DW loan officers worked very closely together. Finally, Eugene Meyer was both the chairman of both the Federal Reserve and the RFC when the RFC began authorizing loans. There were three differences between the RFC and the Federal Reserve’s DW. First, the RFC interest rate was 1.5 to 2 percentage points higher than that of the Federal Reserve’s. The discount rate averaged 3.5 percent across Federal Reserve Districts (FRB, 1932). Second, DW loans were offered for shorter durations than RFC loans. This inconvenience likely directed member banks to the RFC, which was offering longer maturity loans (FRB, 1932). RFC loans were given with maturities up to three years, but banks could roll over their loans for an additional two years (RFC,1932). Third, the RFChad morediscretionwith itscollateralrequirementsthanthe Federal Reserve, but both accepted the same types of collateral which included gold, Treasury securities, and commercial, industrial, and agricultural paper (FRB, 1932; Olson, 1977). By the end of 1932, 6,865 eligible institutions (banks and non(cid:12)nancial (cid:12)rms) had been authorized over $1.6 billion in loans by the RFC (RFC, 1932). At the DW, over $6 billion in loans were authorized in 1932.4 These facts highlight the signi(cid:12)cance of the RFC and DW, their effect on the (cid:12)nancial system, and their functions as LOLRs. For a thorough review of the RFC, see Butkiewicz (1995, 1999), Mason (2001, 2003, 2009), and Calomiris et al. (2013). For more information about the DW during the Great Depression, see Richardson and Troost (2009) and Wheelock (1990). 4WeareunawareofhowmanyeligiblebanksreceivedDWloansbeyondthoseintheFederalReserve6thDistrict. By the end of 1932, there were 6,816 eligible DW member banks (FRB, 1932). The majority of DW member banks were National Banks and therefore were much larger than the country banks that were located in more rural areas. These country banks were the banks that the RFC was intended to support (Calomiris et al., 2015). 8

2.2 Revelation of the First List The main event in this paper is the unexpected publication of banks that con(cid:12)dentially borrowed from the RFC beginning on August 22, 1932. We model banks’ choice of LOLR and subsequent liquidity seeking behavior using a trivariate model with recursive endogeneity. We also analyze the performance of eligible banks in the Federal Reserve Sixth District after the publication of the (cid:12)rst list in a reduced form setting.5 Initially, the identities of all RFC borrowers (banks and non-banks) were kept secret from the public. Since its establishment, the RFC had used elaborate secret codes to transmit messages to its loan agencies and individual banks (Olson, 1977). However, on July 21, 1932, the Emergency Relief and Construction Act of 1932 (ERCA) amended the original RFC Act to expand the RFC’s authority into state and local relief, public works construction, slum clearance, and so on. In this act, Section 201 (b) required that monthly reports of new borrower names be made known to Congress only (RFC, 1932). President Hoover initially planned to veto the bill because of the addition of the last-minute clause but was assured by the Senate majority leader that RFC loans would not be revealed to the public without congressional approval (CFC, 1932). It was decided thatthemonthlyreportsofnewborrowernameswouldbecon(cid:12)dentialandheldbytheclerksofthe Senate and the House of Representatives until Congress resumed session in December (RFC, 1932). Despitethisdecision, onAugust22, 1932, SouthTrimble, theclerkoftheHouseofRepresentatives, took it upon himself to release a partial list of the identities of banks that accepted new loans from the RFC to inform the U.S. public. The list was (cid:12)rst published in the New York Times and the Commercial & Financial Chronicle and coverage of this list was widespread. It is likely that the publication of the list was unexpected given the assurances that no borrower list would be released without congressional approval. The loan authorization date for a bank determined whether the bank identity was revealed. The (cid:12)rst monthly report that was submitted by the RFC to Congress revealed banks that had loans authorized between July 21 and July 31, 1932. Because the ERCA was passed on July 21, this (cid:12)rst monthly report was the only one Mr. Trimble had access to. Banks not revealed had a loan authorized on or before July 20, 1932. Because Mr. Trimble published all names available to 5Our paper is limited to studying banks in the Federal Reserve Sixth District because DW data is only available from this District. In addition, we only study banks that were eligible to approach both the DW and RFC. 9

him on the monthly lists, this suggests he did not choose which banks to reveal in a way that is systematic with bank characteristics. Because Congress was not in session, Mr. Trimble published four additional lists of borrower names following the August 22, 1932 list, (cid:12)nishing on January 26, 1933. The lists included all banks with loans authorized between July 21 and December 31, 1932, and loans over $100,000 authorized between February 2 and July 20, 1932. 2.3 Choice of LOLR Prior to the publication of the (cid:12)rst list on August 22, 1932, banks approached the RFC and DW interchangeably. The interest rate, collateral requirements, and duration of the loan were known at both LOLR facilities which would have affected banks’ choice of LOLR. However, after August 22, a stigma problem was unexpectedly introduced at the RFC, as loan con(cid:12)dentiality could no longer be guaranteed because of the renegade clerk. We (cid:12)nd that the pool of LOLR borrowers ex-post separated into speci(cid:12)c groups after the publication of the (cid:12)rst list: banks that continued borrowing from the RFC (RFC banks), banks that switched away from the RFC (switched banks), and banks that remained only at the DW (DW banks). The separation of banks ex-post revealed information about their liquidity preferences to market participants. In our linear model, we examine the performance of these bank groups based on their choice of LOLR lending facility. We also introduce two additional bank groups into our trivariate model. We include a fourth category of banks that never applied to either the RFC or DW for a loan (non-applicant banks). Finally, we also include a (cid:12)fth category of banks that were revealed on a list in the New York Times. We expect this (cid:12)nal group of banks to endure the largest cost of stigma at the RFC, as the public viewed the news that a bank borrowed from its LOLR as a sign of (cid:12)nancial weakness (Anbil, 2017; Vossmeyer, 2017). Figure 1 shows that there were 325 eligible banks in the Federal Reserve Sixth District, where 127 borrowed from the RFC, 211 borrowed from the DW, and 85 non-applicants did not borrow from any LOLR (as of September 30, 1933). There were 98 RFC banks that borrowed from the RFC after August 22, 1932. During the same period, there were 105 DW banks that borrowed solely from the DW. There were 55 revealed banks on a list in the New York Times that had a loan authorized between February 2, 1932 and December 31, 1932. Finally, there were 67 switched banks that borrowed from the RFC prior to August 22, 1932, but either stopped borrowing from 10

