Bank Supervision and Managerial Control Systems: The Case of Minority Lending

Finance and Economics Discussion Series Federal Reserve Board, Washington, D.C. ISSN 1936-2854 (Print) ISSN 2767-3898 (Online) Bank Supervision and Managerial Control Systems: The Case of Minority Lending Byeongchan An, Robert Bushman, Anya Kleymenova, and Rimmy E. Tomy 2022-036 Please cite this paper as: An,Byeongchan,RobertBushman,AnyaKleymenova,andRimmyE.Tomy(2023). “Bank SupervisionandManagerialControlSystems: TheCaseofMinorityLending,” Financeand Economics Discussion Series 2022-036r1. Washington: Board of Governors of the Federal Reserve System, https://doi.org/10.17016/FEDS.2022.036r1. 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.

Bank Supervision and Managerial Control Systems: The Case of Minority Lending⋆ Byeongchan An1, Robert Bushman2, Anya Kleymenova3, Rimmy E. Tomy4,∗ February 15, 2023 Abstract This paper investigates how bank supervisors’ enforcement decisions and orders (EDOs) influence the allocation of mortgage lending across demographic groups underlying a banks’ borrower base. Specifically, we investigate how banks’ mortgage lending to minority borrowers relative to white borrowers changes following the resolution of severe EDOs. We hypothesize that improvements in management control systems imposed by EDOs serve as channels through which EDOs affect a bank’s borrower base generally, and minority lending specifically. We empirically examine how changes in loan policies and internal governance mechanisms specified in EDOs influence banks’ mortgage lending decisions. We find that relative to white borrowers, mortgage lending to minority borrowers significantly increases following the resolution of EDOs, where this positive effect increases with the strictness of bank supervisors and severity of the EDO. Consistent with management controls serving as channels for this change, there is a more pronounced effect on minority lending when an EDO mandates improvements in lending policies and stronger internal governance over lending decisions. JEL Classification: G21, G28, G38 Keywords: Banking, Bank supervision, Discrimination, Enforcement actions, Internal controls, Internal audit, Loan policy, Mortgage lending ⋆ WethankC.K.Lee(communitybankerdiscussant),DanielBarth,HansChristensen,ChristineCuny(discussant),DougDiamond,Jennifer Dlugosz, Ferdinand Elfers (discussant), Jo˜ao Granja, Sehwa Kim, John Krainer, Benjamin Kay, Dalida Kadyrzhanova, Eva Labro, Christian Leuz,MariaLoumioti,MaureenMcNichols,JoeNichols,AndreaPassalacqua,MatthewPlosser(discussant),RaghuramRajan,DorianaRuffino, Jo˜ao Santos (discussant), Haresh Sapra, Ishita Sen (discussant), Doug Skinner, Abbie Smith, Shasta Shakya (discussant), Andrew Sutherland (discussant), Chad Syverson, Ana-Maria Tenekedjieva, Harald Uhlig, James Vickery (discussant), Cindy Vojtech, Dushyant Vyas (discussant), JamesWang, TengWang, ShuangWu(discussant), LuigiZingales(discussant), andseminarparticipantsattheUniversityofChicagoBanking Workshop,theEarlyInsightsinAccounting,theFederalReserveBoard,UniversityofNorthCarolina,ErasmusUniversityRotterdam,Applied MicroDayAheadConference,R&Slunchseminar,2021EFMA,2021CommunityBankinginthe21stCenturyResearchandPolicyConference, OIG2021FallInstituteResearchConference,NZFM2021,ParisDecemberFinanceMeeting2021,2022AFAAnnualMeeting,2022FARSMidyear Meeting, Berkeley, London Business School, 2022 Haskayne and Fox Accounting Conference, and the 2022 Workshop on Financial Institutions Research St. Louis Fed for their helpful comments and suggestions. We are grateful to Nobuyuki Furuta, Lori Leu, Samuel Shin, and Jizhou Wangforexcellentresearchassistance. WethankJamesKiselikforeditorialassistance. Wegratefullyacknowledgethefinancialsupportofthe Fama-MillerCenterforResearchinFinanceandtheUniversityofChicagoBoothSchoolofBusiness. Theviewsexpressedinthisstudyarethose oftheauthorsanddonotnecessarilyreflecttheviewsoftheFederalReserveBoardortheFederalReserveSystem. ∗ CorrespondingAuthor 1 UniversityofUtah,1655CampusCenterDr,SaltLakeCity,UT84112;Byeongchan.An@utah.edu 2 UniversityofNorthCarolina–ChapelHill,300KenanCenterDrive,ChapelHill,NC27599;RobertBushman@kenan-flagler.unc.edu 3 FederalReserveBoard,20thStreetandConstitutionAvenueNW,Washington,DC20551;Anya.Kleymenova@frb.gov 4 TheUniversityofChicagoBoothSchoolofBusiness,5807SouthWoodlawnAvenue,Chicago,IL60637;Rimmy.Tomy@chicagobooth.edu 1

1. Introduction An important question in banking is how supervision affects bank lending.1 Recent researchexploitsvariationinthestrictnessofbanksupervisiontoexamineitseffectsoncredit supply (Agarwal et al., 2014; Granja & Leuz, 2022). Supervisory activities can influence loan supplynotonlybyscrutinizingbanks’loanlossrecognitionbutalsobyidentifyingdeficiencies in banks’ management practices, including operating procedures and policies and internal governance structures. For example, following a shift to a more rigorous supervisory regime, Granja&Leuz(2022)documentsignificantchangesinbanks’internalmanagementpractices. They find that banks that experienced significant changes in internal practices following a supervisory regime shift exhibit a more pronounced increase in complex lending, such as lending to small businesses.2 That is, in addition to impacting the amount of lending, supervisory pressure can also fundamentally influence the types of borrowers to which banks lend. In this paper, we extend this literature by investigating how U.S. bank supervisors’ enforcement decisions and orders (EDOs) against financial institutions influence the allocation of mortgage lending across demographic groups underlying a banks’ borrower base. Specifically, we investigate how banks’ mortgage lending to minority borrowers relative to white borrowers changes following the resolution of severe EDOs. The intuition for our study is that credit risk assessment is more complex for minority borrowers who may lack credit scores or other standard sources of credit information. Therefore, if stricter bank supervision facilitates more complex lending (as Granja & Leuz, 2022, find) then these benefits of supervision should also extend to other types of borrowers whose credit risk is more challenging to evaluate, such as minority borrowers. We are particularly interested in examining the extent 1A large literature analyzes the impact of bank regulation and supervision on lending (e.g., Eisenbach et al., 2022; Hirtle et al., 2020; Kandrac & Schlusche, 2021; Altavilla et al., 2020). 2Lendingtosmallandmediumenterprises(SME)iscomplexbecausetheircreditriskishardertoevaluate. For example, Schwert (2018) and Cort´es et al. (2020) find that SME borrowers are more likely to borrow fromwell-capitalizedbanksbecausecapital-constrainedbankshaveahardertimeevaluatingSMEborrowers. 2

to which administrative controls, in the form of loan and internal governance policies and adherence to such policies, serve as mechanisms through which EDOs transmit their effect on lending outcomes. Recent research finds some support for this mechanism, showing that racial disparities can derive from the biases of individual loan officers and limitations on the scope of borrowers’ information used in the lending decision (Di Maggio et al., 2022; Jiang et al., 2022). A longstanding and growing literature in accounting considers the role of managerial control systems in shaping firm performance by aligning the behaviors and decisions of employees with an organization’s objectives (Brickley et al., 2015; Grabner & Moers, 2013; Langfield-Smith, 1997; Malmi & Brown, 2008; Tucker et al., 2009). Administrative controls are an important element of managerial control systems that span standard operating procedures and policies and internal governance structures (Abernethy & Chua, 1996; Simons, 1987). Administrative controls direct employee behavior by establishing action protocols, specifying how tasks are to be performed, and ensuring adherence to policies (Malmi & Brown, 2008; Merchant & Van der Stede, 2007; Simons, 1987). However, our understanding of the impact of administrative controls on banks’ lending decisions is limited. Understanding the impact of improvements in banks’ administrative controls on their credit allocation decisions is crucial and has critical socio-economic implications.3 We hypothesize that improvements in banks’ administrative controls that are imposed throughEDOscouldaffectabank’sborrowerbasegenerallyandminoritylendingspecifically. A primary objective of supervisors in issuing EDOs is to correct specific deficiencies by forcing banks to make fundamental changes in operational processes, including changes in administrative controls. Changes to administrative controls can include the introduction of new loan policies that require loan officers to consistently follow best practices in making 3For example, banks’ mortgage lending decisions influence homeownership. Owning a home conveys a numberofsocialandeconomicbenefits,suchastheabilitytoaccumulatewealth,accesstocreditbybuilding home equity, higher educational attainment, and a lower likelihood of incarceration (Aaronson, 2000; Blau & Graham, 1990; Collins & Margo, 2001; Di et al., 2007; Green et al., 1997; Newman & Holupka, 2016; Shapiro, 2006; Wainer & Zabel, 2020). 3

lending decisions, and the strengthening of internal governance mechanisms that monitor regulatory compliance and adherence to internal bank policies. Such changes may increase lending to minorities by limiting the discretion of individual loan officers to bias lending decisions against minority borrowers (Frame et al., 2022).4 Innovations in loan policies could also expand the scope of credit risk assessments to incorporate information beyond metrics like credit scores that may put minority borrowers at a disadvantage. In this regard, recent research on FinTech lending platforms provides evidence that incorporating alternative data into traditional bank lending models can significantly increase credit access to applicants with a low credit score (Di Maggio et al., 2022). EDOs are issued against financial institutions for unsafe or unsound practices; breaches of fiduciary duty; and violations of laws, rules, or regulations. We study EDOs because they often contain provisions that require banks to improve their administrative controls. Regulators bring enforcement actions against problem banks as a measure of last resort and exercise some discretion in issuing EDOs. If a bank fails to satisfy the requirements of the order, regulators can enforce the order in U.S. district courts, terminate deposit insurance, or take further actions that might lead to bank closure.5 While a few EDOs directly reference fairlendingpractices,EDOsaregenerallynotconcernedwithbanks’adherencetofairlending laws. A separate and distinct supervisory process oversees compliance with fair lending laws. We begin by examining the extent to which EDO banks’ lending to minority borrowers changes in the five years following the resolution of the EDO. We find that EDO banks significantly increase their mortgage lending to minority borrowers relative to white borrowers following the termination of an enforcement order. This result also holds if we define 4There are two proposed models for observed discrimination: taste-based (or prejudice-based) and statistical discrimination. Taste-based discrimination presumes some form of animus directed toward members of particular groups. Statistical discrimination presumes stereotyping based on group membership due to imperfectinformation(seeGuryan&Charles,2013,foradetaileddiscussionandsummaryoftheliterature). 5Upon completion of the required actions and improved ratings from bank examiners, supervisors issue a termination order. If a bank fails, a formal termination order is issued. If a bank is acquired or merges with another bank, the EDO remains under the original name of the bank and is only terminated once the regulators are satisfied that the new entity has met the requirements of the original order. Sometimes supervisors modify EDOs to include additional conditions or requirements. 4

