TY - JOUR
AU - Lopez, Luis, A
AB - Abstract We test for pricing disparities in mortgage contracts using a novel data set that allows us to observe the race and ethnicity of both parties to the loan. We find that minorities pay between 3% and 5% more in fees than similarly qualified whites when obtaining a loan through the same white broker. Critically, we find that the premium paid by minorities depends on the race of the broker. We also examine recent policy changes around broker compensation rules that may not only reduce these price disparities but may also limit access to credit for minorities. Economists and policy makers have long observed that economic outcomes are correlated with race and ethnicity in many countries and societies. In the United States, these observations created pressure to pass legislation to remove discriminatory practices, such as the 1964 Civil Rights Act, which ended segregation and sought to create equal opportunities in labor markets, and the 1968 Fair Housing Act, which explicitly targeted discriminatory practices in the housing market. Yet, over half a century later, evidence suggests that discriminatory practices in housing and financial markets may continue. For example, several large mortgage lenders were recently subject to litigation involving charges of disparate treatment (i.e., illegal discrimination based on a protected characteristic such as race or gender) of minorities (New York, Office of the Attorney General, Civil Rights Bureau 2006, 2007, 2008, United States Dept. of Justice 2010, 2011), and academic researchers continue to find evidence of loan pricing disparities (Bartlett et al. 2019; Bhutta and Hizmo 2020; Woodward and Hall 2012; Ghent, Hernández-Murillo, and Owyang 2014; Cheng, Lin, and Liu 2015), suggesting that discrimination in the mortgage industry remains a concern. Earlier studies often relied on observing differences in transaction outcomes (e.g., accept/reject decisions, pricing, or mortgage performance) across race as evidence for discrimination. The seminal study in this line is Munnell et al. (1996), who reported that minorities in Boston had mortgage loan denial rates that were twice as high as white applicants. In addition, Berkovec et al. (1994, 1998) found evidence consistent with discrimination using FHA loan performance (i.e., mortgage default) data.1 Building on these studies, the literature has evolved to consider disparate treatment or differentials across racial groups in loan pricing (boehm2007mortgage, haughwout2009subprime, CourchaseZorn2012, ghent2014differences,bartlett:2018, kau:2019), yield spreads or overages (Courchane and Nickerson 1997,Crawford and Rosenblatt 1999,Black, Boehm, and DeGennaro 2003, Clarke and Rothenberg 2017), incidences of high APR spreads (Bhutta and Ringo 2015, Bayer, Ferreira, and Ross 2017), application rejection rates (Munnell et al. 1996; Ross and Yinger 2002; Horne 1997), and loan performance (Bayer, Ferreira, and Ross 2016; Deng and Gabriel 2006; Berkovec et al. 1994, 1998; Ambrose and Capone 1996; Brueckner 1996; Ross 1996; Yinger 1996). In this paper, we examine racial differences in total broker compensation, which as noted in Courchane and Nickerson (1997) is a key margin through which disparate treatment may occur, and thus is the source of several recent administrative actions against lenders (e.g. New York, Office of the Attorney General, Civil Rights Bureau 2006, 2007, 2008). We use a novel administrative data set containing all underwriting information collected for over 300,000 mortgages originated by 124,736 individual brokers between 2003 and 2007. In doing so, we make four key contributions. First, whereas previous studies have only observed the race and ethnicity of the borrower, we identify the race of the individual loan officer (i.e., broker) using the Bayesian Improved First Name Surname Geocoding (BIFSG) method developed in Voicu (2018). This allows us to examine the racial interactions that lead to minority pricing premiums, thus providing additional insights into the literature examining within and across ethnic and racial group interactions (Agarwal et al. 2019; Li 2014; Wong 2013; Zhang and Zheng 2015; Bertrand, Luttmer, and Mullainathan 2000; Bayer, McMillan, and Rueben 2004). After conditioning on an extensive set of control variables, we find that all minority borrower groups (Hispanic, Black, Asian) pay more than their white counterparts when obtaining loans through white brokers. However, we find heterogeneity in minority premiums across broker race. For example, Hispanic borrowers pay a premium even when the broker is Hispanic. On the other hand, Black brokers do not appear to charge different fees to white and Black borrowers, and Asian borrowers pay lower fees than comparable white borrowers when originating loans through an Asian broker. We also report on a battery of robustness tests in the Internet Appendix to alleviate concerns regarding endogeneity and selection biases. Thus, our results suggest that minority premiums depend critically on the race of the mortgage broker. Second, our data allow us to include individual mortgage broker fixed effects in the regression models. We can thus determine whether a minority pays more than a comparable white borrower when using the same mortgage broker. Minority premiums remain, but are smaller in magnitude, after the inclusion of broker fixed effects, which suggests that minority borrowers tend to systematically select into high-fee brokers. This is consistent with the hypothesis originated by Yezer, Phillips, and Trost (1994), and is related to recent studies showing that the inclusion of lender (institution) fixed effects reduces the magnitude of earlier findings of racial disparities in mortgage outcomes (Bayer, Ferreira, and Ross 2017, Bhutta and Hizmo 2020,Avery, Canner, and Cook 2005,Avery, Brevoort, and Canner 2007). The inclusion of broker fixed effects allows for a similar interpretation, but at a more granular (individual broker) level, allowing us to focus on the differences in fees attributable to the individual mortgage broker, an area previously unexplored. Our analysis also confirms and complements those of Woodward and Hall (2012) for Black and Hispanic borrowers, which were based on a different loan product (Federal Housing Administration [FHA] mortgages) originated during a different time (a 6-week period in 2001). Although including broker fixed effects results in more modest minority fee premiums, we note that the premiums paid by minorities are within the range of pricing differences that triggered legal action against lenders for disparate impact (i.e., the application of a neutral policy that adversely affects a group with protected characteristics). Third, our analysis builds on the work of Woodward and Hall (2012) and Woodward (2008) to provide novel evidence regarding the relation between broker compensation and borrower credit risk.2 Standard pricing theory suggests that broker compensation should be independent of borrower credit risk since the broker is simply the facilitator for originating a mortgage and does not bear default risk (Woodward and Hall 2012).3 Supporting this theory, our results indicate no clear relationship between broker compensation and standard credit risk metrics. We do find some evidence that loans with extreme levels of broker compensation (those in the fifth quintile of the fee distribution) have slightly higher ex post default rates, but this does not explain the observed minority pricing premiums. In fact, we find the surprising result that minority premiums are generally positively correlated with borrower credit quality—that is, minority borrowers with higher credit scores pay higher premiums, all else equal. Finally, we contribute to the debate surrounding the efficiency and effects of financial regulations designed to protect consumers (Campbell et al. 2011). Although the race of the broker or loan officer is immaterial from a legal perspective in prosecuting fair lending laws, we offer policy makers new insights into the source of potential disparate treatment (whether it is the originating institution or individual), which is crucial to designing public policy remedies. We design a test to illustrate and quantify the consequences (intended and unintended) of recent consumer protection regulations that were designed to restrict broker compensation. For example, Title XIV (the Mortgage Reform and Anti-Predatory Lending Act) of the Dodd-Frank Act severely restricts how mortgage brokers may be compensated.4 Since we find significant fee disparity across race, our study shows the importance for continued evaluation of the effectiveness of regulatory oversight