Can the Federal Reserve Effectively Target Main Street? Evidence from the 1970s Recession

Finance and Economics Discussion Series Divisions of Research & Statistics and Monetary Affairs Federal Reserve Board, Washington, D.C. Can the Federal Reserve Effectively Target Main Street? Evidence from the 1970s Recession John Kandrac 2021-061 Please cite this paper as: Kandrac, John (2021). “Can the Federal Reserve Effectively Target Main Street? Evidence from the 1970s Recession,” Finance and Economics Discussion Series 2021-061. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2021.061. 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.

Can the Federal Reserve Effectively Target Main Street? Evidence from the 1970s Recession John Kandrac∗ Federal Reserve Board August 27, 2021 Abstract Moderncentralbankersconfrontachallengeofprovidingeconomicstimulusevenwhen the policy rate is constrained by a lower bound. This challenge has led to substantial innovationbypolicymakersandaproliferationofnewpolicytools. Inthispaper,Ioffer evidence on the efficacy of a new tool known as funding for lending, which provides banks with subsidized funding to make additional loans. I focus on a historical episode from the United States in which the Federal Reserve provided banks with steeply subsidized loans to promote the expansion of credit within their local communities. I showthatthecheapfundingsucceededingeneratingmorelendingbycounteringbanks’ excessive liquidity preference. The additional credit benefited the real economy. Local areas enjoyed higher rates of small business formation and more rapid employment growth. Finally, I show that the cost of the subsidy provided by the government was more than offset by the additional payroll taxes paid out of higher wages and salaries. These results suggest that funding for lending programs deserve consideration for the modern central banker’s toolkit and demonstrate that certain unconventional tools can offer monetary policymakers the means to pursue more targeted objectives. JEL classification: E52, E58, G21, G28 Keywords: monetary policy, funding for lending, bank lending, countercyclical policy, discount window ∗Board of Governors of the Federal Reserve System. E-mail: john.p.kandrac@frb.gov. Tel.: +12029127866. I am grateful for helpful comments from Sriya Anbil, Haelim Anderson, Mark Carlson, Jennifer Dlugosz, Atanas Mihov, Win Monroe, Marcelo Rezende, and Gary Richardson; seminar participants at the Federal Reserve Board; and conference participants at the FRBNY/NYU Stern Conference on Financial Intermediation, the Sciences Po/OFCE Workshop on Empirical Monetary Economics, and the Villanova School of Business Research Conference in Finance. I thank John O’Hare for excellent research assistance. The views expressed in this paper are solely the responsibility of the author and should not be interpreted as reflecting the views of the Board of Governors of the Federal Reserve System or of anyone else associated with the Federal Reserve System.

1 Introduction Providing economic stimulus when the policy rate is constrained by the effective lower bound is a principal challenge for modern central bankers. To meet this challenge and protect the relevance of monetary policy, policymakers have experimented with new tools to speed economic recoveries and limit the damage from deeper and longer downturns (Bernanke, 2020). The most widely adopted alternative policy tools include quantitative easing (QE) and forward guidance, and a large literature now exists to evaluate the efficacy of these policies. However, central banks continue to innovate beyond QE and forward guidance, possibly due to some combination of declining marginal benefits, increasing costs and risks, the inability of unconventional policy to fully overcome the limitations introduced by the effective lower bound (Bernanke, 2020), and the desire to pursue more nuanced objectives. Even before the COVID-19 crisis, central banks adopted policies including negative rates, yieldcurvecontrol,fundingforlending,andnon-sovereignQE.Althoughpolicymakersreport little regret in employing these novel policy tools (Blinder et al., 2017), whether or not they become part of the standard toolkit will ultimately depend on their efficacy. In this paper, I offer evidence on the ability of a funding for lending (FFL) facility to boost credit supply and promote growth in the real economy. FFL programs offer subsidized loans to banks on the condition that banks increase their lending, particularly to bankdependent borrowers that lack access to capital markets and likely face cash and credit constraints. Therefore, FFL can help support an economic recovery even outside of a crisis orfinancialpanic. Suchprogramscanbeparticularlyattractivetopolicymakersiftheyjudge that the social benefits of additional lending and a more rapid recovery exceed the private benefits banks enjoy from the additional loans (English and Liang, 2020). The social benefits of easier credit are likely to be substantial, particularly during steep downturns when small firms could fail and cut employment en masse (Brunnermeier and Krishnamurthy, 2020; Hanson et al., 2020). FFL programs with sufficient subsidies can also be beneficial if banks become too risk averse to pursue many profitable lending opportunities, or if bank funding 1

costs are abnormally high. The subsidy offered to banks is then justified because it helps banks internalize the positive externalities from looser credit, compensates banks for the additional risk, and counteracts high funding costs.1 Lowering the hurdle rate for new loans with a FFL program can transmit additional stimulus through the banking sector even if the main policy rate is pinned at an inefficiently high level because of an effective lower bound on nominal rate. Not only can subsidized lending be offered at increasingly negative rates to help circumvent the lower bound on the policy rate, but the ability to adjust the subsidy at will offers policymakers an effective tool even if a low and flat term structure of rates limits the gains from additional QE and forward guidance. However, the potential benefits of FFL can be negated if unattractive terms restrain total borrowing from the facility, if banks channel credit to nonviable firms, or if banks simply make loans that they would have originated even without subsidized funding. To evaluate the efficacy of FFL programs, I appeal to a historical episode in which the Federal Reserve (Fed) provided highly subsidized funding to banks so that they could expand lending within their local communities. Specifically, I examine the launch of the Seasonal Credit Facility (SCF), which was introduced as a discount window program in 1973 and continues to this day. The SCF is meant to provide funding to banks that lack reliable access to national capital and money markets and that witness pronounced seasonality in deposits and loans. Many of the institutions that meet these criteria are small banks in agricultural areas. These banks witness robust loan demand during the planting and growing seasons. In contrast, rapid deposit growth materializes in the fall and winter when farmers sell their crops and pay off their loans. The inevitable deposit runoffs in the spring and 1Policymakers may also wish to design a program that alleviates bank balance sheet constraints. In this case,structuringalendingprogramthatworksthroughpurchasingloanparticipations—asinthecaseofthe FederalReserve’sMainStreetLendingProgram(MSLP)initiatedin2020—maybethemosteffectiveoption. However, FFL programs can also ease banks’ balance sheet constraints by providing sufficient compensation for the balance sheet growth, and by boosting expected capital through a higher flow of retained earnings. In addition, the limited success of the MSLP suggests that a loan participation structure may be difficult to implement in practice. 2

summer restrain banks’ willingness to invest in long-term illiquid assets when they receive deposit windfalls. Instead, banks subject to strong seasonal patterns carry an inefficiently large share of assets in the form of liquid securities throughout the year. The SCF helps banks meet funding needs during periods of peak loan demand so that they can reduce liquidity buffers to support more loans. As the Fed explained when rolling out the SCF, the explicit goal of the facility was to enhance the ability of member banks to expand credit within their communities. The circumstances and details surrounding the implementation of the SCF reveal that the lending facility was a de facto FFL program. First, SCF credit from the Fed was available to banks outside of a lender of last resort (LOLR) context. Second, SCF credit was offeredattheordinarydiscountrate, whichwasoftensetbelowprevailingshort-terminterest rates before 2003. The discount rate could at times become substantially unmoored from market rates, partly due to a view that increasing the relatively high-profile discount rate would attract unwanted attention and criticism. This disconnect was especially pronounced when the SCF was introduced, which led to a steep subsidy. Over the first 18 months of the program, the subsidy averaged 2.75 percentage points, with a maximum of nearly 5 percentage points. Third, banks were forbidden from using the funds to increase lending to other banks or out-of-market areas, and local Reserve Banks monitored SCF borrowers to ensure compliance with this rule. Fourth, because the terms of SCF advances generally made larger banks ineligible, any additional lending supported by the facility was almost certainly directed to the bank-dependent firms and households that compose the bulk of small bank customers. Another appealing feature of the historical setting is that it offers a compelling strategy to draw causal inference despite the endogentiy of banks’ borrowing decisions. When the SCF was introduced, only Fed member institutions enjoyed the seasonal borrowing privilege. As a result, central bank funding was open to only a subset of institutions operating within a given area. Eligible and ineligible banks faced similar local demand conditions, so a 3

divergence in lending activity after the introduction of the SCF points to a causal effect of the facility. Furthermore, the variation in total SCF draws across geographic areas correlates with the prevalence of eligible institutions. This correlation opens up an instrument for seasonal credit funding, which allows for causal estimates of cheap central bank funding on local economic activity. Isolating the geographic area that reaps the benefit of SCF loans is relatively straightforward because severe branching restrictions at the time resulted in tightly defined loan market areas for most banks. The identification strategy requires a distribution of Fed member banks in rural counties that is as good randomly assigned with respect to local growth prospects in the 1970s. Although all nationally chartered banks must be members of the Fed, state-chartered banks may choose to join. Thus, a selection concern may arise if banks’ membership choice was driven by actual or expected local growth rates. I offer evidence that the membership decision is neither endogenous to 1970s growth rates nor driven by the SCF itself. For example, there is little transition into Fed membership in the years before and quarters after the introduction of the SCF, and in fact the overwhelming majority of banks established their membership status many decades earlier. In addition, the state member banks that selected into Fed membership did not draw on the SCF at higher rates than nationally chartered banks. I also observe a similar balance sheet composition between member and nonmember banks, and I show that the two groups displayed comparable seasonal loan swings. Ifindthatsubsidizedfundingfromthecentralbankboostedloangrowthandsupported real economic activity. Banks that drew on the SCF reduced their liquid asset buffers and increased lending. The increase in loans was split between non-agricultural and agricultural businesses. Theshiftinassetcompositionalsoincreasedinterestincome, becausetheinterest earned on the additional loans exceeded the interest lost from the sale of liquid assets. Communities with banks that took up more SCF loans witnessed faster employment growth, particularly among sole proprietorships. Consistent with that evidence, the additional credit from the central bank helped support an increase in small business establishments. The new 4

establishments were concentrated in industries that were well represented in rural counties, had lower startup costs, and were more dependent on credit from small banks. By contrast, prevalent industries with high regulatory, staffing, and capital hurdles witnessed no net change. In sum, these results are consistent with the notion that credit constraints could restrain productivity growth, in which case a FFL program can boost the supply potential of the economy (Churm et al., 2012). This paper contributes most directly to the nascent literature on the efficacy of the mostnonstandardunconventionalpolicymeasuresintroducedsincetheglobalfinancialcrisis. While a large and expanding literature examines the effects of QE, forward guidance, and negative policy rates, far less evidence has been brought to bear on FFL programs. The dearth of FFL studies reflects the limited examples of such programs. The studies that exist focus on either the BoE’s funding for lending scheme (Churm et al., 2018) or the ECB’s targeted longer-term refinancing operations (TLTROs) (Benetton and Fantino, 2018; Afonso and Sousa-Leite, 2019; Jasova et al., 2020; Laine, 2021), although TLTROs did not embed an explicit subsidy component until March 2020. Consequently, the most comparable analysis using the experience with TLTROs is (Da Silva et al., 2021). As explained in Churm et al. (2018) and Andrade et al. (2019), all of these studies confront a difficult identification challenge stemming from the endogeneity of bank participation. Addressing this challenge by using the restricted eligibility is a key feature of the present study. The focus on the real effects of the FFL program represents another contribution to the existing literature. Churm et al. (2018) conducts a time series analysis in which a measure of bank funding costs is used to indirectly estimate the macroeconomic benefit of the BoE’s FFL program, but other studies focus mainly on financial effects. As explained in Gros et al. (2016), gauging the real effects of FFL programs may be particularly important because banks can originate loans that make them eligible for cheaper funding but serve no real economic purpose. Lastly, I offer the only study of an FFL program within the United States, which also entails a rare analysis of Fed lending outside of a crisis or lender of last resort (LOLR) context. 5

The present study also contributes to the literature on central bank policy objectives (Yellen,2009;Woodford,2014)bydemonstratingthescopeforthemonetarypolicytoolkitto make progress towards more targeted objectives. Most monetary policy tools are characterized as blunt instruments that are best suited to ease or tighten overall financial conditions. However, central banks have recently been at the forefront of debates and policy making aimed toward such varied goals as promoting green energy, mitigating distributional equality, and offering support to specific nonfinancial sectors and geographic areas. Targeting these and other objectives requires policy tools that can be employed with more finesse than traditional policy tools, which may suffer from imprecision. In the present setting, targeted borrowers were identified according to a stated geographic preference because banks drawing subsidized credit were expected to boost lending to their local communities. However, other FFL programs could tie eligibility to a requirement that additional lending be directed towards a particular demographic, industry, or firm type. One potential drawback to the research design presented here is that the lessons may lackgeneralizability. Theexternalvalidityofstudiesexamininghistoricalepisodesisalwaysa question, butsomekeyaspectsofthefacilityandsettingaresimilartomodernconsiderations and enhance the overall relevance. First, the Fed provided funding at an attractive subsidy, and bank participation was robust. Robust participation is central to the success of a program that is intended to support a wide array of nonfinancial companies without ready accesstocapitalmarkets(EnglishandLiang,2020). Second, theSCFisaimedateliminating the financial friction introduced by an excessive liquidity preference among banks. Banks placing too high a premium on liquidity is a typical affliction that policymakers confront when designing monetary policies to counter significant downturns. Third, by coincidence, theFedintroducedtheSCFjustbeforethe1973–1975recession. Policymakersaremostlikely to consider FFL programs when facing a flagging economy. Because the effects of additional credit can change over the course of the business cycle, it is helpful that the results in this paperapplytosubsidizedfundinggrantedduringarecession. Fourth, althoughtheSCFonly 6