minority borrowers as nonwhite borrowers or consider lending to Black or African-American borrowers relative to white male borrowers. Specifically, the share of residential mortgage lending to minority borrowers in EDO banks’ total residential mortgage portfolio, measured at the county level, increases by 2% to 7% after EDO termination. We also estimate changes in the market shares of EDO banks in the counties where they operate. Specifically, we find that EDO banks increase their market shares of mortgage lending to minorities relative to all banks in a given county following EDO termination. Relative to the pre-EDO period, EDO banks’ market share of mortgage lending to minorities increases by 0.58%–0.62%. On average, EDO banks’ market share of lending to minorities in the residential mortgage market is 0.41%, making the increase economically significant. It is important to note that our results are not mechanical, as the EDOs considered in our analyses are not directly associated with fair lending laws. An important concern is that changes in the demand for mortgages or economic changes that affect all banks could be driving the relative increase in lending to minorities. To mitigate this concern, we include the number of loan applications to control for changes in the demand for mortgage loans. We also control for bank-specific characteristics and countylevel employment growth and include year and bank effects to control for any unobserved heterogeneityduetomacroeconomicconditionsandtime-invariantbankcharacteristics. Furthermore, our market share analysis encompasses all banks’ lending to minorities in a county, mitigating concerns that general economic trends drive our findings. Our main analysis uses a staggered difference-in-differences research design to study changes in EDO banks’ portfolio and market shares of lending to minorities relative to all other banks. Recent studies have shown that estimates from the staggered difference-indifferences analysis could be biased due to the combination of staggered treatment timing and dynamic treatment effects (Baker et al., 2022; Barrios, 2021; Goodman-Bacon, 2021; Sant’Anna & Zhao, 2020; Sun & Abraham, 2021). The problem is more severe with a smaller sample size when almost all units are treated and with considerable treatment het- 5

erogeneity (Baker et al., 2022). However, in our sample, most banks are not treated, and there is limited heterogeneity in treatment, making the problem less pronounced.6 Nonetheless, consistent with suggestions in the literature (Baker et al., 2022; Cengiz et al., 2019; Sun & Abraham, 2021), we conduct additional analysis to further mitigate this concern. Specifically, we repeat our analysis in stacked subsamples by matching treated (EDO) banks with control (non-EDO) banks based on size and geography. Our results indicate that, relative to non-EDO banks (all non-EDO banks and a matched sample), EDO banks significantly expand their lending to minority borrowers. Having established that EDO banks increase their mortgage lending following the termination of the EDO, we next examine channels that explain this increase in lending. We conjecture that EDOs force banks to resolve fundamental deficiencies in internal bank management practices that expand minority borrowers’ access to mortgage loans. To test this conjecture, we extract information about the corrective actions specified by supervisors from the text of the EDOs themselves. In particular, we create two variables designed to capture improvements in administrative controls specifically related to changes in lending policies and internal governance. The first variable reflects whether the EDO requires the bank to revise or formally establish a written loan policy. The second captures whether the order requires the bank to develop written internal audit procedures that monitor regulatory compliance and adherence to internal bank policies. We find that increases in minority lending are significantly higher for EDOs that specify revisions of loan policies and or implement more formal internal governance procedures in counties with a higher proportion of subprime borrowers. We conduct additional analyses to further tie our results to the supervisory enforcement process. Specifically, we find that the increase in minority lending is greater for banks with stricter regulators. We also find that banks with more severe EDOs or with low CRA ratings 6Only14%ofoursamplebanksreceiveEDOs. Also,werestrictouranalysistoonlysevereEDOs,limiting the treatment heterogeneity. 6

expand their minority lending more after exiting EDOs.7 These results are consistent with bankshavingmoresevereoperationaldeficienciesalsoexhibitingmorescopeforimprovement in their lending practices. We next explore how corrective actions directly influence loan approval decisions. We find that mortgage loan denial is 9.6% more likely for minority borrowers relative to white borrowers prior to an EDO. However, following EDO termination, the likelihood of denial decreases by five percentage points for minority borrowers. Turning to specific loan denial reasons, we find that, relative to the pre-EDO period, the rejection of minority loan applications due to borrower credit history is 3.4% less likely following EDO termination. Banks use non-price terms, such as credit history, collateral, and debt-to-income ratios, to ration credit (Stiglitz & Weiss, 1981). Minority borrowers are more likely to be constrained by these non-price terms because they are also more likely to have lower wealth (Acolin et al., 2016; Bostic, 1997; Gyourko et al., 1999). Our result shows that banks rely less on non-price terms in determining whether to reject residential mortgage loan applications from minorities following EDO termination. Lower reliance on non-price terms is consistent with EDOs forcing corrective actions that improve loan policies and credit assessment processes that disproportionately benefit minority borrowers. Forexample,adherencetowrittenloanpoliciesandproceduresreducesthediscretion afforded to loan officers, and improved credit assessment may cause banks to process additional sources of information and thus reduce reliance on non-price terms. Bolstering this interpretation, we find no evidence that this increase in minority lending is a consequence of a shift towards riskier lending as we do not find increases in nonperforming assets, the market share of risky loans, or lending by the Federal Housing Administration (FHA).8 Finally, we examine alternative explanations. We find no evidence that banks expand 7Enacted in 1977, the Community Reinvestment Act (CRA) serves to encourage credit availability in low- and moderate-income areas. Regulators rate banks based on their record in meeting the credit needs of communities in which they operate. 8FHA loans have lower down-payment requirements and are generally offered to riskier borrowers. 7

residential mortgage lending to minority borrowers to improve their capital ratios or that increased competition from non-EDO banks drives EDO banks to lend more to minority borrowers. Overall, we provide robust evidence that banks increase lending to minority borrowers relative to white borrowers following the resolution of EDOs. This increase is consistent with corrective actions improving banks’ administrative controls and thus facilitating profitable lending to minority borrowers. Ourpapermakesseveralcontributionstotheliterature. First,wecomplementandextend research on the crucial role played by bank supervisors in improving banks’ lending and risk management decisions (Agarwal et al., 2014; Granja & Leuz, 2022; Hirtle et al., 2020). Specifically, we highlight that improvements in banks’ loan policies due to the supervisory enforcement process can enhance access to credit for borrowers whose credit risk is more challenging to evaluate, such as minority borrowers. Our results complement Granja & Leuz (2022), who find that stricter regulators are associated with more SME lending. Credit risk in both SME and minority lending is more difficult to assess and may benefit from improved internal and managerial controls. Second, we extend the managerial accounting literature by providing large sample empirical evidence consistent with an important aspect ofmanagerialcontrolsystems(administrativecontrols)havingapositive, first-ordereffecton mortgagelendingdecisions(Grabner&Moers,2013;Ittner&Larcker,2001;Langfield-Smith, 1997; Malmi & Brown, 2008; Zimmerman, 2001). Third, we contribute to the literature examining the impact of EDOs on banks (e.g., Delis et al., 2017, 2020; Danisewicz et al., 2018; Kleymenova & Tomy, 2022; Roman, 2016, 2020). To the best of our knowledge, we are the first to investigate the effect of the supervisory enforcement process on changes in banks’ borrower bases and to study the channels through which it manifests. Finally, we contribute to the literature on mortgage lending to minority borrowers. A large body of work in this area finds disparities in credit access. However, this literature has not reached a consensus on whether non-economic factors, such as race and gender, influence lenders’ decisions to extend credit (Holmes & Horvitz, 1994; Munnell et al., 1996; 8

Horne, 1997; Blanchflower et al., 2003; Asiedu et al., 2012). Our findings suggest that EDOs result in greater access to lending for minority communities through improvements in banks’ internal operations, even for enforcement actions unrelated to fair lending laws. Our work has implications for the enforcement of fair lending laws, which often rely on outcomebased measures such as banks’ share of lending to minorities in a region (e.g., Walter, 1995). Specifically,weemphasizethepossibilitythatimprovementsinbanks’administrativecontrols can play an important role in enhancing access to credit for minority borrowers. 2. Data and sample Our data come from various sources. We identify all enforcement actions issued by bank regulators starting from 1997 using the S&P Global SNL Financial database. Several types of enforcement actions exist, and they vary by degree of severity. Similar to other research using EDOs (Delis et al., 2017; Kleymenova & Tomy, 2022), we restrict our analyses to the most common and severe EDO types that require banks to take corrective actions: cease and desist (C&D) orders, formal or supervisory agreements, consent orders, and prompt corrective action (PCA) orders. C&D orders are enforceable, injunction-type orders that may be issued to a bank when it engages, has engaged, or is about to engage in an unsafe or unsound banking practice or violation of the law. Formal agreements prescribe restrictions and remedies that banks must take to return to a safe and sound condition. PCA orders require banks to take measures to protect or raise the level of their regulatory capital. Our main sample consists of 1,350 unique severe EDOs issued by all federal bank regulators for the years 1997 to 2013, and we use the first EDO that a bank receives.9 Our analyses focus on the three years before an EDO is received, the period when a bank is subject to the EDO, 9Among the 1,350 EDO banks in our sample, 981 have only one EDO; 293 have two; 67 have three; seven have four; and only two banks have five. In our sample, C&D orders are the most common, with 769 EDOs, followed by formal agreements and consent orders (537) and PCA orders (44). We use EDOs from theFederalDepositInsuranceCorp. (FDIC),theFederalReserveSystem,andtheOfficeoftheComptroller of the Currency (OCC). 9

and the five years that follow the EDO’s termination.10 We focus our empirical analyses on commercial banks and obtain their financial data from the Federal Financial Institutions Examination Council (FFIEC) call reports. Table 1, Panel A shows the summary statistics for our sample of EDO banks using quarterly call report data. On average, 65.3% of EDO banks’ assets are in total loans, and 10.2% of total assets are residential mortgages. Total loans are, on average, 78.6% funded by deposits. For our main analyses of residential loan mortgage portfolios and their composition, we use the Home Mortgage Disclosure Act (HMDA) data that provides transaction-level disclosureofresidentialmortgageloanapplicationsandunderwrittenloans, aswellasreasons for denial of an application. These data are available annually. Table 1, Panel A also shows that the percentage market share of residential mortgage lending to minorities in a given county is 0.41%. Table 1, Panel B shows the breakdown of the number of loans originated and the number of applications denied by applicants’ race and gender and loan type and purpose. On average, EDO banks deny 33.8% of all applications. However, minority and female borrowers represent a smaller portion of originated loans and a higher portion of denials (34.5% for minorities and 28.4% for females). We use the reported race and gender oftheprimaryapplicantanddefineminorityborrowersasapplicantswhoseracewasspecified in the loan disclosure documents as nonwhite.11 As can be seen from Panel B of Table 1, the majority of originated loans are for nonminority and male borrowers. We winsorize all of the continuous variables at the 1% and 99% tails of their respective distributions in each sample year and provide detailed definitions of all variables used in our analyses in Appendix A. 10We start our sample in 1997 so that the three-year pre-EDO period begins in 1994 when the Summary of Deposits data begins. We stop our EDO sample in 2013 so that the post-termination period is five years for all EDO banks. 11Minorities are defined as reporting the following races on the application: American Indian or Alaska Native,Asian,BlackorAfricanAmerican,orNativeHawaiianorOtherPacificIslander. NonwhiteHispanics are also included in this definition. Among originated loans, 12.7% do not report race, and 9.2% do not report gender. 10

3. EDO banks’ loans to minority borrowers We begin our analyses by examining the extent to which EDO banks’ lending to minority borrowers changes following the resolution of the EDO. Specifically, we estimate variations of the following staggered difference-in-differences model. Portfolio shares = β +β During EDO +β Post EDO +γX itc 0 1 it 2 it i(t−1)c (1) +α +δ +η +ϵ , i t c itc where i indexes the bank, t the year, and c the county. The dependent variable, Portfolio shares, represents residential mortgage loans to minorities as a share of banks’ total residential mortgage loans at the bank-county level. During EDO is an indicator that equals one for the period an EDO is in effect and zero otherwise; Post EDO is an indicator that equals one for the five years after the termination of the EDO and zero otherwise; X is a vector of lagged control variables, and includes size, profitability, liquidity, capital ratio, nonperforming assets, county-level employment growth as a control for local economic conditions, and the number of mortgage loan applications scaled by county population as a control for loan demand; and α , δ and η are bank, year, and county effects, respectively. The benchmark i t c period is three years prior to the issuance of the EDO. The sample includes all EDO and non- EDO banks. For EDO banks, we only retain data for the benchmark period, the duration of the EDO, and five years after the termination of the EDO. We apply this restriction to all of our specifications. If EDO banks increase their portfolio share of lending to minorities following EDO termination, we expect β to be positive and significant. 2 The dependent variable (Portfolio shares) contains many zero values because banks do not lend to minorities in all counties where they operate.12 Prior studies have used Tobit models to analyze data in cases where the dependent variable has many zeros. For example, 12As can be seen in Table OA1 of the online appendix, EDO banks lend to minorities in only 29% (6/21) of the counties where they are active during the EDO. This figure increases to 35% (11/31) in the five years after EDO termination. 11