Figure 1: RFC versus DW Banks This (cid:12)gure displays the timeline illustrating how many eligible banks were in the Federal Reserve Sixth District between February 2, 1932 and September 30, 1933. Non-applicant banks were banks that never approached an LOLR during this time period. RFC banks borrowed from the RFC after August 22, 1932. DW banks borrowed from the DW after August 22, 1932. Revealed banks were revealed on a list in the New York Times on or after August 22, 1932. Finally, switched banks borrowed from the RFC prior to August 22, but then either switched to the DW or stopped borrowing from the LOLR afterwards. an LOLR or borrowed from the DW. Weexpectnon-applicantbankstobethehighest-capitalizedbanksintheSixthDistrictbecause the Federal Reserve Bank of Atlanta was accommodative with LOLR policy in the United States (Richardson and Troost, 2009). The (cid:12)ndings in Richardson and Troost (2009) support our case in that the DW was not stigmatized in this District. The President of the Federal Reserve Bank of Atlanta did not adhere to the Real Bills Doctrine where the LOLR would only lend to banks against \real" loans as collateral, such as trade contracts with merchants. Accordingly, we assume non-applicant banks did not apply for LOLR loans, as they were well-capitalized. Because we can model the choice selection mechanism, we compare switched, RFC, DW, and revealed banks to non-applicant banks. Because stigma is costly and present at the RFC (Anbil, 2017; Vossmeyer, 2017), we expect the performance of switched and DW banks to be the most like non-applicant banks. These banks were unwilling to bear the cost of stigma and not as desperate for funds to 11

remain at the RFC. Next, because revealed banks faced withdrawals and possible (cid:12)re sales due to the publication of the list, we expect these banks to be the most desperate for liquid securities, and their performance to be unlike non-applicant banks. Finally, we expect the performance of RFC banks that continued to borrow from the RFC to be worse than that of DW banks because these banks were willing to risk their identities being revealed at the RFC, suggesting they were desperate for funds.6 Table 1 describes summary statistics of RFC, DW, switched, revealed, and non-applicant banks as of December 31, 1931, prior to the publication of the list. The balance sheets of RFC, DW, switched, and revealed banks, which make up the pool of LOLR borrowers, appear remarkably alike. However, non-applicant banks have considerably smaller loans-and-discounts (scaled by total assets) portfolios to those of RFC, DW, switched, or revealed banks. Furthermore, their cashdue-to-banks and bond-and-securities portfolio levels are much higher compared to the other bank groups, suggesting that non-applicant banks exhibited some hoarding of cash during the Great Depression and were likely the highest-capitalized banks in the Sixth District. Interestingly, many non-applicant banks would approach the RFC by the end of the Depression, particularly after the RFC experienced a regime change and could purchase preferred stock in banks. Finally, Table 1 also con(cid:12)rms the sample selection issues of comparing banks that approached the LOLR to banks that did not, which we are able to control for in our trivariate model. Why would a bank want to borrow from the RFC given the existence of the DW if it was a member bank? First, we can see from RFC loan applications that many banks were encouraged to borrow from the RFC to increase its validity as an LOLR. Second, RFC loans were of longer durationthanDWloanssorolloverriskforRFCloanswouldbeless. Therefore,theRFCmighthave attractedbanksmoreconcernedwithrolloverriskdespitethehigherinterestrateontheloan. Third, anecdotalevidencesuggeststhatRFCexaminersweremorelenientwiththecollateralrequirements for a loan. If a bank had a much weaker balance sheet, the bank may prefer approaching the RFC (cid:12)rst. We are able to control for many of these observable characteristics in the choice framework. However, some of these preferences are latent or unobservable for individual banks. Hence, the joint model will estimate the covariance between the errors of LOLR choice and bank liquidity, so 6We do not observe if banks were rejected from the DW because these data do not exist. Vossmeyer (2016) highlights the importance of modeling declined applications. However, in this case, we observe three banks that approached the RFC that were rejected for loans but then subsequently borrowed from the DW. This suggests that the RFC did not receive all the banks that the DW may have rejected. 12

we can better understand the relationship. After stigma was unexpectedly introduced at the RFC, what might cause a bank to switch to the DW? If the bank was more concerned with its depositors discovering that it received LOLR assistancethanrolloverrisk, thebankwouldseekassistancefromtheDW.However, ifthebankdid not have the collateral required to receive a loan at the DW, it might decide to remain at the RFC. Those banks that switched to the DW (switched banks) were more concerned about stigma than rollover risk, and had the necessary collateral to borrow from the DW. However, those banks that stayed at the RFC were more concerned about rollover risk or did not have the necessary collateral to borrow from the DW. The introduction of stigma and banks’ subsequent choice ex-post revealed their liquidity seeking preferences to market participants. This information about their preferences wasunavailabletomarketparticipantsbeforethepublicationofthelistbecausebankswerepooling andborrowingfrombothLOLRfacilities. Theseliquiditypreferencesaredifficult,ifnotimpossible, to disentangle in a setting where there is a single lending facility. Finally, it is important for our identi(cid:12)cation in our panel linear model that the list of revealed banks be chosen by Mr. Trimble in a way that is uncorrelated with the outcome variables used in the estimation. There should be nothing systematically important about the dates of loan authorizations that he chose to publish implying that the decision of when a bank chose to borrow from the RFC also needs to be uncorrelated with all outcome variables. Otherwise, the revealed, switched,DWandRFCgroupsmaydifferalonganumberofobservabledimensions,therebybiasing the results of the estimation. Because Mr. Trimble published all names available to him on the monthly lists, this suggests he did not choose which banks to reveal in a way that is associated with bank characteristics. Mr. Trimble would have been unable to choose which banks to reveal because it was the RFC that provided the names to the Office of the Clerk (Anbil, 2017). 3 Data and Methodology 3.1 Data RFC loan information and borrower names are from the RFC Card Index to Loans Made to Banks and Railroads 1932-1957 acquired from the National Archives. The cards report the name and address of the borrower; the date, request and amount of the loan; whether the loan was approved or declined; and loan renewals. The names of banks revealed to the public are from the New York 13

Table 1: Summary Statistics of RFC and DW Banks Variable RFC DW Switched Revealed Non-Applicant No. Banks 98 105 67 55 85 Financial Ratios (averages) Cash / Assets 0.13 0.16 0.13 0.12 0.21 Loans / Assets 0.62 0.55 0.64 0.60 0.42 Bonds / Assets 0.19 0.22 0.16 0.20 0.31 Deposits / Liabilities 0.70 0.69 0.67 0.65 0.74 Paid Up Capital / Liabilities 0.10 0.13 0.10 0.10 0.10 County Characteristics (averages) Population ((cid:2)1000) 42.7 58.5 37.8 48.6 54.0 No. Manufact. Est. 51 81 46 56 65 Cropland ((cid:2)1000 acres) 94.0 87.7 96.8 83.9 81.1 Unemp. Rate 0.043 0.047 0.041 0.046 0.048 This table provides summary statistics for RFC, DW, switched, and revealed banks. RFC banks are those that approached the RFC after August 22, 1932. DW banks are those that approached the DW after August 22, 1932. Switched banks are those that borrowed from the RFC prior to August 22, 1932, and then switched to the DW or stopped borrowing from an LOLR altogether afterwards. Revealed banks are those that were revealed on a list published in the New York Times. Non-applicant banks are those that did not approach an LOLR before September 1933. Characteristics of the banks in each subgroup include the cash-due-to-banks, loans-and-discounts, bonds-and-securities, deposits, and paid-up capital, all scaled by total assets. All bank data are as of December 31, 1931 and from the Rand McNally Bankers’ Directory. All county data are from the 1930 census. Times and veri(cid:12)ed in the Commercial & Financial Chronicle. These announcements included the loan amounts and interest rates. All data are hand-collected. The DW data are proprietary, have never been seen before, and are from the Federal Reserve Bank of Atlanta Archives. Therefore, our DW data only include banks from the Sixth District, which are the states of Alabama, Florida, Georgia, and portions of Tennessee.7 The data are from daily ledgers containing loan and collateral amounts outstanding from January 1, 1931 through September 30, 1933. The ledgers report the name and address of the borrower, date, the loan amount outstanding, and the collateral amount outstanding.8 Our data include National and State member banks that were eligible to borrow from both the 7We do not have data on banks from Mississippi or Louisiana because we think those banks went to the New Orleans Federal Reserve Branch. 8SincewedonotobserveDW(cid:13)ows,weassumethatlargeincreasesintheloanamountoutstandingisanewloan. 14