targeted small banks, large institutions typically enjoy access to liquid capital markets that the Fed can exert more influence over using other tools. FFL is typically aimed at bankdependent firms that often maintain relationships with smaller financial institutions that cannot raise funding as readily as large banks. Finally, as outlined above, several features of the SCF’s implementation support the internal validity of the research design. Insofar as internal validity is a prerequisite for external validity (Carlson and Morrison, 2009; Campbell and Stanley, 2015), these features also advance the external validity of the results. 2 Institutional Background: Discount Window Lending, the Seasonal Credit Facility, and Bank Access 2.1 The Discount Window and Preexisting Credit Programs The discount window is the primary lending facility maintained by the Fed. In the years following the establishment of the Fed, discount window lending was the principal tool of monetary policy. By the Great Depression, discount window lending started to wane because of the discovery of open market operations and the emergence of the federal funds market. In general, the prevailing view over subsequent years was that banks should tap the private markettosatisfyfundingneedsinnormaltimes. Overtime, discountloanswereextendedfor two main reasons. First, banks that faced an unexpected shortage of reserves could approach the discount window for a short-term loan that could be used to make payments or meet minimum reserve requirements. Second, the discount window eased the upward pressure on the federal funds rate that could arise when reserves supplied via open market operations fell short of anticipated demand (Clouse, 1994; Madigan and Nelson, 2002). In the early 1970s, the Fed extended loans to its member banks through two discount window programs. The “adjustment credit” program was by far the most common of the two programs. Adjustment credit loans were ordinary discount window loans granted to 7

requesting institutions to meet temporary liquidity needs. Banks initiated loan requests by telephoning their district Reserve Bank, which would ensure that the bank posted adequate collateral and had an appropriate reason to request a discount window loan. To grant the loan, the Reserve Bank would have to make a determination that the borrower exhausted reasonablyavailablealternativesourcesoffunds. Whileatemporaryandunexpectedfunding shortfall was a valid reason to borrow, loans would be disallowed for foreseeable increases in loans or securities and for replacing the anticipated runoff of private funds (Clouse, 1994). The “extended credit” program was intended to meet the needs of banks facing extended periods of illiquidity under exceptional circumstances.2 The terms of extended credit loans were strict, and the Fed made almost no loans through the extended credit program during the years covered in this study.3 Adjustment credit loans were offered at the discount rate, which was usually set below overnight market interest rates. Reserve Banks set the discount rates for member banks in their district, subject to approval by the Board of Governors. This practice did not result in meaningful variation across districts in the early-to-mid 1970s. In fact, differences in discount rates charged by Reserve Banks were almost always no more than 25 basis points for no more than a few days. Below-market rates on discount window loans led to an administrative burden for the Reserve Banks. To prevent an uncontrolled expansion of reserves, discount window credit was rationed via rules that prohibited banks from borrowing too frequently, forbade arbitrage of the spread between the federal funds rate and discount rate, and required banks to exhaust other sources of funding. Administrative rationing of discount window credit therefore required discount officers to review every prospective borrower’s funding situation and monitor borrowing institutions’ federal funds sold position. The judgments necessary to 2In 2003, the adjustment and extended credit programs were replaced by the primary and secondary credit facilities. 3In the decades before 1980, the extended credit facility saw nontrivial use only once when over a billion dollars was provided to aid in the wind-down of the mafia-linked (?!) Franklin National Bank. This bank does not appear in the sample used in the main analysis. 8

grant discount window loans could be subjective, complicating efforts to achieve consistency across the 12 Reserve Banks (Madigan and Nelson, 2002). 2.2 The Design of the Seasonal Credit Facility The Fed announced its intention to offer a seasonal borrowing privilege through the discount window on November 22, 1972. As with the preexisting discount window lending programs, the new Seasonal Credit Facility (SCF) was available only to banks that were members of the Federal Reserve System. The SCF was intended to assist institutions that operated in areas with pronounced seasonality in loan demand and deposits. Such seasonal patterns typicallyresultedfromthelocaleconomicimportanceofasingleseasonallysensitiveindustry such as agriculture.4 Banks with seasonal deposit flows would often carry a large share of liquid assets to guard against the correlated deposit outflows and sudden credit needs of their customers during other times of the year. For example, agricultural banks witnessed rapid deposit growth following the harvest season, but these funds were not used to support loan growth. Instead, banks held large shares of liquid assets in anticipation of deposit runoffs and increased demand for short-term loans through the growing season. The Fed specifically cited such asset and liability management issues as a cause of banks’ inability to provide adequate banking services to their communities. Prior to the introduction of the SCF, discount window credit was not available to help banks address predictable liquidity pressures. Byofferingseasonalcreditonattractiveterms,theFedhopedtotargetthefrictionsthat led banks to hold inefficiently large liquidity buffers so that these banks could extend more loans to their local communities (Clouse, 1994). As stated in the circular requesting public comment on the facility, “the seasonal borrowing privilege now proposed is meant to enhance the ability of member banks to serve the credit needs of their communities and areas.” In 4Meaningful seasonal fluctuations in deposits and loans were not unique to agricultural banks. Among other things, seasonal patterns could also be caused by construction, college, tourism, and municipal financing. 9

a review of the potential benefits of offering seasonal credit, the Fed explicitly identified the ability of such funds to aid in rural banks’ provision of credit to boost the “developmental capital” available to their communities. Because seasonal credit was intended to help banks expand credit within their local communities, banks were disallowed from drawing funds to make loans outside of their normal market areas or to purchase loans from other institutions. Similarly, banks were not permitted to use the funds to increase their liquid asset holdings. The rationale underpinning the SCF thus mirrored that of a funding-for-lending program. In contrast to other discount window programs, the SCF did not require banks to exhaust market sources of funding to receive a loan. Instead, banks were simply expected to submit historical data on loans and deposits to demonstrate persistent seasonal fluctuations. A seasonal credit line was available to a bank in the amount that the bank’s decline in available funds (deposits minus loans) exceeded a “deductible” that the bank had to meet out of its own resources.5 The Fed also required that the seasonal funding strains lasted for at least two months. Unlike other Reserve Bank credit programs at the time, SCF lending was not designed to cover short-term needs. If a bank met these criteria, the Fed stated a willingness to extend credit for the maximum allowable time of 90 days and to issue a new loan in the event that a member bank’s seasonal needs persisted for more than 90 days. In this way, lending through the SCF provided a means for the Fed to offer credit to banks outside of a LOLR context. The SCF favored smaller banks for several reasons. First, more economically diverse metropolitan areas are less likely to generate pronounced seasonal fluctuations. Banks in rural areas are thus more likely to satisfy the eligibility requirements, and rural banks tend to be smaller than urban banks. A Fed report on the needs for seasonal credit assistance emphasized the unique needs and challenges confronted by banks in smaller towns and communities, which operated at a more limited scale. Second, the deductible that banks were required to meet with their own resources could be quite substantial for larger banks, 5The deductible changed over the years, but was initially set at 5% of a bank’s average total deposits in the preceding calendar year. 10

and the seasonal component of their business would have to be very large to secure SCF funding. In later years, the percentage deductible increased with bank size. Third, smaller institutions lacked the capacity to smooth seasonality in their deposit base with wholesale funding from large-denomination CDs and the Eurodollar or federal funds markets. 2.2.1 Borrowing at the Seasonal Credit Facility As shown in Figure 1, borrowing from the SCF commenced immediately upon the facility’s introduction. Because banks were required to go through a review process with their Reserve Banks before receiving funds via the SCF, the rapid take-up indicates that eligible banks began preparing for access in advance of the facility’s launch in April of 1973. The participation also indicates that the credit program was offered on favorable terms and correctly identified a funding need for certain institutions. The SCF’s interest rate was an important factor influencing bank participation in the facility. As demonstrated in Figure 2, the effective subsidy as measured by the difference between the federal funds rate and the discount rate was already over 1 percentage point during the SCF comment period in early 1973. In fact, the multiple-month maturity of SCF loans implies that the true subsidy was somewhat higher than reported in Figure 2. During peak seasonal credit needs later in the year, the subsidy was over 3 percentage points. For the first 18 months of the program, the subsidy averaged 2.75 percentage points, with a maximum of nearly 5 percentage points near the peak borrowing period in 1974. The relatively modest take-up in 1975 and 1976 in part reflects the low or nonexistent effective subsidy in these years, when the federal funds rate fell more rapidly than Reserve Banks’ discount rates. The geographic distribution of SCF credit was disperse in its first two years, as shown in Figure 3. Although banks in popular vacation areas in New England and Southern Florida drew heavily on the facility, rural counties with a large agricultural presence benefited the most from the subsidized lending. Figure 4 demonstrates the relevance of SCF credit to rural 11

counties. Compared with urban counties, the maximum SCF draw by banks in a county as a share of county-wide deposits was nearly twice as high across all rural counties. A more restrictive definition of rural counties increases the gap further. While the terms of the Fed’s Regulation A virtually ensured that the SCF was not a substitute for ordinary adjustment credit discount window borrowing, an analysis in Appendix A confirms that banks’ SCF borrowing did not simply replace non-SCF loans. 2.3 Discount Window Access and Federal Reserve Membership Before the Monetary Control Act was passed in 1980, ordinary access to discount window credit was limited to banks that were members of the Federal Reserve System. Banks with national charters were required to become members, while state-chartered banks could choose to become members. In the years immediately following the creation of the Fed, few state banks opted into membership. Though many state bank executives held a general distaste for federalization, state banks were also put off by the extra clerical expense, red tape, and additional requirements associated with membership, such as the requirement to accept checks at par (Krueger, 1933). After several years, changes to the Federal Reserve Act made membership more desirable and a patriotic desire to assist the government in financing World War I sparked an increase in membership (Krueger, 1933).6 Anderson et al. (2018) document that larger state banks with a respondent banking network were among the first to join the Fed, followed by banks that were attracted by the potential for the discount window to limit liquidity risk resulting from runs and fluctuations in loan demand. Member institutionsweresubjecttominimumcapitalrequirements, althoughstate-levelrequirements were usually similar in rural areas. As a result of these factors, membership rates among state banks rose for a few years before leveling off off by the mid-1920s with around 8% of state banks eventually joining (Anderson et al., 2018). 6As detailed in Anderson et al. (2018), amendments to the Federal Reserve Act were aimed at assuaging fears related to regulation and oversight, as well as relaxing rules related to branching, issuing loans against improved real estate, and permissible directors. The changes also permitted the withdrawal from membership. 12

Membership transitions were relatively few in the decades before the introduction of the SCF, due in part to a combination of economic and financial development, the reduction in depositor run risk following the creation of the FDIC, and regulatory communication from the Fed emphasizing the limited and short-term nature of discount window credit. Of the limited membership transitions that did occur during this period, most banks gave up their membership, which allowed them to avoid the effective tax imposed by zero-interest reserve requirements.7 While limited reporting in the National Information Center database prevents an exhaustive analysis before 1959, I find that 97% of banks did not change their membership status at all between 1959 and 1970, and only 0.7% of banks transitioned into Fed membership. The average time since membership transition for these banks was five years. The low rate of transition into membership continued in the quarters following the introduction of the SCF. Most banks’ membership decision was made many decades prior to the 1970s, and the majority of banks’ membership status was determined by the nature of their charter (Board, 1930). Not only was the charter decision often unrelated to the Fed, but most of the banks in the sample were formed before the creation of the Fed.8 As of 1970, the median bank in the sample was over 60 years old, and three-quarters of banks were at least 40 years old. Despite the temporal separation between banks’ membership decision and the introduction of the SCF, a selection concern may arise if the state-chartered banks that opted into membership did so for reasons that also led them to draw more heavily on the SCF. Though there was a clear shift in focus towards macro stabilization by the 1950s, a key role of the Fed in its early years was to facilitate seasonal swings caused by agricultural flows (Miron, 1986). This role may have caused banks to select membership if they were subject to pronounced 7Althoughregulatoryregimescoulddifferbetweenstateandnationallycharteredbanks,FDICinsurance subjected all covered banks to nationally applicable standards. 8As reported in Board (1930), national banks commonly selected their charter because of the prestige and goodwill connoted by the national label in the banks’ communities. State charters were frequently chosen because of permissible fiduciary and trust powers and the ability to lend against real estate, which were addressed through future amendments to the Federal Reserve Act and in some cases even deferred to the applicable state powers. 13

seasonality and attracted by discount window access (Anderson et al., 2018). Despite the significant changes in discount window philosophy and operation in the intervening decades, these historical factors could influence SCF borrowers in the 1970s. However, it does not appear that state membership rates were driven by the same type of seasonal needs addressed by the SCF. For instance, membership rates among state-chartered banks were high in states such as New York and New Jersey that were most seasonally affected by relationships with country banks. State-chartered membership rates in agricultural states were far lower, and the statewide correlation between the state bank membership share and SCF borrowing as a share of total deposits is 0.04 (p = 0.77). In addition, if the membership decision was driven by the same factors that led banks to borrow from the SCF, state-chartered members should exhibit the highest incidence of borrowing. However, national banks were a bit more likely to draw on the SCF compared with state member banks (7.8% versus 6.0%). Evidently, State banks did not opt into membership for reasons that also led to high SCF usage. 3 The Financial Effects of Seasonal Credit In this section, I focus on evaluating whether observed financial outcomes are consistent with the rationale behind the SCF. The goal of the SCF was to expand the lending capacity of banks that maintained inefficiently high liquidity buffers because of steep seasonal liquidity pressuresandcorrelateddepositflows. Iftheadditionalbankloansreachedcreditconstrained households and businesses, seasonal communities could enjoy more rapid growth. Table1reportssummarystatisticsfortheanalyticsampleofbanksinruralareas,which receivedtheoverwhelmingmajorityofSCFloans. AsexplainedinSection2,discountwindow loans were ordinarily available to Federal Reserve member banks only. Nonmember banks that operated in the same areas as member banks may be used to form a counterfactual, so it is useful to split the sample accordingly. 14