Yermack (1995) uses a Tobit specification to analyze CEO stock option awards because, in close to 45% of firm-years, there is no CEO stock option award resulting in a mass of observations at zero. Rosen & Wu (2004) model the portfolio shares of investment in certain asset classes using a random-effects Tobit estimator. Poterba & Samwick (2003) also use a Tobit specification to model portfolio shares of financial assets held by households.13 Followingtheliterature, weestimateEquation1usingaTobitregressionmodel(Tobin,1958; Boulton & Williford, 2018; Keele & Miratrix, 2019). We present our results from this estimation in Table 2, Panel A. Column (1) shows that the share of residential mortgage loans to minorities in banks’ total residential mortgage portfolio increases by a relative 2% following EDO termination. While column (1) shows changes in the portfolio shares for all minority borrowers, column (2) focuses on Black or African American borrowers. Consistent with the result for all minorities, EDO banks increase their portfolio shares of residential mortgage loans to Black or African American borrowers by 2.4% following EDO termination. Column (3) presents the results for portfolio shares of loans to Black or African American borrowers relative to white males and shows a 7% increase in lending to this group following the termination of the enforcement action. Our results are robust to excluding enforcement actions issued specifically for violations of fair lending laws.14 We conduct additional analyses to mitigate concerns that the increase in lending to minorities may be driven by changes in loan demand or underlying local economic conditions that affect all commercial banks, including those that did not receive an EDO. In our main analysisdescribedabove, weuseastaggereddifference-in-differencesresearchdesigntostudy changes in EDO banks’ portfolio shares of lending to minorities relative to all other banks. Our analysis includes the total number of loan applications scaled by the population at the 13ForotherexamplesofstudiesthatusearandomeffectsTobitspecification,pleaseseeBorokhovichetal. (2000); Haigh & List (2005); Edwards (2008) and Chay & Suh (2009). Also, a Tobit specification assumes that the zero and positive observations are generated by the same mechanism (Silva et al., 2015). 14In our sample, only 18 EDOs relate to the lack of compliance with fair lending laws. Fair lending laws are examined and enforced through a separate mechanism. 12

county-year level as a control for changes in loan demand. In additional robustness tests, we study changes in the market shares of residential mortgage loans to minorities. Specifically, we create a variable Market shares, which is loans to minority borrowers granted by EDO banks as a share of total loans to minority borrowers made by all banks in a given county. We reestimate Equation 1 using Market shares as the dependent variable. This approach allows us to estimate changes in lending to minorities by EDO banks relative to all other banks operating in a county.15 As before, given many zeros in Market shares, we employ a Tobit regression model in our estimations. Column (4) of Table 2, Panel A, presents the results of this estimation. The sample in this column includes all counties where EDO banks lend at least once in the sample period. The table shows that EDO banks significantly expand lending to minorities in the years following EDO termination. Relative to the pre-EDO period, the market share in mortgage lending to minorities increases by 0.62%. On average, as reported in Panel A of Table 1, EDO banks have a market share of 0.41% in mortgage lending to minorities over our sample period, suggesting that the changes in market shares are economically significant. In column (5), Market shares is redefined to include only Black or African American borrowers. The column shows that EDO banks’ market shares of loans to Black or African American borrowers increase by 0.58% following the termination of the enforcement action. These results mitigate concerns that macroeconomic changes in the local market could have driven EDO banks’ increase in lending to minorities because, relative to non-EDO banks operating in the county, EDO banks disproportionately expand their lending to minority communities. One concern with the market share analysis is that the counties are equally weighted, which mayoverweightsmallercountiesandobscuretheeconomicsignificance. Therefore, weweight 15An alternative approach to account for local economic conditions is to use transaction-level data and county × year fixed effects (Buchak et al., 2018; Fuster et al., 2019). A drawback of this approach in our settingisthatmultiplebanksinacountycouldreceiveEDOsduringoverlappingtimeperiods. Thisapproach wouldresultinalltransactionsofnon-EDObanks(atthecountylevel)beingrepeatedmultipletimesinthe dataset, quickly inflating our sample. Therefore, we believe that in our setting, our current approach using market shares is a better-suited and clearer way to account for changes in local economic conditions. 13

our regressions by county size using county-level population as a robustness check. Results from this estimation are presented in Table OA2 of the online appendix and show that our inferences continue to hold. Recent work has pointed out several issues with estimates from staggered difference-indifferences analyses (Baker et al., 2022; Barrios, 2021; Goodman-Bacon, 2021; Sant’Anna & Zhao, 2020; Sun & Abraham, 2021). Specifically, the combination of staggered treatment timing and dynamic treatment effects can result in biased estimates due to a “bad comparisons” problem. The issue is particularly severe for smaller sample sizes in which predominantly all units are treated and when there is considerable heterogeneity in treatment (Baker et al., 2022). Therefore, this concern is less pronounced in our analyses because the majority of banks in our sample (86%) are not treated (i.e., did not receive an enforcement action). Furthermore, we have limited heterogeneity in treatment because we restrict our sample of enforcement actions to only severe EDOs. Nonetheless, we conduct additional analyses to allay this concern. One of the recommendations to deal with the issue of bias in staggered difference-indifferences regressions is to create stacked cohorts of separate subsamples of treated and control units by events (Cengiz et al., 2019; Sun & Abraham, 2021). In additional analyses, we follow this approach by creating subsamples for each treated (EDO) bank matched to a control sample of non-EDO banks. We match the control banks on size and geography (county), stack the subsamples, and estimate the following specification: Portfolio shares = β +β During EDO +β Post EDO itc 0 1 it 2 it +β During EDO ×Treatment +β Post EDO ×Treatment (2) 3 it i 4 it i +γX +α +δ +η +ϵ , i(t−1)c i t c itc where Treatment is an indicator that takes the value of 1 for EDO banks and 0 otherwise. The remaining variables are as defined before. If EDO banks increase lending to minorities following EDO termination, we expect β to be positive and significant. Table 2, Panel B 4 14

shows the results from estimating Equation 2. Consistent with our main findings, EDO banks significantly increase lending to minorities relative to the matched sample of control banks. 4. Economic channels Having established that EDO banks increase mortgage lending following EDO termination, we next examine the economic channels that explain this increase in lending. Similar to Granja & Leuz (2022), one plausible explanation is that EDOs force banks to resolve fundamental deficiencies in internal bank management practices that expand minority borrowers’ access to mortgage loans. For example, enforcement actions may require loan policies that specify standards for assessing credit risk, require an internal review of loans, establish a loan committee, or spell out the committee members’ responsibilities. Such changes in loan policies could improve credit assessment as banks follow established standards and procedures. Improvements in credit assessment may also lead to EDO banks better analyzing alternative sources of information and thereby reducing their reliance on a single metric, such as a credit score. Minority borrowers are more likely to be denied a loan based on credit scores because they tend to have lower wealth and are more prone to income shocks. These factors impede their ability to build a strong credit history, which is an important determinant of credit scores. Enforcement actions may also improve internal audit procedures that require compliance with applicable statutes and regulations and with policies prescribed by the management or board. Such changes improve the internal governance at EDO banks as they increase compliance with regulation and internal bank policy. Finally, written loan and internal audit procedures may also reduce the subjectivity afforded to individual loan officers, which may disproportionately benefit minority borrowers. 15

To evaluate our conjecture, we estimate variations of the following model: Portfolio shares = β +β During EDO +β Post EDO +β Treatment itc 0 1 it 2 it 3 i +β During EDO ×Treatment +β Post EDO ×Treatment (3) 4 it i 5 it i +γX +α +δ +η +ϵ , iτ−1 i t c itc where Treatment represents variables associated with greater improvements in internal governance and administrative controls following the receipt of an enforcement action. The remaining variables are as defined before. WeconstructourfirsttwomeasuresofTreatment byanalyzingEDOs’textandidentifying the specific details of the corrective actions supervisors require banks to take. Using the textual content of enforcement orders, we identify EDOs that explicitly require a bank to establish or revise a loan policy or develop written internal audit procedures. We create two variables to reflect such improvements. The first, Loan policy, is an indicator of whether the enforcement order requires revising or establishing a loan policy. The second, Internal audit, is an indicator if the order requires the affected bank to develop written internal audit procedures.16 Column (1) of Table 3 shows no change in the portfolio share of loans to minorities following EDO termination for enforcement orders that require changes in loan policy. However, in column (2), we interact Post EDO × Treatment with Subprime share, which is the percentage of borrowers in the county with FICO scores of 619 or below.17 The results in column (2) indicate that loan policy-related improvements are associated with an increase in lending to minority borrowers located in regions with a greater share of low credit scores. 16In Appendix B.2 of the online appendix, we provide excerpts from an enforcement order that required changes to loan policy and internal audit procedures. 17A FICO score is a credit score created by the Fair Isaac Corporation. We source FICO scores from the CoreLogicLoan-LevelMarketAnalyticsdataset. WeaggregatetheloanoriginationdatatotheZIPcodeand origination year level. We then convert ZIP-code-level FICO scores to the county level by using a crosswalk file from the Department of Housing and Urban Development, which contains the fraction of all addresses in a given ZIP code belonging to a county. Our definition of subprime is based on Keys et al. (2010). 16

In terms of the economic magnitude, at the 75th percentile of Subprime share, EDO banks with loan policy changes experience a 6.2% increase in lending to minorities relative to EDO banks without loan policy changes. We find similar results based on our second measure of whether the enforcement order requiredwritteninternalauditprocedures. Theseresultsarepresentedincolumns(3)and(4) of Table 3. Column (4) shows that EDO banks that had to implement written internal audit procedures increased lending to minorities in counties with a greater share of borrowers with low credit scores. In terms of the economic magnitude, at the 75th percentile of Subprime share, EDO banks with internal audit changes experience a 16.2% increase in lending to minorities relative to EDO banks without internal audit changes. In cross-sectional analyses, we further investigate the lending behavior of banks that are likely to have witnessed greater improvements in their internal governance due to the enforcement process. In our first set of tests, we reestimate Equation 3 with Treatment representingthestrictnessoftheregulator. WeexpectthatEDObanksinstateswithstricter regulators are likely to improve more as a result of receiving an EDO. We use the measure developed by Agarwal et al. (2014), who find that, due to institutional differences, varying incentives, and resource constraints, state and federal banking regulators are inconsistent in implementing the same supervisory rules. Specifically, based on regulatory ratings, Agarwal et al. (2014) find that federal regulators are generally stricter than state regulators, and there is variation across states in their level of strictness. Although this measure pertains to state regulators, federal and state regulators collaborate in issuing enforcement actions to state-chartered banks. We present our results from this analysis in column (1) of Table 4. The sample only includes state-chartered banks, as the Agarwal et al. (2014) measure applies only to statechartered banks by construction. Our results indicate that EDO banks with stricter regulators expand their portfolio shares of lending to minorities by 7.6% following EDO termination. Next, we estimate Equation 3 with Treatment representing the severity of the 17