versus the reliance on enforcement of existing laws to combat disparate treatment in mortgage markets. We first focus on the Dodd-Frank regulation meant to increase pricing transparency: the proposed elimination of dual compensation (broker compensation from both the borrower and the lender as discussed in footnote 2). We continue to find fee differences on transparently priced loans, that is, mortgages without dual compensation. This suggests that a regulation banning dual compensation, per se, is unlikely to eliminate racial price disparities. Next, we consider whether differences in fees arise from borrower heterogeneity with respect to broker loan production costs, and whether the Dodd-Frank regulations may result in credit rationing disparities. Based on a quantile regression framework, we estimate at the 30th quantile that over 25% of Hispanic and Black borrowers (and 6% of Asian borrowers) with loans originated by white brokers would have been at risk of being credit rationed as a result of fee caps imposed by Dodd-Frank. In contrast, only about 16% of white borrowers with loans originated by white loan officers would be at risk of credit rationing. Thus, although the restrictions may reduce pricing disparities, they may also result in credit rationing to borrowers needing extra effort by mortgage brokers to originate loans as suggested by Yezer, Phillips, and Trost (1994) and Yezer (2017). As others have noted (Bayer et al. 2012, Anwar and Fang 2006, Ayers and Siegelman 1995), it is difficult to determine whether racial disparities are driven by statistical (Arrow 1971, Phelps 1972) or taste-based discrimination (Becker 1957). Although our results do not provide definitive proof in support of one theory over the other, we offer some observations regarding the interpretation of our results in light of these two leading economic theories of discrimination. First, following the logic developed in Antonovics and Knight (2009), we note that if statistical discrimination is driving observed disparities, then the fees charged to borrowers should be independent of the broker’s race. Clearly, our results do not support this hypothesis, and thus are inconsistent with a traditional statistical discrimination theory. Furthermore, we rule out a variety of alternative causal mechanisms, such as differences in borrower credit risk, broker effort, borrower contract selection, nonrandom borrower-broker matching, language, or ethnic enclaves. Taste-based discrimination, on the other hand, predicts that brokers charge premiums to borrowers that do not share the same race. Under this theory, our results would suggest that white and Asian brokers exhibit taste-based discrimination, while Black brokers do not. Our results would also point to preference-based discrimination by Hispanic brokers against Hispanic borrowers. Although this seems counterintuitive, this is possible under certain assumptions.5 Ultimately, we are unable to ascertain the exact causal mechanism driving price disparities; however, we reiterate that our results do not appear to be driven by statistical discrimination as the term is commonly defined in the economics literature. In contrast to economics, the legal literature often focuses on whether discrimination arises from disparate treatment or disparate impact. In our context, disparate treatment implies that the broker uses race in making pricing decisions, while disparate impact occurs from the application of a race-neutral policy that has a disproportionately adverse effect on borrowers of a certain race (Clarke and Rothenberg 2017). Our results show that minorities encounter a different front-end/back-end fee trade-off, a finding that suggests brokers treat minorities differently. Although our results point to disparate treatment, we cannot entirely rule out disparate impact. But, regardless of whether our results are defined as disparate treatment or disparate impact, the observed discrimination would likely be deemed illegal by the courts. 1. Data We use data on loan applications for brokered, first-lien, residential mortgages that were approved and funded by New Century Financial Corporation between January 2003 and March 2007.6 Although the loans were funded by a single lender, our analysis focuses on mortgages that were originated by 124,736 independent mortgage brokers who had access to a variety of lenders, thus reducing concerns that our results are idiosyncratic to one particular lender. Ambrose, Conklin, and Yoshida (2016) discuss the data in greater detail and provide comparisons to mortgages in other studies that indicate that the New Century loans are representative of the overall subprime market. Nonetheless, we also discuss below that our sample is representative of the subprime market before the Great Recession. We use the New Century data because each loan file contains elements central to our analysis: the borrower’s Home Mortgage Disclosure Act (HMDA) race code and the broker’s name and office location.7 The data set also contains borrower, property, and loan characteristics as well as broker fees. Based on property location, we merge the New Century data with Census 2000 data to gain geographic controls. The Census variables are similar to those used in Bayer, Ferreira, and Ross (2017). Table A1 in the Internet Appendix lists and describes the variables. 1.1 Sample specification and representativeness Following Ambrose, Conklin, and Yoshida (2016) and Conklin (2017), we exclude loan applications with missing data or when (1) the borrower’s and coborrower’s combined monthly income is negative or greater than |${\$}$|26,900; (2) the combined loan-to-value ratio is negative or larger than 125%; (3) the borrower’s FICO credit score is less than 450; (4) the debt-to-income ratio is negative or larger than 60%; and (5) the borrower’s age is reported as younger than 18 years or older than 99 years. We also winsorize the 1% tails of the combined monthly income and broker fees. Furthermore, we keep loans originated by white, Hispanic, Black, or Asian or Pacific Islander brokers to borrowers in those same racial/ethnic groups. The final sample includes 323,846 originated loans. As noted in Section A.1 of the Internet Appendix, the typical principal borrower is a 40-year-old, married male with a credit score of 619 and an annual income of approximately |${\$}$|68,500.8 The average loan is an adjustable rate mortgage with a loan amount of |${\$}$|172,800 on a 30-year term with a prepayment penalty.9 Forty-two percent of these loans were originated to purchase a residential property, and the rest to refinance an existing mortgage. Among refinances, 85% are cash-out mortgages having loan amounts that exceed the outstanding balance of debt being refinanced. To ensure that the sample is representative of the subprime market from 2003 to 2007, in the Internet Appendix we compare the New Century data with the subprime loan sample in Demyanyk and Van Hemert (2009). Their loan sample comprises loans across many subprime lenders and covers roughly half of the subprime mortgage market (85% of the securitized subprime market). As we detail in Section A.1, the descriptive statistics across the two samples are quite similar. In Section A.1, we also compare the New Century data to the Home Mortgage Disclosure Act (HMDA) loan application register data. We note that the minority share of subprime originations in HMDA for New Century (51%) is nearly identical to the share in the rest of the subprime market (52%); thus, alleviating concerns that the New Century data suffer from selection issues based on borrowers’ minority status. 