RFC and DW. Because the RFC only began authorizing loans on February 2, 1932, we include only those DW loans made after this date in the linear analysis. After February 2, all banks in the sample were eligible to borrow from either LOLR. We end the loan sample at September 30, 1933, as that is when our DW data end. Bank balance sheet data are from Rand McNally Bankers’ Directory, which was published every six months. We collect the amounts of paid-up capital, surplus and pro(cid:12)ts, deposits, other liabilities, loans and discounts, bonds and securities, miscellaneous, cash due from other banks, and the name of the president for each bank. The data are hand-collected from eight books beginning December 31, 1930 and continuing to September 30, 1934, resulting in eight observations per bank. We also collect bank balance sheet data from the Office of the Comptroller of the Currency. These yearly data include the amount of U.S. Treasury government securities versus other securities on eachbank’sbalancesheetforDecember1931,December1932,andDecember1933. Othersecurities do not include government securities and are likely corporate bonds. For failed banks, we assume total assets and liabilities are zero. We (cid:12)lter out observations where the balance sheet data are identicalfromperiodtoperiod, approximately11percentofthedata. Weobserveifthebankfailed from the Rand McNally Bankers’ Directory and verify the failure in the Moody’s Directory. To account for differing macroeconomic trends and business environments across each county, we include several additional control variables as of December 30, 1930 in the estimation. These variables also capture the broader health of the banking system. We use the dollar amount of total deposits and the total number of banks in each state to account for the size, organization, and resources of the banking system. Next, we use the dollar amount of suspended deposits and the total number of suspended banks in each state to account for the health of the banking system. Suspended banks include both banks that closed their doors to depositors for at least one business day and later resumed operations, and banks that ceased operations, surrendered their charters, and repaid creditors under a court-appointed receiver (Heit(cid:12)eld et al., 2017). The data are from the FDIC Bank Deposit Data, 1920-1936 (Inter-university Consortium for Political and Social Research). Finally, we also include data from the 1930 census of population, manufacturing, and agriculture at the county level to capture cross-sectional changes in a bank’s business environment. 15

3.2 Methodology We employ two methodological approaches in this paper: a trivariate analysis and a reduced form approach. The trivariate framework employs a cross-sectional sample, which includes all National banks operating in the Sixth Federal Reserve District. This is a total of 270 banks that faced an LOLRchoice: receivebothDWandRFCassistance,receiveDWassistance,receiveRFCassistance, or receive no assistance. The analysis jointly models this choice set along with bank performance. In our reduced form approach, the main source of identi(cid:12)cation is the unexpected publication of banks that con(cid:12)dentially borrowed from the RFC beginning on August 22, 1932. We analyze the performance of eligible banks in the Federal Reserve Sixth District after the publication of the (cid:12)rst list in a panel data set from December 31, 1930 to September 30, 1934, where these panel data do not include non-applicant banks. The combination of these methodological approaches and unique data allows us to answer the following questions. Did the publishing of bank names that borrowed from the RFC shift more banks to the DW and away from the RFC? How did banks that continued borrowing from the RFC after August 22, 1932 use their loans in comparison to banks that shifted to the DW? Did the revealing of bank names make the DW a more effective LOLR facility than the RFC? With the results, we can cleanly identify whether liquidity-seeking behavior after receiving an LOLR loan should be an important consideration when designing future LOLR lending facilities. 3.2.1 Trivariate Model As discussed in Section 2.1, the RFC and DW were similar in many aspects, with slight differences in interest rates, maturity, and collateral requirements. These differences, along with other unobservables (for example, bank management, encouragement to borrow from the RFC over the DW, (cid:12)nancial contagion, fundamentals), may lead to different decision-making processes for each program. Furthermore, as the DW was established before the RFC, participation in the RFC program could be driven by previous access to the DW. To accommodate these concerns, we employ a trivariate model with recursive endogeneity and jointly examine the determinants of LOLR choice 16

and bank liquidity after the disbursements. The model is as follows: ′ z = x (cid:12) +" (1) i1 i1 1 i1 ′ z = x (cid:12) ++x (cid:12) +" (2) i2 i2 21 i2;endog 22 i2 ′ ′ z = x (cid:12) +x (cid:12) +" (3) i3 i3 31 i3;endog 32 i3 for banks i = 1;:::;n and " (cid:17) (" ;" ;" ) (cid:24) N (0;Ω), where i i1 i2 i3 3 0 1 1 ! ! 12 13 @ A Ω = ! 1 ! : (4) 21 23 ! ! ! 31 32 33 The observed choices fy ;y g′ are related to the latent data fz ;z g′ through i1 i2 i1 i2 { 1 if z > 0 y = ij (5) ij 0 if z (cid:20) 0 ij for j = 1;2, i.e., equations (1) and (2). For equation (3), the latent data are the observed data y = z . The (cid:12)rst observed outcome y takes the value 1 if the bank received assistance from i3 i3 i1 the DW and 0 otherwise. The second outcome y takes the value 1 if the bank received assistance i2 from the RFC and 0 otherwise. Thus, the set of all possible outcomes for equations (1) and (2) (LOLR choice) is: 8 > > (1;1) ′ if the bank received both DW and RFC assistance < ′ (1;0) if the bank received DW assistance and no RFC assistance y = (6) i > > (0;1) ′ if the bank received RFC assistance and no DW assistance : ′ (0;0) if the bank received no assistance. Equation (3), bank liquidity, is jointly modeled with LOLR choice. The bank liquidity outcome is measured in 1933 and is calculated as the amount of U.S. securities held as a ratio of total assets. The covariates that enter x include 1931 balance sheet information and correspondent i1 information. The covariates that enter x include 1932 balance sheet information and county i2 information. Also included is x which is an indicator of whether the bank borrowed from the i2;endog DW prior to the establishment of the RFC. This is endogenous because it is a function of y . The i1 covariates that enter x include county information and lagged balance sheet information. The i3 covariates selected for each equation follow the (cid:12)ndings in Vossmeyer (2016), where the exclusion restrictions are based on information excluded from the RFC applications. The RFC applications do no include information on correspondents and bank age, so this information is excluded from 17