enforcement action, measured as the length of time it takes a bank to exit an EDO from its issuance to resolution. Banks with more severe enforcement actions have problems on several fronts that must be resolved before the regulator will terminate the enforcement action. Therefore EDO banks with more severe enforcement actions are more likely to improve their operations following EDO termination, relative to the pre-EDO period. Column (2) of Table 4 shows that banks with more severe EDOs significantly increase lending to minorities after the EDO. Specifically, for these banks, lending to minorities increases by 3.1% following EDO termination. We also conduct tests using banks’ CRA ratings. The CRA was enacted by Congress in 1977 to encourage credit availability in low- and moderate-income areas. Regulators rate banks based on their record in meeting the credit needs of communities in which they operate. These ratings are used to evaluate banks’ applications for deposit facilities which include new charters, deposit insurance, mergers or acquisitions, opening a new branch, or the relocation of a branch or home office. Therefore, banks need to maintain a satisfactory CRA rating if they plan to expand or make any substantial changes to their operations. Furthermore, if banks’ failure to comply with the CRA is correlated with the racial makeup of underserved neighborhoods, intentional discrimination can be inferred (Schwemm, 1994). If the supervisory process improves banks’ internal processes then banks with low pre-EDO CRA ratings should show greater improvements in lending to minorities following the enforcement action. It is important to note that noncompliance with the CRA and low CRA ratings do not result in formal enforcement actions.18 CRA rating changes are relatively infrequent and take one of four possible values: outstanding,satisfactory,needstoimprove,andsubstantialnoncompliance. Themajority(75%) of bank-year observations in our sample of EDO banks have a rating of outstanding or satisfactory. Column (3) of Table 4 shows that banks with a low CRA rating (needs to improve 18In 1994, the Department of Justice issued an opinion that formal EDOs, such as C&D or civil money penalties, do not fall into the scope of CRA (for more details, please see “Community Reinvestment Act: Challenges Remain to Successfully Implement CRA” (Chapter Report, 11/28/95, GAO/GGD-96-23)). 18

or substantial noncompliance) in the pre-EDO period expand their lending to minority borrowers by 9.8% in the post-EDO period relative to EDO banks that had an outstanding or satisfactory rating. Overall, our findings suggest that improvements in banks’ operations due to enforcement are associated with increased lending to minorities. These results are consistent with banks having more severe operational deficiencies also exhibiting more scope for improvement in lending practices. Our findings thus far allow us to tie the increase in lending to minorities to the supervisory enforcement actions which require changes in administrative controls such as lending policies and internal governance improvements. To further support our findings of increased mortgage lending to minorities due to the supervisory enforcement process, we next explore how corrective actions directly influence loan approval decisions. 5. Changes in mortgage application denials for minorities We evaluate changes in the loan approval process by investigating changes in denials of mortgage loan applications and the reasons banks list for denying an application from minority borrowers. In particular, we estimate the following OLS model: Denial = β +β During EDO +β Post EDO +β Minority it 0 1 it 2 it 3 i +β During EDO ×Minority +β Post EDO ×Minority (4) 4 it i 5 it i +γX +δ +α ×η +ϵ , i(t−1)c t i c itc where Denial is an indicator variable if a loan application is denied or if it is denied for a specified reason. The remaining variables are as described before. We include year and bank × county fixed effects and therefore account for local economic conditions faced by the same bank lending in different counties. In our sample of mortgage loan applications, 33.8% get denied (Table 1, Panel B). The more frequent reasons for denial include a lack of collateral (32.2% of all cases), a poor credit history (17.8%), and a high debt-to-income ratio (8.4%) (untabulated). Mortgage 19

application requirements, such as collateral, credit history, and debt-to-income ratios, are nonpricetermsthatlendersusetorationcreditandtolimitmoralhazardoradverseselection (Stiglitz & Weiss, 1981). Borrowers who do not meet the thresholds for these terms may not receive credit, even if they are willing to pay higher interest rates. Minority borrowers are more likely to be constrained by nonprice terms because they are more likely to have lower wealth (Acolin et al., 2016; Bostic, 1997; Gyourko et al., 1999). For example, Bostic (1997) finds that minority applicants are rejected more often if debt-to-income ratios are used in credit assessment because they have lower incomes and are, therefore, prone to default in case of income shocks. Table 5 presents the results from estimating Equation 4. Consistent with prior studies (Black et al., 1978; Duca & Rosenthal, 1993; Munnell et al., 1996; Wheeler & Olson, 2015), the coefficient on Minority in column (1) indicates that minorities are 9.6% more likely to be denied loans relative to white borrowers in the pre-EDO period. However, following EDO termination, loan denials for minority borrowers decline by a relative five percentage points. Much of this decline is driven by lower denials due to credit history (a nonprice term). Specifically, EDO banks are 3.4% less likely to deny loans to minorities relative to white borrowers due to their credit history following EDO termination. These results are consistentwithEDObankschangingtheircreditassessmentprocessestorelylessonnonprice termsfollowinganenforcementaction. Forexample,improvementsinloanpoliciesandcredit risk assessment may allow banks to process additional sources of hard information better to assess borrowers’ creditworthiness, as opposed to relying solely on their credit scores. The lack of a strong credit history is reflected in borrowers’ credit scores. For example, FICO scores consider various aspects of individuals’ credit history—the length of their credit history as well as how long they have gone without negative credit events, such as bankruptcies, foreclosures, or delinquencies. Building a credit history requires access to a line of credit, which minority borrowers may find harder to get because they are more likely to have less wealth than white borrowers. Minority borrowers are also more likely to face income 20

shocks and therefore negative credit events. If, following enforcement actions, EDO banks can better process and use alternative sources of hard information, such as utility payments, rental histories, and remittance histories (Brevoort et al., 2016; Schneider & Schutte, 2007), they may deny fewer loan applications based on nonprice terms. Therefore, the decline in denials should be concentrated among borrowers with low credit scores. We do not have information on borrowers’ credit scores; therefore, we proxy for it using theaveragetransaction-matchedcreditscoresatthecensustractlevel. Specifically, wecreate a subprime indicator (Subprime) using FICO scores for originated loans from CoreLogic’s Loan-Level Market Analytics dataset. We calculate average FICO scores from the CoreLogic dataset at the level of the census tract, loan origination year, loan type, loan purpose, and occupancy status of the property. Based on these characteristics, we merge the average FICO scores with the transactions in our sample. Subprime takes a value of 1 if the average transaction-matched FICO score is 619 or below and 0 otherwise.19 We lose 9% of our sample by including the subprime measure because the CoreLogic data does not cover all census tracts for which we have transaction-level data from HMDA. Our results (presented in Table OA3 of the online appendix) indicate that minority borrowers in subprime regions are 5.4% less likely to be denied a mortgage loan application based on nonprice terms, such as collateral requirements. These results are consistent with changes in EDO banks’ loan policies and credit risk assessment leading to less reliance on nonprice terms and, as a result, higher access to residential mortgage loans by minorities. Our analyses offer insights into why lending to minorities increases following EDO termination. We find that EDO banks are less likely to deny loans to minority applicants based on nonprice terms, indicating changes in credit assessment procedures. Reduced reliance on nonprice terms, such as collateral requirements and credit histories, disproportionately affects lending to minorities because this category of borrowers is more likely to be constrained 19In additional robustness tests, we define Subprime as FICO scores of 669 and below and find consistent results. 21

by such terms (Acolin et al., 2016; Bostic, 1997; Gyourko et al., 1999). 6. Changes in risk Next, we investigate whether increased lending to minority borrowers is associated with a rise in risky lending along several dimensions of risk, including nonperforming assets, the market share of risky loans, and changes in FHA lending. If EDO banks were to increase lendingtolesscreditworthycustomers, suchanincreasewouldresultinhighernonperforming assets. Accordingly, we study the changes in EDO banks’ nonperforming assets in the years followingEDOterminationrelativetothepre-EDOperiodbyestimatingthefollowingmodel: NPA = β +β During EDO +β Post EDO +γX +α +δ +ϵ , (5) it 0 1 it 2 it it−1 i t it where NPA is the total and residential nonperforming loans scaled by total loans. The remaining variables are as defined before. Table 6, Panel A, presents our findings from estimating Equation 5. Columns (1) and (2) show changes in total nonperforming assets during and following the termination of an EDO relative to the period prior to the EDO. Column (1) does not include bank-level controls, while column (2) does. Total nonperforming assets increase during an EDO, consistent with regulators inducing banks to recognize previously hidden nonperforming loans. However, nonperforming assets revert to their pre-EDO levels following EDO termination. In column (3), the dependent variable is nonperforming assets for residential mortgages. Due to data restrictions, we can only analyze NPAs for residential mortgages starting from 2001. Consistent with the results for total nonperforming assets, column (3) shows that NPAs for residential mortgages do not increase following EDO termination. Overall, these findings suggest that EDO banks do not witness an increase in their nonperforming assets in the years following EDO termination. Next, we study changes in the market shares of risky mortgage loans originated by EDO banks at the county level. Specifically, we reestimate Equation 5, where the dependent vari- 22

able (Market shares of risky loans) is defined as EDO banks’ share of higher-priced, closedend mortgages as a percentage of such residential mortgage loans made by all commercial banks at the county level. Loans are classified as higher priced if the annual percentage rate (APR) exceeds the average prime offer rate (APOR) for loans of a similar type by at least 1.5 percentage points for first-lien loans or 3.5 percentage points for junior-lien loans. Given data limitations, this analysis starts from 2004. Panel B of Table 6 presents the results from these analyses. Column (1) includes the full sample, whereas column (2) uses the sample conditional on whether the EDO bank makes at least one such risky loan in the county. The dependent variable in column (1) consists of many zeros because EDO banks do not make such loans in all counties where they operate. Accordingly, we use a Tobit specification in estimating column (1). The dependent variable in column (2) contains only positive values for the market share of risky loans. Therefore we estimate column (2) using OLS. Our results indicate a decrease or no change in the market shares of risky loans following EDO termination, suggesting that the increase in lending to minority borrowers is not associated with an increase in risky lending. We also assess whether EDO banks grant fewer FHA loans. FHA loans have lower down-payment requirements and may be offered to borrowers with pre-existing high debt or low credit scores. Therefore, FHA loans tend to be given to riskier borrowers than conventional mortgages (Fuster et al., 2019). As further evidence of banks decreasing risky lending following EDO termination, we find a decline in FHA loans to minorities originated by these banks. Specifically, Panel C of Table 6 shows that FHA loans to minorities decline by 4.7%–5.6% following EDO termination. Overall, our findings in this section suggest that theincreaseinminoritylendingfollowingEDOterminationisnotassociatedwithanincrease in risky lending. This result is consistent with the enforcement process improving credit risk assessment processes at EDO banks. 23

7. Alternative mechanisms Next, we investigate two alternative mechanisms for the increase in lending to minorities following EDO termination. First, EDO banks may have expanded residential mortgage lending to improve their capital ratios, and this expansion would be possible by only lending to previously underserved borrowers. Second, increased competition from non-EDO banks may have resulted in EDO banks expanding their lending to minority borrowers. 7.1. Improving capital ratios Because secured loans have relatively lower risk weights, EDO banks could increase their capital ratios by expanding residential mortgage lending. However, an increase in this kind of lending may be possible only if EDO banks expand lending to previously underserved categories of borrowers, such as minorities. To test this hypothesis, we reestimate Equation 3 where Treatment represents low capital, measured as an indicator for EDO banks in the lowest tercile of regulatory capital in the period prior to receiving an EDO. We present our findings from this estimation in column (1) of Table 7. The results do not suggest that EDO banks expand lending to minorities to manage their capital following the termination of their enforcement actions. 7.2. Competition from non-EDO banks We also investigate whether competition from banks that did not receive enforcement actions leads EDO banks to expand their lending to minorities. Increased competition could result in greater lending to minority borrowers for two reasons. First, because EDO banks losedepositsandlikelyfacereputationalcostsduetothepublicdisclosureofEDOs, theymay lose their more profitable customers to their competitor non-EDO banks. This might force EDO banks to expand their reach to new borrowers who previously did not qualify for a loan. Second, because competition erodes excess margins, it increases the cost of discriminating. If banks were previously engaged in taste-based discrimination (Becker, 1957), they would have had to pay a cost for the utility derived from not lending to specific groups of borrowers. An 24

increase in competition reduces banks’ ability to pay this cost, resulting in greater lending to minority borrowers. This argument is consistent with prior work that finds increased competition results in a more equitable distribution of rents (Ashenfelter & Hannan, 1986; Black & Brainerd, 1999; Black & Strahan, 2001). To evaluate whether competition from non-EDO banks drives the increase in lending to minorities, we study the impact of market concentration in the deposits and residential mortgage markets on EDO banks’ lending. If, driven by competition from non-EDO banks, EDO banks were to increase their lending to minorities, the increase should be higher in counties where EDO banks face greater competition for deposits and loans. Accordingly, we reestimate Equation 3, where Treatment represents a highly competitive environment for an EDO bank. Our proxy for higher competition is a measure of the deposit or loan market concentration based on the lowest tercile of the Herfindahl-Hirschman index (HHI) measured in the year prior to the EDO issuance in a given county. Wepresenttheresultsfromthisanalysisincolumns(2)and(3)ofTable7. Thecoefficient for Treatment indicates that lending to minorities forms a greater share of banks’ lending portfoliosinhighlycompetitivecounties,supportingthevalidityofourmeasures(Ashenfelter & Hannan, 1986; Black & Brainerd, 1999; Black & Strahan, 2001). However, we do not find that EDO banks in high-competition counties increase lending to minorities more following the termination of their enforcement actions, suggesting that an increase in competition from non-EDO banks does not drive our results. Our findings in Section 6 that banks do not experience an increase in the riskiness of loans following EDO termination are also inconsistent with the competition channel. If, driven by a loss of better customers to competitors, EDO banks were to increase lending to less creditworthy customers, the increase should result in higher nonperforming assets or an increase in risky lending. Overall, our results suggest that competition from non-EDO banks is unlikely to drive our findings. 25