Although nonmember banks were smaller than member banks on average, the two groups of banks held a similar amount of total assets because nonmembers were more numerous. While member banks are about 65% larger than nonmembers on average, this difference is far smaller than the roughly 400% differential observed in urban areas.9 The size differential likely stems from the historical size requirement for Fed membership. Members and nonmembers held a similar share of assets in loans, and the composition of the banks’ loan portfolios was very similar. The funding structure was also similar across membership status, with each group deriving about 90% of funding from deposits. Interest income was about the same, but the slightly smaller liquid asset portfolio share among member banks— possibly reflecting the liquidity insurance offered by the discount window—made a smaller contribution to return on assets. At the bottom of Table 1, I report the average of member banks’ maximum SCF draw between 1973 and 1974 as a share of 1972 deposits. Conditional on using the SCF, banks averaged a peak SCF draw of about 5% of deposits. Table 2 reveals that membership status is not easily determined based on observable characteristics. The pseudo R2 from a logit regression of membership status on the 11 characteristics in Table 1 is only 7.5%. Only two of these characteristics—the capital and interest income ratios—achieve statistical significance. While there is evidence that state banks that opted into membership were observably different from nonmember institutions at the time of the membership decision (Anderson et al., 2018), membership status in the 1970s was evidently not easy to predict based simply on balance sheet characteristics.10 Members and nonmembers were also similar in their exposure to seasonal swings. Figure 5 displays the regular seasonal patterns in banks’ loan portfolio shares using the semiannual Call Reports for banks operating in communities with seasonal patterns sufficient to 9The average bank size for both groups in these metropolitan areas is an order of magnitude larger than that of rural banks reported in Table 1. 10The current analysis is not strictly comparable with that of Anderson et al. (2018). The sample in the present paper includes many nationally chartered banks which were automatically placed into Fed membership, and is not limited to New York state only. In supplemental analysis, I find that observable differences between members and nonmembers in urban areas are much more pronounced. 15

qualify for the SCF. Specifically, I plot the β s from the following regression: j (cid:18) (cid:19) Loans (cid:88) bt = γ +φ ·t+ β ·1 (t)+ε (1) b ZIP3 j j bt Assets bt j where j ∈ {1,...,T}, and the interaction of a time trend with 3-digit ZIP location dummies accounts for any drift in each area’s average portfolio shares. The timing of the Call Reports does not generally line up with the seasonal lending peaks and troughs, but there is a clear pattern of increasing loan issuance through the growing season and pay-downs following the harvest. In sum, rural banks are quite similar regardless of membership status. Although a high degree of similarity between the groups of banks is not necessarily required by the difference-in-differences identification strategy used below, such similarity may increase the plausibility of the parallel trends assumption. Figure5alsosuggestsaneffectoftheSCFoncreditprovision, becausetheloan-to-asset ratio for member banks moves up relative to nonmember institutions after the introduction of the lending facility. To more formally examine the financial effects of the SCF, I estimate the following two-way fixed effects regression with two groups of banks: Y = γ +φ + (cid:88) β (cid:2)1 (t)×1 (b) (cid:3) +ε (2) bt b ZIP3,t j j SCF bt j where Y is either the bank-quarter liquid asset share or loan share of total assets, and bt 1 (b) is an indicator function that identifies banks that drew on the SCF by the end SCF of 1974. Because the decision to request seasonal credit may be endogenous to lending opportunities, equation (2) may offer more of a descriptive rather than causal interpretation. Nevertheless, the ZIP-time fixed effects help control for local economic and credit demand conditions that might otherwise give rise to endogeneity concerns. As shown in Figure 6, the response of SCF borrowers accords with the theory behind the facility. Following the introduction of the facility, SCF borrowers reduced their liquid 16

asset holdings in favor of higher loan shares relative to other banks. The relative increase in loan-to-asset ratios reached more than 4 percentage points in 1974 after deviating by less than one percentage point in the three and a half years before the SCF’s introduction. Because SCF credit is extended only after a consultation with and review by a Reserve Bank, the noticeable divergence of lending and liquidity positions so soon after the comment period suggests that institutions anticipated their eventual use of the facility. This interpretation alignswiththeevidencepresentedinSection2.2.1thatmemberbankswereincommunication with their Reserve Banks throughout the comment period. To facilitate the interpretation of the financial effects, Table 3 reports the average post-implementation effects for banks that tapped the SCF from an estimation of: Y = γ +φ +β (cid:2) Post ×1 (b) (cid:3) +ε . (3) bt b ZIP3,t t SCF bt Theloan-to-assetratiofordrawingbanksincreasedbyabout3percentagepoints. Thenearly exactly offsetting estimate for the liquid asset share indicates that the additional lending cannot be attributed to interference with banks that did not draw on the SCF. Liquid assets including Treasuries and agency debt are either allowed to mature or sold using brokers into national markets. If SCF borrowers simply cannibalized loans from nearby banks, the simultaneous reduction in credit by SCF non-borrowers would result in a much larger estimate for loan shares relative to the decline in liquid asset shares. While it could also be the case that SCF non-borrowers’ liquidity ratios rose mechanically because of a passive decrease in size as SCF borrowers poached loans, this does not appear to be the case. In unreported results, I find that using the natural logarithm of total assets as the outcome variable yields a coefficient estimate of -0.01 (p = 0.17). This pattern of results suggests the absence of interference across banks and therefore supports the validity of the stable unit treatment value assumption.11 11If loans were simply switching between banks operating in the same area, this would likely limit the real effects of the SCF. An assessment of the real effects of the SCF is taken up later in the paper. 17

Theincreaseintheloan-to-assetratioresultedinloangrowththatwasabout6percentage points higher, as seen in the third column of panel A. An increase in loan growth that is twice the size of the increase in the loan-to-asset ratio follows from the average loan-to-asset share of about 50% (Table 1) with no meaningful difference in asset growth. Decomposing the increase in the loan share of assets in panels B and C of Table 3 reveals that commercial and industrial, commercial real estate, and agricultural loan shares all increased by about 10% compared with the member bank averages reported in Table 1.12 The final column in panel C shows that the substitution from liquid assets to loans boosted banks’ interest income margins by about 11 basis points on average. To further address the endogenous decision to borrow from the SCF, I proceed with two additional exercises. First, I estimate intent-to-treat effects with the specification: Y = γ +φ +β (cid:2) Post ×1 (b) (cid:3) +ε (4) bt b ZIP3,t t Mem50 bt where 1 (b) is an indicator function that identifies banks that the Fed expected to Mem50 be the most likely beneficiaries of the SCF. Specifically, the initial SCF proposal included an explanation that the terms of the facility would be particularly attractive to member institutions with less than $50 million in total assets, and a maximum seasonal funding swing that exceeded 5% of deposits. The results in Table 4 broadly confirm those reported in Table 3, although there is now some evidence of an increase in residential real estate loans in place of the commercial real estate lending increase observed earlier. The parameter estimates in Table 4 will be attenuated somewhat because the indicator function includes member banks that would never draw on the SCF. Asasecondexercisetoaddresstheendogenousborrowingdecision,Iestimateaninstrumented difference-in-differences (DiD-IV) system of equations using 2SLS. In this system, the targeted member bank dummy instruments for the seasonal credit borrower dummy, 12Using SCF funds to support additional loans to other financial institutions was prohibited, and the accounts of drawing banks were monitored by the Fed. Consistent with this rule, I achieve a precisely estimated zero (not shown) when I use loans to financial institutions as the outcome variable. 18

both of which are interacted with a post-SCF time indicator: Post ×1 (b) = γ +φ +µ (cid:2) Post ×1 (b) (cid:3) +η (5) t SCF b ZIP3,t t Mem50 bt Y = γ +φ +β (cid:2) Post (cid:92) ×1 (b) (cid:3) +ε . (6) bt b ZIP3,t t SCF bt For this exercise, I collapse the time dimension for each bank to a single observation before and after the implementation of the SCF as in Bertrand et al. (2004). While the results and conclusions are nearly identical without collapsing, the two-period DID-IV with a binary instrument and treatment corresponds to a special case of DiD-IV in which the β coefficient identifies a local average treatment effect as in Imbens and Angrist (1994) so long as familiar difference-in-differences and instrumental variable assumptions are met (de Chaisemartin, 2010; Hudson et al., 2017). The results in Table 5 reveal qualitatively similar results to those obtained above. According to the 2SLS results, however, SCF loans from the Fed helped boost residential real estate loans in addition to C&I and agricultural loans. Because sole proprietorships entangle the personal and business financial positions of the owner, it is possible that residential real estate loans helped support small business activity.13 As I show in the next section, areas exposed to more SCF funding witnessed higher proprietorship growth. The preceding analysis supports the original theory rationalizing the SCF. Geographic areasreliantonhighlyseasonalindustriesproducelargeswingsinliquiditydemandsonbanks that could lead to inefficient liquidity hoarding. These areas also feature highly correlated depositor behavior that can contribute to banks’ precautionary liquidity holdings (Diamond and Dybvig, 1983). The SCF helped eligible banks draw down their liquidity buffers in favor of additional credit that would not have otherwise been issued. The benefits of offering lending facilities so that banks do not need to hold liquidity buffers to safeguard against all contingencies has been a motivating principle in central banking at least since Bagehot 13Estimating DiD-IV regressions using all time periods also yields a statistically significant increase in consumer loans of about 1 percentage point (p=0.019). 19

(1873). The SCF was similarly motivated, with the goal of supporting economic activity by fostering easier credit conditions. Whether such growth materialized following the expansion of credit identified above is an empirical question that I address in the following section. 4 The Real Effects of Seasonal Credit 4.1 Data and Empirical Methods Tomeasuretheeconomiceffectsoftheseasonalcreditfacilityonlocalcommunities,Ifocuson county-level outcomes for a few reasons. First, regulators frequently use county boundaries to demark banking markets for rural areas. Second, branching restrictions in the mid-1970s were often tied to the county or city of a bank’s headquarters. Such stringent branching restrictions at that time increase the likelihood that a bank’s lending activity was directed within its home county. In fact, many of the Midwestern states that received the bulk of SCF loans were unit banking states that prohibited branching entirely (Amel, 1993).14 Third, data availability constraints preclude a finer geographic unit of observation. Table 6 reports summary statistics for rural counties as of December 1972, just prior to the introduction of the SCF. The sample includes only rural counties in the contiguous U.S., excluding counties in the few states that never witnessed SCF draws. Employment and population totals are reported annually by the BEA. To gauge the real effects of subsidized Fedlending, Ifocusonbothproprietors’employmentandtotalemployment. Proprietorships are the most common form of business organization in the U.S. by a wide margin, not least because they require no formal registration with states, are easy to start, and face few of the operational hurdles associated with other business types such as corporations or LLCs. Because proprietorships entwine individuals’ personal legal and financial conditions with 14About 65 percent of states tied branching restrictions to the local county or municipality of a bank’s headquarters. Over 71 percent of states had finer than state-wide branching restrictions. Midwestern states with unit banking included Colorado, Illinois, Iowa, Kansas, Nebraska, North Dakota, Montana, Minnesota, Missouri, Oklahoma, Texas, and Wyoming. Indiana and Wisconsin tied branching restrictions to a bank’s home county. 20

that of their business, proprietors are often dependent on different forms of credit from small banks. This dependence was particularly pronounced during the early 1970s, because little credit was extended via credit cards and nonbanks.15 Establishment counts from the County Business Patterns (CBP) data published by the Census Bureau are reported in the middle of Table 6. The selected industries reported in the table correspond to the most prevalent business types in rural counties, and businesses in these industries account for about 90% of all county establishments on average. The ubiquity of building material/hardware stores, bars and restaurants, and miscellaneous retail shops in ruralcountiesensurethatmanyindividualspossessindustry-specificknowledge, connections, and experience, which can lower staffing and startup hurdles. Many of the building materials establishments—the most common establishments in rural counties by far—supply materials for housing and building construction. Miscellaneous retail stores include book stores, used merchandise purveyors, jewelry stores, hobby shops, and tobacco stores. The disparate businesstypesthatcomposethemiscellaneousretailindustryoperateanabove-averageshare of small establishments. By contrast, I can also conduct falsification or “placebo outcome” tests by examining certain industries that are unlikely to be affected by small bank loans because of high regulatory and capital requirements associated with starting such businesses. For this purpose, I focus on gas stations, banks, and lodging services, all of which are among the most prevalent establishment types in rural counties. The items at the bottom of Table 6 report the average number of banks in each county, along with both the average proportion of those banks that are Fed members and the share of assets controlled by member banks. Appendix B describes some special considerations regarding the CBP establishment data. 15One notable exception is thrift institutions, which were far fewer in number than commercial banks, but accounted for a meaningful share of residential mortgage credit outstanding. At this time, thrifts were largely prohibited from extending all but residential real estate loans. 21

The structural rtrelationship of interest uncovers the effect of SCF loan intensity on employment and establishment growth: SCF (cid:2) c (cid:3) ln(Y ) = γ +φ +λX +β Post × +ε (7) ct c t ct t ct Deposits c where Y is the total employment or establishment count in county c in year t, γ are county ct c fixed effects, φ are year fixed effects, X is a vector of county characteristics, and SCFc t ct Depositsc sums each county’s bank-level maximum draw on the SCF during the recession and divides this amount by total county deposits just prior to the introduction of the SCF. An OLS estimate of β will be biased if SCFc is not orthogonal to ε . The SCF borrowing decision Depositsc ct is likely to be endogenous because banks may face higher credit demand in faster-growing counties. If banks’ lending opportunities are unobservably different across counties, the orthogonality assumption is likely violated. To identify the effect of SCF funding on local economic outcomes, I extract exogenous variation in the intensity of Fed borrowing by using the within-county presence of Fed member banks as an instrument for SCF funding during the recession. As seen in Figure 7, member bank presence in a county is positively related to the county-wide SCF borrowing intensity. Thus, I supplement equation (7) with the first-stage regression: SCF c (cid:2) (cid:3) Post × = γ +φ +κX +δ Post ×MemberAssetShare +η , (8) t c t ct t c ct Deposits c whereMemberAssetShare istheproportionoftotalbankassetsinthecountythataremanc aged by Fed members, measured just before the introduction of the SCF.16 While variation in treatment intensity may help strengthen the case for a causal interpretation, continuous treatments define more treatment effect parameters than binary treatments (Callaway et al., 16The empirical strategy outlined above is methodologically similar to Nguyen (2019), while the use of membership status to identify the effects of Fed policy has been previously used by Park and Van Horn (2015) and Carlin and Mann (2020). As demonstrated in a previous draft of this study, all of the results are robust to using the share of banks in a county that are Fed members in place of Fed member asset shares. 22

