Does Competition Reduce Racial Discrimination in Lending?

Transcription

1 Does Competition Reduce Racial Discrimination in Lending? Greg Buchak University of Chicago Adam Jørring University of Chicago June 27, 2016 Abstract This paper examines whether increases in bank competition reduce discriminatory practices in mortgage lending. It demonstrates that lenders are significantly less likely to approve black applicants loan applications despite facing similar credit risk. However, following the relaxation of interstate bank branching laws in the 1990s, increases in local lending competition reduced the approval differential between potential white and black borrowers by roughly one quarter. The reduction was driven partially by incumbent banks altering their lending policies, and partly by the entry of new banks. Incumbent banks that were less likely to lend to black applicants lost more market share than other incumbents. While both entry and competitive pressures significantly reduced racial discrimination in lending, the results suggest that neither force completely eliminates it, suggesting that there is room for both market-based and direct legal approaches targeting discrimination.

2 1 Introduction A policymaker seeking to reduce racial discrimination in a market faces two choices: Use legal sanctions to force discriminators to change their behavior, or reduce barriers for non-discriminating competing businesses to enter the market and thus pressure the discriminators to leave or reform their policies. The former approach directly raises the cost of discriminating through the threat of direct litigation; the latter indirectly raises the cost of discrimination through the threat of market share losses to competitive entrants. Skeptics of the direct approach may fear behavioral distortions, costly compliance, and difficulty in enforcement. Skeptics of the market-based approach may question its strength and empirical relevance. This paper draws on a natural experiment to examine the potential strength of market forces in reducing discrimination and explore the channel by which they do so. It begins by examining racial bias in acceptances and rejections of home mortgage applications during a time period in which changes to interstate bank branching laws led to exogenously increased competition in local lending markets. In particular, it compares the impact of a potential borrower s race in whether economically similar loan applications are accepted or rejected before and after the relaxation of interstate bank branching restrictions. When lenders face similar ex-ante expected losses, black borrowers in the early 1990s were roughly eight percent less likely to have their loan applications accepted than white borrowers. This lending differential closed by roughly one quarter as lending markets became more competitive when states relaxed interstate bank branching laws during the mid-1990s and early 2000s. Two factors led to the gap narrowing: First, previously discriminating incumbent banks lost more market share than non-discriminating banks and subsequently changed their behavior to lend to black applicants. Second, the banks that entered new markets tended to discriminate less than the incumbents. These findings confirm and quantify economic predictions that increased competition will reduce but not necessarily eliminate racial discrimination, and that a taste for discrimination is costly to the discriminator. 2

3 Additionally, they suggest that policy makers seeking to reduce discrimination should take seriously the role that policies designed to increase competition can play in achieving their goals. Discrimination based on an mortgage applicant s race has been illegal since the Fair Housing Act of Additionally, the Community Reinvestment Act of 1977 makes discrimination based on the racial makeup of the applicant s neighborhood illegal. Despite its illegality, numerous case studies conducted after the passage of these laws, documented by Benston (1979), indicate that the practice continued to be widespread. To facilitate monitoring, the Home Mortgage Disclosure Act of 1975 (HMDA) requires lenders to report acceptances and rejections of all mortgage applications, along with the applicants race, location, and income. Empirical evidence following HMDA has been mixed on whether widespread discrimination actually occurs. Holmes and Horvitz (1994) study lending in Houston between 1998 and 1991 and find little evidence for discrimination against areas with high minority populations. Tootell (1996) find a similar result in Boston in The conclusions are that after controlling for economic conditions in the area, the area s racial makeup has little explanatory power. Instead, what appears to be discrimination against minority areas is actually due to poor economic conditions that are correlated with an area being a high-minority area. There is more evidence for discrimination against individuals. Charles and Hurst (2002) find that black mortgage applicants are 73 percent more likely to be rejected than whites after controlling for economic observables. This paper also looks for evidence of discrimination against individuals, and finds that it is widespread, particularly when lending markets were less competitive before interstate branching was allowed. Moreover, while Holmes and Horvitz (1994) and Tootell (1996) studied discrimination at particular places and times, this paper looks for whether incidences of discrimination vary across times and locations. The paper s primary focus is on how competition impacts discrimination. After measuring discrimination, it tests whether discrimination changes with changes in lending competition within an area. The theory that increased 3