the RFC equation. These characteristics, however, affect bank performance, so they are included in the other equations. Apparent from the RFC Paid Loan Files and Declined Loan Files, the RFC examiners often commented on the county in which the bank operated, which is why this information enters the RFC equation. This variable selection framework is formally tested via model comparison in Vossmeyer (2016). The endogenous covariate vector x is a set of indicator variables de(cid:12)ned by y and y i3;endog i1 i2 and the timing of the loan authorization. The indicator groups follow along with Section 3.2.2 which aligns the multivariate and linear model results. The groups are: (1) RFC Bank, (2) DW Bank, (3) Switched, (4) Revealed, and (5) Non-Applicant. The (cid:12)rst four groups are the same as the linear model section and group (5) is introduced for the trivariate model to capture the population of national banks. Estimation of the trivariate model relies on simulation techniques due to the discrete outcomes in the (cid:12)rst two equations, endogenous covariates, and restricted covariance matrix, where the restrictions are normalizations for iden(cid:12)cation. This paper implements a Bayesian framework for the model described by equations (1)-(5). The model is completed by specifying the prior distributions for the parameters. It is assumed that (cid:12) has a joint normal distribution with mean b 0 and variance B and (independently) ! (cid:24) N((cid:26) ;R )1f! 2 Sg, where S is the set of parameters 0 0 0 that produce the positive de(cid:12)nite matrix Ω. The complete-data posterior is given by: 0 2 3 1 ∏n ∏2 (cid:25)((cid:12);Ω;zjy) / @ 4 1fz > 0g5 N(z jX (cid:12);Ω) A(cid:2)N((cid:12)jb ;B )N(!j(cid:26) ;R )1f! 2 Sg: ij i i 0 0 0 0 i=1 j=1 The above posterior gives rise to a Markov chain Monte Carlo (MCMC) estimation algorithm. The novel algorithm is designed particularly for this application and is inspired by other work on multivariate discrete data models (Jeliazkov et al., 2008) and models with restricted covariance matrices (Chan and Jeliazkov, 2009). Furthermore, the algorithm features data augmentation for thesamplingofz,whichfollowsfromTannerandWong(1987)andAlbertandChib(1993). Details on the sampler are below, where as a matter of notation, we use \nk" to represent all elements in a set except the kth one. Details on the sampler are as follows: Algorithm 1 MCMC Estimation Algorithm 18

( ) 1. Sample [(cid:12)jz;Ω] (cid:24) N ^b;B^ , where ^b and B^ are given by ( ) ( ) ∑n ∑n (cid:0)1 ^b = B^ B (cid:0)1b + X ′ Ω (cid:0)1z and B^ = B (cid:0)1+ X ′ Ω (cid:0)1X : 0 0 i i 0 i i i=1 i=1 2. Sample Ωjy;(cid:12);z using the Metropolis-Hastings algorithm (use ! to produce Ω) 3. For equations k = 1;2, sample z ik jy;(cid:12);Ω;z nk (cid:24) TN A i ((cid:22) kjnk ;V kjnk ) where (cid:22) kjnk and V kjnk are the usual conditional mean and conditional variance, respectively. If y = 0, A is ((cid:0)1;0), ik i and if y = 1, A is (0;1). ik i The sample for the multivariate model is slightly different than that of Section 3.2.2, which will be discussed shortly. The sample includes all 270 National banks in Alabama, Florida, Georgia, and Tennessee operating in the Sixth district in 1931. The difference between this sample and that of the next section is that these data only include National banks. Section 3.2.2 has the additional group of State banks that are Federal Reserve member banks. The purpose of using only National banks is to better assess bank liquidity. If we narrow the sample to National banks, we can gather balance sheet data from the OCC’s Individual Statements of Condition of National Banks. These data are more detailed and provide more balance sheet categories. Much of Mason’s work on the RFC (Mason, 2001; Calomiris et al., 2013) utilized the National bank sample for the reason of better balance sheet information and has been successful in assessing the RFC’s effectiveness. For our purposes, the OCC data separately measure U.S. government securities from other bonds and securities,whichallowsustobetterunderstandbankliquidity. Furthermore,thissampleisdifferent from Section 3.2.2 because it also includes non-applicant banks, 85 of the 270 banks. Thus, with these data, we can model the two choice equations and control for the endogenous treatment in the liquidity behavior equation. 3.2.2 Reduced-Form Speci(cid:12)cation We run the following bank-level ordinary least squares regression (OLS) from December 31, 1930 through September 30, 1934: Y = (cid:11)+(cid:12) RFCBank (cid:2)1ft (cid:21) Listg+(cid:13)X (cid:2)1ft (cid:21) Listg+(cid:17) +(cid:14) +ϵ (7) i;t 1 i i t i i;t whereY istheoutcomeofinterestmeasuredeverysixmonthst forbanki. RFCBank isadummy i;t equal to 1 if the banks borrowed from the RFC after August 22, 1932 (RFC banks). 1ft (cid:21) Listg 19

is a dummy equal to 1 following the start of list publications on August 22, 1932. The coefficient of interest is (cid:12) , which measures the change in Y following the publication of the (cid:12)rst list for RFC 1 i banks in comparison with DW banks.9 We use three main outcome variables of interest: bondsand-securities at time t divided by total assets from t-1; loans-and-discounts at time t divided by total assets from t-1; and cash-due-from-banks at time t divided by total assets from t-1. We use these outcome variables as proxies for the performance of each bank. For failed banks, we record zero for these ratios. We scale bonds-and-securities, loans-and-discounts, and cash-and-exchanges by total assets from t-1 to account for the bank’s size, and to ensure the size of the balance sheet is not confounding Y contemporaneously. Finally, we run two additional versions of equation (7) i to examine the performance of switched banks in comparison with DW banks, and revealed banks in comparison with DW banks. A key issue that prevents both speci(cid:12)cations from identifying the effect of the revelation on Y is that Y may be correlated with unexplained macroeconomic conditions, bank borrower i;t i;t characteristics in the error term ϵ , or both. Therefore, we include controls, X (cid:2)1ft (cid:21) Listg to i;t i mitigate this bias where the controls only enter into the speci(cid:12)cation after the (cid:12)rst list is published on August 22, 1932 to ensure the covariates do not confound Y (Barrot, 2016). i;t X is a vector of controls measured at December 31, 1930 and includes the following covariates i at the state level: employment rate, per capita income, total deposits, total deposits at suspended banks, the number of banks, the number of suspended banks. X also includes the following coi variates at the county level: the total population, the number of manufacturing establishments, the total dollar sales of wholesale establishments, the total dollar sales of retail establishments, the amount of crop land, the number of unemployed persons, and the unemployment rate. These covariates are intended to capture observable proxies for macroeconomic conditions and bank characteristics that might explain Y , and only enter the speci(cid:12)cation after 1ft (cid:21) Listg equals 1. i;t This feature is to ensure that the controls do not confound Y prior to the publication of the i;t list. However, the speci(cid:12)cation may still be biased if some bank characteristics are unobservable. Therefore, we rely on bank (cid:12)xed effects, (cid:14) to exclude biases that could result from time-invariant i bank characteristics and to capture the extent to which each bank affects Y . Additionally, we i;t include half-year (cid:12)xed effects, (cid:17) to account for time trends in Y eliminating the concern that t i;t 9Notethatwedonotincludea1ft(cid:21)ListgdummynoraRFCBank dummybecausetheyarenotidenti(cid:12)edonce we include half-year and bank (cid:12)xed effects. 20