2021). However, Callaway et al. (2021) show that δ can be interpreted as a familiar average causal response in the style of Angrist and Imbens (1995) under a parallel trends assumption whereby the evolution of potential outcomes would be similar for counties in the absence of treatment and across treatment doses.17 A key identifying assumption is that county-level distribution of Fed member banks is as good as randomly assigned with respect to local growth prospects, conditional on the controls. Several factors support the validity of this assumption. First, as discussed in Section 2.3, many banks’ membership status was established decades before the time period in question, there were few membership transitions before or during the sample period, and membership rates among state-chartered institutions is not geographically correlated with the degree of SCF borrowing. This indicates that the membership choice would be unrelated to any consideration for a lending facility akin to the SCF. Not only was the request for comment on the seasonal borrowing privilege announced without advance warning in late 1972, but the facility was fundamentally different from other discount window and lender-oflast-resort programs, as detailed in Section 2. The proposed terms of the facility also differed somewhat from those outlined in the Fed’s 1968 reappraisal of discount window credit, which discussed the design of a potential seasonal borrowing program. Second, county-level fixed effects control for any unobserved heterogeneity at the county level that remained constant over time and happens to be correlated with both the historical membership mix and the local economic performance in the 1970s.18 Finally, as shown in Figure 8, Fed membership rates display a wide degree of geographic variation. Channels through which Fed membership status could affect economic outcomes outside of SCF borrowing are few. It is particularly unlikely that the outcome variables are 17AtraditionalparalleltrendsassumptionisstillsufficienttorecoverstandardATT-typeparametersand I recover such a parameter in Appendix A. Also in Appendix A, I conduct an exercise similar to that of Cengiz et al. (2019) to show that treatment effects are not much correlated with observed member bank assetshares,whichsuggestshomogeneityoftreatmentdose/responsefunctionsacrosscountieswithdifferent membership shares. 18Using an instrument for counties’ economic performance in the 1970s recession, I demonstrate in subsection 4.3 that the effect did not materially differ by growth prospects. 23

directly affected by Fed membership. Nevertheless, institutions did face differences based on membership status, even when comparing state-chartered members to state-chartered nonmembers. For example, member institutions enjoyed access to the Fed’s payment services, but this access was constant through time and it is unlikely to explain differences in economic performance in any case. Member institutions were also subject to zero-interest reserve requirements, but these requirements scarcely changed following the announcement of the SCF. Reserve requirements on demand deposits did not change by more than 0.5 percentage points and the reserve requirement on savings deposits did not change at all. One potential concern with equation (9) is that the administration of the discount window through the District Federal Reserve Banks described in Section 2 means that the mapping from membership shares to SCF borrowing can vary systematically across Fed districts over time. If a Reserve Bank is more or less assiduous in its communications with member institutions, this could affect the likelihood that its members draw on the SCF, all else equal. To account for this potential variation across districts, I replace the simple year fixed effects (φ ) in equations (7) and (9) with finer district-year fixed effects (φ ). These t FRS,t district-specific time dummies not only help explain more of the variation in county-level SCF borrowing, but they also serve as better controls for year-to-year changes in regional economic conditions. Another potential concern regarding the equations above is the possibility of a bad control problem stemming from the county-level controls X . To avoid this issue, I measure ct the controls as of 1972 Q4 and interact them with a Post dummy. This method helps t ensure robustness to cross-sectional differences across counties that may be correlated both withmembershipsharesandwitheconomicoutcomesinthepost-treatmentperiod. However, the preferred specifications exclude the county controls entirely. In the following analysis, I highlight the reduced form for two reasons. First, member banks that did not draw on the SCF may still be affected by the establishment of a seasonal borrowing privilege if those banks viewed the SCF as an effective backstop against liquidity 24

risks. Second, thecoefficientestimateiseasiertointerpretastheeffectofmovingfromalocal area where no Fed members supply credit to an area where businesses rely exclusively on Fed members for their credit needs. Alternatively, the coefficients of the reduced form estimates can be divided by 3 to achieve the approximate effect of a one standard deviation change in the member asset share across counties, which roughly translates to one additional member bank on average. To confirm the findings from the reduced form estimates, I additionally report the results from the DiD-IV described in equation (7). Appendix A includes a more thorough discussion of the treatment effect parameters identified under different parallel trends assumptions, and additionally computes ATT-type parameters to demonstrate the robustness of the conclusions from the main analysis. 4.2 The Economic Effects of Seasonal Credit I first show that SCF funding had a positive effect on local employment outcomes in Table 7. Although the principal focus of this analysis is on the qualitative lessons for FFL programs, it is useful to verify that the results are economically significant. In fact, relative to counties with no Fed member banks and thus no SCF loans, counties with only Fed member banks witnessed total employment growth that was about 1.5 percentage points higher in the years following the introduction of the SCF on average. For reference, I present the time pattern of the total employment effect in Figure 9. To facilitate the interpretation of the parallel trends assumption, Figure 9 compares the counties in the highest membership share tercile against those in the lowest tercile. The two groups of counties exhibited nearly identical employment growth in the years before the announcement of the SCF. After the provision of subsidized credit through the recession, however, more exposed counties enjoyed gradually increasing employment growth for a few years before leveling off. Evidently, the additional loans were directed to productive uses and banks did not draw on the SCF merely to lend to nonviable business during the recession. 25

The middle columns of Table 7 reveal that proprietors’ employment expanded at an even more rapid rate in counties that were more likely to receive SCF credit. To verify that the increase in economic activity spurred by the post-SCF credit expansion was not driven solely by the agricultural sector, the rightmost columns show similar effects for proprietors’ employment within non-farm businesses. Anincreaseinproprietors’employmentimpliesanincreaseinbusinessformationwithin counties that were more exposed to SCF loans. The first two columns of Tables 8-10 show that the SCF promoted net establishment growth among building materials and hardware stores, bars and restaurants, and retail shops. The middle and final columns demonstrate that the growth in total establishments within these industries was driven by the smallest establishments. Establishments with few employees are more reliant on loans from the small banks targeted by the SCF. Conversely, larger establishments are less likely to be unincorporated proprietorships and are more likely to have access to internal financing and to have relationships with larger banks. The null results and smaller point estimates for larger establishments also suggest that the effects among small businesses reflect the flow of small business credit and are not simply an artifact of an unrelated improvement in economic conditions in counties with high member bank presence. In total, the results in Tables 8-10 are consistent with expected patterns if the SCF had its intended effect of promoting real activity by boosting credit supply to the customers of small banks. The relatively modest increase in places of business in operation implied by the average effects reported in these tables accords with the total credit provided via the SCF, which was not immense. Because the CBP data include only nonfarm establishments, employment growth may be a better gauge of the real effects of the lending facility. As an additional exercise to help rule out the possibility that stronger establishment growth among SCF-intensive counties reflects a shortcoming in the identification strategy, Table 11 presents several falsifications tests. In these tests, I report the results for three well represented industries within rural counties that should be less responsive to changes 26

in credit provision by smaller banks because of heavy regulations and steep initial capital requirements. In contrast to the earlier results, establishments in these industries show no clear response to the introduction of the SCF. As with the firm size results, the distributional patternsacrossindustrieshelpruleoutthepossibilitythatthestrongerperformanceisdriven by some unobserved factor correlated with both the predetermined county-level membership shares and economic activity in the mid-1970s. In Appendix A I show that all of these results hold for a sample of only the most rural counties. Further, I compute an ATT parameter for each outcome by retaining only counties with the lowest and highest member bank asset shares and assigning a binary treatment status to each group. Appendix A also reports the results from the DID first introduced in de M Chaisemartin and d’Haultfoeuille (2020), which computes treatment effects that are robust toheterogeneouseffectsandpermitscontinuousmeasuresoftreatment(deChaisemartinand D’Haultfœuille, 2021). All permutations of the main results reveal similarly robust effects of the SCF on real activity at the county level. 4.3 The Cyclical Asymmetry of the Effects of Seasonal Credit An important question regarding the efficacy of any monetary policy tool is whether the tool’s potency depends on the state of the economy. This question is particularly salient for unconventional policy tools introduced during recent recessions because central banks maintained their reliance on these tools well into the subsequent recoveries. Recent literature highlights the state-dependence of monetary stabilization (Jorda` et al., 2020) as well as the concern that certain unconventional policies may be ineffective after non-financial shocks (Karadi and Nakov, 2020). TheforegoingevidenceregardingtheefficacyofaFFLprogrammayonlyapplytoweak economic environments because the introduction of the SCF was followed by a long recession and slow recovery. Conditioning on the dependent variable could invalidate inference based 27

on a naive sub-sample analysis that simply divides counties by growth rates, which I have already shown depend upon SCF take-up. To assess the efficacy of FFL programs across the economic cycle, I sort counties using an instrument for their post-1973 growth prospects. OAPEC’s oil embargo in October of 1973 triggered a dramatic rise in the price of oil that choked off a robust expansion and helped tip the economy into a recession that lasted nearly 18 months. However, counties with a petroleum extraction industry stood to benefit from higher oil prices. As shown in Table 12, these counties witnessed much stronger employment and income growth compared with counties that had no oil extraction industry. InTable13,Ishowthereducedformestimates(forbrevity)ofthekeyoutcomevariables when including an interaction of a petroleum production dummy (Petro ) with the key c regressor. All specifications additionally include a Post × Petro interaction term. In t c general, the coefficient estimate on the triple interaction term is a noisy zero, indicating that the economic benefits of the SCF did not vary systematically across different growth environments. The point estimates alone mostly indicate that, if anything, the SCF had somewhat stronger effects in counties experiencing growth booms. 4.4 The Net Effect on Consolidated Government Finances Potential financial losses associated with unconventional policies can result in a reluctance to pursue programs that would otherwise benefit the economy. For example, the possibility of losses was cited by policymakers as a potential cost of QE (Bernanke, 2012). Although economic benefits such as those identified above should be the key determining factor in the decision to adopt a policy, the potential for financial losses may weigh on policymakers who also consider risks to the central bank’s independence and reputation. For the SCF, the subsidy offered by the Fed relative to market rates may be viewed as a cost of the facility, even if the loans are extended through the issuance of unremunerated reserves. 28

For this reason, I evaluate the narrow fiscal implications of the SCF to test whether a successful FFL program can generate tax revenues that surpass the “costs” of the subsidy. In Table 14, I report the results from equation (7) as well as its reduced form. To more easily compare tax revenues against the dollar cost of the subsidy offered by the Fed, dependent variables are expressed in levels. PanelsAandBrevealthattheSCFboostedtotalwageandsalaryemployment, withan associated increase in total wages and salaries within the counties. Focusing on the second specification, both the reduced form and the 2SLS results point to an average salary per additional job of about $17,500. Dividing the coefficient estimates for contributions to social insurance programs by that for total wages and salaries shows that about 14% was passed on in the form of payroll taxes. This figure conforms with the total payroll taxes of 11.7% at the time, plus the average state and federal unemployment insurance taxes of about 2.5%. A full accounting of phaseouts and lower rates for certain types of workers would yield a more precise number, but it is clear that the estimates in Table 14 are consistent with prevailing tax rates and average salaries of the time. The total cost of the subsidy (calculated using the values from Figures 1 and 2) point to a total subsidy of about $5 million between the introduction of the SCF and the end of the recession in 1975. The value of the contributions to government social insurance funds implied by the final coefficient estimate in panel B of Table 14 suggests that the jobs created by the SCF in the average affected county resulted in about $1 million per year in payroll taxes. Even in the event of bank defaults, the additional economic activity supported by the SCF-fueled credit expansion would more than offset the subsidy provided by the Fed. 5 Policy Implications and Discussion Directingmonetarysupporttobusinessposesasubstantialchallengeforcentralbanks, which may not be able to legally or practically offer either direct grants or loan guarantees. This 29

challenge is multiplied when the goal is to direct credit towards small and medium sized businesses, because these firms lack access to financial markets and will not benefit from LOLR backstop facilities that can ensure credit continues to flow through capital markets. One option, which the Fed employed during the COVID-19 crisis, is to purchase loan participations from banks if the loans meet certain criteria. Another possible option is to subsidize bank loans through a FFL program. Both options necessitate a reliance on financial institutions as a practical matter, because central banks do not have the capacity to make credit decisions on a large scale basis, particularly for more opaque small business. FFLprogramscanensurethatmonetarystimulustransmitsthroughthebankingsector even if the main policy rate remains at its effective lower bound. In contrast to the effects of negative policy rates on banks’ assets, negative rates on loans from the central bank should improve the health of the banking sector and thus promote credit growth. Besides opening up additional policy space at the lower bound, FFL also allows the central bank to incur contingent losses during severe downturns. When adverse shocks generate real economic losses that must be borne regardless of financial market conditions—for example, as in the COVID-19 crisis (Hanson et al., 2020)—FFL facilities offer the central bank the ability to shoulder a share of those losses by setting the lending rate well below the rate paid on reserves. To be successful, a FFL program should be offered on terms that are sufficient to induce meaningful participation by banks. The main factor involved in banks’ decision to request subsidized credit from the central bank is the size of the subsidy. For this reason, FFL programs can be particularly successful in offsetting abnormally high bank funding costs (English and Liang, 2020). However, sufficient compensation via funding subsidies can also work to overcome banks’ balance sheet constraints and excessive risk aversion. Because banks originate the loans and keep them on their balance sheets, FFL programs face fewer complications related to asymmetric information and adverse selection. Consequently, FFL programs do not necessarily require that borrowers meet a litany of eligibility criteria, which 30

in turn increases the likelihood of broad participation and benefits to a variety of small businesses.19 As with other programs that encourage lending to small businesses, FFL programs should result in loans that would not have otherwise been originated. Because of the unobserved counterfactual, assessing the extent of additional lending spurred by an FFL facility can pose a challenge to policymakers, absent some randomization embedded in the design of the facility. The results above deliver an encouraging message. SCF-eligible banks boosted credit relative to ineligible banks operating in the same communities, and an increase in SCF funding corresponded to better economic outcomes. The results also offer no evidence to support the concern that highly subsidized credit might be directed towards unproductive loans that serve to temporarily prop up nonviable businesses. Ensuring that banks direct the additional lending towards institutions or areas that the central bank wishes to support can be a key consideration for an FFL program. If sufficiently attractiveFFLtermsareconditionedonlendingtoparticulartypesofbusinessesorlocations, banks will almost certainly devise a method to certify compliance with the requirements. Some types of targets, including specific geographic areas or demographic characteristics of business owners may be easier to verify than other targets, such as businesses most affected by a pandemic, but a certification protocol is still feasible. In terms of the SCF, policymakers hopedtotargetcredit-starvedbusinessesandindividualsinareassubjecttolargeseasonality. IfindthattheSCFwassuccessfullyabletotargettheseareas,andcreditwasdirectedtowards non-agricultural industries including restaurants and bars. The ability to fine-tune the targets of a FFL program presents policymakers with a policy tool that is far less blunt than both conventional tools and more common unconventional tools. As mentioned above, FFL programs may be used to target particular geographic areas, demographic groups, or industries. Although efforts to promote distributional equality 19Another example of FFL program’s relative flexibility is the fact that they may be structured through thecentralbank’sexistingdiscountwindowauthority,whicharesubjecttoestablishedregulationsthatmay be flexibly adjusted. In contrast, loan participation programs such as the Fed’s MSLP may be subject to additional constraints imposed by its emergency (Section 13(3)) lending classification. 31