1.2 Observable race and ethnicity While we observe borrower race and ethnicity (because of HMDA reporting requirements), we do not directly observe the race and ethnicity of brokers. However, we are able to infer their race and ethnicity using a Bayesian-based classifier approach, which is similar in spirit to the methodology used by regulators to determine consumer race and ethnicity (Consumer Financial Protection Bureau 2014b). In addition, various courts have relied on Bayesian-based classification methods in cases in which it was necessary to infer an individual’s race or ethnicity (e.g., Guardians Ass’n of N.Y.C. Police Dep’t v. Civil Serv. Comm’n 1977, ¶ 32).10 We infer mortgage broker race using the Bayesian Improved First Name Geocoding (BIFSG) method developed in Voicu (2018) (see Section A.2 in the Internet Appendix for a detailed discussion). The intuition of the approach is to calculate the probability (Bayesian score) that a person falls into a certain race and ethnicity based on the individual’s surname, first name, and location. A Bayesian score for each race is calculated for every broker in the sample. To categorize a broker’s race discretely, we apply a “maximum a posteriori” (MAP) classification scheme that sets an individual’s race to that of the group with the highest Bayesian score.11 Relative to other classification schemes, MAP is more accurate, minimizes bias, and maximizes data coverage (Voicu 2018). Table 1 reports the number of unique brokers in the sample by the number of loans they originated. We identify the race and ethnicity of 124,736 individual brokers. Sixteen percent are identified as Hispanic, 8% as Black, 4% as Asian or Pacific Islander, and the rest as white. To our knowledge, the only other source of demographic information on mortgage loan officers is Hanson et al. (2016). Whereas our sample covers 2003 to 2007, the Hanson et al. (2016) sample is from 2012, a period when subprime mortgage lending was virtually nonexistent. But, consistent with Hanson et al. (2016), we find that the overwhelming majority of loan officers are white. The similarity between our loan officer demographics and those reported in Hanson et al. (2016) suggest that brokers that worked with New Century are representative of loan officers in the broader mortgage market. Table 1 Unique brokers by race Loans per broker . White . Hispanic . Black . Asian or Pacific Islander . 53,941 11,946 5,682 3,130 2 15,057 3,294 1,486 719 3 6,959 1,580 740 368 4 3,888 870 391 181 5 2,333 582 255 124 6 1,687 424 176 74 7 1,163 332 132 58 8 900 217 108 39 9 698 180 60 26 10+ 3,409 979 394 154 Total 90,035 20,404 9,424 4,873 Sample share (%) 72 16 8 4 Loans per broker . White . Hispanic . Black . Asian or Pacific Islander . 53,941 11,946 5,682 3,130 2 15,057 3,294 1,486 719 3 6,959 1,580 740 368 4 3,888 870 391 181 5 2,333 582 255 124 6 1,687 424 176 74 7 1,163 332 132 58 8 900 217 108 39 9 698 180 60 26 10+ 3,409 979 394 154 Total 90,035 20,404 9,424 4,873 Sample share (%) 72 16 8 4 This table reports the number of unique brokers by race and the number of loan originations they arranged in our sample. Open in new tab Table 1 Unique brokers by race Loans per broker . White . Hispanic . Black . Asian or Pacific Islander . 53,941 11,946 5,682 3,130 2 15,057 3,294 1,486 719 3 6,959 1,580 740 368 4 3,888 870 391 181 5 2,333 582 255 124 6 1,687 424 176 74 7 1,163 332 132 58 8 900 217 108 39 9 698 180 60 26 10+ 3,409 979 394 154 Total 90,035 20,404 9,424 4,873 Sample share (%) 72 16 8 4 Loans per broker . White . Hispanic . Black . Asian or Pacific Islander . 53,941 11,946 5,682 3,130 2 15,057 3,294 1,486 719 3 6,959 1,580 740 368 4 3,888 870 391 181 5 2,333 582 255 124 6 1,687 424 176 74 7 1,163 332 132 58 8 900 217 108 39 9 698 180 60 26 10+ 3,409 979 394 154 Total 90,035 20,404 9,424 4,873 Sample share (%) 72 16 8 4 This table reports the number of unique brokers by race and the number of loan originations they arranged in our sample. Open in new tab We partition the data into three subsamples: (1) Hispanic and white brokers and borrowers (HW); (2) Black and white brokers and borrowers (BW); and (3) Asian and Pacific Islander and white brokers and borrowers (AW). Note that the white borrower/white broker observations are the same in each subsample and serve as the reference group. Performing the analysis by individual minority groups, with whites as the reference group, may shed light on the channels through which “minority premiums” emerge. Table 2 provides observation counts by broker and borrower race for each subsample. We observe that brokers tend to originate loans to borrowers who share the same race or ethnicity. Nineteen percent of loans arranged by white brokers in the HW sample were to Hispanic borrowers. In contrast, 81% of those originated by Hispanic brokers were to Hispanic borrowers. We observe similar patterns in the BW and AW subsamples. Table 2 Summary statistics of broker fees and underwriting factors A. HW . White broker . Hispanic broker . . White . Hispanic . White . Hispanic . . borrower . borrower . borrower . borrower . Broker fees 5,116 6,184 5,796 6,435 Stated income 36 49 41 59 Debt-to-income 39 41 40 41 CLTV 85 86 84 86 Credit score 616 624 624 636 Annual income 82,305 80,979 89,019 80,365 Age 42 40 43 40 Obs. 142,539 33,415 11,117 47,342 B. BW White broker Black broker White Black White Black borrower borrower borrower borrower Broker fees 5,116 5,578 5,086 5,255 Stated income 36 33 37 36 Debt-to-income 39 40 39 40 CLTV 85 86 86 87 Credit score 616 602 614 610 Annual income 82,305 71,549 78,741 70,767 Age 42 44 43 43 Obs. 142,539 46,709 5,788 18,539 C. AW White broker API broker White API White API borrower borrower borrower borrower Broker fees 5,116 6,619 6,777 7,106 Stated income 36 51 42 59 Debt-to-income 39 41 40 41 CLTV 85 88 85 88 Credit score 616 638 630 653 Annual income 82,305 102,131 101,385 115,475 Age 42 41 43 41 Obs. 142,539 7,394 4,017 6,986 A. HW . White broker . Hispanic broker . . White . Hispanic . White . Hispanic . . borrower . borrower . borrower . borrower . Broker fees 5,116 6,184 5,796 6,435 Stated income 36 49 41 59 Debt-to-income 39 41 40 41 CLTV 85 86 84 86 Credit score 616 624 624 636 Annual income 82,305 80,979 89,019 80,365 Age 42 40 43 40 Obs. 142,539 33,415 11,117 47,342 B. BW White broker Black broker White Black White Black borrower borrower borrower borrower Broker fees 5,116 5,578 5,086 5,255 Stated income 36 33 37 36 Debt-to-income 39 40 39 40 CLTV 85 86 86 87 Credit score 616 602 614 610 Annual income 82,305 71,549 78,741 70,767 Age 42 44 43 43 Obs. 142,539 46,709 5,788 18,539 C. AW White broker API broker White API White API borrower borrower borrower borrower Broker fees 5,116 6,619 6,777 7,106 Stated income 36 51 42 59 Debt-to-income 39 41 40 41 CLTV 85 88 85 88 Credit score 616 638 630 653 Annual income 82,305 102,131 101,385 115,475 Age 42 41 43 41 Obs. 142,539 7,394 4,017 6,986 Panel A reports the mean values of broker fees and underwriting factors of loans originated by white and Hispanic brokers for white and Hispanic borrowers. Panel B reports the mean values of broker fees and underwriting factors of loans originated by white and Black brokers for white and Black borrowers. Panel C reports the mean values of broker fees and underwriting factors of loans originated by white and API brokers for white and API borrowers. The variable combined loan-to-value is the nominal combined loan amount to collateral value ratio. The variable credit score is the borrower’s nominal FICO score. See Table A1 for definitions of the other variables. Open in new tab Table 2 Summary statistics of broker fees and underwriting factors A. HW . White broker . Hispanic broker . . White . Hispanic . White . Hispanic . . borrower . borrower . borrower . borrower . Broker fees 5,116 6,184 5,796 6,435 Stated income 36 49 41 59 Debt-to-income 39 41 40 41 CLTV 85 86 84 86 Credit score 616 624 624 636 Annual income 82,305 80,979 89,019 80,365 Age 42 40 43 40 Obs. 142,539 33,415 11,117 47,342 B. BW White broker Black broker White Black White Black borrower borrower borrower borrower Broker fees 5,116 5,578 5,086 5,255 Stated income 36 33 37 36 Debt-to-income 39 40 39 40 CLTV 85 86 86 87 Credit score 616 602 614 610 Annual income 82,305 71,549 78,741 70,767 Age 42 44 43 43 Obs. 142,539 46,709 5,788 18,539 C. AW White broker API broker White API White API borrower borrower borrower borrower Broker fees 5,116 6,619 6,777 7,106 Stated income 36 51 42 59 Debt-to-income 39 41 40 41 CLTV 85 88 85 88 Credit score 616 638 630 653 Annual income 82,305 102,131 101,385 115,475 Age 42 41 43 41 Obs. 142,539 7,394 4,017 6,986 A. HW . White broker . Hispanic broker . . White . Hispanic . White . Hispanic . . borrower . borrower . borrower . borrower . Broker fees 5,116 6,184 5,796 6,435 Stated income 36 49 41 59 Debt-to-income 39 41 40 41 CLTV 85 86 84 86 Credit score 616 624 624 636 Annual income 82,305 80,979 89,019 80,365 Age 42 40 43 40 Obs. 142,539 33,415 11,117 47,342 B. BW White broker Black broker White Black White Black borrower borrower borrower borrower Broker fees 5,116 5,578 5,086 5,255 Stated income 36 33 37 36 Debt-to-income 39 40 39 40 CLTV 85 86 86 87 Credit score 616 602 614 610 Annual income 82,305 71,549 78,741 70,767 Age 42 44 43 43 Obs. 142,539 46,709 5,788 18,539 C. AW White broker API broker White API White API borrower borrower borrower borrower Broker fees 5,116 6,619 6,777 7,106 Stated income 36 51 42 59 Debt-to-income 39 41 40 41 CLTV 85 88 85 88 Credit score 616 638 630 653 Annual income 82,305 102,131 101,385 115,475 Age 42 41 43 41 Obs. 142,539 7,394 4,017 6,986 Panel A reports the mean values of broker fees and underwriting factors of loans originated by white and Hispanic brokers for white and Hispanic borrowers. Panel B