aggregate changes in Y and the publication of the list occurred together. i;t Finally, standard errors are clustered at the bank level according to Bertrand et al. (2004). The results are robust to including Y i;t(cid:0)1 as a control variable to account for autocorrelation in the dependent variable (Petersen, 2009). Furthermore, all continuous variables are winsorized at the 1 percent level to avoid outliers driving the estimation results. 4 Results 4.1 Trivariate Model Table2displaystheresultsforthetrivariatemodelwithrecursiveendogeneity. Columns2-4display the results for the DW selection, RFC selection, and bank liquidity, respectively. The results are based on 11,000 MCMC draws with a burn in of 1,000. Inefficiency factors were computed for the estimated parameters and all are low, implying excellent mixing of the Markov chain. The priors are centered at 0 with a variance of 25. The (cid:12)rst purpose to the multivariate modeling is to understand the determinants of LOLR choice. The coefficients in the DW and RFC columns demonstrate that the loans-and-discounts portfolio has a positive effect on receiving both DW and RFC assistance. Interestingly, the other securities to assets ratio is not statistically different from 0 for DW assistance, and it is positively associated with RFC assistance. The RFC mainly took bonds and securities as collateral and had more discretion with regard to collateral than the DW, so this is possibly being re(cid:13)ected in that result. Column3 alsodisplaystheresultsfor thecountyinformation. As discussedinSection3.2.1, the RFC Paid Loan Files and Declined Loan Files provide the examiners’ reports on each application decision. The examiners often discussed information about the applicant’s county and business environment, which is why these are being controlled for in the RFC assistance equation. The results demonstrate that county population has a positive effect on RFC assistance, and cropland and manufacturing have a negative effect. The results align with Calomiris and Mason (2003) and Richardson (2007) who (cid:12)nd that bank distress is a continuation of agricultural distress. The endogenous covariate in the RFC equation is \DW, Pre-RFC". The variable is an indicator that takes the value of 1 if the bank accessed the DW prior to the RFC’s establishment in 1932. The result is positive and statistically different from 0. Thus, accessing the DW in 1931 has a 21

Table 2: Results for the Trivariate Model with Recursive Endogeneity. DW RFC Bank Liquidity Intercept 0.883 (0.672) -0.436 (0.687) 0.112 (0.034) [-0.44, 2.15] [-1.82, 0.90] [0.05, 0.18] Loans / Assets 2.868 (0.675) 3.233 (0.633) [1.57, 4.21] [2.03, 4.49] Other Securities / Assets 0.880 (1.10) 3.398 (1.152) 0.199 (0.095) [-1.31, 3.01] [1.20, 5.72] [0.01, 0.38] Deposits / Liabilities -3.046 (0.710) -3.868 (0.681) [-4.43, 1.63] [-5.23, 2.60] No. Correspondents 0.025 (0.044) [-0.05, 0.11] Bank Age -0.105 (0.055) [-0.21, -0.00] County Population 0.513 (0.249) [0.019, 0.99] Manufact. Est. -0.005 (0.002) [-0.01, -0.00] Cropland -0.336 (0.140) [-0.59, -0.06] Unemployment rate 1.430 (0.521) [0.40, 2.43] Endog: DW, Pre-RFC 0.593 (0.234) [0.00, 1.18] Endog: RFC Bank -0.085 (0.029) [-0.14, -0.03] Endog: Non-Applicant 0.069 (0.026) [0.02, 0.12] Endog: Switched -0.073 (0.061) [-0.19, 0.05] Endog: Revealed -0.122 (0.029) [-0.18, -0.07] Posterior means, standard deviations (in parentheses), and 95% credibility intervals (in brackets, calculated using quantiles) are based on 11,000 MCMC draws with a burn-in of 1,000. positive effect on accessing the RFC. Interest remains in the change in the probability of receiving RFC assistance, between cases when banks did and did not receive DW assistance prior to 1932. y z Thus, the probability difference is described below, where the two vectors X and X differ only i i in the value of X i;DW Pre(cid:0)RFC and (cid:18) is all model parameters. To understand the magnitude of this result, the covariate effect is averaged over the sample and MCMC draws and is calculated as 22

follows: ∫ ( ) ( ) (cid:14) DW;Pre(cid:0)RFC = [Pr y i = 1jX y i ;(cid:18) (cid:0)Pr y i = 1jX z i ;(cid:18) ]f(X)(cid:25)((cid:18)jy)dXd(cid:18): The covariate effect is 0.146. The histogram of the probability distribution is displayed in Figure 2. Thus, after controlling for a banks health, balance sheet, and business environment, receiving DW assistance increases the probability of receiving RFC assistance by 14.6 percentage points. The result implies banks are viewing the LOLRs similarly and the choice is entering the banks’ random utility function as they maximize.10 Figure 2: Covariate effect of DW assistance on RFC assistance. Focusing now on the bank liquidity equation, the results show that the unemployment rate in a county has a positive effect on the U.S. government securities held at banks. Thus, banks in areas with higher unemployment rates will increase their holdings of safe assets. The results for the endogenouscovariatesshowthatrelativetoDW Banks,(1)banksthatstayedattheRFCdecreased their holdings of U.S. government securities, (2) banks that switched are not statistically different in U.S. government securities, (3) revealed banks decreased their holdings of U.S. government securities, and (4) non-applicant banks increased their holdings of U.S. government securities. Therefore, revealed and RFC banks reduced their positions of safe assets during a (cid:12)nancial crisis, inconsistentwithliquidity-seekingbehavior. Becausethepublicationofthelistforcedpooledbanks toseparate, byreducingtheirpositionsofsafeassetsRFCbanksrevealedtheirliquiditypreferences to market participants by continued borrowing from the RFC. This information would have been 10This is per McFadden’s (1974) initial discussion of the latent utility speci(cid:12)cation for discrete choice models. See Train (2003) for a review. 23

impossible to determine prior to the publication of the list when banks were pooling by borrowing from both LOLR facilities. Table 3 presents the posterior means, standard deviations, and implied correlation form for Ω. The results show a positive correlation between applying for DW and RFC funding, however, the 95 percent credibility interval overlaps zero. This implies that unobservables are not driving the results of the relationship between the DW and RFC (recall that the endogenous covariate was positive and statistically different from 0). Variables controlled for in the equation, which include balance sheet characteristics, county characteristics, and borrowing from the DW before the RFC, adequately represent the joint determinants for LOLR choice. Also note that there is a positive correlationbetweenLOLRassistanceandholdingsofU.S.securities. Thecorrelationsareofsimilar size and sign for both ! and ! , implying the unobservables are analogous for both LOLR choice 13 23 options. Table 3: Results for Ω in the Trivariate Model. Ω ! ! ! ! ! ! 11 12 22 13 23 33 Mean 1 0.159 1 0.031 0.033 0.021 Standard Deviation . 0.236 . 0.011 0.010 0.004 Implied Correlation 1 0.159 1 0.214 0.228 1 Posterior means, standard deviations, and implied correlation form for Ω. Posterior means and standard deviations are based on 11,000 MCMC draws with a burn-in of 1,000. 4.2 Reduced Form For clarity and simplicity, we use a linear panel data model to examine the response of banks to the publication of the list on August 22, 1932 where all banks in our sample were eligible to approach both the DW and the RFC. Using our trivariate model, we are able to effectively model banks’ choice of LOLR and (cid:12)nd that banks separate their LOLR choice into several groups: DW banks, RFC banks, switched banks, and non-applicant banks. Based on this choice, we analyze the performance and balance sheet composition of banks in these groups after the publication of the list. A bene(cid:12)t of this approach is that we can utilize the panel structure of the data and capture the time dimension of the revelation. First, we determine the probability that a revealed bank continued borrowing from the RFC 24