may be best addressed through targeted fiscal policies, central banks can introduce some measure of specificity in their objectives through careful design of a FFL program. Some of the most vociferous criticisms of FFL programs concern the quasi-fiscal component of the stimulus. Ensuring a separation between fiscal and monetary policy can be an integral aspect of central bank independence, so these concerns should not be minimized. However, considerations surrounding credit allocation and fiscal encroachment are not completely foreign to other unconventional policies. Even simple interest rate adjustments can disproportionately affect rate-sensitive industries. Purchases of mortgage-backed debt have long been criticized by policymakers that disfavor the credit-allocation effects of such purchases. Purchasing corporate debt involves taking credit risk in addition to the interest rate risk introduced with all QE programs, and such purchases can generate backlash if they are seen to favor some firms or industries. In fact, many central bank policies may favor larger and safer borrowers over smaller borrowers insofar as the former can access capital markets that benefit most from LOLR backstop facilities, direct purchases, and portfolio balance effects. Borrowers that depend on bank credit can be cut off from the support and stimulus offered by monetary actions if credit spreads at banks remain high and credit provision low, even as benchmark rates are depressed by QE, forward guidance, and other measures. Even with limited ability to further adjust the level or path of the main policy rate, an FFL program can ensure that monetary stimulus continues to transmit to needy borrowers through the banking system. Efforts to improve the flow of credit to small, bank-dependent businesses can help ensure that unconventional policies are not predominantly benefiting borrowers with access to securities markets. There is a pervasive view that QE and other central bank policies are primarily directed at large corporations and large financial institutions on Wall Street. FFL programs and other facilities that target small- and medium-sized businesses can help countervail that perception because they improve credit access for bank-dependent firms with productive investment opportunities. 32

6 Conclusion Asmoderncentralbanksfacetheeffectivelowerboundwithmoreregularity, itisincreasingly important to assess new policy tools that may help manage downturns and foster robust recoveries. Some unconventional tools—such as QE and forward guidance—saw widespread adoption during the global financial crisis, and many central banks actively employed such policies throughout the decade-long recovery that followed. A large literature sprung up to evaluate the effficacy of these tools, but far more work remains to understand the potency and transmission channels of less tested unconventional tools. In this paper, I evaluate the success of a de facto funding-for-lending program wherein the Fed offered some banks subsidized funding with the goal of promoting credit growth within the banks’ local communities. I find that eligible banks increased their credit supply to small businesses after receiving the cheap central bank funding. The additional credit supported employment growth, entrepreneurship, and net business creation across a variety ofsectors. Theeconomicbenefitsofthecreditmaterializedacrossabroadsetofcommunities, and was not limited to areas with the best growth environment. In sum, the foregoing analysis suggests that funding-for-lending programs warrant serious consideration for inclusion in the modern central banker’s toolkit. Not only can these programs flexibly target “Main Street” businesses that may benefit less directly from other unconventional policies, but they can be creatively adapted to aid in an effort to attain other targeted objectives. 33

References Afonso, Anto´nio, andJoanaSousa-Leite, 2019, Thetransmissionofunconventionalmonetary policy to bank credit supply: evidence from the TLTRO, REM Working Paper 065–2019. Amel, Dean F, 1993, State laws affecting the geographic expansion of commercial banks . Anderson, Haelim, Charles W Calomiris, Matthew Jaremski, and Gary Richardson, 2018, Liquidity risk, bank networks, and the value of joining the federal reserve system, Journal of Money, Credit and Banking 50, 173–201. Andrade, Philippe, Christophe Cahn, Henri Fraisse, and Jean-St´ephane M´esonnier, 2019, Can the provision of long-term liquidity help to avoid a credit crunch? evidence from the eurosystem’s ltro, Journal of the European Economic Association 17, 1070–1106. Angrist, Joshua D, and Guido W Imbens, 1995, Two-stage least squares estimation of average causal effects in models with variable treatment intensity, Journal of the American statistical Association 90, 431–442. Bagehot, Walter, 1873, Lombard street (Henry S. King & Co.). Benetton, Matteo, and Davide Fantino, 2018, Competition and the pass-through of unconventional monetary policy: evidence from TLTROs, Bank of Italy Temi di Discussione (Working Paper) No 1187. Bernanke, Ben S, 2012, Opening remarks: monetary policy since the onset of the crisis, in Proceedings: Economic Policy Symposium Jackson Hole, volume 1, 22. Bernanke, Ben S, 2020, The new tools of monetary policy, American Economic Review 110, 943–83. Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, 2004, How much should we trust differences-in-differences estimates?, The Quarterly Journal of Economics 119, 249– 275. Blinder, Alan, Michael Ehrmann, Jakob De Haan, and David-Jan Jansen, 2017, Necessity as the mother of invention: Monetary policy after the crisis, Economic Policy 32, 707–755. Board, Federal Reserve, 1930, The dual banking system in the united states, Technical report, Board of Governors of the Federal Reserve System (US). Brunnermeier, Markus, and Arvind Krishnamurthy, 2020, Corporate debt overhang and credit policy, in BPEA conference. Callaway, Brantly, Andrew Goodman-Bacon, and Pedro HC Sant’Anna, 2021, Difference-indifferences with a continuous treatment, arXiv preprint arXiv:2107.02637 . 34

Campbell, Donald T, and Julian C Stanley, 2015, Experimental and quasi-experimental designs for research (Ravenio Books). Carlin, Bruce, and William Mann, 2020, The real effects of fed intervention: Revisiting the 1920-1921 depression, Working Paper . Carlson, Melissa DA, and R Sean Morrison, 2009, Study design, precision, and validity in observational studies, Journal of Palliative Medicine 12, 77–82. Cengiz, Doruk, Arindrajit Dube, Attila Lindner, and Ben Zipperer, 2019, The effect of minimum wages on low-wage jobs, The Quarterly Journal of Economics 134, 1405–1454. Churm, Rohan, Michael Joyce, George Kapetanios, and Konstantinos Theodoridis, 2018, Unconventional monetary policies and the macroeconomy: The impact of the uk’s qe2 and funding for lending scheme, The Quarterly Review of Economics and Finance . Churm, Rohan, Amar Radia, Jeremy Leake, Sylaja Srinivasan, and Richard Whisker, 2012, The funding for lending scheme, Bank of England Quarterly Bulletin Q4. Clouse, James A, 1994, Recent developments in discount window policy, Fed. Res. Bull. 80, 965. Da Silva, Emilie, Vincent Grossmann-Wirth, Benoit Nguyen, and Miklos Vari, 2021, Should we pay banks to lend? evidence from the Eurosystem’s TLTRO and the euro area credit registry, Working Paper, Banque de France . de Chaisemartin, Clement, 2010, A note on instrumented difference in differences, Unpublished Manuscript, University of Warwick . de Chaisemartin, Cl´ement, and Xavier d’Haultfoeuille, 2020, Two-way fixed effects estimators with heterogeneous treatment effects, American Economic Review 110, 2964–96. de Chaisemartin, Cl´ement, and Xavier D’Haultfœuille, 2021, Difference-in-differences estimators of intertemporal treatment effects, Available at SSRN 3731856 . de Chaisemartin, Cl´ement, and Xavier d’Haultfoeuille, 2018, Fuzzy differences-in-differences, The Review of Economic Studies 85, 999–1028. Diamond, Douglas W, and Philip H Dybvig, 1983, Bank runs, deposit insurance, and liquidity, Journal of Political Economy 91, 401–419. English, William B, and J Nellie Liang, 2020, Designing the main street lending program: Challenges and options, Technical report, Hutchins Center Working Paper. Gros, Daniel, Diego Valiante, and Willem De Groen, 2016, The ecb’s latest gimmick: Cash for loans, CEPS Policy Brief . 35

Hanson, Samuel, Jeremy Stein, Adi Sunderam, and Eric Zwick, 2020, Business credit programs in the pandemic era, Brookings Papers of Economic Activity . Hudson, Sally, Peter Hull, and Jack Liebersohn, 2017, Interpreting instrumented differencein-differences, Metrics Note, Sept . Imbens, Guido W,andJoshuaDAngrist, 1994, Identification andestimationoflocalaverage treatment effects, Econometrica 467–475. Jasova, Martina, Caterina Mendicino, and Dominik Supera, 2020, Policy uncertainty, lender of last resort and the real economy, Journal of Monetary Economics . ` Jorda`, Oscar, Moritz Schularick, and Alan M Taylor, 2020, The effects of quasi-random monetary experiments, Journal of Monetary Economics 112, 22–40. Karadi,Peter,andAntonNakov,2020,Effectivenessandaddictivenessofquantitativeeasing, Journal of Monetary Economics . Krueger, Leonard Bayliss, 1933, History of Commercial Banking in Wisconsin (Ayer Publishing). Laine, Olli-Matti, 2021, The effect of targeted monetary policy on bank lending, Journal of Banking and Financial Economics . Madigan, Brian F, and William R Nelson, 2002, Proposed revision to the federal reserve’s discount window lending programs, Federal Reserve Bulletin 88, 313. Miron, Jeffrey A, 1986, Financial panics, the seasonality of the nominal interest rate, and the founding of the fed, The American Economic Review 76, 125–140. Nguyen, Hoai-Luu Q, 2019, Are credit markets still local? Evidence from bank branch closings, American Economic Journal: Applied Economics 11, 1–32. Park, Haelim, and Patrick Van Horn, 2015, Did the reserve requirement increases of 1936–37 reduce bank lending? evidence from a quasi-experiment, Journal of Money, Credit and Banking 47, 791–818. Woodford, Michael, 2014, Monetary policy targets after the crisis, What Have We Learned?: Macroeconomic Policy after the Crisis 55. Yellen, Janet L., 2009, US monetary policy objectives in the short and long run, Federal Reserve Bank of San Francisco. 36

Appendix Appendix A Supplemental Analysis and Robustness In this Appendix, I present supplemental analysis to that reported in the main text. Appendix A.1 Discount Window Substitution Figure A1 shows that banks borrowing from the SCF did not merely substitute out of other discount window credit supplied by the Fed. Banks drawing on the SCF did not differ materially in their adjustment credit borrowing either before or after the introduction of the SCF. Besides the nearly universal null estimates reported in Figure A1, the point estimates also reveal negligible effects. For context, the maximum difference of about 0.1% of assets compares with the 5% average draw among SCF borrowers reported in Table 1. Appendix A.2 Ultra-Rural County Subsample Table A1 reports the key results for a sample of ultra-rural counties only. Table A1 excludes certain ancillary results for the sake of brevity, but all patterns and conclusions from the main results hold for this subsample. For example, larger establishments exhibit much less sensitivity to SCF exposure and none of the coefficient estimates are statistically different from zero. Similarly, establishment growth in placebo industries demonstrates no response to SCF exposure. Appendix A.3 Randomization Inference In Figure A2, I depict p-values for the main reduced form results from Tables 7-11 and Table 14 computed via randomization inference. Each point represents a p-value for a coefficient estimate reported in the main tables, with the cluster robust derived p-value on the horizontal axis and the corresponding p-value obtained via randomization inference on the vertical axis. I resample from the actual realizations of member asset share 1,000 times. In most instances, the randomization inference p-value lies below the 45-degree line, suggesting tighter confidence intervals. Sampling membership asset shares from a random 37

uniform distribution yields nearly the same results, though I do not present these results for brevity and because of the strong restriction that this sampling procedure imposes on the rerandomization process. All but one of the points in the second panel corresponds to the null results from the falsification/placebo outcome tests of large establishment growth or growth of total establishments within high-barrier industries. None of the randomized inference p-values in the second panel are less than 0.05. The lone null result for a nonfalsification test (the second specification for total bar and restaurant establishment growth in Table 9) achieves a p-value of 0.087 using randomization inference. Appendix A.4 Within-County Evaluation of Financial Effects In Table A2, I report estimates of the financial effects using exposed counties as the geographic unit of interest. The exercise in Table A2 corresponds to that carried out at the ZIP3-level in Table 4. The results are similar across both tables, though the county-level estimates reported here are generally larger in absolute value. However, the ZIP3 geographic unit is more appropriate for this exercise, which expresses a comparison between treated and untreated units within a given area at each point in time. The ZIP3 level of analysis is sufficiently large to ensure that most areas include a number of untreated banks, yet small enough to capture a similar lending environment (see, for example, Figure 5). In contrast, about 23% of exposed counties do not host a single nonmember institution, and 21% of counties host only a single nonmember institution. Appendix A.5 Variation in Treatment Intensity, Treatment Effect Heterogeneity, and Computing Average Treatment Effects As Callaway et al. (2021) discuss at length, continuous measures of treatment can identify multiple treatment effect parameters, including the average treatment effects and average causal responses (Angrist and Imbens, 1995), which are not distinct concepts for a binary treatment. The main text invokes a “strong” parallel trends assumption whereby the evolution of potential outcomes is similar even for counties with different treatment intensities. Under a traditional parallel trends assumption and heterogenous treatment dose/response 38

functions across counties, the TWFE estimation method in the main text can mix together the average causal responses with the average treatment effects for counties treated with different exposure to the SCF (i.e. membership shares). However, the average treatment effect on the treated (ATT) counties with similar member bank asset shares is nonparametrically identified under a traditional parallel trends assumption on untreated potential outcomes. In Table A3 I compute the ATT by retaining only those counties with member bank asset shares in the highest (avg. MemberAssetShare = 86%) and lowest (avg. c MemberAssetShare = 0%) terciles, assigning a binary treatment/control treatment indic cator to the respective groups, and estimating: (cid:2) (cid:2) (cid:3) ln(Y ) = γ +φ +β Post ×Treat ]+κ Post ×X +ε . (9) ct c FRS,t t c t c ct The results reported in Table A3 confirm the positive effects of local SCF credit exposure on real economic activity. Although the interpretation of this causal parameter differs from the parameter described in the main text, the primary goal of this study is to establish generalizable lessons about the relationship between targeted central bank credit and real outcomes. Retrieving precise elasticities or average causal responses is less important in this setting because such estimates would not likely generalize, and the precise estimates could not be used to form the basis of policy. Instead the main purpose of the analysis is to evaluate whether the mechanisms crucial to the success of FFL programs are operational in the United States. On this score, both the average causal responses identified under appropriate assumptions in the main text and the ATTs computed here support the rationale underlying FFL. To assess treatment effect heterogeneity across SCF exposure, I carry out an exercise similar to Cengiz et al. (2019). Specifically, Figure A3 reports the estimate of the causal response parameter (κ) from reduced form regressions of employment (panel a) and building material establishments (panel b) for different bins of MemberAssetShare . The evidence c presented in Figure A3 suggests that the causal response of employment declines slightly in 39