4 competition should decrease racial discrimination dates to Becker (1971). In short, those with a preference for prejudice must pay for it. Becker focuses on racial discrimination in the labor market. In a market with discriminatory and non-discriminatory agents, the non-discriminatory agents face lower costs because they are willing to hire minority, whom the discriminatory agents are less willing to hire. Because there is less demand for minority workers, their wages are lower, so the non-discriminator faces lower labor costs on average and is at a competitive advantage relative to the discriminator. Wage differences among races can persist in equilibrium, but entry of new non-discriminatory agents will reduce the differences. Moreover, in markets with large wage differences, the reward to entry for a non-discriminator is high. These effects suggest that markets that previously saw large wage differences should see those differences diminish as competition increases. Similar arguments apply in the context of lending. Discriminatory lenders forgo profitable lending opportunities by not lending to African Americans or by charging higher rates. Non-discriminatory lenders who are willing to make these loans therefore have access to better lending opportunities than discriminatory lenders. To the extent that non-discriminatory lenders are allowed to enter the market, it will be profitable to do so, and their entry should reduce the wedge between minority and non-minority borrowers. Moreover, they should put downward pressure on lending rates to local whites and upward pressure on deposit rates for local depositors, leading to worsening business conditions for incumbents who continue to discriminate. Therefore, lenders demonstrating a larger preference for discrimination should disproportionately lose lending market share as competition increases, even relative to incumbents who discriminate less. Peterson (1981), and more recently, Zenou and Boccard (2000) consider a theoretical formalization of this argument. Under particular assumptions, a corollary of Becker (1971) is that if there is racial discrimination, loans made to minorities should perform better on average than those made to whites because conditional on receiving a loan, the minority borrower must have been a better borrower ex-ante. 1 Berkovec et al. (1994) find that on 1 Yinger (1996) spells out and criticizes these assumptions. In particular, one must 4

5 the contrary, for a sample of loans originated between 1987 and 1989, default rates were higher among minorities than whites after conditioning on other observable risk factors. 2 This paper s test for discrimination is more direct. Using applicationlevel data collected after the passage of the Home Mortgage Disclosure Act (HMDA), it tests whether conditional on observables, the race of an applicant impacts her probability of loan approval. The paper then asks whether an increase in competition reduces the impact of race on the lending decision, and find that it does. Using the passage of Interstate Banking and Branching Efficiency Act as a source of exogenous variation in banking competition, the paper shows that following its passage, treated states saw approximately a one-quarter reduction in disparities in loan acceptances between similar black and white borrowers. Moreover, a one standard deviation increase in the level of a state s lending competition causes nearly a sixty percent reduction in the amount of discrimination taking place. Having provided evidence on racial discrimination and how increased competition leads to reduced discrimination, the paper conducts an analysis of what happens to banks racial lending practices and what happens to discriminating banks market shares following the implementation IBBEA. In particular, the paper finds that discriminating banks lose greater market share than banks that do not discriminate. This is consistent with the predictions of Becker (1971). Next, the paper shows that as local banking markets become more competitive, banks in those markets become change their behavior assume (1) unobserved credit characteristics are not correlated with race, (2) black and white applicants receive equal treatment in foreclosure proceedings, and (3) the losses incurred when a black borrowers defaults are at least equal to those incurred when a white borrower defaults. 2 Berkovec et. al. examine FHA-insured loans between 1980 and They find that loans made to black applications perform worse than loans made to white applicants with similar observable characteristics. If lenders were discriminating, they argue, loans made to black applicants should perform better because the marginal black applicant needs to be a better ex-ante credit risk than the marginal white applicant. They examine performance of FHA loans for data-availability purposes. However, because the loans are insured by the FHA, the lender has little exposure to loan performance and hence it is not clear why the ex-post default probability is a major concern for the lender. See Yinger (1996) for more discussion of the Berkovec/default-rate methodology. 5

6 and discriminate less. Finally, the paper quantifies the extent to which overall reductions in discrimination have been driven by incumbent banks changing behavior and the extent to which reductions in discrimination have been driven by new entrants that discriminate less. The data show that both that incumbents discriminate less, and that the new entrants discriminate slightly less than the reformed incumbents. The paper s findings suggest that policy makers looking to reduce non-economic discrimination should consider marketbased approaches and direct-sanctions as complementary in eliminating discrimination. The paper proceeds as follows: Section 2 details the empirical design and data sources. Section 3 presents the main results regarding discrimination and competition. Section 4 investigates what happens at discriminating banks. Section 5 discusses the implications of the paper s findings, and Section 6 concludes. 2 Empirical Design and Data 2.1 Empirical Design Defining and measuring racial discrimination is central to this paper. Here, it defines racial discrimination as occurring when a borrower s race impacts the lender s decision of whether or not to lend even though the loan offers the lender the same ex-ante expected return. Quantification of racial discrimination presents a challenge for two reasons: First, there is a data problem in that loan officers deciding whether to issue a loan observe more about applicants ability to repay than is recorded in HMDA or is otherwise available. For instance, while HMDA records an applicant s income, it does not record an applicant s credit score, which would be available to a loan officer. Second, an applicant s race may serve as a proxy for economically relevant factors impacting the applicant s ability to repay the loan that are unobservable to both loan officer and researcher. That is, a loan officer may refuse a loan to a black applicant who is otherwise observationally equivalent to a white appli- 6