reports the mean values of broker fees and underwriting factors of loans originated by white and Black brokers for white and Black borrowers. Panel C reports the mean values of broker fees and underwriting factors of loans originated by white and API brokers for white and API borrowers. The variable combined loan-to-value is the nominal combined loan amount to collateral value ratio. The variable credit score is the borrower’s nominal FICO score. See Table A1 for definitions of the other variables. Open in new tab 2. Analysis of Broker Fees We examine differences in log broker fees, calculated as the natural logarithm of the sum of front- and back-end fees. Front-end fees include the application fee, underwriter fee, mortgage brokerage firm fee, and points.12 Borrowers generally incur these fees during the loan origination process and pay them at closing.13 Back-end fees include the yield spread premium and correspondence premium. The yield spread premium is the total rebate that the lender provides to the broker at closing for locking a contract rate above the minimum rate the borrower qualifies to receive (par).14 The correspondence premium is analogous to a yield spread premium; it compensates the broker for originating a loan at an interest rate above par. 2.1 Distribution of broker fees Table 2 displays the mean value of broker fees and several key underwriting factors by borrower-broker race across the three race subsamples. Broker fees vary considerably across broker and borrower race groups. Minorities pay more, on average, than white borrowers within a given broker race in each subsample. Panels A and C offer evidence that white borrowers pay higher fees when obtaining a loan through a minority broker. However, underwriting factors also differ across groups with large differences existing in the share of stated income loans and average annual income across groups, which suggests that borrowers and loan products may vary systematically with broker and borrower race.15 Thus, unconditional mean fee differences may be uninformative. We address this more formally in Section 2.2. Figure 1 depicts the kernel density of log broker fees by borrower-broker race or ethnicity for each subsample. For Hispanic borrowers, the distribution of fees sits to the right of white borrowers regardless of the race of the mortgage broker. In the BW subsample, the picture is less clear as the right tail of the Black borrower fee distribution for both white and Black brokers appears to have an additional mass. Finally, in the AW subsample, the minority fee distribution sits to the right of the white fee distribution when the broker is white and when the broker is Asian or Pacific Islander (API). Overall, the unconditional fee distributions in Figure 1 suggest that minority premiums exist, regardless of the broker’s race or ethnicity. Figure 1 Open in new tabDownload slide Kernel density of log broker fees by group This figure displays the distribution of log broker fees in the three main samples used in our analysis (Hispanic/white, Black/white, and API/white). Distributions are separated by borrowers’ and mortgage brokers’ race. Figure 1 Open in new tabDownload slide Kernel density of log broker fees by group This figure displays the distribution of log broker fees in the three main samples used in our analysis (Hispanic/white, Black/white, and API/white). Distributions are separated by borrowers’ and mortgage brokers’ race. 2.2 Empirical model Table 2 indicates that significant mean differences exist in underwriting factors across broker and borrower race combinations. These differences in borrower characteristics and loan products across groups could drive the observable variation in broker fees. Indeed, Bayer, Ferreira, and Ross (2017) note that a limitation of recent studies on mortgage pricing is that some of the key loan attributes associated with high cost loans are unobservable in standard data sets, making it impossible to determine whether demand for these product types explains the minority premiums. In contrast, our administrative data set contains all information collected by the lender at origination and the characteristics of the originated mortgage. This allows us to account for observable differences across borrowers, brokers, and product features. Thus, we test the impact of borrower’s and broker’s minority status on broker fees with the following ordinary least squares (OLS) regression: $$\begin{equation}\label{eq:interaction} P_{imt} = \delta_1 B^M_i + \delta_2L^M_i + \delta_3 B^M_i\times L^M_i + X_{imt}'\beta + \tau_t + \kappa_m + \varepsilon_{imt}, \end{equation}$$(1) where |$P_{imt}$| is the natural logarithm of broker fees paid by borrower |$i$|⁠, in metropolitan statistical area (MSA) |$m$|⁠, at time |$t$|⁠. |$B^M_i$| is a dummy variable that equals one when the borrower is a minority, and zero otherwise. |$L^M_i$| is a dummy variable that equals one when the broker is a minority, and zero otherwise. |$X_{imt}$| denotes the matrix of control variables (described in Table A1), |$\tau_t$| denotes origination year-quarter fixed effects, and |$\kappa_m$| denotes MSA fixed effects. The origination year-quarter fixed effects account for variation in broker fees that arise from temporal changes in the economic environment. The MSA fixed effects account for geographic-specific differences. The error term |$\varepsilon_{imt}$| is clustered at the MSA level. We classify control variables into four broad categories: borrower, loan, property type, and area/geography. As noted above, these variables represent virtually all information collected at the time of origination thereby allowing the regression framework to estimate the effect of differences in borrowers’ race or ethnicity holding constant all observable factors that might affect origination fees. Borrower controls include variables that describe demographic attributes (i.e., gender, age, and marital status), and underwriting risk factors (i.e., credit score, a subprime indicator if the FICO score is less than 620, income, debt-to-income, and employment status). Property-type controls indicate whether the collateral is owner occupied, a second home, an investment property, a condominium, a two-to-four unit multifamily, or a single-family residence. Loan controls include variables that describe features specific to the loan contract, such as the loan purpose (i.e., purchase, rate-term refinancing, or cash-out refinancing), loan type (i.e., adjustable rate, interest only, or fixed rate), log loan amount, combined loan-to-value ratio (CLTV), loan term, spread between the contract interest rate and the 2-year Constant Maturity Treasury, prepayment penalty presence, stated-income documentation, and loan arrangement settings (i.e., coborrower presence and meeting face-to-face).16 As in Haughwout, Mayer, and Tracy (2009), we allow the loan-to-value to affect the cost of credit nonlinearly by using dummy variable bins. Finally, area controls include variables that influence the competitive setting and economic environment at the property location. This category includes the MSA/quarter broker Herfindahl-Hirschman index (HHI) that proxies for market competition among brokers. Ambrose and Conklin (2014) show that broker HHI affects the costs of obtaining a mortgage. The area controls also include the Pahl index that provides a measure of mortgage broker regulations and occupational licensing requirements across states (Pahl 2007). The effect of regulation on fees is ambiguous as increased monitoring of broker activities could decrease fees, while increased costs of broker compliance could increase fees. We include the share of college educated adults in the county to control for the effect observed by Woodward and Hall (2012) that borrower education (proxied by area education level) affects the cost of credit.17 We capture variation in area wealth levels by including the per capita income at the ZIP code level, county-level median income, and the county poverty share in the year of loan origination. We also include the share of the county adult population that is unmarried. To capture geographic differences in housing markets, we include the county rent to price ratio and the county share of housing that is owner occupied. To control for the possibility that brokers and borrowers are operating within ethnic enclaves, we include additional county demographic controls measured as a fraction of county population: percentage Hispanic, percentage Black, percentage Asian or Pacific Islander, percentage foreign born, only the English-speaking share, and only the Spanish-speaking share.18 Finally, we include the monthly MSA unemployment rate from the Bureau of Labor