memberbankassetshares, butthatthecausalresponseofestablishmentgrowthisessentially uncorrelated with member asset shares. Finally, Table A4 reports the main results using the DID estimator proposed by de M Chaisemartin and d’Haultfoeuille (2018) and de Chaisemartin and d’Haultfoeuille (2020), which accommodate non-binary treatments in unstaggered designs by comparing groups switching treatment between periods to a control group whose treatment does not change, but with the same treatment as the switching groups in the pre-treatment period. In the context of the SCF, counties without Fed member institutions compose the control groups, andthecontrolgroupcanbeexpandedtoincludetherelativelyfewcountieswithlowmember asset shares. As described in de Chaisemartin and D’Haultfœuille (2021), these estimators can be used with a multivalued treatments even when treatment adoption is not staggered. Under a parallel trends assumption, the DID is an unbiased and consistent estimate of the M average treatment effect among switchers.20 To ease computational burden for this exercise, I collapse each county’s data into pre- and post-SCF averages. Consistent with the conclusions from the approaches above, Table A4 reports statistically significant and positive effects of subsidized Fed credit access on local outcomes. Appendix B Data Appendix In this Appendix I describe data sources, variable definitions, and miscellaneous considerations relevant to the data used in the main analysis in order to promote replication. Table B1 lists key bank-level variables with their item numbers from the Call Reports (FR 105 and FR 107, cross-listed as FFIEC 010 and 011). Bank borrowing at the discount window is reported daily and derived from the Fed’s Transmission of Edited Deposits System, form FR 414a. For each bank, item number 2867 reports seasonal credit borrowing, item 2865 reports total borrowings from Federal Reserve Banks, and item 2200 provides total deposits. The sample includes commercial banks only. The analysis makes no adjustment for mergers to no practical effect: only about a dozen banks were involved in mergers during the 20All of these estimators can be computed by the did multiplegt Stata package, available from the SSC repository. 40

sample period. Whenever a bank’s location is missing in the NIC database, I used the 1975 Summary of Deposits data to identify the bank’s zip code. County-level variables aside from establishment counts were sourced from the BEA Regional Economic Accounts database. Rural counties are identified according to the Rural- Urban continuum codes produced by the USDA’s Economic Research Service using the 1975 vintage. In particular, the definitions produced in “Social and Economic Characteristics of the Population in Metro and Nonmetro Counties: 1970,” by Hines, Brown, and Zimmer (Agricultural Economic Report No. 262). Rural counties exclude those in metropolitan (SMSA) areas, and urbanized SMSA-adjacent counties, most of which transitioned to SMSA counties by the time of the release of the next vintage. I define ultra-rural counties as those with a population of less than 2,500 or counties with fewer than 20,000 inhabitants that are not adjacent to SMSAs. EstablishmentcountsarereportedbytheU.S.CensusviatheCountyBusinessPatterns (CBP) data. I favor establishment count data (over, e.g. employment data) in the CBP, because establishment/reporting unit counts are not considered disclosure information and are thus never suppressed. Data suppression in the CBP data can be prevalent among ultra-rural counties where other CBP series may be concealed for confidentiality reasons. The CBP data collection and reporting underwent a change in 1974, which generated two issues pertinent to the analysis in Section 4. First, the Building Materials and Hardware Stores (SIC 5200) industry was subject to administrative shuffling so that certain sub-SICs were excluded as of 1974 (5220, 5240, 5252) while others were included (5260, 5270). To harmonize the parent 5200 SIC, I exclude all of the constituent sub-SICs that transitioned as a result of the reclassification. This change is made to little effect. By way of example, the average establishment count for this industry in rural counties reported in Table 6 is only 1.2 units (1%) higher using the unadjusted CBP data. Second, prior to 1974, CBP data were tabulated in terms of reporting units rather than establishments. For manufacturing companies, each individual location of a company is counted as a separate reporting unit. Consequently, reporting units are the same as the establishments reported as of 1974. For nonmanufacturing industries, multi-location employers are counted once in each county for 41

each industry in which they operate, and multi-location employers in a given county may control more establishments in a given county compared to pre-1974 reporting units. It is unclear whether this poses an issue for the main analysis, because most firms operate only oneestablishment. Amongmulti-locationfirms, thereissomechancethattheestablishments are in separate counties. The focus on rural counties also likely demotes the relevance of the reporting change because of the lower overall economic activity. However, if (small) multilocation employers are distributed across counties in a manner that is correlated with bank membership rates, the resulting increase in reporting units may incorrectly be attributed to the incidence of Fed membership (and by extension, the rate of takeup at the SCF). It is unclear why multi-location employers would not be randomly distributed with respect to membership shares, and the inclusion of county-level fixed effects offer some measure of control for unobserved factors that influence the rates of both. Nevertheless, I perform a test that can help identify whether counties with high membership shares also host more multi-location employers. Specifically, I regress the 1973-1974 change in reporting units on a second-order polynomial of the consistently reported employment total from the BEA and save the residuals. Counties with disproportionately high rates of multi-location employers will thus have positive residuals, and vice versa. The correlation of residuals from this regressionwithcounty-levelmembershipsharesisonly0.0155(p = .53)and0.0181(p = 0.46) for member asset-shares. This exercise supports the notion that multi-location employers are distributed randomly with respect to Fed membership rates. 42

200 180 SCF Introduced 160 140 120 100 80 60 40 20 0 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7 7 7 7 7 7 7 7 7 7 7 7 9 9 9 9 9 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 1 1 1 1 1 / / / / / / / / / / / / / 2 6 2 6 2 6 2 6 2 6 2 6 2 1 0 1 0 1 0 1 0 1 0 1 0 1 )snoillim ,$( snaoL ytilicaF tiderC lanosaeS Figure 1: Total Seasonal Credit Facility Lending. Notes: This figure plots the weekly volume of discount window loans granted to member banks as part of the Seasonal Credit Facility. Source: Board of Governors of the Federal Reserve System (Transmission of Edited Deposits System, form FR 414a). 43

5 SCF Introduced 4 3 2 1 0 −1 −2 0 1 1 2 2 3 3 4 4 5 5 6 6 7 7 7 7 7 7 7 7 7 7 7 7 7 9 9 9 9 9 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 1 1 1 1 1 / / / / / / / / / / / / / 2 6 2 6 2 6 2 6 2 6 2 6 2 1 0 1 0 1 0 1 0 1 0 1 0 1 )tnecrep( etaR tnuocsiD etaR sdnuF laredeF evitceffE Figure 2: Subsidy offered through the Seasonal Credit Facility. Notes: This figure plots the effective subsidy offered to banks that borrowed from the Seasonal Credit Facility for each month. The subsidy is defined as the effective federal funds rate (averaged over days within each month) minus the adjustment credit (discount) rate established by the Federal Reserve Bank of New York. Source: Annual Statistical Digest; Board of Governors of the Federal Reserve System (H.15). 44

Figure 3: Seasonal Credit Facility loans by county as a percentage of total countywide deposits. Notes: This figure depicts the peak county-level discount window borrowing between 1973 and 1974, expressed as a percentage of total county-wide deposits. Counties are binned accordingtothevaluesshowninthelegend. County-levelborrowingiscomputedbysumming the maximum borrowing across banks within each county and year, and dividing by the beginning-of-year deposits. Source: Board of Governors of the Federal Reserve System (Transmission of Edited Deposits System, form FR 414a), (form FR-105); National Information Center 45

8 7 6 5 4 3 2 1 0 Urban Rural Ultra-Rural Counties Counties Counties (H :∆=0; t=4.16) (H :∆=0; t=1.52) (H :∆=0; t=3.27) 0 0 0 )%( stisopeD latoT / gniworroB deF Discount Window Borrowing Seasonal Credit Borrowing Figure 4: Discount Window Borrowing Relative to Total Deposits. Notes: This figure plots peak county-level discount window borrowing between 1973 and 1974, expressed as a percentage of total county-level deposits. Solid blue circles represent Adjustment Credit borrowing, and open squares represent borrowing from the Seasonal Credit Facility. County-level borrowing is computed by summing the maximum borrowing across banks within each county and year, and dividing by the beginning-of-year deposits. A t test of the equality of means between Adjustment Credit and Seasonal Credit borrowing intensityisreportedinparenthesesbeneatheachcountygrouping. Ruralcountiesaredefined as non-MSA counties with a total urban population of fewer than 20,000 people unless the county is not adjacent to an MSA. Ultra-rural counties are those with a total urban population of fewer than 2,500 people or counties with a total urban population of fewer than 20,000 that are not adjacent to an MSA. Source: Board of Governors of the Federal Reserve System (Transmission of Edited Deposits System, form FR 414a); United States Department of Agriculture. 46

SCF Comment 0.04 Period 0.02 0 −0.02 −0.04 0 0 1 1 2 2 3 3 4 4 7 7 7 7 7 7 7 7 7 7 9 9 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 1 1 / / / / / / / / / / 6 2 6 2 6 2 6 2 6 2 0 1 0 1 0 1 0 1 0 1 stessA latoT fo erahS naoL Member banks Nonmember banks Figure 5: Seasonality of Banks’ Loan Portfolios. Notes: This figure plots the β coefficients from an estimation of the following specification j for zip codes with sufficient seasonality to justify SCF loans: (cid:0)Loans bt (cid:1) = γ +φ ·t+ (cid:80) β ·1 (j)+ε . Assets b ZIP3 j t bt bt j Source: Board of Governors of the Federal Reserve System (Transmission of Edited Deposits System, form FR 414a); National Information Center. 47

0.05 0.04 0.03 0.02 0.01 0 −0.01 −0.02 −0.03 SCF Comment −0.04 Period −0.05 0 0 1 1 2 2 3 3 4 4 7 7 7 7 7 7 7 7 7 7 9 9 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 1 1 / / / / / / / / / / 6 2 6 2 6 2 6 2 6 2 0 1 0 1 0 1 0 1 0 1 stessA latoT fo erahS Liquid Assets Total Loans Figure 6: Bank Balance Sheet Allocation Over Time. Notes: This figure plots the β parameters from an estimation of the following specification j for zip codes with sufficient seasonality to justify SCF loans: Y = γ +φ + (cid:80) β (cid:2)1 (t)×1 (b) (cid:3) +ε . bt b ZIP3,t j j SCF bt j Y is either the bank-quarter liquid asset share of assets (solid blue dots) or loan share of bt assets (open red squares). 95% confidence intervals accompany each point estimate. Source: Board of Governors of the Federal Reserve System (form FR-105); National Information Center. 48

1.6 1.4 1.2 βˆ=1.17∗∗∗ SE =0.10 β 1 0.8 0.6 0.4 0.2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 County Member Asset Share )%( stisopeD latoT / tiderC lanosaeS (a) All rural counties 8 βˆ=5.37∗∗∗ SE =1.11 β 6 4 2 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 County Member Asset Share )%( stisopeD latoT / tiderC lanosaeS (b) Rural counties with Seasonal Credit > 0 Figure 7: The Relationship Between Seasonal Credit Use and Asset Share of Fed Member Banks. Notes: This figure shows the relationship between county-level SCF borrowing and the share of assets held by Federal Reserve member banks in each county using a binned scatter plot. Each point represents an average across counties within each bin. Panel (a) includes all rural counties, and Panel (b) includes only the subset of counties that received at least one SCF loan. County-level borrowing is computed by summing the maximum borrowing across banks within each county between 1973 and 1974, and dividing by the beginning-of-period deposits. 49 Source: Board of Governors of the Federal Reserve System (Transmission of Edited Deposits System, form FR 414a); National Information Center.