7 cant that he would grant the loan to. The refusal may not be due to the loan officer s disliking lending to minorities, but rather because he believes (rightly or wrongly) that race is correlated with the applicant s ability to repay. For instance, race may be correlated with differences in future economic prospects not accounted for by other observables. Under this paper s definition, this would not be racial discrimination, and therefore this paper seeks to avoid including lending differences between races resulting from these differences in its measure of discrimination. 3 This paper overcomes this challenge in two ways: First, it looks for racial disparity among loans whose expected losses to lenders are as similar as possible. To do this, it first restricts the sample to loans applied for under the Federal Home Administration (FHA) insurance program. To qualify for FHA insurance, loans must conform to certain underwriting standards including credit score and income requirements. 4 Once approved by the FHA, the lender has the final say in whether to make the loan. An FHA-approved borrower pays an insurance premium to the FHA, and in exchange, the FHA insures the lender in case of default. As per Berkovec et al. (1994), FHA loans are not primarily rationed based on price, but rather through qualification standards imposed by the FHA. Hence, FHA-loans, conditional on approval by the FHA, offer the same expected return to lenders because lenders do not bear credit risk upon the borrower s default. 5 HMDA does not record whether a 3 The lender s using a race as a proxy for economically important unobservables to make a lending decision is often termed statistical discrimination. It is illegal in lending markets. however, because it may genuinely reflect diffferent economic risks in making the loan, it is not the type of discrimination this paper attempts to identify. Instead, this paper attempts to isolate pure non-economic discrimination. 4 See Appendix 7.1 for details. 5 Note that expected losses to the insurer the FHA may differ significantly if race is a predictor of economically relevant outcomes. While denying loans on this basis may be socially desirable, from the economic calculus of the lender, this is still discrimination in the sense that the lender is using race as a factor in rejecting loans that offer him the same expected return. The existance of this potential moral hazard problem is a broader policy question regarding the implementation of the FHA-insurance program and while interesting, is beyond the scope of this paper. This paper paper avoids making broad welfare statements of whether the relaxation of branching laws exascerbated this moral hazard problem, and simply measures discrimination occuring at the level of the economic decision-maker. 7

8 loan applying for FHA insurance indeed qualifies for FHA insurance but only whether the borrower applied for the loan under the FHA program. The FHA may deny the loan if the borrower fails to meet its minimum lending standards. 6 Thus FHA applications in the sample may be denied either due to failure to meet the minimum requirements of the FHA or due to the lender s decision to not extend credit even though the FHA has approved the loan. FHA requirements, however, are mostly objective and quantitative, and most of the requirements are observable in the data set. 7 Hence, by controlling for a rich set of observables, including most observables relevant to the FHA s objective standards, the paper is able to compare loans that are more-or-less equally likely to be approved by the FHA but only differ on race, which is not a factor in the FHA s decision. Thus, when comparing loan application A with a white applicant and loan application B with a black applicant that are otherwise identical on observables, if A was accepted and B was not, the inferences is that both loans met FHA standards and hence the lender for B would have been insured. 8 The paper therefore identifies B s rejection as racially motivated discrimination. Table 1 illustrates the intuition: A Outcome B Outcome Interpretation (1) Accepted Accepted Both approved at FHA and bank (2) Rejected Rejected Both denied at either FHA or bank (3) Accepted Rejected A approved at both; B Denied at Bank Table 1: Illustration of empirical design. Identification comes from differenes in outcomes for observationally similar loans that differ on race. If loan A with a white borrower and loan B with a black borrower are observationally identical and A is accepted, the inference is that A met FHA standards and therefore so did B. Because the lender would be insured on B, the paper identifies B s rejection as being racially motivated. Still, measures of FHA compliance may be noisy, particularly because 6 The minimums are fairly leinient. See Appendix Section 7.1 for details. 7 The notable exception is the applicant s credit score, which the paper proxies for using census-tract level economic information and zip-level credit information. 8 As an even more conservative robustness check, Section constructs a sample of applications indexed by borrower, where the data is conditioned on the borrower applying for multiple FHA loans and getting approved for at least one loan. This insures that this subsample of borrowers are indeed approved for FHA insurance. Among this subset of borrowers, the results are similar. 8

9 HMDA does not record the individual s credit score. The paper s proxy for individual credit score the credit conditions inside the applicant s ZIP Code are imperfect. Moreover, even if FHA compliance is well-measured, lenders still bear the risk of losses from default due to reputational concerns, due to fear of an FHA audit if too many insured loans default, or due to defaults before insurance coverage begins. Therefore, this strategy may still overstate discrimination to the extent that applicant s race is predictive of these residual losses. The primary concern with this paper, however, is not the level of discrimination per se, but rather how discrimination changes in response to increased competition. 9 This motivates the second prong of the paper s empirical strategy. It is typically difficult to identify the causal link between increases in competition and endogenous economic outcomes because changes in competition in an area are often themselves endogenous responses to economic conditions. In the context of lending, suppose that a positive economic shock is anticipated in an area that will disproportionately benefit minorities. In response, banks are more likely than before to lend to minorities. Simultaneously, anticipating better lending conditions in the future, outside banks enter the market and increase local competition. Regressing changes in proportional minority lending on changes in competition will yield a positive coefficient even though both effects were caused by the (unobservable) changes in local economic outlook. Of particular concern in this paper is whether competition is correlated as 9 Suppose that the true degree of discrimination in state s at time t, d st, is mismeasured ˆd st = d st + η st Where η st is (non-necessarily mean-zero) mismeasurement coming from the concerns raised above. If γ is the true causal effect of how discrimination d st varies with competition c st, regressing ˆd st on a measure of competition will yield the following estimator in the limit: ˆγ p γ + cov(η st, c st ) var(c st ) ˆγ will be a consistent estimator for γ so long as the mismeasurement η st is uncorrelated with the measure of competition. A standard instrumental-variables approach detailed below will provide exogenous variation in c st so that the the measurement error in discrimination is uncorrelated with the variation in competition. 9