Statistics and the log distance in miles between the borrower’s and broker’s ZIP codes reported by New Century. We estimate Equation (1) separately for the HW, BW, and AW subsamples. The white borrowers that work with white brokers are the same in each subsample and serve as the reference group. The parameters |$\delta_j$|⁠, where |$j\in\{1,2,3\}$|⁠, represent the coefficients of interest as they reveal whether minority premiums exist and to what extent they vary with broker race. Since brokers had significant discretion over the fees negotiated on each loan, broker heterogeneity may explain the observed pricing differentials. For example, if minority borrowers select into “high-fee” white brokers, while white borrowers select into “low-fee” white brokers, then the observed differences may simply reflect that the two borrower groups use different mortgage brokers. To address this issue, we expand Equation (1) to include individual mortgage broker fixed effects. The broker fixed effects models exploit within broker variation in borrower race to identify minority pricing premiums. The intent is to isolate variation in fees and borrower minority status from variation in unobserved broker attributes. These models are similar in spirit to those of Bayer, Ferreira, and Ross (2017) and Munnell et al. (1996); however, we control for potential unobserved heterogeneity at a more granular (individual) level. Our broker fixed effects models also closely approximate the identification strategy used in experimental paired-audit studies (e.g. Ayers and Siegelman 1995). By including broker fixed effects along with a rich set of control variables, we ask whether a minority borrower pays more than a comparable white borrower when obtaining a loan from the same mortgage broker. Additionally, we observe whether within broker minority premiums vary across broker race.19 2.3 Are mortgage brokers compensated for borrower credit risk? As mentioned above, standard theories of pricing suggest that mortgage broker compensation should not vary systematically with borrower credit risk because brokers do not bear default risk on the loans they originate. We provide evidence in support of this hypothesis by comparing the estimated coefficients from a linear probability model of mortgage default with the estimated coefficients on the log fee model. If broker compensation is directly related to risk, then we would expect the coefficient estimates to follow the same pattern. Figure 2 shows the coefficient estimates (with 95% confidence intervals) for a set of risk characteristics that are commonly used in the mortgage default literature. In the right panels, we plot the corresponding coefficient estimates from the broker fee model.20 No clear relationship between default risk and broker compensation emerges. For example, although high CLTV loans (⁠|$>$|95%) increase the likelihood of default, they are not associated with greater broker compensation.21 FICO score is inversely related to default, but the same relationship does not hold with respect to broker fees. Finally, although second homes and stated income loans increase the likelihood of default, broker compensation is actually lower on these loans.22 Taken together, Figure 2 offers compelling evidence that broker fees are not directly tied to credit risk and thus alleviates concerns that observable minority premiums are due to credit risk factors that are unobservable to the econometrician. Figure 2 Open in new tabDownload slide Default risk characteristics and broker fees This figure shows that no clear relationship exists between mortgage risk characteristics and mortgage broker compensation. The left panels plot coefficient estimates and 95% confidence intervals from the mortgage default regression reported in column 1 of Table A6 in the Internet Appendix. The right panels plot coefficient estimates from a regression with log broker fees as the dependent variable as reported in column 2 of Table A6. The dependent variable in the regression takes a value of one if the mortgage becomes 60 or more days delinquent within 2 years of origination and zero otherwise, and the control variables are those from Equation (1). Figure 2 Open in new tabDownload slide Default risk characteristics and broker fees This figure shows that no clear relationship exists between mortgage risk characteristics and mortgage broker compensation. The left panels plot coefficient estimates and 95% confidence intervals from the mortgage default regression reported in column 1 of Table A6 in the Internet Appendix. The right panels plot coefficient estimates from a regression with log broker fees as the dependent variable as reported in column 2 of Table A6. The dependent variable in the regression takes a value of one if the mortgage becomes 60 or more days delinquent within 2 years of origination and zero otherwise, and the control variables are those from Equation (1). 2.4 Main results We now turn to our main results. Table 3 contains the OLS estimates from Equation (1).23 Columns 1 through 3 correspond to the Hispanic/white (HW) subsample. In addition to the borrower and broker race variables, we include the natural logarithm of loan amount as a control in all models. Column 1 shows that Hispanic borrowers that obtain a loan through a white broker pay 9% more than white borrowers that also use a white broker.24 In dollar terms, this premium translates into an additional |${\$}$|548 in fees on the average loan, which is nearly identical to the Hispanic premium (⁠|${\$}$|489) estimated by Woodward (2008) in a study of 7,560 Federal Housing Administration (FHA) insured loans originated in 2001 (table 3a). The Hispanic premium exists even when the broker is Hispanic; a Hispanic borrower pays 11% (or |${\$}$|624) more than a white borrower that receives a loan through a Hispanic broker.25 The minority/minority premium is not significantly different from the minority premium with a white mortgage broker. In other words, Hispanic borrowers pay a significant premium relative to white borrowers regardless of the broker’s race. Column 1 also shows that white borrowers pay a small premium (2%, or |${\$}$|102) when obtaining a loan through a Hispanic broker. Table 3 OLS regressions of ln(Broker fees) . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . Dep. var.: ln(Broker fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.09*** 0.05*** 0.04*** 0.14*** 0.08*** 0.05*** 0.04*** 0.03*** 0.03*** (0.01) (0.01) (0.00) (0.02) (0.01) (0.00) (0.01) (0.01) (0.01) Minority broker 0.02* 0.06*** 0.07*** (0.01) (0.01) (0.02) Minority borrower |$\times$| Minority broker 0.01 0.00 0.01 -0.08*** -0.06*** -0.05*** -0.14*** -0.08*** -0.08** (0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.02) (0.03) (0.03) Observations 234,413 175,670 175,660 213,575 158,543 158,535 160,936 114,303 114,293 Adjusted |$R^2$| .32 .51 .56 .33 .54 .58 .30 .51 .56 log loan amount Y Y Y Y Y Y Y Y Y Broker FE N Y Y N Y Y N Y Y Other controls N N Y N N Y N N Y Year-quarter FE N N Y N N Y N N Y MSA FE N N Y N N Y N N Y Minority/minority premium 0.11 0.06 0.05 0.06 0.02 0.00 -0.10 -0.05 -0.04 p-value .00 .00 .00 .00 .13 .79 .00 .08 .14 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . Dep. var.: ln(Broker fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.09*** 0.05*** 0.04*** 0.14*** 0.08*** 0.05*** 0.04*** 0.03*** 0.03*** (0.01) (0.01) (0.00) (0.02) (0.01) (0.00) (0.01) (0.01) (0.01) Minority broker 0.02* 0.06*** 0.07*** (0.01) (0.01) (0.02) Minority borrower |$\times$| Minority broker 0.01 0.00 0.01 -0.08*** -0.06*** -0.05*** -0.14*** -0.08*** -0.08** (0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.02) (0.03) (0.03) Observations 234,413 175,670 175,660 213,575 158,543 158,535 160,936 114,303 114,293 Adjusted |$R^2$| .32 .51 .56 .33 .54 .58 .30 .51 .56 log loan amount Y Y Y Y Y Y Y Y Y Broker FE N Y Y N Y Y N Y Y Other controls N N Y N N Y N N Y Year-quarter FE N N Y N N Y N N Y MSA FE N N Y N N Y N N Y Minority/minority premium 0.11 0.06 0.05 0.06 0.02 0.00 -0.10 -0.05 -0.04 p-value .00 .00 .00 .00 .13 .79 .00 .08 .14 This table reports OLS estimates. The dependent variable is the natural log of broker fees. Each reported covariate is a dummy variable that equals one when true, and zero otherwise. Borrower controls include age, gender, marital status, income, credit (FICO) score, a subprime mortgage indicator (FICO|$<$|620), debt-to-income ratio, and employment status. Property-type controls include owner-occupancy versus investor status and single-family versus condominium or multiple unit structure. Loan controls include indicators for broker face-to-face interaction, purchase versus refinance, coborrowing, combined debt loan-to-value (CLTV), loan type (adjustable rate, fixed rate, interest only), interest rate spread, and borrower documentation. For the loan-to-value ratio, we use five CLTV categories: CLTV |$<$| 80%, 80% |$\le$| CLTV |$<$| 85%, 85% |$\le$| CLTV |$<$| 90%, 90% |$\le$| CLTV |$\le$| 95%, and CLTV |$\ge$| 95%. Area controls include the distance between borrower and broker, broker competition, area unemployment rate, regulatory environment, education, income, share of housing that is owner occupied, price-to-rent ratio, and county population share that is married, foreign born, Hispanic, Black, Asian or Pacific Islander, English speaking, and Hispanic speaking. The stand-alone broker race term (⁠|$L^M$|⁠) in Equation (1) is absorbed by the broker fixed effects. See Table A1 for a complete description of the variables. Robust standard errors clustered by MSA are in parentheses. *|$p$| < .1; **|$p$| < .05; ***|$p$| < .01. Open in new tab Table 3 OLS regressions of ln(Broker fees) . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . Dep. var.: ln(Broker fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.09*** 0.05*** 0.04*** 0.14*** 0.08*** 0.05*** 0.04*** 0.03*** 0.03*** (0.01) (0.01) (0.00) (0.02) (0.01) (0.00) (0.01) (0.01) (0.01) Minority broker 0.02* 0.06*** 0.07*** (0.01) (0.01) (0.02) Minority borrower |$\times$| Minority broker 0.01 0.00 0.01 -0.08*** -0.06*** -0.05*** -0.14*** -0.08*** -0.08** (0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.02) (0.03) (0.03) Observations 234,413 175,670 175,660 213,575 158,543 158,535 160,936 114,303 114,293 Adjusted |$R^2$| .32 .51 .56 .33 .54 .58 .30 .51 .56 log loan amount Y Y Y Y Y Y Y Y Y Broker FE N Y Y N Y Y N Y Y Other controls N N Y N N Y N N Y Year-quarter FE N N Y N N Y N N Y MSA FE N N Y N N Y N N Y Minority/minority premium 0.11 0.06 0.05 0.06 0.02 0.00 -0.10 -0.05 -0.04 p-value .00 .00 .00 .00 .13 .79 .00 .08 .14 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . Dep. var.: ln(Broker fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.09*** 0.05*** 0.04*** 0.14*** 0.08*** 0.05*** 0.04*** 0.03*** 0.03*** (0.01) (0.01) (0.00) (0.02) (0.01) (0.00) (0.01) (0.01) (0.01) Minority broker 0.02* 0.06*** 0.07*** (0.01) (0.01) (0.02) Minority borrower |$\times$| Minority broker 0.01 0.00 0.01 -0.08*** -0.06*** -0.05*** -0.14*** -0.08*** -0.08** (0.01) (0.01) (0.01) (0.01) (0.02) (0.02) (0.02) (0.03) (0.03) Observations 234,413 175,670 175,660 213,575 158,543 158,535 160,936 114,303 114,293 Adjusted |$R^2$| .32 .51 .56 .33 .54 .58 .30 .51 .56 log loan amount Y Y Y Y Y Y Y Y Y Broker FE N Y Y N Y Y N Y Y Other controls N N Y N N Y N N Y Year-quarter FE N N Y N N Y N N Y MSA FE N N Y N N Y N N Y Minority/minority premium 0.11 0.06 0.05 0.06 0.02 0.00 -0.10 -0.05 -0.04 p-value .00 .00 .00 .00 .13 .79 .00 .08 .14 This table reports OLS estimates. The dependent variable is the natural log of broker fees. Each reported covariate is a dummy variable that equals one when true, and zero otherwise. Borrower controls include age, gender, marital status, income, credit (FICO) score, a subprime mortgage indicator (FICO|$<$|620), debt-to-income ratio, and employment status. Property-type controls include owner-occupancy versus investor status and single-family versus condominium or multiple unit structure. Loan controls include indicators for broker face-to-face interaction, purchase versus refinance, coborrowing, combined debt loan-to-value (CLTV), loan type (adjustable rate, fixed rate, interest only), interest rate spread, and borrower documentation. For the loan-to-value ratio, we use five CLTV categories: CLTV |$<$| 80%, 80% |$\le$| CLTV |$<$| 85%, 85% |$\le$| CLTV |$<$| 90%, 90% |$\le$| CLTV |$\le$| 95%, and CLTV |$\ge$| 95%. Area controls include the distance between borrower and broker, broker competition, area unemployment rate, regulatory environment, education, income, share of housing that is owner occupied, price-to-rent ratio, and county population share that is married, foreign born, Hispanic, Black, Asian or Pacific Islander, English speaking, and Hispanic speaking. The stand-alone broker race term (⁠|$L^M$|⁠) in Equation (1) is absorbed by the broker fixed effects. See Table A1 for a complete description of the variables. Robust standard errors clustered by MSA are in parentheses. *|$p$| < .1; **|$p$| < .05; ***|$p$| < .01. Open in new tab The results in column 1 highlight market-level disparities in the cost of mortgage credit. In column 2, we introduce mortgage broker fixed effects.26 The large increase in the adjusted |$R^2$| moving from column 1 to column 2 is consistent with individual mortgage brokers having considerable discretion in pricing. After accounting for broker heterogeneity, Hispanic borrowers pay a 5% (or |${\$}$|317) premium relative to white borrowers when obtaining a loan through the same white broker. The minority/minority premium is 6%, indicating that Hispanic borrowers pay 6% (or |${\$}$|335) more than whites when they obtain a loan through the same Hispanic broker, providing evidence that a within broker minority premium exists regardless of the race of the loan officer.27 The inclusion of broker fixed effects significantly reduces the magnitude of the minority premium from 9% to 5% for white brokers, and 11% to 6% for Hispanic brokers. The fact that broker fixed effects account for a large portion of the minority premium documented in column 1 suggests that Hispanic borrowers systematically select into high-fee brokers as hypothesized by Yezer, Phillips, and Trost (1994). This is closely related to, and consistent with, the findings of Bayer, Ferreira, and Ross (2017) that lender fixed effects reduce racial differences in the likelihood of receiving high-cost loans. Finally, in column 3, we add the full set of control variables and fixed effects. The magnitude of the minority premium slightly declines in column 3 to about 4% (or |${\$}$|277); however, it remains statistically significant at the 1% level. Columns 4 through 6 report the estimates using the Black/white (BW) subsample. Column 4 shows that Black borrowers that obtain a loan through a white mortgage broker pay 14% more, on average, than white borrowers that work with a white broker. This translates into a |${\$}$|837 premium, which is statistically and economically reduced when using a Black broker. The minority/minority premium is 6% (or |${\$}$|347) when the mortgage broker is Black. White borrowers obtaining a loan through a Black broker pay 6% (or |${\$}$|338) more than white borrowers working with a white broker. Consistent with the results in the HW subsample, the inclusion of broker fixed effects in column 5 significantly reduces the magnitude of the minority premium. After accounting for broker heterogeneity, Black borrowers pay an 8% (or |${\$}$|468) premium relative to white borrowers when obtaining a mortgage through the same white broker. A key departure from the HW results, however, is that we find no evidence that the same Black broker treats white and Black borrowers differently. To see this, notice that the minority/minority premium is economically small (2%, or |${\$}$|137) and not significantly different from zero in column 5. Although the magnitude of the minority borrower coefficient declines from column 5 to column 6, it is still statistically significant.28 A Black borrower pays 5% more than a comparable white borrower when obtaining a loan through the same white broker, which translates to a |${\$}$|294 premium. This estimate is somewhat smaller than the Black premium of |${\$}$|563 reported in Woodward (2008). We also note that the minority/minority premium is estimated as zero with a p-value of .79. Thus, we find no evidence that the same Black broker treats a Black borrower differently from a comparable white borrower. We repeat the analysis in columns 7 through 9 using the Asian/white (AW) subsample. Here again, we see in column 7 that minorities pay a premium (4%, or |${\$}$|236) relative to white borrowers when obtaining a loan through a white broker. However, the minority/minority premium results are quite different in the AW subsample relative to the HW and BW subsamples. The minority/minority premium of |$-10$|% indicates that Asian borrowers actually pay |${\$}$|529 less, on average, than their white counterparts when obtaining a loan through an Asian broker. Consistent with the HW and BW results, white borrowers on average pay more (7%, or |${\$}$|385) to obtain a loan through an Asian broker. Column 8 shows that Asian borrowers still pay a premium of 3% (or |${\$}$|165) relative to white borrowers who receive a loan through the same white broker. Also, we still see evidence that the same Asian broker treats Asian borrowers