after its identity was revealed in the New York Times. Table 4 presents the results. From the OLS regression in Column (1), revealed banks were 52 percent more likely to continue borrowing from the RFC. This result suggests that revealed banks may have continued borrowing from the RFC because their identities were already revealed and they did not need to worry about \additional" stigma. Furthermore, as deposit withdrawals followed after the publication of the list, they likely needed more funds (Anbil, 2017). Table 4: Probability of RFC Bank Given Revealed Identity (1) (2) (3) OLS Logit Probit main (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) (cid:3)(cid:3)(cid:3) Revealed Bank 0.518 2.455 1.488 (6.98) (4.72) (5.02) Controls Yes Yes Yes Observations 230 230 230 R2 0.3316 ThistablepresentstheresultsofOLS,logit,andprobitcross-sectionalregressionsontheprobability ofbeinganRFCbank. AnRFCbankisabankthatborrowedfromtheRFCafterAugust22, 1932, the publication of the (cid:12)rst list. A revealed Bank is a bank that was revealed on a list on or after August 22. Controls is a vector of bank-level, state-level, and county-level controls. Bank-level controls include the average log of total assets. State-level controls include per capita income, total dollar deposits, total dollar deposits at suspended banks, the number of banks, and the number of suspended banks. County-level controls include the total population, the number of manufacturing establishments, the dollar amount of wholesale sales, the dollar amount of retail sales, the amount of crop land, the number of unemployed persons, and the unemployment rate. Standard errors are calculated robustly and presented in parentheses. All continuous variables are winsorized at the 1% level. ***, **, and * indicate signi(cid:12)cance at the 1%, 5%, and 10% levels respectively. Next, we compare the performance of switched banks to DW banks after the publication of the list. Because stigma is costly and remaining at the RFC increases the probability that the loan would be revealed, the performance of switched and DW banks should be similar. Table 5 presents the results. Switched banks experienced a small drop in their bonds-and-securities portfolio of 4.1 percentage points in comparison with DW banks. This result is possibly by construction because switchers had to pledge collateral to the RFC and then possibly more collateral to the DW. Next, from columns (2) and (3), we observe no differences in the estimates of the loan-and- 25

Table 5: Switched to DW versus DW Borrowing After (1) (2) (3) bonds/assets(t-1) loans/assets(t-1) cash/assets(t-1) SwitchedBank (cid:2)1ft = List(cid:0)1g -0.008 -0.022 0.007 i (-0.46) (-0.93) (0.61) SwitchedBank (cid:2)1ft (cid:21) Listg -0.042 (cid:3)(cid:3) -0.052 -0.018 i (-2.29) (-1.58) (-1.29) Time FE Yes Yes Yes Bank FE Yes Yes Yes Controls (cid:2)1ft (cid:21) Listg Yes Yes Yes i Observations 822 832 832 R2 0.8265 0.6852 0.6374 This table presents the reduced form estimates of the effect of list publications of RFC borrowers beginning on August 22, 1932 on bonds-and-securities, loans-and-discounts, and cash-due-frombanks. SwitchedBankisadummythatequals1ifthebankborrowedfromtheRFCpriortoAugust 22,andthenborrowedfromtheDWornotatallafterwards. SwitchedBank (cid:2)1ft (cid:21) Listgequals1 i if the bank switched to the DW on or after August 22. SwitchedBank (cid:2)1ft = List(cid:0)1g equals 1 if i thebankswitchedtotheDWbeforethe(cid:12)rstlistwaspublished. Controls (cid:2)1ft (cid:21) Listgisavector i of bank-level, state-level, and county-level controls that turn on when 1ft (cid:21) Listg equals 1, and are measured as of December 31, 1930. Bank-level controls include the log of total assets. State-level controls include per capita income, total dollar deposits, total dollar deposits at suspended banks, the number of banks, and the number of suspended banks. County-level controls include the total population, the number of manufacturing establishments, the dollar amount of wholesale sales, the dollar amount of retail sales, the amount of crop land, the number of unemployed persons, and the unemployment rate. Standard errors are clustered at the bank level and presented in parentheses. All continuous variables are winsorized at the 1% level. ***, **, and * indicate signi(cid:12)cance at the 1%, 5%, and 10% levels respectively. discountsandcash-due-from-banksportfoliosbetweenswitchedandDWbanksafterthepublication of the list. This result con(cid:12)rms our earlier conjecture that switched and DW banks would have similarbalancesheettrendsafterthepublicationofthelistbecausebothgroupsofbankswantedto avoid stigma and were less concerned with rollover risk. Because the publication of the list forced pooled banks to separate, switched and DW banks revealed their liquidity-seeking preferences to market participants by borrowing only from the DW. This information would have been impossible to determine prior to the publication of the list when banks were pooling by borrowing from both 26