10 with mis-measurement in discrimination. For instance, consider the case of an urban zip code with many affluent white borrowers and many poor black borrowers. Credit score is only observed at the zip level, but the poorer black applicants will tend to have lower credit scores than the affluent white borrowers. A white borrower here will be more likely to comply with FHA standards than a black borrower because of the credit score difference, but this difference is not observable in the HMDA data. The paper will misidentify this as discrimination even though it is due to measurement error. Simultaneously, this measurement error occurs in an urban area with significant lending competition. To the extent that these within-zip credit-score disparities occur systematically in areas with high banking competition, the measured impact of competition in reducing discrimination would be biased towards zero because the the paper mismeasures discrimination as being high precisely in areas where competition is high. To address these concerns, this paper uses the staggered implementation of the Riegle-Neal Interstate Banking and Branching Efficiency Act of 1994 (IBBEA). Before this act, banks were not allowed to branch across state lines. Following its passage, states were able to relax branching laws into their states. When states allowed branching, local banking markets were opened to nationwide competition. Many papers have used the IBBEA as an instrument for increases in bank competition to study subsequent lending outcomes. Favara and Imbs (2015) show that branching deregulation significantly increased the supply of mortgage credit and consequently housing prices. Closer to this paper, Tewari (2014) finds that the relaxation of these laws increased consumer credit access, particularly to low-income households and minorities. This paper differs from Tewari in that it focuses on changes in discriminatory lending to minorities rather than credit access per se. That is, while Tewari finds increased credit availability to minorities following the branching deregulation, this paper provide a mechanism for this increased access a provides in racial discrimination. The use of the IBBEA as an instrument for banking competition completes the primary identification strategy. Recall the concern that omitted 10

11 variables may bias upwards the estimate of discrimination because FHA decisions and residual losses to lenders even conditional on FHA approval may be correlated with factors prediction default particularly credit score. Because the paper is chiefly concerned with changes in discrimination, however, the law change will be a valid instrument for how competition impacts discrimination so long as the law change is uncorrelated with measurement error in lending discrimination. This identifying assumption sidesteps the issue of consistently estimating the level of discrimination by allowing a consistent estimate the impact of competition on discrimination. 2.2 Data Before proceeding to the analysis, this section summarizes the data used in the paper. The unit of observation for all following regressions is a loan application from HMDA. HMDA records, among other things, the outcome of the application, the borrower s race, income, loan amount, year, census tract, and bank at which the lender applies. 10 It also has an indicator for whether the application was done as an FHA loan. The following table summarizes the individual-level used in the regressions. 11 are accepted; roughly 16% of applicants are black. Importantly, roughly 80% of loans The study merges the individual HMDA data with data at the censustract and zip-code level. census conducted in 1990 and The census-tract level data is from the decennial Data for intermediate years uses the most recent census. The zip-level data are updated yearly and are consumercredit data from equifax. Importantly, this data contains measurements of 10 Regarding loan acceptance: HMDA codes outcomes as (1) Loan Originated, (2) Application approved but not accepted by borrower, (3) Application denied by the financial institution, (4) Application withdrawn by applicant before a credit decision, (5) File closed for incompleteness, (6) Loan purchased by the institution, (7) Preapproval denied, and (8) Preapproval granted but not accepted by the applicant. This study counts (1), (2), and (8) as acceptances; it counts (3), (4), (5), and (7) as denials. Regarding race: The study compares only white and black borrowers. 11 Table 12 in the appendix has the full set of covariates. 12 At the census-tract level: % Black Residents, median household income, % adults with less than a high-school education, % adults with a college education, % of the census living in poverty. 11

12 Table 2: Summary Statistics. See Table 12 for full covariates. Variable N Mean St. Dev. Accepted 4,367, Black 4,367, Income 4,367, Loan Amount 4,367, Post 4,367, Herf 4,367, credit scores and defaults within the zip code. 13 Finally, at the state-year level, the paper uses information on the relaxation of IBBEA from Rice and Strahan (2010), which includes both the first time an IBBEA reform was implemented in the state and an index measuring the degree of deregulation. Competition is measured as a Herfindahl index of all (including non-fha) accepted loans in a county. Because the exogenous variation is at the state level, the Herfindahl index is aggregated from a county to a zip level, weighted by county share of loans. 14 The main specifications are a linear probability model of acceptance as follows. 13 At the zip-year level: % with credit scores below 620, % with credit scores below 640, % with credit scores below 660 (subprime), % of total credit accounts in default, % of mortgages in default, % consumers with mortgages, % consumers with home equity loans, % consumers with credit cards, and % consumers with auto loans. 14 In particular, within state-county-year: ACCEP T ED bsct SHARE bsct = b s,c,t ACCEP T ED bsct HERF sct = SHAREbsct 2 [0, 1] b s,c,t Exogenous variation is at state level; The paper calculates a weighted Herfindahl in state: State Herf st = c s ACCEP T ED sct c s ACCEP T ED HERF sct sct 12