differently from white borrowers (minority/minority discount of 5%, or |${\$}$|286 less). A similar pattern appears in the saturated regression model in column 9, however, the minority/minority discount of 4% (or |${\$}$|241) is slightly outside traditional statistical significance thresholds. To summarize, we first note that minority borrowers pay a premium relative to white borrowers when they obtain loans through white brokers. This holds even after controlling for an extensive set of control variables and broker fixed effects, suggesting that minority borrowers receive different treatment than comparable white borrowers when obtaining loans through the same white broker. Second, a significant portion of the racial mortgage pricing disparity is explained by broker fixed effects, which suggests that minorities tend to systematically select into high-fee brokers. Third, some evidence indicates that white borrowers pay more on average when obtaining a loan through a minority broker. Fourth, minority/minority premiums vary across minority groups with Hispanic brokers charging a premium to Hispanic borrowers, while Black brokers do not appear to treat white and Black borrowers differently. For Asian brokers, evidence points to Asian borrowers receiving more favorable treatment relative to white borrowers.29 Finally, with respect to whether the statistically significant premiums are economically meaningful, we note that the magnitude of the observed premium paid by minority borrowers is near the threshold cutoff established in recent consent decrees agreed to by various financial institutions accused of disparate treatment of minorities in residential mortgage origination fees (New York, Office of the Attorney General, Civil Rights Bureau 2007, 2006, 2008).30 For example, using estimates from our saturated regression models, along with the 20-bp threshold cutoff agreed to by HCI Mortgage in its settlement with the New York Attorney General for determining incidences of disparate treatment, we identified 9,290 minority borrowers that paid premiums (conditional) above the level that indicates potential disparate treatment.31 Thus, although one might argue that the minority premiums in our study are small in magnitude, they are clearly within the range that drew the attention of regulators in recent years. 2.5 Robustness checks We perform a number of additional empirical exercises to confirm that our primary results are robust to different specifications and methodologies. We briefly describe these tests here, with more detailed discussion in the Internet Appendix. First, we examine whether our analysis is sensitive to the method used in inferring broker race. In the majority of our analysis, we infer the broker’s race using the MAP BIFSG classification scheme discussed in Sections 1.2. However, we also used different classification schemes based on the BIFSG scores (see Internet Appendix Section A.2) and report the baseline regression results for different threshold classification schemes in Table A15. Regardless of the classification threshold, results are similar in sign, significance, and magnitude to those reported in the saturated regression models of Table 3. We also used BIFSG scores directly, as opposed to a binary classification scheme, to estimate the effect of loan officer race on minority premiums. Again, this methodology produced estimates similar to those using the MAP BIFSG scheme (see Table A16). Thus, we conclude that the results are not sensitive to the method or criteria used to infer broker race. A second concern is whether the fee differential across borrower race reflects differences in broker effort required to generate a successful loan application. Brokers bear the risk that a loan application does not result in a funded loan (funding uncertainty). Generally, brokers are only compensated on applications that result in funded loans. If funding uncertainty comoves with race and broker fees, then the coefficient estimates in the previous section may be biased. To address this concern, we report the results for the estimation of a Heckman model that accounts for funding uncertainty in Internet Appendix section A.3 and find no evidence that the minority pricing premiums are driven by funding uncertainty. A third concern is that observed pricing differentials reflect borrower contract selection rather than disparate treatment. To fix ideas, suppose a white broker offers two distinct contracts to every applicant, regardless of borrower race: (1) a high front-end fee/low rate contract and (2) a low front-end fee/high rate contract. Also assume, consistent with existing evidence, that the broker earns greater revenue on the latter (Woodward and Hall 2012, Ambrose and Conklin 2014). If minority borrowers tend to select the contract that generates greater revenue for the broker, while white borrowers select into the lower revenue contract, then we should observe minority fee premiums even though there is no differential treatment by the mortgage broker.32 We investigate this possibility by examining the trade-off between front-end and back-end fees across borrowers’ and brokers’ race. The results, discussed in detail in Internet Appendix section A.4, point to differential treatment by brokers rather than borrower contract selection. In Internet Appendix section A.5, we formally examine the borrower’s choice of mortgage broker race to ensure that pricing disparities are not driven by borrower selection into broker race. Results using a propensity score matching technique, which are reported in Tables A18 and A19, are consistent with our primary results presented in Table 3. Finally, we examine whether the premiums uncovered in our main analysis are driven by language differences (see Internet Appendix section A.6). For example, perhaps Hispanic brokers charge a premium for providing bilingual services. Although we do not observe the language in which the mortgage application was taken, we do observe whether the borrower is a U.S. citizen, and use this information as a proxy for the use of a foreign language. We exclude non-U.S. citizens in Table A20 in the Internet Appendix and find that the results reported in Table 3 are unchanged. 2.6 Does credit risk explain minority premiums? In Section 2.3, we provided evidence that broker compensation is not directly tied to standard measures of credit risk. However, the potential remains that minority pricing premiums reflect broker compensation for credit risk. To investigate this possibility, we examine whether minority pricing premiums vary with an observable measure of credit quality (FICO score), after conditioning on the rich set of control variables and fixed effects used in our saturated regression models. We interact borrower minority status with credit score quartiles to see if the minority pricing premium varies across credit scores. Figure 3 displays the estimated minority fee premiums (and 95% confidence intervals) across credit quartiles. Each panel represents a separate regression. For example, the regression used to create the top-left graph includes loans originated to white and Hispanic borrowers by white brokers. The top-right panel, on the other hand, includes loans originated to Hispanic and white borrowers by Hispanic brokers. Figure 3 Open in new tabDownload slide Minority premium across credit scores This figure displays the minority premium estimates and 95% confidence intervals across FICO score quartiles separated by borrower and mortgage broker race. Each graph represents a separate regression model and corresponds to a single column in Table A21 in the Internet Appendix. Figure 3 Open in new tabDownload slide Minority premium across credit scores This figure displays the minority premium estimates and 95% confidence intervals across FICO score quartiles separated by borrower and mortgage broker race. Each graph represents a separate regression model and corresponds to a single column in Table A21 in the Internet Appendix. A clear pattern emerges from the panels in Figure 3. The minority pricing premium is generally larger at higher credit scores. For example, in the top-left panel, Hispanics with low credit scores (bin 1) pay a 2% premium relative to comparable white borrowers with low credit scores. However, this minority premium increases significantly to 7% when comparing high credit score Hispanics to comparable high credit score whites (bin 4). This pattern holds in 5 of the 6 panels in Figure 3. If minority fee premiums reflect additional credit risk, then we should observe higher default rates on loans to minorities. Additionally, given the positive relationship between credit score and minority premium documented in