8. Supplemental analyses: Lending to women To further support our hypothesis, we explore lending to another category of borrowers whose credit risk was historically difficult to evaluate: women who are primary or solo mortgage borrowers. Women, in general, tend to have lower wealth and shorter credit histories, putting them at a disadvantage if banks rely on summary measures of credit risk, such as the credit history. Similar to our analyses for minority borrowers, we explore whether EDO banks expand their lending to women. Specifically, we reestimate Equation 1 with the dependent variables representing lending to female borrowers. Table 8 presents the results from this analysis. The dependent variable in column (1) represents lending to women as a share of banks’ portfolio of residential mortgage lending at the bank-county level. Consistent with our results for minority borrowers, EDO banks expand their portfolio share of lending to women by 6.1% following EDO termination. We also find an increase of 3.4% in mortgage lending to women during the time the EDO is in effect. Column (2) of Table 8 shows the market shares results. Banks significantly expand lending to women following EDO termination. Relative to the pre-EDO period, EDO banks’ market share in mortgage lending to women increases by 0.72%. The results in Table 8 indicate that similar to our findings for minority borrowers, EDO banks also expand lending to women who are primary or solo borrowers. Our findings are consistent with improvements atthebankduetotheenforcementprocessdrivingaccesstocreditforborrowerswhosecredit risk is more difficult to evaluate. 9. Conclusion Recent research exploits variation in the strictness of bank supervision to examine the effects of supervision on credit supply (Agarwal et al., 2014; Granja & Leuz, 2022). Supervisory activities can influence loan supply by identifying deficiencies in banks’ management practices, including operating procedures and policies and internal governance structures. We hypothesize that improvements in management control systems imposed by EDOs serve 26

as channels through which EDOs affect a bank’s borrower base generally, and minority lending specifically. Management controls can serve an important role in aligning the behavior of employees with an organization’s objectives by establishing action protocols and directing employees to adhere to policies (Abernethy & Chua, 1996; Malmi & Brown, 2008; Merchant & Van der Stede, 2007; Simons, 1987). We empirically examine how changes in loan policies and internal governance mechanisms specified in EDOs influence banks’ mortgage lending decisions. Our focus on minority lending builds on research showing that racial disparities can arise from the biases of individual loan officers and limitations of the scope of borrowers’ information used in lending decisions (Di Maggio et al., 2022; Jiang et al., 2022). In this regard, changes in management controls that improve loan policies, operating procedures and employees’ adherence to such policies and procedures, could disproportionately benefit minority borrowers by reducing discretion in lending decisions. Furthermore, improvements increditassessmentprocedurescoulddirecttheuseofadditionalsourcesofhardinformation, reducing reliance on single metrics such as credit scores. Minority borrowers are less likely to have a line of credit for building a credit history and are more prone to income shocks; therefore, they are more likely to be disadvantaged by banks’ reliance on metrics, such as credit scores (Acolin et al., 2016; Bostic, 1997; Brevoort et al., 2016; Gyourko et al., 1999; Schneider & Schutte, 2007). We find that, following the termination of enforcement actions, banks significantly increase residential mortgage lending to minorities and increase their market share of lending to this group of borrowers within the counties where they operate. Our results are robust to excluding enforcement actions received for violating fair lending laws. We identify EDOs that focus on lending policies and internal governance improvements and find stronger results, particularly in regions with a greater proportion of subprime borrowers. The effect is also stronger for banks likely to have experienced greater improvements—–those with stricter bank supervisors, more severe EDOs, and low CRA ratings in the pre-EDO period. 27

In studying how such corrective actions can influence the loan approval process, we find that the rejection of minority loan applications due to non-price terms is less likely following termination of the enforcement order. Non-price terms such as credit history, collateral, and debt-to-income ratios are more likely to constrain minority borrowers because these borrowers are more likely to have lower wealth. Therefore, improvements in credit assessment couldallowbankstoprocessadditionalsourcesofhardinformationandrelylessonnon-price terms, disproportionately benefiting minority borrowers. We find no evidence that the increase in mortgage lending to minorities is associated with diminished loan performance risk. We also find no support for the alternative explanations that low capital or competition from non-EDO banks may be driving our results. Finally, we findsimilarincreasesinlendingtoanotherclassofborrowerswhosecreditriskwashistorically difficult to evaluate—–women who are primary or solo borrowers—–consistent with process improvements at EDO banks increasing access to credit for marginalized borrowers. While previous literature considers the effects of bank supervision on credit supply (e.g., Agarwal et al., 2014; Granja & Leuz, 2022), we extend this literature by investigating how EDOs influence the allocation of mortgage lending across demographic groups underlying a banks’ borrower base. We also highlight the critical role of improvements in management control systems in shaping banks’ lending behavior. We show that supervisory enforcement, through its impact on banks’ internal management procedures, results in greater access to credit for minority borrowers, even for enforcement actions unrelated to fair lending laws. Although our analysis focuses on poorly managed banks relative to the general population of banks, our sample of banks receives enforcement actions as a measure of last resort, our findings highlight important policy implications. Specifically, we underscore the importance of proper administrative controls at the bank as a critical factor in enhancing access to credit for minority borrowers. We study extreme examples of banks’ weak administrative controls, allowing us to identify improvements in banks’ lending policies and internal operating procedures. We look forward to future research studying the impact of prudential bank 28

supervision on minority borrowers in other settings. 29

References Aaronson, D.(2000). Anoteonthebenefitsofhomeownership. Journal of Urban Economics, 47, 356–369. Abernethy, M. A., & Chua, W. F. (1996). A field study of control system “redesign”: The impact of institutional processes on strategic choice. Contemporary Accounting Research, 13, 569–606. Acolin, A., Bricker, J., Calem, P., & Wachter, S. (2016). Borrowing constraints and homeownership. The American Economic Review, 106, 625–29. Agarwal, S., Lucca, D., Seru, A., & Trebbi, F. (2014). Inconsistent regulators: Evidence from banking. The Quarterly Journal of Economics, 129, 889–938. Altavilla, C., Boucinha, M., Peydro´, J.-L., & Smets, F. (2020). Banking supervision, monetary policy and risk-taking: big data evidence from 15 credit registers. Monetary Policy and Risk-Taking: Big Data Evidence from, 15. Ashenfelter, O., & Hannan, T. (1986). Sex discrimination and product market competition: The case of the banking industry. The Quarterly Journal of Economics, 101, 149–173. Asiedu, E., Freeman, J. A., & Nti-Addae, A. (2012). Access to credit by small businesses: How relevant are race, ethnicity, and gender? The American Economic Review, 102, 532–37. Baker, A. C., Larcker, D. F., & Wang, C. C. (2022). How much should we trust staggered difference-in-differences estimates? Journal of Financial Economics, 144, 370–395. Barrios, J. M. (2021). Staggeringly problematic: A primer on staggered DiD for accounting researchers. Working Paper. Becker, G. S. (1957). The Economics of Discrimination. University of Chicago Press. Black, H., Schweitzer, R. L., & Mandell, L. (1978). Discrimination in mortgage lending. The American Economic Review, 68, 186–191. Black, S.E., &Brainerd, E.(1999). Importingequality? Theeffectsofincreasedcompetition on the gender wage gap. FRB of New York Staff Report, 74. Black, S.E., &Strahan, P.E.(2001). Thedivisionofspoils: Rent-sharinganddiscrimination in a regulated industry. The American Economic Review, 91, 814–831. Blanchflower, D. G., Levine, P. B., & Zimmerman, D. J. (2003). Discrimination in the small-business credit market. Review of Economics and Statistics, 85, 930–943. Blau, F. D., & Graham, J. W. (1990). Black-white differences in wealth and asset composition. The Quarterly Journal of Economics, 105, 321–339. 30

Borokhovich, K. A., Bricker, R. J., & Simkins, B. J. (2000). An analysis of finance journal impact factors. The Journal of Finance, 55, 1457–1469. Bostic, R. W. (1997). The role of race in mortgage lending: Revisiting the Boston Fed study. Board of Governors of the Federal Reserve System Finance and Econ. Disc. Series, 97–2. Boulton, A. J., & Williford, A. (2018). Analyzing skewed continuous outcomes with many zeros: A tutorial for social work and youth prevention science researchers. Journal of the Society for Social Work and Research, 9, 721–740. Brevoort, K. P., Grimm, P., & Kambara, M. (2016). Credit invisibles and the unscored. Cityscape, 18, 9–34. Brickley, J., Smith, C., & Zimmerman, J. (2015). Managerial economics and organizational architecture. McGraw-Hill Education. Buchak, G., Matvos, G., Piskorski, T., & Seru, A. (2018). Fintech, regulatory arbitrage, and the rise of shadow banks. Journal of Financial Economics, 130, 453–483. Cengiz, D., Dube, A., Lindner, A., & Zipperer, B. (2019). The effect of minimum wages on low-wage jobs. The Quarterly Journal of Economics, 134, 1405–1454. Chay, J.-B., & Suh, J. (2009). Payout policy and cash-flow uncertainty. Journal of Financial Economics, 93, 88–107. Collins, W. J., & Margo, R. A. (2001). Race and home ownership: A century-long view. Explorations in Economic History, 38, 68–92. Cort´es, K. R., Demyanyk, Y., Li, L., Loutskina, E., & Strahan, P. E. (2020). Stress tests and small business lending. Journal of Financial Economics, 136, 260–279. Danisewicz, P., McGowan, D., Onali, E., & Schaeck, K. (2018). The real effects of banking supervision: Evidence from enforcement actions. Journal of Financial Intermediation, 35, 86–101. Delis, M. D., Iosifidi, M., Kokas, S., Xefteris, D., & Ongena, S. (2020). Enforcement actions on banks and the structure of loan syndicates. Journal of Corporate Finance, 60, 101527. Delis, M. D., Staikouras, P. K., & Tsoumas, C. (2017). Formal enforcement actions and bank behavior. Management Science, 63, 959–987. Di, Z. X., Belsky, E., & Liu, X. (2007). Do homeowners achieve more household wealth in the long run? Journal of Housing Economics, 16, 274–290. Di Maggio, M., Ratnadiwakara, D., & Carmichael, D. (2022). Invisible primes: Fintech lending with alternative data. NBER Working Paper No. w29840. Duca, J. V., & Rosenthal, S. S. (1993). Borrowing constraints, household debt, and racial discrimination in loan markets. Journal of Financial Intermediation, 3, 77–103. 31