13 Ordinary Least Squares accept it = β 1 Race i + Γ A X it + Γ C X ct + Γ Z X zt + γ s + γ t + γ b + ɛ it (1) (1) measures the extent to which an applicant s race impacts her lending outcome. X it, X ct and X zt are individual, census-tract, and zip-level controls, respectively. γ s is a state-county fixed effect; γ t is a year fixed effect; γ b is a bank fixed effect. OLS Difference in Difference: accept it = β 1 Race i + β 2 HERF st + βc OLS Race i HERF st + Γ A X it + Γ C X ct + Γ Z X zt + γ s + γ t + γ b + ɛ it (2) (2) estimates a (likely correlative) association between state-level competition and the degree to which race impacts the lending decision. HERF st is the state-level Herfindahl index of size-weighted accepted loans by bank. A larger Herfindahl index is associated with less competition, so if βc OLS is negative it indicates that in areas with greater competition, the race of the applicant is relatively less important. The interpretation of βc OLS is not causal because the degree of competition in an area may be correlated with unobserved reasons related to race for denying the loan and measurement error in discrimination. Reduced Form Difference in Difference: accept it = β 1 Race i + β 2 POST st + βc RF Race i POST st + Γ A X it + Γ C X ct + Γ Z X zt + γ s + γ t + γ b + ɛ it (3) (3) estimates the causal effect of the IBBEA relaxation on how race matters in lending under the assumption that the timing of the relaxation is unrelated to unobserved reasons for denying loans to black applicants. A positive coefficient on βc RED implies that post-ibbea, black applications became more likely to be approved. 13

14 Instrumental Variables Difference in Difference: accept it = β 1 Race i + β 2 HERF st + β IV Race i HERF st C + Γ A X it + Γ C X ct + Γ Z X zt + γ s + γ t + γ b + ɛ it (4) Where the first stage is as follows: Race i HERF st = η 1 POST_IBBEA st + η 2 POST_IBBEA st Race i + ζ st HERF st = η 1 POST_IBBEA st + η 2 POST_IBBEA st Race i + ζ st (4) measures the causal link between state-level lending competition and the degree to which race impacts the lending decision. A negative coefficient β IV C implies that increased competition reduces racial discrimination. As the experiment occurs at the state-year level, standard errors are clustered by State Year. With the empirical framework and data described, the paper s main results follow. 3 Discrimination and Competition This section presents the main results regarding lending discrimination and the impact of competition and the passage of the IBBEA on discrimination. Section 3.1 gives the main results; section 3.2 provides a placebo test to rule out the change reflects a decrease in lending standards broadly. Appendix Section 7.2 contains a number of robustness checks on these results. 3.1 Main Results Table 3 presents the main results of the paper. 15 Column (OLS) shows that a black applicant is 7% less likely to have his loan application accepted than an observationally equivalent white applicant, which given the empirical setup, 15 For readability, the in-text tables exclude the controls that are not of interest. Full tables of coefficients are given in the appendices. 14

15 provides evidence of non-economic discrimination. Given that the base rate of rejection is a relatively low 21%, this corresponds to a rejection rate roughly 35% for a black applicant relative to an observationally equivalent white applicant. Table 3: Results for specifications (1), (2), (3), (4). See Table 13 for full covariates. Dependent variable: Application Accepted (OLS) (OLS DiD) (RED DiD) (IV DiD) Black (0.006) (0.007) (0.009) (0.017) Herf Black (0.085) - (0.394) Post Black (0.010) - State-County, Year, Bank FE Y Y Y Y State Year Clusters Y Y Y Y Observations 4,367,944 4,367,944 4,367,944 4,367,944 R Adjusted R Note: p<0.1; p<0.05; p<0.01 More central to the objective of this paper, columns (OLS DiD), (RED DiD), and (IV DiD) show that the degree to which race matters in the lending decision declines in the face of more competition. In particular, from the OLS regression, a one standard deviation increase in lending competition as measured by a state-level Herfindahl index decreases the average disparity a black applicant faces from a mean value of -7.2% to -5.4% a difference of roughly 25%. Similarly, after the implementation of the IBBEA, the reducedform regression points to roughly a 25% change. The IV regression suggests that the causal effect takes the mean racial disparity from -7.5% to -3.4% a 54% difference. The upshots from Table 3 are as follows: (1) The paper finds clear evidence for racial discrimination. (2) Increased competition, coming through the implementation of the IBBEA, reduces but does not fully eliminate racial 15