LOLR facilities. Next, we compare the performance of revealed banks to DW banks after the publication of the list. Table 6 presents the results. Revealed banks experienced large drops of 9.8 and 15.3 percentage points drop in their bonds-and-securities and loans-and-discounts portfolios, respectively, in comparison with DW banks. Table 6: Revealed versus DW Borrowing After (1) (2) (3) bonds/assets(t-1) loans/assets(t-1) cash/assets(t-1) Revealed (cid:2)1ft = List(cid:0)1g -0.026 -0.060 -0.001 i (-0.80) (-1.58) (-0.05) Revealed (cid:2)1ft (cid:21) Listg -0.098 (cid:3)(cid:3)(cid:3) -0.153 (cid:3)(cid:3)(cid:3) -0.024 i (-2.94) (-3.04) (-1.37) Time FE Yes Yes Yes Bank FE Yes Yes Yes Controls (cid:2)1ft (cid:21) Listg Yes Yes Yes i Observations 728 734 734 R2 0.8391 0.6568 0.6263 ThistablepresentsthereducedformestimatesoftheeffectoflistpublicationsofRFCborrowersbeginningon August 22, 1932 on bonds-and-securities, loans-and-discounts, and cash-due-from-banks. Revealed (cid:2)1ft(cid:21) i Listgequals1ifthebankwaspublishedonalistonorafterAugust22. Revealed (cid:2)1ft=List(cid:0)1gequals1 i forrevealedbankspriortothepublicationofthelist. Controls (cid:2)1ft(cid:21)Listgisavectorofbank-level,statei level,andcounty-levelcontrolsthatturnonwhen1ft(cid:21)Listgequals1,andaremeasuredasofDecember31, 1930. Bank-levelcontrolsincludethelogoftotalassets. State-levelcontrolsincludepercapitaincome, total dollardeposits,totaldollardepositsatsuspendedbanks,thenumberofbanks,andthenumberofsuspended banks. County-level controls include the total population, the number of manufacturing establishments, the dollar amount of wholesale sales, the dollar amount of retail sales, the amount of crop land, the number of unemployed persons, and the unemployment rate. Standard errors are clustered at the bank level and are presented in parentheses. All continuous variables are winsorized at the 1% level. ***, **, and * indicate signi(cid:12)cance at the 1%, 5%, and 10% levels respectively. Althoughtherevelationwascostlytothesebanks,theywerefarmorelikelytoapproachtheRFC suggesting they were desperate for funds.11 The performance of these banks was much worse than switched and DW banks because they were unable to maintain the same trends of their bond-andsecurities and loans-and-discounts portfolios. Overall, these results imply that the RFC attracted 11We do not observe which banks were rejected from the DW. However, from our trivariate model, we (cid:12)nd that borrowing from the DW in 1931 increased the probability that a bank received RFC assistance by 14.6 percentage points. 27

more desperate banks after the publication of the list. It is also likely that these banks were more concerned with rollover risk due to their shrinking bond-and-securities portfolios, and preferred the longer-duration loans of the RFC over their fear of stigma. Prior to the publication of the list, market participants would have been unable to determine these distinct liquidity preferences from switched or DW banks. The publication forced banks to separate into groups that ex-post revealed their liquidity preferences. Table 7: RFC Borrowing After versus DW Borrowing After (1) (2) (3) bonds/assets(t-1) loans/assets(t-1) cash/assets(t-1) RFCBank (cid:2)1ft = List(cid:0)1g -0.023 -0.036 0.001 i (-1.22) (-1.33) (0.13) RFCBank (cid:2)1ft (cid:21) Listg -0.052 (cid:3)(cid:3)(cid:3) -0.055 (cid:3) -0.016 i (-2.92) (-1.83) (-1.32) Time FE Yes Yes Yes Bank FE Yes Yes Yes Controls (cid:2)1ft (cid:21) Listg Yes Yes Yes i Observations 937 947 947 R2 0.8453 0.6714 0.6287 This table presents the reduced form estimates of the effect of list publications of RFC borrowers beginning on August 22, 1932 on bonds-and-securities, loans-and-discounts, and cash-due-from-banks. RFCBank (cid:2) i 1ft (cid:21) Listg equals 1 if the bank borrowed from the RFC after the (cid:12)rst list was published on August 22, 1932. RFCBank (cid:2) 1ft = List (cid:0) 1g equals 1 for RFC banks prior to the publication of the list. i Controls (cid:2)1ft (cid:21) Listg is a vector of bank-level, state-level, and county-level controls that turn on when i 1ft(cid:21)Listgequals1, andaremeasuredasofDecember31, 1930. Bank-levelcontrolsincludethelogoftotal assets. State-levelcontrolsincludepercapitaincome,totaldollardeposits,totaldollardepositsatsuspended banks, the number of banks, and the number of suspended banks. County-level controls include the total population, the number of manufacturing establishments, the dollar amount of wholesale sales, the dollar amount of retail sales, the amount of crop land, the number of unemployed persons, and the unemployment rate. Standard errors are clustered at the bank level and presented in parentheses. All continuous variables arewinsorizedatthe1%level. ***,**,and*indicatesigni(cid:12)canceatthe1%,5%,and10%levelsrespectively. Finally, we compare the performance of RFC banks to DW banks after the publication of the list. Table 7 presents the results. RFC banks experienced drops of 5.2 and 5.8 percentage points in their bonds-and-securities and loans-and-discounts portfolios (albeit at the 10% level), respectively, in comparison with DW banks. These banks were willing to approach the RFC despite the chance their identities would be revealed on a subsequent list. This behavior suggests that RFC banks 28

werealsodesperateforLOLRfunds, andtheypreferredlonger-durationloansoverthecostofbeing revealed to the public (the stigma problem). The drop in their bonds-and-securities and loans-anddiscounts portfolio was far less than revealed banks, re(cid:13)ecting this cost. That RFC and revealed banks continued to borrow from the RFC despite its stigma problem suggests that banks that approached the RFC preferred longer-duration loans despite the cost of stigma. This information about their liquidity preferences was revealed to market participants only after the publication of the list. Furthermore, interestingly, RFC banks experienced no drop in their cash-due-to-banks portfolios. This might suggest that RFC banks continued to support their correspondent network, although qualitatively less than DW banks. 5 Implications for LOLR Facilities In this paper, we examine which banks borrow from the LOLR and when they do so. We shed light onhowlendingfacilitiescanbedesignedthatachievethreeobjectives: (1)easefundingconstraints; (2) are least subject to a stigma problem; and (3) attract banks with liquidity concerns. We use a unique setting of an unexpected disclosure of partial bank lists that introduced stigma at one of two nearly identical LOLRs during the Great Depression: the RFC and the DW. Using a unique hand-collected data set of balance sheet, DW, and RFC loan information for banks in the Federal Reserve Sixth District, we implement a novel trivariate model with recursive endogeneity to model each banks’ choice of LOLR and its subsequent liquidity preferences. We (cid:12)nd that the pool of LOLR borrowers ex-post separated into speci(cid:12)c groups of banks that revealed information about the liquidity preferences to market participants. Prior to the publication of the (cid:12)rst list, this information would have been unavailable because banks were pooling by borrowing from both LOLR facilities. After the separation, for clarity, we also use a linear panel data model to highlight the differences in balance sheet composition across the groups. Altogether, our results imply that a facility that guarantees anonymity might attract banks that value a more liquid balance sheet. We (cid:12)nd that the DW attracted banks that purchased more government securities onto their balance sheet. These banks were more concerned with stigma over rollover risk, and this information was revealed to market participants only after banks separated. Moreover, these banks did not contract their lending in comparison with RFC banks. Because a 29