(a) Share of member banks in each county (b) Share of member bank assets in each county Figure 8: Federal Reserve membership shares by county. Notes: This figure depicts the percent of commercial banks in each county that are members of the Federal Reserve System as of 1973 Q1 in panel (a). Panel (b) depicts the percent of bank assets in each county that are controlled by member banks. Counties with no commercial banks are coded as zero. Source: National Information Center. 50

0.03 SCF Introduced 0.02 0.01 0 −0.01 9 0 1 2 3 4 5 6 6 7 7 7 7 7 7 7 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 htworG tnemyolpmE (a) Total employment SCF 0.04 Introduced 0.03 0.02 0.01 0 −0.01 −0.02 9 0 1 2 3 4 5 6 6 7 7 7 7 7 7 7 9 9 9 9 9 9 9 9 1 1 1 1 1 1 1 1 htworG tnemyolpmE (b) Nonfarm proprietors’ employment Figure 9: Employment Growth Effects Over Time. Notes: This figure plots the β parameters from an estimation of the following county-level j regression: ln(employment ) = γ +φ + (cid:80) β (cid:2)1 (t)×TopMemberShareTercile (cid:3) +ε . ct c FRS,t j j c ct j for both total employment (panel a) and nonfarm proprietors’ employment (panel b). 51

0.5 SCF 0.4 Introduced 0.3 0.2 0.1 0 −0.1 −0.2 −0.3 −0.4 −0.5 0 1 2 3 4 7 7 7 7 7 9 9 9 9 9 1 1 1 1 1 )stessa fo %( gniworroB tiderC tnemtsujdA Figure A1: Adjustment Credit Borrowing Over Time. Notes: This figure plots the β parameters from an estimation of the following specification j for zip codes with sufficient seasonality to justify SCF loans: (cid:0)AdjustmentCreditLoans bt (cid:1) = γ +φ + (cid:80) β (cid:2)1 (t)×1 (b) (cid:3) +ε . Assets b ZIP3,t j j SCF bt bt j Source: Board of Governors of the Federal Reserve System (Transmission of Edited Deposits System, form FR 414a); National Information Center. 52

0.02 0.015 0.01 0.005 0 0 5 1 5 2 0 0 1 0 0. 0 0. 0. 0 0. Reported Cluster Robust p–value eulav–p ecnerefnI noitazimodnaR (a) Original p ≤ 0.10 0.75 0.5 0.25 0 0 5 5 5 0. 2 0. 0. 7 Reported Cluster Robust p–value eulav–p ecnerefnI noitazimodnaR Falsification/Placebo Outcome (b) Original p > 0.10 Figure A2: Randomization Inference. Notes: This figure shows the relationship between p-values reported in the main reduced form results (Tables 7-11 and Table 14) on the horizontal axis and p-values obtained via randomization inference on the vertical axis, with resampling from actual realizations of the member asset share for each county. Points expressed as triangles correspond to falsification tests of growth among large establishments or growth of total establishments within highbarrier industries. 53

Full Sample Est. 0.03 Slope = –0.0001 0.02 0.01 0 0 10 20 30 40 50 60 70 80 90 100 County Member Asset Share (%) )htworG .pmE latoT( etamitsE .ffeoC (a) Dependent Variable: ln(Employment) 0.1 Full Sample Est. 0.08 Slope = –0.00008 0.06 0.04 0.02 0 0 10 20 30 40 50 60 70 80 90 100 County Member Asset Share (%) )htworG .batsE .taM .gdlB( etamitsE .ffeoC (b) Dependent Variable: ln(Small Building Material and Hardware Stores) Figure A3: Heterogeneity in the Reduced Form Estimate Across the Member Asset Share Distribution. Notes: This figure reports the average causal responses for subsamples of counties based on their member asset shares. Each point estimate indicated by a solid circle is computed using a sample of counties with member asset shares within +/- 10 percentage points of value reported on the horizontal axis. Panel (a) uses the natural logarithm of total employment as the dependent variable, and panel (b) reports the results using the natural logarithm of small building material establishments as the dependent variable. The dashed horizontal line corresponds to the point estimate using the full sample reported in Tables 7 and 8. The solid 54 line fits the average causal responses. The bandwidth for the first point (county member asset share of 20%) doubles the subsample bandwidth given the limited observations at the bottom of the distribution.

Table 1: Rural Commercial Bank Summary Statistics (Dec. 1972) Nonmembers Members Assets ($mil) 8.7 14.7 (27.3) (15.0) Liq. Assets 0.43 0.40 Assets (0.15) (0.14) Loans 0.46 0.47 Assets (0.13) (0.12) C&ILoans 0.08 0.09 Assets (0.06) (0.06) CRELoans 0.02 0.02 Assets (0.03) (0.03) Cons. Loans 0.09 0.11 Assets (0.06) (0.07) RRELoans 0.05 0.05 Assets (0.05) (0.06) AgLoans 0.22 0.18 Assets (0.13) (0.13) Deposits 0.90 0.90 Assets (0.04) (0.03) Equity 0.08 0.08 Assets (0.03) (0.03) Int. Income(%) 5.20 5.07 Assets (0.73) (0.69) max{SCF} – 0.01 Deposits (0.03) (cid:12) max{SCF}(cid:12)1 (b) – 0.05 Deposits (cid:12) SCF (0.03) N 1,583 1,061 Notes: This table reports descriptive for rural commercial banks operating in areas with sufficient seasonal variation in loans and deposits to qualify for the SCF. Average values for each variable are reported as of 1972 Q4, with the standard deviation in parentheses. The indicator function 1 (b) takes a value of one for all SCF banksthatdrewontheSCFatanypointinitsfirsttwo years. The maximum number of observations for each group is reported in the final line of the table. 55

Table 2: Membership Prediction Based on Observable Characteristics Coeff. Estimate Assets ($mil) 39.4 Liq. Assets –2.30* Assets Loans –2.96 Assets C&ILoans 0.48 Assets CRELoans –4.04 Assets Cons. Loans 3.19 Assets RRELoans 2.26 Assets AgLoans 0.97 Assets Deposits –4.25 Assets Equity -12.29** Assets Int. Income(%) -0.47*** Assets Pseudo R2 0.075 N 2,635 Notes: This table reports results from a logitregressionofruralbanks’membership status on balance sheet and income characteristics as of 1972 Q4. Statistical significance: ∗∗∗ p ≤ 0.01,∗∗ p ≤ 0.05,∗ p ≤ 0.10. 56

Table3: BalancesheetandincomechangesforbanksthatborrowedfromtheSeasonalCredit Facility (OLS estimates) Panel A Liq. Assets Loans ln(Loans) Assets Assets Post ×1 (b) -0.031*** 0.032*** 0.060*** t SCF (0.004) (0.004) (0.014) Adj. R2 0.83 0.81 0.97 N 39,602 39,602 39,599 Panel B C&ILoans CRELoans Cons. Loans Assets Assets Assets Post ×1 (b) 0.009*** 0.003** 0.002 t SCF (0.002) (0.001) (0.002) Adj. R2 0.79 0.85 0.88 N 29,022 29,022 29,022 Panel C RRELoans AgLoans Int. Income(%) Assets Assets Assets Post ×1 (b) 0.000 0.017*** 0.111*** t SCF (0.002) (0.003) (0.031) Adj. R2 0.92 0.91 0.66 N 29,022 29,022 15,762 Bank FEs (cid:88) (cid:88) (cid:88) ZIP3×Time FEs (cid:88) (cid:88) (cid:88) Notes: This table reports difference-in-differences estimates from regressionsoftheindicateddependentvariablesontheinteraction term,Post ×1 (b),whichequalsoneforpost-treatment(1973 t SCF Q2 and later) observations of banks that ever drew on the SCF, bank fixed effects, and ZIP3-time fixed effects: Y =γ +φ +β (cid:2) Post ×1 (b) (cid:3) +ε . bt b ZIP3,t t SCF bt Standard errors (in parentheses) are clustered at the bank level. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 57

Table 4: Balance sheet and income changes for member banks that the Federal Reserve identified as likely beneficiaries (Intent-to-treat/Reduced form estimates) Panel A Liq. Assets Loans ln(Loans) Assets Assets Post ×1 (b) -0.017*** 0.018*** 0.025** t SCF (0.004) (0.004) (0.012) Adj. R2 0.83 0.81 0.97 N 39,602 39,602 39,599 Panel B C&ILoans CRELoans Cons. Loans Assets Assets Assets Post ×1 (b) 0.005*** 0.000 0.002 t SCF (0.002) (0.001) (0.002) Adj. R2 0.79 0.85 0.88 N 29,022 29,022 29,022 Panel C RRELoans AgLoans Int. Income(%) Assets Assets Assets Post ×1 (b) 0.002* 0.011*** 0.159*** t SCF (0.002) (0.002) (0.027) Adj. R2 0.92 0.91 0.66 N 29,022 29,022 15,762 Bank FEs (cid:88) (cid:88) (cid:88) ZIP3×Time FEs (cid:88) (cid:88) (cid:88) Notes: This table reports difference-in-differences estimates from regressionsoftheindicateddependentvariablesontheinteraction term, Post × 1 (b), which equals one for post-treatment t Mem50 (1973Q2andlater)observationsofbanksthattheFederalReserve identified as the most likely to benefit from the SCF given the terms at the time, bank fixed effects, and ZIP3-time fixed effects: Y =γ +φ +β (cid:2) Post ×1 (b) (cid:3) +ε . bt b ZIP3,t t Mem50 bt Standard errors (in parentheses) are clustered at the bank level. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 58

Table 5: Balance sheet and income changes for member banks that the borrowed from the Seasonal Credit Facility (2SLS DID-IV estimates) Panel A Liq. Assets Loans ln(Loans) Assets Assets (cid:92) Post ×1 (b) -0.081*** 0.084*** 0.114** t SCF (0.019) (0.019) (0.057) 1st Stage F 82.7 82.7 82.7 AR Wald (p) 0.000 0.000 0.040 N 5,284 5,284 5,284 Panel B C&ILoans CRELoans Cons. Loans Assets Assets Assets (cid:92) Post ×1 (b) 0.023*** 0.001 0.011 t SCF (0.009) (0.004) (0.008) 1st Stage F 82.3 82.3 82.3 AR Wald (p) 0.006 0.881 0.180 N 5,272 5,272 5,272 Panel C RRELoans AgLoans Int. Income(%) Assets Assets Assets (cid:92) Post ×1 (b) 0.012** 0.054*** 0.785*** t SCF (0.006) (0.013) (0.164) 1st Stage F 82.3 82.3 82.2 AR Wald (p) 0.040 0.000 0.000 N 5,272 5,272 5,248 Bank FEs (cid:88) (cid:88) (cid:88) ZIP3×Time FEs (cid:88) (cid:88) (cid:88) Notes: This table reports difference-in-differences estimates from DID-IVregressionsoftheindicateddependentvariablesontheinstrumented interaction term, Post (cid:92) ×1 (b), which equals one t SCF forpost-treatment(1973Q2andlater)observationsofbanksthat ever drew on the SCF, bank fixed effects, and ZIP3-time fixed effects: Y =γ +φ +β (cid:2) Post (cid:92) ×1 (b) (cid:3) +ε . bt b ZIP3,t t SCF bt The Kleibergen-Paap (2006) first stage F-stat, the p-value of the Anderson-Rubin (1949) weak-instrument robust inference test of thestatisticalsignificanceoftheendogenousregressorinthestructural equation, and the number of observations are reported for each specification. Standard errors (in parentheses) are clustered at the bank level. Statistical significance: ∗∗∗ p ≤ 0.01,∗∗ p ≤ 0.05,∗ p≤0.10. 59

Table 6: Rural County Summary Statistics (Dec. 1972) Mean Std. Dev. N 1,631 – Employment and Population Total Employment 7,863 (7,395) Proprietors’ Employment 2,002 (1,287) Nonfarm Proprietors’ Employment 1,139 (888) Population 18,792 (16,044) Establishment Counts Building Materials 120.3 (96.0) Bars and Restaurants 20.7 (21.6) Misc. Retail 18.4 (18.3) Gas Stations 27.4 (24.1) Banks 23.5 (28.7) Hotels & Lodging 78.3 (81.9) Commercial Bank Presence (Mar. 1973) Banks 3.7 (2.6) Fed membership share 0.37 (0.34) Fed membership asset share 0.42 (0.37) Notes: This table reports descriptive statistics for rural counties as of 1972 Q4, except for bank membership variables, which are reported as of 1973 Q1, just before the implementation of the Seasonal Credit Facility. Sources: Bureau of Economic Analysis; Census Bureau; National Information Center. 60

Table 7: Employment outcomes for counties exposed to the Seasonal Credit Facility Panel A: Reduced Form Estimates ln(Proprietors’ ln(Non-farm Prop. ln(Employment) Employment) Employment) (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.016*** 0.014** 0.025*** 0.018*** 0.026*** 0.018** t c (0.006) (0.006) (0.007) (0.006) (0.008) (0.007) Adj. R2 0.99 0.99 0.99 0.99 0.99 0.99 N 11,417 11,417 11,417 11,417 11,417 11,417 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Panel B: 2SLS Estimates ln(Proprietors’ ln(Non-farm Prop. ln(Employment) Employment) Employment) (1) (2) (1) (2) (1) (2) (cid:92) Post × SCFc 1.130** 0.979** 1.843*** 1.282*** 1.846*** 1.276** t Depositsc (0.445) (0.46) (0.513) (0.466) (0.583) (0.578) 1st Stage F 61.9 47.1 61.9 47.1 61.9 47.1 AR Wald (p) 0.008 0.020 0.000 0.003 0.001 0.014 N 11,382 11,382 11,382 11,382 11,382 11,382 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct in panel A, where Y corresponds to employment totals within each county as indicated. In panel ct B, Post × MemAssetShare is used to instrument for county-wide seasonal credit borrowing as a t c share of total deposits, Post t × De S p C os F i c tsc . Standard errors (in parentheses) are clustered at the county level. The Kleibergen-Paap (2006) first stage F-stat, the p-value of the Anderson-Rubin (1949) weakinstrumentrobustinferencetestofthestatisticalsignificanceoftheendogenousregressorinthestructural equation,andthenumberofobservationsarereportedforeach2SLSspecification. Statisticalsignificance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 61

Table 8: Building materials and hardware establishment growth for counties exposed to the Seasonal Credit Facility Panel A: Reduced Form Estimates ln(Building Material and Hardware Stores) All ≤4 employees >20 employees (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.049*** 0.035** 0.062*** 0.046*** 0.049 0.031 t c (0.015) (0.015) (0.016) (0.015) (0.032) (0.033) Adj. R2 0.97 0.98 0.96 0.96 0.91 0.91 N 11,222 11,222 11,222 11,222 9,200 9,200 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Panel B: 2SLS Estimates ln(Building Material and Hardware Stores) All ≤4 employees >20 employees (1) (2) (1) (2) (1) (2) (cid:92) Post × SCFc 3.520*** 2.473** 4.457*** 3.287*** 3.529 2.279 t Depositsc (1.177) (1.145) (1.219) (1.182) (2.370) (2.493) 1st Stage F 61.8 47.3 61.8 47.3 54.9 50.2 AR Wald (p) 0.001 0.023 0.000 0.006 0.129 0.357 N 11,187 11,187 11,187 11,187 9,172 9,172 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct in panel A, where Y corresponds to establishment counts within each county as indicated. In panel ct B, Post × MemAssetShare is used to instrument for county-wide seasonal credit borrowing as a t c share of total deposits, Post t × De S p C os F i c tsc . Standard errors (in parentheses) are clustered at the county level. The Kleibergen-Paap (2006) first stage F-stat, the p-value of the Anderson-Rubin (1949) weakinstrumentrobustinferencetestofthestatisticalsignificanceoftheendogenousregressorinthestructural equation,andthenumberofobservationsarereportedforeach2SLSspecification. Statisticalsignificance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 62