16 discrimination. The paper now proceeds to run tests to rule out an alternative explanation of these findings namely, that the applicant s race is a proxy for unobserved but economically relevant variables, and that increased competition depresses lending standards across the spectrum of economically relevant variables. 3.2 Competition and Applicant Income This section attempts to rule out an alternative explanation of Table 3. In particular, suppose that an applicant s race proxies for economically relevant but unobserved variables, like credit score, and that increased competition forces banks to lowers standards on all economically-relevant variables. If the impact of competition on race merely measures declining lending standards broadly, running a similar test with an interaction of IBBEA and income should yield significant results indicating that applicant income becomes less important over time. To rule out this explanation end, the paper again runs the regression described in equation (3), this time including an interaction term of Post log(income). If income has become less important due to broadly declining lending standards, the coefficient on the interaction term should be negative; on the other hand, a well-estimated zero coefficient on this term will indicate that lending standards for FHA loans at least those relating to applicant income did not decline across the board, and that observed change in Table 3 was indeed specific to the applicant s race and non-economic discrimination. Table 4 shows the results. Columns (Placebo 1) and (Placebo 2) show that the interaction coefficient on Post log(income) is a well-estimated zero, as desired. Because the coefficients on Post log(income) are close to zero, this provides evidence that the changes in the race coefficient do not simply reflect lower lending standards broadly. Instead, the applicant s race is indeed important in whether the loan is accepted, and the fact that similar effects do not show up with clearly economically relevant variables like the applicant s 16

17 Table 4: Results for specification (3) with different or additional interractions. Column (Baseline) is (3) as before. Column (Placebo 1) adds a Post log(income) term. Column (Placebo 2) replaces the race interraction term with the income interraction term. See Table 15 for full covariates. Dependent variable: Application Accepted (Baseline) (Placebo 1) (Placebo 2) Black (0.009) (0.009) (0.006) log(income) (0.006) (0.008) (0.008) Post Black (0.010) (0.010) - Post log(income) (0.008) (0.008) State-County, Year, Bank FE Y Y Y State Year Clusters Y Y Y Observations 4,367,944 4,367,944 4,367,944 R Note: p<0.1; p<0.05; p<0.01 income support the contention that there was racial discrimination and that it decreased after the passage of IBBEA. Appendix 7.2 provides further robustness tests by (1) investigating the plausible importance of omitted variable bias, and (2) concentrating on a sample of loans known to be FHA-compliant. The results are consistent with the main results of this section. Namely, (1) discrimination occurs and (2) competition reduces it. Having established that discrimination occurs and that increased competition reduces it, the paper now investigates the channel through which this change happens. 4 Bank-Level Mechanism The paper has so far demonstrated that lending discrimination occurs and that the amount of discrimination decreases in more competitive lending markets and following the implementation of IBBEA. It now focuses on the mechanisms 17

18 by which the changes in discrimination occur by examining discrimination at the bank level. In particular, the objective is to study whether the overall level of discrimination decreases because the relaxation of bank branching laws allows incumbents to enter who discriminate less, or because discrimination is indeed costly, and the threat of entry induces discriminatory banks to change behavior. To answer this question, the paper first identifies, on a bank-by-bank basis, how much the applicant s race impacts the bank s lending decision. The strategy for doing so is similar in spirit to the main empirical design and is discussed in detail in Section 4.1. It next tests whether discriminating banks pay for their discrimination by disproportionately losing market share relative to less discriminatory competitors when competition increases. As the paper shows in Section 4.2, they do. Finally, the paper tests whether competition pressures discriminating banks to discriminate less, and whether new entrants are on average less discriminatory. As shown in Section 4.3, discriminating banks do discriminate less in the face of competitive pressures, but there is weak evidence that the entering banks are even less discriminatory than the reformed incumbents. 4.1 Empirical Strategy The objective is to identify which banks are more discriminatory. To that end, the paper runs the following regression each year: accept i = β bt Bank b Race i + Γ A X i + Γ C X c + Γ Z X z + γ s + γ b + ɛ i (5) For each bank b and each year y regression (5) recovers β bt, which represents how much the bank considers race in making its lending decision. A more negative β bt means the bank tends to lend less to black applicants. These β bt estimates then become left-hand side and right-hand side variables in the analysis to follow. 18

19 4.2 Changes in Market Share The first objective is to test whether banks pay for discrimination in the face of more competition. To do so, the paper considers the following scenario. Consider bank D and ND in a given market; D discriminates, ND does not. Suppose that lending competition in the market increases. The expectation is that both D and ND will lose market share, but if D is engaged in costly discrimination, D will lose more market share than ND because with more competition, it is more difficult for D to pay for its discriminatory practices. To test this hypothesis, the paper regresses: market share sbt = γ 1 β st + γ D β bst Comp st + η s + η t + ɛ sbt Where Comp st is either State-Level Herfindahl (OLS), Post IBBEA (RED), or the first stage of State-Level Herfindahl on IBBEA (IV). γ D is the coefficient of interest. For the OLS and IV, a negative γ D means that the loss of market share in high-competition environments is large for banks banks with more negative β bst those who discriminate more. For the reduced form, a positive γ D means that after IBBEA, banks that discriminate more lost more market share. Table 5 shows the results. The sign of the coefficients are consistent with discriminating banks losing market share in high competition environments. This confirms the hypothesis that discriminating banks pay for their discrimination in the face of more competition. Note further that while all types of incumbents lose market share when new banks enter, the results above imply that the incumbent banks who do not discriminate grow relative to the incumbent banks who do discriminate. Thus, the overall change in discrimination is partly driven by a within-incumbent composition change, where the share of loans made by non-discriminators grows relative to the share of loans made by discriminators. An alternate explanation is that it is not discriminatory banks that lose market share, but rather high-standards banks that lose market share. That is, to the extent that race proxies for unobserved but economically relevant variables, banks that are more conservative with respect to these variables 19