Edwards, R. D. (2008). Health risk and portfolio choice. Journal of Business & Economic Statistics, 26, 472–485. Eisenbach, T. M., Lucca, D. O., & Townsend, R. M. (2022). Resource allocation in bank supervision: Trade-offs and outcomes. The Journal of Finance, 77, 1685–1736. Frame, W. S., Huang, R., Mayer, E. J., & Sunderam, A. (2022). The impact of minority representation at mortgage lenders. NBER Working Paper No. w30125. Fuster, A., Plosser, M., Schnabl, P., & Vickery, J. (2019). The role of technology in mortgage lending. The Review of Financial Studies, 32, 1854–1899. Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing. Journal of Econometrics, 225, 254–277. Grabner, I., & Moers, F. (2013). Management control as a system or a package? Conceptual and empirical issues. Accounting, Organizations and Society, 38, 407–419. Granja, J., & Leuz, C. (2022). The death of a regulator: Strict supervision, bank lending and business activity. European Corporate Governance Institute (ECGI) - Law Working Paper No. 447/2019, LawFin Working Paper No. 4. Green, R. K., White, M. J. et al. (1997). Measuring the benefits of homeowning: Effects on children. Journal of Urban Economics, 41, 441–461. Guryan, J., & Charles, K. K. (2013). Taste-Based or Statistical Discrimination: The Economics of Discrimination Returns to its Roots. The Economic Journal, 123, F417–F432. Gyourko, J., Linneman, P., & Wachter, S. (1999). Analyzing the relationships among race, wealth, and home ownership in America. Journal of Housing Economics, 8, 63–89. Haigh, M. S., & List, J. A. (2005). Do professional traders exhibit myopic loss aversion? An experimental analysis. The Journal of Finance, 60, 523–534. Hirtle, B., Kovner, A., &Plosser, M.(2020). Theimpactofsupervisiononbankperformance. The Journal of Finance, 75, 2765–2808. Holmes, A., & Horvitz, P. (1994). Mortgage redlining: Race, risk, and demand. The Journal of Finance, 49, 81–99. Horne, D. K. (1997). Mortgage lending, race, and model specification. Journal of Financial Services Research, 11, 43–68. Ittner, C.D., &Larcker, D.F.(2001). Assessingempiricalresearchinmanagerialaccounting: A value-based management perspective. Journal of Accounting and Economics, 32, 349– 410. Jiang, E. X., Lee, Y., & Liu, W. S. (2022). Disparities in consumer credit: The role of loan officers in the FinTech era. Working Paper. 32

Kandrac, J., & Schlusche, B. (2021). The effect of bank supervision and examination on risk taking: Evidence from a natural experiment. The Review of Financial Studies, 34, 3181–3212. Keele, L., & Miratrix, L. (2019). Randomization inference for outcomes with clumping at zero. The American Statistician, 73, 141–150. Keys, B. J., Mukherjee, T., Seru, A., & Vig, V. (2010). Did securitization lead to lax screening? Evidence from subprime loans. The Quarterly Journal of Economics, 125, 307–362. Kleymenova, A., & Tomy, R. E. (2022). Observing enforcement: Evidence from banking. Journal of Accounting Research, 60, 1583–1633. Langfield-Smith, K. (1997). Management control systems and strategy: a critical review. Accounting, organizations and society, 22, 207–232. Malmi, T., & Brown, D. A. (2008). Management control systems as a package—opportunities, challengesandresearchdirections. Management Accounting Research, 19, 287–300. Merchant, K.A., &VanderStede, W.A.(2007). Management control systems: Performance measurement, evaluation and incentives. Pearson Education. Munnell, A. H., Tootell, G. M., Browne, L. E., & McEneaney, J. (1996). Mortgage lending in Boston: Interpreting HMDA data. The American Economic Review, (pp. 25–53). Newman, S. J., & Holupka, C. S. (2016). Is timing everything? Race, homeownership and net worth in the tumultuous 2000s. Real Estate Economics, 44, 307–354. Poterba, J. M., & Samwick, A. A. (2003). Taxation and household portfolio composition: US evidence from the 1980s and 1990s. Journal of Public Economics, 87, 5–38. Roman, R. A. (2016). Enforcement actions and bank loan contracting. Economic Review (Kansas City), 101, 69–101. Roman, R. A. (2020). Winners and losers from supervisory enforcement actions against banks. Journal of Corporate Finance, 60, 101516. Rosen, H. S., & Wu, S. (2004). Portfolio choice and health status. Journal of Financial Economics, 72, 457–484. Sant’Anna, P. H., & Zhao, J. (2020). Doubly robust difference-in-differences estimators. Journal of Econometrics, 219, 101–122. Schneider, R., & Schutte, A. (2007). The predictive value of alternative credit scores. Chicago: Center for Financial Services Innovation. Schwemm, R. G. (1994). Introduction to mortgage lending discrimination law. The John Marshall Law Review, 28, 317. 33

Schwert, M. (2018). Bank capital and lending relationships. The Journal of Finance, 73, 787–830. Shapiro, T. M. (2006). Race, homeownership and wealth. Wash. UJL & Pol’y, 20, 53–74. Silva, J. M. S., Tenreyro, S., & Windmeijer, F. (2015). Testing competing models for nonnegative data with many zeros. Journal of Econometric Methods, 4, 29–46. Simons, R. (1987). Accounting control systems and business strategy: An empirical analysis. Accounting, Organizations and Society, 12, 357–374. Stiglitz, J. E., & Weiss, A. (1981). Credit rationing in markets with imperfect information. The American Economic Review, 71, 393–410. Sun, L., & Abraham, S. (2021). Estimating dynamic treatment effects in event studies with heterogeneous treatment effects. Journal of Econometrics, 225, 175–199. Tobin, J. (1958). Estimation of relationships for limited dependent variables. Econometrica, 26, 24–36. Tucker, B., Thorne, H., & Gurd, B. (2009). Management control systems and strategy: What’s been happening? Journal of Accounting Literature, 28, 123. Wainer, A., & Zabel, J. (2020). Homeownership and wealth accumulation for low-income households. Journal of Housing Economics, 47, 101624. Walter, J. R. (1995). The fair lending laws and their enforcement. FRB Richmond Economic Quarterly, 81, 61–77. Wheeler, C. H., & Olson, L. M. (2015). Racial differences in mortgage denials over the housing cycle: Evidence from US metropolitan areas. Journal of Housing Economics, 30, 33–49. Yermack, D. (1995). Do corporations award CEO stock options effectively? Journal of Financial Economics, 39, 237–269. Zimmerman, J. L. (2001). Conjectures regarding empirical managerial accounting research. Journal of Accounting and Economics, 32, 411–427. 34

Appendix A. Variable definitions Variable Definition Source Code DependentVariables Denial Indicator variable, which takes HMDA ActionTaken=3 the value of 1 if a mortgage applicationisdeniedbyafinancial institutionand0otherwise Marketshares Total residential mortgage loans HMDA and authors’ to minorities (women) for EDO calculations banks in a county / Total residentialmortgageloanstominorities(women)forallbanksinthe county Portfolioshares Total residential mortgage loans HMDA and authors’ tominorities(women)foragiven calculations bank/Totalresidentialmortgage loans IndependentVariables ConventionalLoans Indicator variable, which takes HMDA LoanType=1 the value of 1 if the loan type is conventional and 0 otherwise. Conventionalloansareanyloans other than FHA, VA, FSA, or RHSloans DuringEDO Indicator variable, which takes SNLandauthors’calthevalueof1fromtheyearEDO culations wasissuedtotheyearEDOwas terminatedand0otherwise. EDOLength EDOlengthinyears SNL FHA-insuredLoans Indicator variable which takes HMDA LoanType=2 the value of 1 if loan type is FHA(Federal Housing Administration)-insuredloansand0otherwise. FSA/RHSLoans Indicator variable, which takes HMDA LoanType=4 thevalueof1iftheloantypeis FSA/RHS(FarmServiceAgency orRuralHousingService)and0 otherwise. 35

HighCompetition Indicator variable, which takes Summary of Deposits the value of 1 for the lowest and authors’ calculadeposit or residential mortgage tions market HHI tercile in a given countyand0otherwise. Home Improvement, Non- Indicator variable, which takes HMDA Loan Purpose = 2 & Owner- Owneroccupied thevalueof1iftheloanpurpose Occupancy=2 is a home improvement and the property is not owner-occupied and0otherwise. Home Improvement, Owner oc- Indicator variable, which takes HMDA Loan Purpose = 2 & Ownercupied thevalueof1iftheloanpurpose Occupancy=1 is a home improvement and the property is owner-occupied as a principal dwelling and 0 otherwise. HomePurchase,Non-Owneroc- Indicator variable, which takes HMDA Loan Purpose = 1 & Ownercupied thevalueof1iftheloanpurpose Occupancy=2 is home purchase and the propertyisnotowner-occupiedand0 otherwise. Home Purchase, Owner occu- Indicator variable, which takes HMDA Loan Purpose = 1 & Ownerpied thevalueof1iftheloanpurpose Occupancy=1 is home purchase and the propertyisowner-occupiedasaprincipaldwellingand0otherwise. InternalAudit Indicator variable, which takes SNL,FDIC,OCC,and the value of 1 if the text of an FRB EDO requires improvements in internalauditprocedures. LoanPolicy Indicator variable, which takes SNL,FDIC,OCC,and the value of 1 if the text of FRB anEDOrequiresimprovementin loanpolicy. LowCapital Indicator variable, which takes CallReports RCFD3210/RCFD2170 the value of 1 if an EDO bank isinthelowesttercileofcapital ratiointheperiodpriortoreceivinganEDO. LowCRA Indicator variable, which takes FFIEC Interathevalueof1ifanEDObankre- gency CRA Ratings ceivesaCRAratingof3(Needs Database to Improve) or 4 (Substantial Noncompliance) at least once in the 3 years of pre-EDO period and0otherwise. Male Indicator variable, which takes HMDA Sex=1 the value of 1 if a mortgag3e6applicantismaleand0otherwise.

Minority Indicator variable, which takes HMDA Race=1,2,3,or4 the value of 1 if a mortgage applicantisnon-whiteand0otherwise. PostEDO Indicator variable, which takes SNLandauthors’calthevalueof1forthefiveyearsaf- culations tertheEDOwasterminatedand 0otherwise. Refinancing,Non-Owneroccupied Indicator variable, which takes HMDA Loan Purpose = 3 & Ownerthevalueof1iftheloanpurpose Occupancy=2 isrefinancingandthepropertyis notowner-occupiedand0otherwise. Refinancing,Owneroccupied Indicator variable, which takes HMDA Loan Purpose = 3 & Ownerthevalueof1iftheloanpurpose Occupancy=1 is refinancing and the property is owner-occupied as a principal dwellingand0otherwise. RegulatoryStrictness Indicator variable, which takes Agarwaletal. (2014) the value of 1 for the lowest regulatoryleniencytercileinthe year before EDO and 0 otherwise. Regulatory leniency measure of Agarwal et al. (2014) measured as the difference betweentheaverageregulatoryratings of federal and state regulators. Subprimeshare Percent of borrowers at the CoreLogic and aucountylevelwithFICOscoresof thors’calculations 619andbelow. VA-guaranteedLoans Indicator variable, which takes HMDA LoanType=3 the value of 1 if the loan type isVA(VeteransAdministration)guaranteed loans and 0 otherwise. ControlVariables CapitalRatio Total equity as a proportion of CallReports RCFD3210/RCFD2170 totalassets. EmploymentGrowth Thegrowthofemploymentlevel Bureau of Economic (Total Employment - Lagged (Total employment is defined as Analysis Total Employment) / Lagged thenumberofjobs) TotalEmployment LiquidityRatio Ratio of cash and cash equiva- CallReports (RCFD0071 + RCFD0081) / lents to total assets, where cash RCFD2170 isdefinedasthesumofinterestbearing balances, noninterestbearing balances, and cu3rr7ency andcoin.