crucial concern when designing a lending facility is to reduce stigma and attract banks that are simply illiquid rather than riskier, it seems that designing a facility that guarantees anonymity will reduce moral hazard concerns only when another facility with a stigma problem is present. Then, the anonymous facility would attract banks that would likely continue lending to the real economy andreducetheex-anteconcernoflendingtoriskierbanks. Itisnotusualtohaveasettingwithtwo facilities where one is stigmatized, as this was a feature of the recent crisis with the Term Auction Facility (TAF { not stigmatized) and DW (stigmatized). Our paper highlights implications as to what market participants can learn from these settings. References Acharya,V.,Pierret,D.,andSteffen,S.(2016),\LenderofLastResortversusBuyerofLastResort: Evidence from the European Sovereign Debt Crisis," Working Paper. Albert, J. and Chib, S. (1993), \Bayesian Analysis of Binary and Polychotomous Response Data," Journal of the American Statistical Association, 88, 669{679. Alves, N., Bon(cid:12)m, D., and Soares, C. (2016), \Surviving the perfect storm: the role of the lender of last resort," Working Paper. Anbil, S. (2017), \Managing Stigma During a Financial Crisis," Journal of Financial Economics, forthcoming. Armantier, O., Ghysels, E., Sarkar, A., and Shrader, J. (2015), \Discount Window Stigma During the 2007-2008 Finanial Crisis," Journal of Financial Economics, 118, 317{335. Bagehot, W. (1873), Lombard Street: A Description of the Money Market, London, NY: H.S. King. Bajaj, A. (2017), \Adverse Selection and Liquidity in Asset Markets," working paper. Barrot, J.-N. (2016), \Trade Credit and Industry Dynamics: Evidence from Trucking Firms," Journal of Finance, 71, 1975{2016. Benmelech,E.,Meisenzahl,R.R.,andRamcharan,R.(2017),\TheRealEffectsofLiquidityDuring the Financial Crisis: Evidence from Automobiles*," The Quarterly Journal of Economics, 132, 317. Bernanke, B. (2009), \The Federal Reserve’s Balance Sheet: An Update," in Speech at the Federal Reserve Board Conference on Key Developments in Monetary Policy. Bernanke, B. S. (1983), \Nonmonetary Effects of the Financial Crisis in the Propagation of the Great Depression," American Economic Review, 73, 257{276. Bertrand, M., Du(cid:13)o, E., and Mullainathan, S. (2004), \How Much Should We Trust Differences- In-Differences Estimates?" The Quarterly Journal of Economics, 119, 249{275. 30

Butkiewicz,J.(1995),\TheImpactofLenderofLastResortduringtheGreatDepression: TheCase of the Reconstruction Finance Corporation," Explorations in Economic History, 32, 197{216. Butkiewicz, J. (1999), \The Reconstruction Finance Corporation, the Gold Standard, and the Banking Panic of 1933," Southern Economic Journal, 66, 271{293. Calomiris, C. and Mason, J. (2003), \Fundamentals, Panics, and Bank Distress during the Depression," American Economic Review, 93, 1615{1645. Calomiris, C. W., Mason, J. R., Weidenmier, M., and Bobroff, K. (2013), \The Effects of the Reconstruction Finance Corporation Assistance on Michigan’s Banks’ Survival in the 1930s," Explorations in Economic History, 50, 525{547. Calomiris, C. W., Jaremski, M., Park, H., and Richard, G. (2015), \Liquidity Risk Bank Networks and Value of Joining Fed," NBER Working Paper. Carpinelli, L. and Crosignani, M. (2017), \The Effect of Central Bank Liquidity Injections on Bank Credit Supply," Working Paper. CFC (1932), \Publicity to be given to loans made by RFC," Commercial & Financial Chronicle, August 20, 1932. Chan, J. and Jeliazkov, I. (2009), \MCMC Estimation of Restricted Covariance Matrices," Journal of Computational and Graphical Statistics, 18, 457{480. Drechsler, I., Drechsel, T., Marques-Ibanez, D., and Schnabl, P. (2016), \Who Borrows from the Lender of Last Resort?" The Journal of Finance, 71, 1933{1974. Ennis, H. and Weinberg, J. (2013), \Over-the-counter Loans, Adverse Selection, and Stigma in the Interbank Market," Review of Economic Dynamics, 16, 601{616. FRB (1932), Nineteenth Annual Report of the Federal Reserve Board. FRB (1959), All-Bank Statistics United States 1896-1955. Heit(cid:12)eld, E., Richardson, G., and Wang, S. (2017), \Contagion During the Initial Banking Panic of the Great Depression," Working Paper. Jeliazkov, I., Graves, J., and Kutzbach, M. (2008), \Fitting and Comparison of Models for Multivariate Ordinal Outcomes," Advances in Econometrics, 23, 115{156. Madigan, B. F. (2009), \Bagehot’s Dictum in Practice: Forumating and Implementing Policies to Combat the Financial Crisis," in Speech for Federal Reserve Bank of Kansas City’s Annual Economic Symposium, Jackson Hole, Wyoming. Mason, J. R. (2001), \Do Lender of Last Resort Policies Matter? The Effects of Reconstruction Finance Corporation Assistance to Banks During the Great Depression," Journal of Financial Services Research, 20, 77{95. Mason, J. R. (2003), \The Political Economy of Reconstruction Finance Corporation Assistance During the Great Depression," Explorations in Economic History, 40, 101{121. 31

Mason, J. R. (2009), The Evolution of the Reconstruction Finance Corporation as a Lender of Last Resort in the Great Depression., In Bailouts: Public Money, Private Pro(cid:12)t. McFadden, D. (1974), \Conditional Logit Analysis of Qualitative Choice Behavior," Frontiers in Econometrics, pp. 105{142. Mitchener, K. J. and Richardson, G. (2016), \Network Contagion and Interbank Ampli(cid:12)cation during the Great Depression," NBER Working Paper. Olson, J. S. (1977), Herbert Hoover and the Reconstruction Finance Corporation 1931-1933, The Iowa State University Press. Petersen, M. (2009), \Estimating Standard Errors in Finance Panel Data Sets: Comparing Approaches," Review of Financial Studies, 22, 435{480. RFC(1932), Reconstruction Finance Corporation Act as amended and Provisions of the Emergency Relief and Construction Act of 1932. Richardson,G.(2007),\CategoriesandCausesofBankDistressDuringtheGreatDepression,1929- 1933: The illiquidity Versus Insolvency Debate Revisited," Explorations in Economic History, 44, 588{607. Richardson, G. and Troost, W. (2009), \Monetary Intervention Mitigated Banking Panics During the Great Depression: Quasi-Experimental Evidence from a Federal Reserve District Border 1929-1933," Journal of Political Economy, 117, 1031{1073. Rose, J.(2016), \TheResolutionofaSystemicallyImportantInsuranceCompanyduringtheGreat Depression," Finance and Economics Discussion Series, 5. Sumit, A., Chomsisengphet, S., Mahoney, N., and Stroebel, J. (2015), \Do Banks Pass Through Credit Expansions? The Marginal Pro(cid:12)tability of Consumer Lending During the Great Recession," Working Paper. Tanner, M. A. and Wong, W. H. (1987), \The Calculation of Posterior Distributions by Data Augmentation," Journal of the American Statistical Association, 82, 528{549. Train, K. E. (2003), Discrete Choice Methods with Simulation, Cambridge University Press, Cambridge. Vossmeyer, A.(2016), \SampleSelectionandTreatmentEffectEstimationofLenderofLastResort Policies," Journal of Business and Economic Statistics, 34, 197{212. Vossmeyer, A. (2017), \Analysis of Stigma and Bank Behavior," Working Paper. Wheelock, D. (1990), \Member Bank Borrowing and the Fed’s Contractionary Monetary Policy during the Great Depression," Journal of Money, Credit, and Banking, 22, 409{426. 32