Table 9: Bar and restaurant establishment growth for counties exposed to the Seasonal Credit Facility Panel A: Reduced Form Estimates ln(Bar and Restaurant Establishments) All ≤4 employees >20 employees (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.060*** 0.028 0.115*** 0.085*** 0.063 0.027 t c (0.022) (0.023) (0.029) (0.029) (0.044) (0.045) Adj. R2 0.95 0.95 0.86 0.86 0.80 0.80 N 9,604 9,604 9,556 9,556 6,161 6,161 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Panel B: 2SLS Estimates ln(Bar and Restaurant Establishments) All ≤4 employees >20 employees (1) (2) (1) (2) (1) (2) (cid:92) Post × SCFc 4.119*** 1.910 7.853*** 5.714*** 4.167 1.861 t Depositsc (1.567) (1.518) (2.165) (2.028) (2.867) (2.775) 1st Stage F 53.4 49.3 53.3 49.2 39.8 40.0 AR Wald (p) 0.006 0.207 0.000 0.003 0.129 0.496 N 9,574 9,574 9,526 9,526 6,147 6,147 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct in panel A, where Y corresponds to establishment counts within each county as indicated. In panel ct B, Post × MemAssetShare is used to instrument for county-wide seasonal credit borrowing as a t c share of total deposits, Post t × De S p C os F i c tsc . Standard errors (in parentheses) are clustered at the county level. The Kleibergen-Paap (2006) first stage F-stat, the p-value of the Anderson-Rubin (1949) weakinstrumentrobustinferencetestofthestatisticalsignificanceoftheendogenousregressorinthestructural equation,andthenumberofobservationsarereportedforeach2SLSspecification. Statisticalsignificance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 63

Table 10: Miscellaneous retail establishment outcomes for counties exposed to the Seasonal Credit Facility Panel A: Reduced Form Estimates ln(Misc. Retail Establishments) All ≤4 employees >20 employees (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.155*** 0.099*** 0.133*** 0.085*** 0.068 0.042 t c (0.029) (0.028) (0.030) (0.030) (0.062) (0.064) Adj. R2 0.94 0.94 0.89 0.89 0.61 0.61 N 9,413 9,413 9,372 9,372 3,186 3,186 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Panel B: 2SLS Estimates ln(Misc. Retail Establishments) All ≤4 employees >20 employees (1) (2) (1) (2) (1) (2) (cid:92) Post × SCFc 11.814*** 7.582*** 10.100*** 6.488*** 4.945 3.379 t Depositsc (2.759) (2.321) (2.667) (2.398) (4.503) (5.016) 1st Stage F 43.6 43.3 43.6 43.3 24.3 21.2 AR Wald (p) 0.000 0.000 0.000 0.004 0.259 0.490 N 9,382 9,382 9,342 9,342 3,175 3,175 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct in panel A, where Y corresponds to establishment counts within each county as indicated. In panel ct B, Post × MemAssetShare is used to instrument for county-wide seasonal credit borrowing as a t c share of total deposits, Post t × De S p C os F i c tsc . Standard errors (in parentheses) are clustered at the county level. The Kleibergen-Paap (2006) first stage F-stat, the p-value of the Anderson-Rubin (1949) weakinstrumentrobustinferencetestofthestatisticalsignificanceoftheendogenousregressorinthestructural equation,andthenumberofobservationsarereportedforeach2SLSspecification. Statisticalsignificance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 64

Table 11: High-barrier-industry establishment growth for counties exposed to the Seasonal Credit Facility Panel A: Reduced Form Estimates Lodging Banking Gas Stations (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.022 0.016 0.002 –0.007 0.022 0.003 t c (0.020) (0.020) (0.021) (0.021) (0.018) (0.018) Adj. R2 0.97 0.97 0.96 0.96 0.96 0.96 N 11,198 11,198 11,221 11,221 10,268 10,268 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Panel B: 2SLS Estimates Lodging Banking Gas Stations (1) (2) (1) (2) (1) (2) (cid:92) Post × SCFc 1.568 1.102 0.173 –0.507 1.658 0.227 t Depositsc (1.412) (1.454) (1.467) (1.496) (1.408) (1.351) 1st Stage F 61.9 47.5 61.8 47.4 54.7 49.8 AR Wald (p) 0.259 0.442 0.906 0.734 0.235 0.867 N 11,163 11,163 11,186 11,186 10,237 10,237 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct in panel A, where Y corresponds to the establishment counts within each county for the indicated ct industries. In panel B, Post ×MemAssetShare is used to instrument for county-wide seasonal credit t c borrowing as a share of total deposits, Post t × De S p C os F i c tsc . Standard errors (in parentheses) are clustered at the county level. The Kleibergen-Paap (2006) first stage F-stat, the p-value of the Anderson-Rubin (1949) weak-instrument robust inference test of the statistical significance of the endogenous regressor in the structural equation, and the number of observations are reported for each 2SLS specification. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 65

Table 12: Rural County Summary Statistics (Dec. 1972) No Petro Industry Petro Industry Difference Employment Growth (%, 1973–1974) 1.74 2.64 0.90*** Income Growth (%, 1973–1974) 4.82 8.87 4.05*** Employment Growth (%, 1973–1975) 1.33 3.49 2.16*** Income Growth (%, 1973–1975) 15.7 21.8 6.08*** Notes: Thistablereportsaverageemploymentandincomegrowthratesforruralcountiesseparatedbased on the presence of a petroleum industry. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 66

Table13: EmploymentandestablishmentgrowthforcountiesexposedtotheSeasonalCreditFacilityandthepetroleumindustry Reduced Form Estimates ln(Proprietors’ ln(Non-farm Prop. ln(Employment) Employment) Employment) (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.017*** 0.016** 0.025*** 0.019*** 0.025** 0.019** t c (0.006) (0.006) (0.007) (0.006) (0.008) (0.008) Post ×MemAssetShare ×Petro –0.005 –0.007 0.016 0.007 0.021 0.010 t c c (0.016) (0.015) (0.017) (0.016) (0.019) (0.018) Adj. R2 0.99 0.99 0.99 0.99 0.99 0.99 N 11,417 11,417 11,417 11,417 11,417 11,417 ln(Small Establishments) Building Bars & Misc. Materials Restaurants Retail (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.064*** 0.050*** 0.119*** 0.092*** 0.139*** 0.100*** t c (0.017) (0.016) (0.031) (0.031) (0.033) (0.032) Post ×MemAssetShare ×Petro 0.034 0.000 0.082 0.041 0.077 –0.028 t c c (0.033) (0.032) (0.074) (0.076) (0.069) (0.063) Adj. R2 0.96 0.96 0.86 0.86 0.89 0.89 N 11,222 11,222 9,556 9,556 9,372 9,372 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameters from an estimation of the following regression: (cid:2) (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+β Post ×MemAssetShare ×Petro ]+κ Post ×X +ε ct c FRS,t 0 t c 1 t c c t c ct where Y corresponds to the dependent variables as indicated. Petro denotes a dummy variable that takes a value ct c of one for counties with a petroleum industry. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 67

Table 14: Employment, wages, and contributions to government social insurance programs for counties exposed to the Seasonal Credit Facility Panel A: Reduced Form Estimates Wage & Salary Total Wages Contrib. to Gov. Employment & Salaries ($000s) Social Insurance ($000s) (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 364*** 135*** 8,882*** 2,378*** 1,432*** 331*** t c (57.4) (50.3) (1,120) (706) (180) (108) Adj. R2 0.99 0.99 0.96 0.97 0.92 0.96 N 11,417 11,417 11,417 11,417 11,417 11,417 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Panel B: 2SLS Estimates (×10−3) Wage & Salary Total Wages Contrib. to Gov. Employment & Salaries ($000s) Social Insurance ($000s) (1) (2) (1) (2) (1) (2) (cid:92) Post × SCFc 26.7*** 9.7** 649.6*** 170.9*** 104.5*** 23.5** t Depositsc (5.3) (3.9) (116.1) (57.5) (18.8) (8.65) 1st Stage F 61.9 47.1 61.9 47.1 61.9.9 47.1 AR Wald (p) 0.000 0.007 0.000 0.001 0.000 0.007 N 11,382 11,382 11,382 11,382 11,382 11,382 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) Y =γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct in panel A, where Y corresponds to the dependent variables indicated in column headers. In panel B, ct Post ×MemAssetShare is used to instrument for county-wide seasonal credit borrowing as a share of t c total deposits, Post t × De S p C os F i c tsc . Standard errors (in parentheses) are clustered at the county level. The Kleibergen-Paap(2006)firststageF-stat,thep-valueoftheAnderson-Rubin(1949)weak-instrumentrobust inference test of the statistical significance of the endogenous regressor in the structural equation, and the numberofobservationsarereportedforeach2SLSspecification. Statisticalsignificance: ∗∗∗ p≤0.01,∗∗ p≤ 0.05,∗ p≤0.10. 68

Table A1: Employment and establishment growth for ultra-rural counties Reduced Form Estimates ln(Proprietors’ ln(Non-farm Prop. ln(Employment) Employment) Employment) (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.018** 0.017** 0.020*** 0.020*** 0.023** 0.023** t c (0.007) (0.007) (0.008) (0.007) (0.010) (0.009) Adj. R2 0.99 0.99 0.99 0.99 0.99 0.99 N 7,854 7,854 7,854 7,854 7,854 7,854 ln(Small Establishments) Building Bars & Misc. Materials Restaurants Retail (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.056*** 0.051*** 0.095*** 0.074** 0.178*** 0.157*** t c (0.020) (0.018) (0.040) (0.036) (0.043) (0.041) Adj. R2 0.95 0.95 0.82 0.83 0.86 0.87 N 7,708 7,708 6,182 6,182 5,968 5,968 County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×MemAssetShare ]+κ Post ×X +ε ct c FRS,t t c t c ct where Y corresponds to the dependent variables as indicated. Standard errors (in parentheses) are clustered ct at the county level. Ultra-rural counties are those with a total urban population of fewer than 2,500 people or counties with a total urban population of fewer than 20,000 that are not adjacent to an MSA. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 69

Table A2: Balance sheet and income changes for member banks that the Federal Reserve identified as likely beneficiaries (county-level geographic unit, intent-to-treat/reduced form estimates) Panel A Liq. Assets Loans ln(Loans) Assets Assets Post ×1 (b) -0.020*** 0.022*** 0.044** t SCF (0.006) (0.006) (0.019) Adj. R2 0.83 0.82 0.97 N 12,974 12,974 12,972 Panel B C&ILoans CRELoans Cons. Loans Assets Assets Assets Post ×1 (b) 0.009*** 0.000 0.005* t SCF (0.003) (0.002) (0.003) Adj. R2 0.80 0.84 0.89 N 9,483 9,483 9,483 Panel C RRELoans AgLoans Int. Income(%) Assets Assets Assets Post ×1 (b) 0.003 0.008** 0.203*** t SCF (0.002) (0.004) (0.047) Adj. R2 0.93 0.93 0.64 N 9,483 9,483 5,154 Bank FEs (cid:88) (cid:88) (cid:88) FIPS×Time FEs (cid:88) (cid:88) (cid:88) Notes: This table reports difference-in-differences estimates from regressionsoftheindicateddependentvariablesontheinteraction term, Post × 1 (b), which equals one for post-treatment t Mem50 (1973Q2andlater)observationsofbanksthattheFederalReserve identified as the most likely to benefit from the SCF given the termsatthetime,bankfixedeffects,andcounty-timefixedeffects: Y =γ +φ +β (cid:2) Post ×1 (b) (cid:3) +ε . bt b FIPS,t t Mem50 bt Standard errors (in parentheses) are clustered at the bank level. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 70

Table A3: Employment and establishment growth with a binary measure of treatment for high membership share counties Reduced Form Estimates ln(Proprietors’ ln(Non-farm Prop. ln(Employment) Employment) Employment) (1) (2) (1) (2) (1) (2) Post ×Treat 0.013** 0.012** 0.018*** 0.011** 0.017** 0.011* t c (0.005) (0.005) (0.006) (0.006) (0.007) (0.007) ln(Small Establishments) Building Bars & Misc. Materials Restaurants Retail (1) (2) (1) (2) (1) (2) Post ×Treat 0.049*** 0.033** 0.087*** 0.062** 0.110*** 0.062** t c (0.015) (0.014) (0.026) (0.027) (0.028) (0.028) County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter from an estimation of the following regression: (cid:2) (cid:2) (cid:3) ln(Y )=γ +φ +β Post ×Treat ]+κ Post ×X +ε ct c FRS,t t c t c ct where Y corresponds to the dependent variables as indicated. Standard errors (in parentheses) are clustered ct at the county level. Treat is a dummy variable that identifies counties with a MemAssetShare value in c c the top tercile. Banks with a MemAssetShare in the bottom tercile compose the control group. Statistical c significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 71

Table A4: Employment and establishment growth using the DID estimator M Reduced Form Estimates ln(Proprietors’ ln(Non-farm Prop. ln(Employment) Employment) Employment) (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.016** 0.016** 0.021*** 0.018** 0.022** 0.020** t c (0.007) (0.007) (0.008) (0.008) (0.010) (0.009) ln(Small Establishments) Building Bars & Misc. Materials Restaurants Retail (1) (2) (1) (2) (1) (2) Post ×MemAssetShare 0.079*** 0.062*** 0.062 0.096** 0.108*** 0.100** t c (0.019) (0.017) (0.045) (0.042) (0.042) (0.041) County FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) District×Time FEs (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) (cid:88) County Controls – (cid:88) – (cid:88) – (cid:88) Notes: This table reports the β parameter using the DID estimator described in de Chaisemartin and M D’Haultfœuille (2021), which can accommodate continuous treatment and non-staggered designs. Y correct sponds to the dependent variables as indicated. Statistical significance: ∗∗∗ p≤0.01,∗∗ p≤0.05,∗ p≤0.10. 72

Table B1: Item Numbers for Bank-Level Financial Information (Forms FR 105, FR 107) Description Item # Description Item # Total Assets 2170 Consumer Loans 1975 Cash and Due From Banks 0010 Residential RE Loans 1430, 1460 Treasury Securities 0400 Agricultural Loans 1420, 1590 GSE/Agency Securities 0600 Loan Interest & Fees 4010 Fed Funds Sold 1350 Income on Fed Funds Sold 4020 Total Loans 1400 Interest on Treasury Securities 4030 C&I Loans 1600 Interest on GSE/Agency Securities 4040 CRE Loans 1480 Interest on Muni Securities 4060 Notes: This table lists the item numbers from each bank’s FR 105 and FR 107 forms that were used to generate each reported series. Prefix mnemonics are RCON for all balance sheet items and RIAD for all income items. 73