Nonperforming Assets Ratio The sum of nonaccruing loans CallReports (RCFD1403 + RCFD1407) (NPA) and accruing loans past 90 days / (RCFD1400 - RCFD3123 dividedbynettotalloans. RCFD2123) Numberofloanapplications Thenumberofmortgageloanap- HMDA plicationsinagivencounty. ReturnonAssets(ROA) Net income divided by average CallReports RIAD4340/RCFD2170 totalassets Size Naturallogarithmoftotalassets CallReports log(RCFD2170) 38

Table 1: Descriptive statistics This table presents the summary statistics for the variables we use in our analyses. Panel A shows bank-level variables using quarterly call report data and county-bank-level portfolio and market shares using annual HMDA data. Panel B shows the breakdown of loans originated and applications declined. To mitigate the effects of extreme observations, all continuous variables arewinsorizedatthe1%and99%tailsoftheirrespectivedistributionsineachsampleyear. AllvariablesaredefinedinAppendix A. PanelA:Bankandcounty-leveldata N Mean Std P1 P25 Median P75 P99 Bank-Level Variables Totalloans/Assets 41,015 0.653 0.137 0.259 0.573 0.673 0.753 0.891 Residentialmortgages/Assets 41,015 0.179 0.106 0.004 0.102 0.165 0.237 0.500 Deposits/Assets 41,015 0.837 0.077 0.567 0.804 0.854 0.889 0.939 Totalloans/Deposits 41,012 0.786 0.181 0.319 0.676 0.794 0.902 1.225 Size 41,015 11.917 1.268 9.363 11.056 11.825 12.628 15.767 ReturnonAssets 41,015 0.001 0.011 -0.043 0.000 0.003 0.006 0.022 LiquidityRatio 41,015 0.067 0.064 0.008 0.027 0.045 0.083 0.328 CapitalRatio 41,015 0.103 0.042 0.036 0.082 0.096 0.114 0.265 NonperformingAssetsRatio 41,015 0.029 0.034 0.000 0.006 0.017 0.040 0.168 County-Level Variables ResidentialMortgagePortfolioShares(ofloanstominorities) 162,769 6.542 19.871 0.000 0.000 0.000 0.000 100.000 ResidentialMortgageMarketShares(ofloanstominorities) 497,594 0.408 3.936 0.000 0.000 0.000 0.000 9.721 39

Table 1: Descriptive statistics, continued PanelB:Thenumberofloansoriginatedordenied Number of Loans Number of Applications % denied Originated Denied Total 2,772,382 1,414,587 33.8% Race Majority 2,156,439 621,376 22.4% Minority 264,161 139,329 34.5% Gender Male 1,883,706 567,325 23.1% Female 632,973 250,883 28.4% Loan Type Conventional 2,401,190 1,330,381 35.7% FHA-insured 251,607 61,429 19.6% VA-guaranteed 100,965 18,203 15.3% FSA/RHS 18,620 4,574 19.7% Loan Purpose & Owner-occupancy HomePurchase: Owner-occupied 885,538 275,244 23.7% HomePurchase: Not-owner-occupied 233,856 74,891 24.3% HomeImprovement: Owner-occupied 194,062 169,741 46.7% HomeImprovement: Not-owner-occupied 24,440 10,029 29.1% Refinancing: Owner-occupied 1,244,578 826,978 39.9% Refinancing: Not-owner-occupied 187,144 57,271 23.4% Others 2,764 433 13.5% 40

Table 2: Lending to minorities for EDO banks This table shows changes in EDO banks’ lending to minorities. Panel A presents a county-level analysis of banks’ portfolio allocation and market shares of residential mortgage lending to minorities. Columns (1)–(3) present the results from a staggered difference-in-differences analysis, whereas columns (4) and (5) analyze changes in EDO banks’ market shares at the county level. The dependent variable in column (1) is banks’ residential mortgage loans to minorities as a share of their total residential mortgage portfolios. In column (2), it is banks’ residential mortgage loans to Black or African American borrowers as a share of their total residential mortgage portfolios, whereas in column (3) it is banks’ residential mortgage loans to Black or African American borrowers scaled by residential mortgage loans to white males. In column (4), the dependent variable is EDO banks’ market shares of residential mortgage loans to minority borrowers, whereas in column (5) it is EDO banks’ market share of residential mortgage loans to Black or African American borrowers. Panel B presents the county-level analysis of banks’ portfolio shares of residential mortgage loans to minorities using a control sample of non-EDO banks, matched on size and geography (county). The dependent variable is banks’ residential mortgage loans to minorities as a share of their total residential mortgage portfolios. Treatment is an indicator variable that takes a value of 1 for EDO banks and 0 otherwise. In both panels, the indicator During EDO refers to the actual time a bank is subject to an EDO, and Post EDO is an indicator variable for the five years after an EDO’s termination. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio, and NPA) andcounty-levelvariables(employmentgrowthandthenumberofloanapplications). Tomitigatetheeffects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in each sample year. All variables are defined in Appendix A. Standard errors are calculatedusingabootstrap. Thez-statisticsarepresentedinparentheses;∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). PanelA:ChangesinEDObanks’portfolioandmarketshares Portfolio Portfolio Portfolio Marketshares Marketshares shares shares(Black shares(Black (Minorities) (Blackor (Minorities) orAfrican orAfrican African American) American American) relativeto whitemales) (1) (2) (3) (4) (5) DuringEDO -0.684* -0.520 1.991*** -0.005 -0.035 (-1.905) (-1.412) (5.249) (-0.146) (-0.792) PostEDO 1.950*** 2.389*** 7.373*** 0.616*** 0.584*** (5.600) (6.584) (20.940) (19.151) (14.029) Observations 1,721,997 1,721,997 1,416,949 690,864 596,203 Waldχ2 5733*** 9225*** 17330*** 9666*** 6986*** Estimationmethod RETobit RETobit RETobit RETobit RETobit Controls Yes Yes Yes Yes Yes Year,County,BankRE Yes Yes Yes Yes Yes Years 1994–2018 1994–2018 1994–2018 1994–2018 1994–2018 41

Table 2: Lending to minorities for EDO banks, continued PanelB:Matchedsampleanalysis: LendingtominoritiesbyEDObanks Portfolio shares (1) DuringEDO -0.037 (-0.071) PostEDO -0.146 (-0.288) DuringEDO×Treatment -0.051 (-0.069) PostEDO×Treatment 1.429** (2.012) Observations 316,133 Waldχ2 1230*** Estimationmethod RETobit Controls Yes Year,Bank,CountyRE Yes Years 1994–2018 42

Table 3: Improvements at EDO banks and minority borrowers Thistablepresentsacounty-levelanalysisforEDObanks’portfolioallocationofresidentialmortgagelending to minorities. The dependent variable, Portfolio shares, is banks’ allocation of credit to minorities within their county-level residential loan portfolios, and Treatment is an indicator variable associated with process improvements at EDO banks. Subprime share is the percent of borrowers at the county level with FICO scores of 619 and below. The table shows the impact of requiring a written loan policy (Columns (1)–(2)) and written internal audit procedures (Columns (3)–(4)). The indicator During EDO refers to the actual time a bank is subject to an EDO, and Post EDO is an indicator variable for the five years after an EDO’s termination. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio,andNPA)andcounty-levelmacrovariables(employmentgrowthandthenumberofloanapplications). To mitigate the effects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in each sample year. All variables are defined in Appendix A. Standard errors are calculated using a bootstrap. The z-statistics are presented in parentheses; ∗p < 0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). Treatment= Treatment= Treatment= Treatment= Loanpolicy Loanpolicy Internalaudit Internalaudit Portfolioshares Portfolioshares Portfolioshares Portfolioshares (1) (2) (3) (4) DuringEDO×Treatment -0.727 3.363* -2.629** 2.319 (-0.631) (1.775) (-2.077) (1.063) PostEDO×Treatment -0.359 -0.033 0.839 -3.305 (-0.312) (-0.018) (0.660) (-1.538) During×Treament×Subprimeshare -32.064** -88.765*** (-2.116) (-5.431) PostEDO×Treatment×Subprimeshare 51.911*** 85.510*** (3.403) (5.118) Observations 151,748 151,559 151,748 151,559 Waldχ2 537*** 590*** 618*** 708*** RegType RETobit RETobit RETobit RETobit Controls Yes Yes Yes Yes Year,County,BankRE Yes Yes Yes Yes Years 1994–2018 1994–2018 1994–2018 1994–2018 43

Table 4: Scope for improvements at EDO banks and minority borrowers Thistablepresentsacounty-levelanalysisforEDObanks’portfolioallocationofresidentialmortgagelending to minorities. The dependent variable, Portfolio shares, is banks’ allocation of credit to minorities within their county-level residential loan portfolios, and Treatment is an indicator variable associated with process improvements at EDO banks. Subprime share is the percent of borrowers at the county level with FICO scores of 619 and below. The table shows changes at banks with stricter regulators (column (1)), longer EDOs (column (2)), and low CRA ratings (column (3)). The indicator During EDO refers to the actual time a bank is subject to an EDO, and Post EDO is an indicator variable for the five years after an EDO’s termination. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio,andNPA)andcounty-levelmacrovariables(employmentgrowthandthenumberofloanapplications). To mitigate the effects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in each sample year. All variables are defined in Appendix A. Standard errors are calculated using a bootstrap. The z-statistics are presented in parentheses; ∗p < 0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). Treatment= Treatment= Treatment= Regulatory EDOLength LowCRA Strictness Rating Portfolioshares Portfolioshares Portfolioshares (1) (2) (3) Treatment 3.436** -1.528*** 5.277* (1.996) (-4.554) (1.846) DuringEDO 0.457 -5.950*** -0.916 (0.440) (-5.311) (-1.492) PostEDO -6.268*** -5.172*** 1.458** (-5.969) (-4.559) (2.501) DuringEDO×Treatment -2.517 2.312*** -2.681 (-1.465) (5.329) (-0.924) PostEDO×Treatment 7.589*** 3.074*** 9.750*** (4.169) (7.136) (3.400) Observations 77,379 162,769 162,769 Waldχ2 276*** 519*** 497*** RegType RETobit RETobit RETobit Controls Yes Yes Yes Year,County,BankRE Yes Yes Yes Years 1994–2018 1994–2018 1994–2018 44

Table 5: Loan denials by EDO banks This table presents coefficient estimates from a linear probability model for the reasons EDO banks give when they deny a loan application. The dependent variableincolumn(1)isanindicatorofwhetheraloanapplicationisdenied. Thedependentvariablesincolumns(2)–(10)areindicatorsforreasonsfordenial, conditional on a loan application being denied. The indicator During EDO refers to the actual time a bank is subject to an EDO, Post EDO is an indicator variable for the five years after an EDO’s termination, and Minority is an indicator taking the value of one if an application is by a minority borrower. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio, and NPA) and a county-level macro variable (employment growth). To mitigate the effects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in each sample year. All variables are defined in Appendix A. The t-statistics are presented in parentheses; ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). Denial Denial: Debt Denial: Denial: Denial: Denial: Denial: Denial: Denial: Denial: toincome Employment Credit Collateral Insufficient Unverifiable Incomplete Mortgage Unspecified history history cash information application insurance denied (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) DuringEDO 0.000 -0.008 -0.003 0.018 -0.018 -0.006** -0.022* 0.016 0.003 0.003 (0.020) (-0.366) (-1.125) (0.951) (-1.011) (-1.975) (-1.804) (0.876) (1.409) (0.089) AfterEDO 0.026 0.018 -0.006 0.014 0.031 -0.000 -0.008 0.001 0.002 -0.064* (1.217) (0.481) (-1.275) (0.577) (0.841) (-0.068) (-0.777) (0.030) (1.190) (-1.779) Minority 0.096*** -0.003 0.000 0.049*** -0.001 0.005* -0.002 -0.013** 0.000 -0.016 (4.453) (-0.290) (0.207) (6.085) (-0.028) (1.868) (-0.770) (-2.195) (-0.162) (-1.514) Minority×DuringEDO 0.015 -0.003 -0.002 -0.001 -0.028 -0.000 0.001 0.019 -0.001 0.017 (1.349) (-0.340) (-0.833) (-0.103) (-1.482) (-0.052) (0.279) (1.360) (-0.723) (1.173) Minority×AfterEDO -0.050* 0.011 0.000 -0.034*** -0.005 -0.006** 0.004 0.015** 0.000 0.004 (-1.798) (1.019) (0.175) (-3.104) (-0.215) (-2.135) (1.501) (2.227) (0.089) (0.417) Observations 3,084,846 629,789 629,789 629,789 629,789 629,789 629,789 629,789 629,789 629,789 AdjustedR2 0.170 0.134 0.052 0.357 0.151 0.026 0.035 0.340 0.031 0.303 Estimationmethod OLS OLS OLS OLS OLS OLS OLS OLS OLS OLS Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes YearFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Bank×CountyFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Cluster Bank Bank Bank Bank Bank Bank Bank Bank Bank Bank Years 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 45