20 Table 5: Does a banks coefficient on applicant race impact their market share loss following more competition or the relaxation of the IBBEA? The top rows are the test; the bottom rows are the placebo test on income coefficient. See 16 for full table. Dependent variable: Market Share (OLS) (RED) (IV) Herf Race Coeff (0.053) - (0.010) Post Race Coeff (0.002) - Herf Income Coeff (0.014) - (0.020) Post Income Coeff (0.0005) - State, Year FE Y Y Y State Post Clusters Y Y Y Observations 10,514 10,514 10,514 R Note: p<0.1; p<0.05; p<

21 lose market share because when new banks enter, the low-standards banks respond by lending to lower-quality applicants. To rule out this explanation, the bottom half of Table 5 tests how banks coefficient on income predicts their loss in market share. That is, do banks that are relatively more conservative with respect to the applicant s income also lose market share in the face of increased competition? As before, a well-estimated zero rules out this story. Indeed, both the OLS and reduced form estimates are precisely close to zero, and the IV estimator, though it has a somewhat larger standard error, is not significant at the 95% confidence level. This test suggests that the changes in market share do not simply reflect more aggressive banks gaining market share relative to less aggressive banks; rather, banks that discriminate do lose market share relative to those that do not. 4.3 Changes in Behavior Having shown that discriminatory banks pay for their discrimination by losing market share, the paper next tests whether increases in competition induce incumbent banks to modify their behavior. The behavior change when the decision-maker has a preference for discrimination could arise from a simple agency problem: Suppose that the bank s management is purely profitmaximizing, but that the bank s loan officers making day-to-day lending decisions do have a preference for discrimination. The bank s management, understanding that they will pay for their employees discriminatory practices, may choose to implement tighter controls more automated screening techniques or auditing in order to rein in the employees discriminatory tastes. Before competition increased, it may not have been worthwhile for the bank mangers to implement these controls, but facing potential market share losses when competition does increase, the policies become worthwhile. While HMDA does not have loan officer-level data, the paper attempts to test for changes at the bank level by testing whether the bank-specific coefficients on race obtained above change when competition increases. To 21

22 that end, the regression is: β bst = γ B Comp st + η sb + ɛ bst Where η sb is a state-bank fixed effect. The coefficient of interest is γ B. The regression compares how a bank within a state changes its behavior when the state becomes more competitive or the IBBEA is implemented in the state. A negative coefficient for γ B for (OLS) and (IV) means that higher competition leads to less discriminatory behavior; a positive coefficient for γ B for (RED) means that after IBBEA, banks discriminate less. Table 6 confirms the prediction that increases in competition reduce bank discrimination. Table 6: How do banks coefficients on applicant race change with competition or passage of the IBBEA? See Table 18 for full table. Dependent variable: Race Coeff (OLS) (RED) (IV) Herf (0.274) - (0.354) Post (0.006) - State Bank FE Y Y Y State Post Cluster Y Y Y Observations 4,753 4,753 4,753 R Note: p<0.1; p<0.05; p<0.01 The regression identifies how an incumbent bank in a given state takes race into account in the low- and high-competition periods. The large and significant coefficients on the reduced form and IV indicate that increases in competition cause discriminating banks to modify their behaviors, and that the overall change identified in Section 3.1 is not merely the result of less discriminatory banks entering. To show that banks do not adjust their lending behavior along other margins such as income, Table 7 shows how banks modify their coefficient 22

23 on income using the same specification as above with the income coefficient replacing the race coefficient on the left-hand side. A negative and significant value for the reduced form would indicate that banks lend to lower-income individuals after the implementation of the IBBEA. On the contrary, the table shows that there is almost no change along this margin. Table 7: Placebo test: How do banks coefficient on income change with competition and the relaxation of the IBBEA. See Table 19 for full table. Dependent variable: Income Coeff (OLS) (RED) (IV) Herf (0.385) - (0.523) Post (0.012) - State Bank FE Y Y Y State Post Cluster Y Y Y Observations 4,753 4,753 4,753 R Note: p<0.1; p<0.05; p<0.01 Finally, the paper attempts to measure how the overall reduction in discrimination in an area comes from incumbents altering their behavior or new entrants entering who have different policies regarding discrimination. To that end, the paper regresses bank coefficients on (1) whether the deregulation has taken place, and (2) whether the bank in question is an incumbent in the market. The regression is as follows: β bst = γ P P ost st + γ E Entrant b + η s + ɛ bst γ P measures how an incumbent bank s behavior changes. An incumbent bank s coefficient on race is greater by γ P relative to before. γ E measures how an entrant in the post period differs by an incumbent in the post period that is, an entrant s coefficient on race in the post period, is γ E greater than an incumbent s coefficient on race in the post period; relative to an incumbent in 23

24 the pre-period, an entrant s coefficient on race is greater by γ P + γ E. Table 8 shows the results for the race coefficient and the income coefficient. Table 8: How does the coefficient on applicant race vary pre and post between incumbents and entrants in the market? Column (1) is the test, column (2) is a placebo on the income on coefficient. See Table 20 for full table. Dependent variable: Race Coeff Income Coeff (1) (2) Post (0.004) (0.007) Entrant (0.006) (0.009) State FE Y Y State Clust. Y Y Observations 10,288 10,288 R Note: p<0.1; p<0.05; p<0.01 To interpret the numbers, the coefficient on post means that relative to before, after IBBEA, incumbent banks are 2.1% more likely to lend to black applicants. The coefficient on entrants, though not significant, means that in addition to the 2.1% change after post, new entrants are additionally more likely to lend to black applicants by 0.4%, meaning that overall, a new entrant after IBBEA is 2.5% more likely to lend to a black borrower than an incumbent in the same market would have been before IBBEA. The result of this test is that (1) both incumbents and entrants discriminate less after the passage of the law, and (2) entrants are slightly but not significantly less discriminatory than incumbents after the law. This section explored what happens at the bank level when competition increases. The results support the predictions of the theory that discrimination is expensive and that competition causes discriminatory banks to lose market share. Moreover, the results show that competition causes discriminating banks to discriminate less. This change in behavior could be caused, for example, by tighter controls or accountability standards given to low-level dis- 24

25 criminatory loan officers from high-level non-discriminatory managers. With the empirical findings presented, the paper now discusses their implications. 5 Discussion This paper established (1) racial discrimination occurs in mortgage lending, (2) competition decreases it, (3) discriminating banks pay for their practices when competition increases, and (4) discriminating banks change their behaviors to become less discriminatory when competition increases. This section discusses the legal implications of these findings. First, it highlights some reasons why, despite racial discrimination being illegal, it stills exists. Second, it discusses how policies promoting competition can be both complements and substitutes for direct legal sanctions. Third, it flags lessons for addressing discrimination in other markets and along other demographic characteristics. 5.1 Why Discrimination Still Happens Racial discrimination is illegal, yet this paper finds evidence of significant differences in outcomes between otherwise similar white and black applicants. Borrowers who have been discriminated against have a cause of action, so what is allowing banks to discriminate? Besides the typical problems related to costs of bringing cases to low-income borrowers, the methodology of this paper illustrates a systematic proof problem in individual discrimination cases seeking to use statistical evidence to prove discrimination. In short, this paper can tease out subtle differences in treatment of black and white borrowers because it has a large N a large number of observations an advantage that an individual plaintiff does not have. In particular, this paper finds discrimination against black borrowers on a nationwide level a systemic problem. To do this, it uses millions of mortgage applications made to the entire US banking system. A discriminated plaintiff, however, cannot sue the banking system but rather must sue a particular bank. While statistical evidence using HMDA is sufficient to prove a bank s intention 25

26 to discriminate in an FHA action, 16 as a statistical matter it is more difficult to establish with high confidence that a particular bank treats black and white applicants differently because estimates of differences are noisier when there are far fewer observations. That is, the confidence with which one can state that based on mortgage acceptance patterns that a bank is discriminating is much lower than the confidence with which this paper can state that discrimination occurs on a systemic level. Even if all banks discriminate based on race, such discrimination may not be observable on the individual bank level with enough confidence for a successful FHA action. 5.2 Competition as a Policy Tool Direct legal action may fail to completely eliminate systemic discrimination for the reasons discussed above. This paper has shown, however, that increases in competition have the effect reducing systemic discrimination without the need for direct legal action against a particular bank. While it does not offer an ex-post remedy to a harmed borrower, competition does provide ex-ante incentives for banks to correct their discriminatory behavior. A direct lawsuit requires proof and funding for the lawsuit. Both of these may be difficult for a low-income borrower to produce. Knowing this, banks have little incentive to correct their behaviors. Reducing discrimination through competition, on the other hand, relies only on the threat of entry by a large player. So long as there are business opportunities in a market and banks willing and able to enter the market, entry and the threat of entry serves to reduce discrimination both by pressuring incumbents and by the direct entry of less discriminatory banks. The benefit of the competition-induced punishment of discrimination is that it puts the onus of enforcement on likely well-funded agents banks with strong incentives to take action. The entering banks entry (or entry threats) rely only on a profit motive, and as a side effect of 16 The elements necessary to se under the Fair Housing Act are (1) the property must be covered, (2) the transaction must be covered, (3) there must be an illegal basis of discrimination (e.g., race), and (4) the denial must be because of the protected basis. The fourth element is usually the element in dispute; evidence that the bank treats otherwise equal white and black applicants differently is sufficient to establish it. See Schwemm (1994). 26