Table 6: Loan portfolio quality for EDO banks This table presents banks’ loan portfolio quality changes during an EDO and after its termination. The dependentvariablesinPanelArefertobank-levelnonperformingassets. ThedependentvariableinPanelB is risky mortgages (defined as higher-priced closed-end mortgages) as a share of total residential mortgages atthebank-county-level,andinPanelCisanindicatorforwhethertheoriginatedloanisFHA-insured. The indicator During EDO refers to the actual time a bank is subject to an EDO, and Post EDO is an indicator variable for the five years after an EDO’s termination. All regressions include lagged bank-level control variables (size, profitability, liquidity, and capital ratio) and a county-level macro variable (employment growth). PanelBalsoincludesthecounty-levelnumberofloanapplications. Inaddition,model(3)ofPanel Aincludeslaggedbank-levelNPAscaledbytotalloans. Column(1)PanelBincludesyear,county,andbank random effects, whereas column (2) of Panel B includes year and bank × county fixed effects. To mitigate the effects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tailsoftheirrespectivedistributionsineachsampleyear. AllvariablesaredefinedinAppendix A.Standard errors in column (1) of Panel B are calculated using a bootstrap. The t-statistics for the OLS models and z-statisticsfortheTobitmodelsarepresentedinparentheses; ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01(two-tailed). PanelA:NonperformingassetsofEDObanks TotalNPA/ TotalNPA/Total NPAfor Totalloans loans residential mortgages/Total loans (1) (2) (3) DuringEDO 0.016*** 0.011*** -0.001* (13.732) (10.743) (-1.929) PostEDO 0.002 0.002 -0.000 (1.241) (1.186) (-0.713) Observations 41,010 41,010 37,322 AdjustedR2 0.552 0.612 0.851 RegType OLS OLS OLS Controls No Yes Yes Year-QuarterFE Yes Yes Yes BankFE Yes Yes Yes Cluster Bank Bank Bank Years 1994–2018 1994–2018 2001–2018 46

Table 6: Loan portfolio quality for EDO banks, continued PanelB:County-levelshareofriskylendingbyEDObanks Market Market sharesof sharesof riskyloans riskyloans (1) (2) DuringEDO -2.473*** -0.198 (-14.324) (-0.304) PostEDO -1.780*** -1.135 (-10.740) (-1.162) Observations 105,860 24,688 AdjustedR2 0.589 Waldχ2 2374*** RegType RETobit OLS Controls Yes Yes Year,County,Bankeffects Yes Yes Years 2004–2018 2004–2018 47

Table 6: Loan portfolio quality for EDO banks, continued PanelC:ChangesinFHAloansofEDObanks FHAloan FHAloan (1) (2) DuringEDO 0.003 0.004 (0.344) (0.672) AfterEDO -0.004 0.000 (-0.266) (0.017) Minority 0.071*** 0.066*** (4.865) (3.823) Minority×DuringEDO -0.037 -0.024 (-1.376) (-1.066) Minority×AfterEDO -0.056*** -0.047** (-3.261) (-2.420) Observations 2,356,796 2,356,796 AdjustedR2 0.207 0.273 Estimationmethod OLS OLS Controls Yes Yes YearFE Yes Yes BankFE No Yes Bank×CountyFE No Yes Cluster Bank Bank Years 1994-2018 1994-2018 48

Table 7: Alternative explanations: Low capital and local market competition ThistablepresentschangesinEDObanks’residentialmortgageloanstominorityborrowers. Thedependent variable is banks’ allocation of credit to minorities within their county-level residential loan portfolios. The indicator During EDO refers to the actual time a bank is subject to an EDO, Post EDO is an indicator variable for the five years after an EDO’s termination, Low capital is an indicator variable for the banks in the lowest tercile of regulatory capital before an EDO, High Competition (deposits) corresponds to the lowestdepositmarketHHItercileinagivencounty,andHigh Competition (loans) correspondstothelowest residential mortgage loan market HHI tercile in a given county. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio, and NPA) and county-level macro variables (employment growth and the number of loan applications). To mitigate the effects of extreme observations, allcontinuousbank-levelvariablesarewinsorizedatthe1%and99%tailsoftheirrespectivedistributionsin eachsampleyear. AllvariablesaredefinedinAppendix A.Standarderrorsarecalculatedusingabootstrap. The z-statistics are presented in parentheses; ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). Treatment=Lowcapital Treatment=High Treatment=High competition(deposits) competition(loans) Portfolioshares Portfolioshares Portfolioshares (1) (2) (3) Treatment 10.045*** 18.672*** 27.441*** (9.371) (21.084) (25.280) DuringEDO -0.370 -2.159** -1.086 (-0.490) (-2.483) (-1.122) PostEDO 0.536 0.561 2.762*** (0.731) (0.653) (2.860) DuringEDO×Treatment -0.298 2.063** 0.196 (-0.275) (2.034) (0.176) PostEDO×Treatment 0.451 0.590 -0.928 (0.398) (0.567) (-0.797) Observations 156,913 156,808 156,874 Waldχ2 430*** 1610*** 1464*** Estimationmethod RETobit RETobit RETobit Controls Yes Yes Yes Year,County,BankRE Yes Yes Yes Years 1994–2018 1994–2018 1994–2018 49

Table 8: Supplemental analysis: EDO banks and loans to women This table presents a county-level analysis of EDO banks’ portfolio allocation and market shares of lending towomen. Column(1)showsEDObanks’allocationofcredittowomenwithintheircounty-levelresidential loan portfolios, whereas column (2) shows EDO banks’ county-level market shares of residential mortgage lendingtowomen. TheindicatorDuringEDO referstotheactualtimeabankissubjecttoanEDO,andPost EDO is an indicator variable for the five years after an EDO’s termination. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio, and NPA) and county-level macro variables (employment growth and the number of loan applications). To mitigate the effects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in each sample year. All variables are defined in Appendix A. Standard errors are calculated using a bootstrap. The z-statistics are presented in parentheses; ∗p < 0.1; ∗∗p < 0.05; ∗∗∗p < 0.01 (twotailed). Portfolio Marketshares shares (Women) (Women) (1) (2) DuringEDO 3.391*** 0.035 (6.090) (1.399) PostEDO 6.068*** 0.724*** (11.792) (30.424) Observations 162,769 521,313 Waldχ2 939*** 17168*** RegType RETobit RETobit Controls Yes Yes Year,County,BankRE Yes Yes Years 1994–2018 1994–2018 50

Appendix B. Online Appendix to “Bank Supervision and Managerial Control Systems: The Case of Minority Lending” Appendix B.1. Additional Tables Table OA1: Number of counties with lending to minorities This table presents a county-level analysis for the number of counties covered by EDO banks in which they lend to minorities. The indicator During EDO refers to the actual time a bank is subject to an EDO; Pre EDO (year) and Post EDO (year) correspond to indicator variables for the years before an EDO and after EDO termination. Averagenumberof Averagenumberof Ofwhich: minority distinctcountieswhere distinctcountieswhere populationgreaterthan EDObanksareactive EDObankslendto 50%ofcounty (perbank) minorities(perbank) population (1) (2) (3) PreEDO(year-3) 22 6 3 PreEDO(year-2) 22 7 3 PreEDO(year-1) 22 7 3 DuringEDO(annualized,onaverage) 21 6 3 PostEDO(year1) 25 8 3 PostEDO(year2) 27 9 3 PostEDO(year3) 29 9 4 PostEDO(year4) 31 10 4 PostEDO(year5) 31 11 4 51

Table OA2: Robustness: Lending to minorities by EDO banks (county population-weighted estimation) This table presents a county-level analysis for EDO banks’ market shares of residential mortgage lending to minorities. The dependent variable is EDO banks’ county-level market shares of residential mortgage loans to minorities. In column (1), the bank-county-level regressions are weighted by the natural logarithm of the county population, whereas in column (2), the regressions are weighted by the county’s share of the total U.S. population. The indicator During EDO refers to the actual time a bank is subject to an EDO, whereas Post EDO corresponds to indicator variables for the one to five years after EDO termination. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio, and NPA) and acounty-level macrovariable(employmentgrowth). To mitigatethe effects ofextremeobservations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in eachsampleyear. AllvariablesaredefinedinAppendix Aofthemanuscript. Thez-statisticsarepresented in parentheses; ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). Marketshares Marketshares (1) (2) DuringEDO -0.026** -0.026 (-2.109) (-0.612) PostEDO 0.979*** 0.407*** (87.381) (10.342) Observations 489,709 489,709 Waldχ2 106165*** 2005*** RegType RETobit RETobit Controls Yes Yes Year,County,BankRE Yes Yes Years 1994–2018 1994–2018 52

Table OA3: Loan denials by EDO banks (interaction with subprime) This table presents coefficient estimates from a linear probability model for the reasons EDO banks give when they deny a loan application. The dependent variables in columns (1)–(9) are indicators for a reason for denial, conditional on a loan application being denied. The indicator During EDO refers to the actual time a bank is subject to an EDO, Post EDO corresponds to an indicator variable taking the value of one for the five years after an EDO’s termination, Minority is an indicator taking the value of one if an application is by a minority borrower, and Subprime is an indicator taking a value of one if the average transaction-matched FICO score at the level of the census tract, loan origination year, loan type, loan purpose, and occupancy status of the property is 619 or below. All regressions include lagged bank-level control variables (size, profitability, liquidity, capital ratio, and NPA) and a county-level macro variable (employment growth). To mitigate the effects of extreme observations, all continuous bank-level variables are winsorized at the 1% and 99% tails of their respective distributions in each sample year. All variables are defined in Appendix A. The t-statistics are presented in parentheses; ∗p<0.1; ∗∗p<0.05; ∗∗∗p<0.01 (two-tailed). Denial: Debtto Denial: Denial: Credit Denial: Denial: Denial: Denial: Denial: Denial: income Employment history Collateral Insufficientcash Unverifiable Incomplete Mortgage Unspecified history information application insurance denied (1) (2) (3) (4) (5) (6) (7) (8) (9) DuringEDO×Minority -0.010* -0.002 -0.002 -0.010 -0.001 -0.001 0.024 -0.001 0.016 (-1.829) (-0.918) (-0.193) (-1.469) (-0.308) (-0.397) (1.522) (-0.694) (1.035) PostEDO×Minority 0.001 -0.001 -0.031*** 0.020** -0.008*** 0.002 0.018** -0.000 0.001 (0.092) (-0.886) (-3.066) (2.349) (-3.589) (0.772) (2.582) (-0.104) (0.116) DuringEDO×Minority -0.013 -0.001 -0.059*** 0.025 0.020 -0.005 0.000 0.002 0.018 ×Subprime (-0.527) (-0.119) (-2.875) (1.008) (1.024) (-0.411) (0.008) (1.215) (0.574) PostEDO×Minority -0.010 0.003 0.023 -0.054** 0.003 -0.018*** -0.012 -0.001 0.056** ×Subprime (-0.437) (0.351) (0.968) (-2.358) (0.391) (-2.945) (-1.368) (-0.435) (2.209) Observations 571,655 571,655 571,655 571,655 571,655 571,655 571,655 571,655 571,655 AdjustedR2 0.133 0.0568 0.321 0.150 0.0271 0.0332 0.342 0.0320 0.302 Estimationmethod OLS OLS OLS OLS OLS OLS OLS OLS OLS Controls Yes Yes Yes Yes Yes Yes Yes Yes Yes YearFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Bank×CountyFE Yes Yes Yes Yes Yes Yes Yes Yes Yes Cluster Bank Bank Bank Bank Bank Bank Bank Bank Bank Years 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 1994-2018 53

Appendix B.2. Excerpts from an enforcement order requiring changes to internal audit and loan policy 54