Figure 3, we would expect that the difference in default rates between minorities and whites also would be positively related to credit score. To test these two predictions, we estimate default regression models with the same controls as the fee regressions in Figure 3. Figure 4 plots the marginal effect of minority status on the likelihood of default across credit score quartiles. Across all panels, there is no evidence that minorities are more likely to default at any point in the credit score distribution. Thus, borrower credit risk does not explain the minority pricing premiums or the positive relationship between these premiums and credit scores. Ultimately, the question of why minority premiums are generally positively correlated with borrowers’ credit quality is a puzzle left to future research. Figure 4 Open in new tabDownload slide Minority effect on the probability of default across credit scores This figure displays the minority effect on mortgage default and 95% confidence intervals across FICO score quartiles separated by borrowers’ and mortgage brokers’ race. Each graph represents a separate regression model and corresponds to a single column in Table A22 of the Internet Appendix. Figure 4 Open in new tabDownload slide Minority effect on the probability of default across credit scores This figure displays the minority effect on mortgage default and 95% confidence intervals across FICO score quartiles separated by borrowers’ and mortgage brokers’ race. Each graph represents a separate regression model and corresponds to a single column in Table A22 of the Internet Appendix. 3. Policy Implications In the wake of the 2007–2008 financial crisis, regulators and lawmakers focused on curbing perceived abuses in mortgage lending, with mortgage brokers garnering significant attention. Although policy makers recognized the importance of brokers in helping consumers choose loans, they held concerns that the way brokers were being paid motivated them to steer borrowers into risky and expensive loan products (Consumer Financial Protection Bureau 2014a). In response to these concerns, the Board of Governors of the Federal Reserve System (the Board) proposed rules in 2009 governing mortgage broker compensation that were later incorporated into the Dodd-Frank Act. The Consumer Financial Protection Bureau (CFPB) subsequently issued regulations (effective January 1, 2014) reconciling the Board’s broker compensation rules (Kider and Kamensky 2015). Thus, in this section, we analyze how the new broker compensation rules could have affected broker fees and borrower access to credit. 3.1 Dual compensation In the run-up to the financial crisis, mortgage brokers could be compensated through direct fees from the borrower, rebates (YSP) from the lender, or a combination of the two (dual compensation). In response to the crisis, regulators proposed a rule that would ban dual compensation.33 By eliminating (or restricting) dual compensation, policy makers aimed to increase transparency in the loan origination process. Supporters of this restriction argue that dual compensation leads to borrower confusion and suboptimal shopping behavior. Indeed, two recent studies document that borrowers pay significantly higher fees on dual compensation loans (Woodward and Hall 2012, Ambrose and Conklin 2014). Proponents of dual compensation, on the other hand, argue that it provides valuable flexibility for consumers by allowing borrowers to choose lower out-of-pocket fees in exchange for a higher interest rate. At this time, dual compensation is not prohibited, but the CFPB has indicated an interest in further investigating dual compensation to determine how it affects borrower confusion and ultimately mortgage choice.34 Thus, the results of this investigation will aid the CFPB in determining whether it should proceed with its initial proposal to ban dual compensation. We reestimate our baseline regressions with three separate subsamples within each set of HW, BW, and AW loans: (1) loans where brokers received upfront fees and compensation from the lender (yield spread premiums), (2) loans where brokers were entirely compensated through upfront fees, and (3) loans where all broker compensation was in the form of yield spread premiums. Columns 1, 2, and 3 show the regression results in the HW subsample for dual compensation loans, front-end-fee-only loans, and back-end (YSP) only loans, respectively. Columns 3 though 6 repeat the exercise in the BW subsample, while columns 7 through 9 do the same in the AW subsample. Note that in all three subsamples, approximately 70% of the loans have both front- and back-end (Dual) broker compensation. Thus, we are careful not to interpret our results too strongly as we move across pricing schemes because sample sizes (and power) are significantly reduced when we focus on front-end-fee-only loans (Front) or back-end-fee-only loans (Back). Column 1 of Table 4 reports coefficient estimates on dual compensation loans for the HW subsample. Minority borrowers pay a within-broker premium regardless of the race of the loan officer. The results are similar on the front-end-fee-only loans in column 2. In fact, none of the minority coefficient estimates in column 2, including the minority/minority premium, is statistically different from those reported in column 1. Turning to the back-end-fee-only loans in column 3, the minority borrower coefficient adjusts to 0.01, which is statistically different from the coefficient in column 1. This provides some evidence that Hispanic pricing gaps are less prevalent in back-end-fee-only loans. However, none of the other minority coefficients in column 3, including the minority/minority premium, is different from those in column 1. Table 4 Proposed ban on dual compensation . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . . Dual . Front . Back . Dual . Front . Back . Dual . Front . Back . Dep. var.: ln(Broker Fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.04*** 0.03*** 0.01 0.04*** 0.05*** 0.04** 0.03*** 0.01 -0.02 (0.01) (0.01) (0.02) (0.00) (0.01) (0.02) (0.01) (0.03) (0.03) Minority borrower |$\times$| Minority broker 0.00 0.02 0.05 -0.04** -0.04 0.05 -0.06* -0.07* -0.09 (0.01) (0.02) (0.05) (0.02) (0.03) (0.08) (0.04) (0.04) (0.07) Observations 106,147 39,697 7,981 97,976 32,998 7,209 70,766 21,108 6,191 Adjusted |$R^2$| .63 .59 .69 .65 .63 .69 .62 .62 .68 log loan amount Y Y Y Y Y Y Y Y Y Broker FE Y Y Y Y Y Y Y Y Y Other controls Y Y Y Y Y Y Y Y Y Year-Quarter FE Y Y Y Y Y Y Y Y Y MSA FE Y Y Y Y Y Y Y Y Y Minority/minority premium 0.04 0.05 0.06 0.00 0.01 0.09 -0.03 -0.06 -0.12 p-value .00 .00 .18 .87 .56 .26 .35 .19 .10 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . . Dual . Front . Back . Dual . Front . Back . Dual . Front . Back . Dep. var.: ln(Broker Fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.04*** 0.03*** 0.01 0.04*** 0.05*** 0.04** 0.03*** 0.01 -0.02 (0.01) (0.01) (0.02) (0.00) (0.01) (0.02) (0.01) (0.03) (0.03) Minority borrower |$\times$| Minority broker 0.00 0.02 0.05 -0.04** -0.04 0.05 -0.06* -0.07* -0.09 (0.01) (0.02) (0.05) (0.02) (0.03) (0.08) (0.04) (0.04) (0.07) Observations 106,147 39,697 7,981 97,976 32,998 7,209 70,766 21,108 6,191 Adjusted |$R^2$| .63 .59 .69 .65 .63 .69 .62 .62 .68 log loan amount Y Y Y Y Y Y Y Y Y Broker FE Y Y Y Y Y Y Y Y Y Other controls Y Y Y Y Y Y Y Y Y Year-Quarter FE Y Y Y Y Y Y Y Y Y MSA FE Y Y Y Y Y Y Y Y Y Minority/minority premium 0.04 0.05 0.06 0.00 0.01 0.09 -0.03 -0.06 -0.12 p-value .00 .00 .18 .87 .56 .26 .35 .19 .10 This table reports OLS estimates. The dependent variable is the natural log of broker fees. Columns labeled Dual include only loans that have both front- and back-end fees. Columns labeled Front include loans that only have front-end fees. Columns labeled Back include loans that only have back-end fees. Each reported covariate is a dummy variable that equals one when true, and zero otherwise. Each column restricts the observations to the subgroup specified in the header. Borrower controls include age, gender, marital status, income, credit (FICO) score, a subprime mortgage indicator (FICO|$<$|620), debt-to-income ratio, and employment status. Property-type controls include owner-occupancy versus investor status and single-family versus condominium or multiple unit structure. Loan controls include indicators for broker face-to-face interaction, purchase versus refinance, coborrowing, combined debt loan-to-value, loan type (adjustable rate, fixed rate, interest only), interest rate spread, and borrower documentation. Area controls include distance between borrower and broker, broker competition, area unemployment rate, regulatory environment, education, income, share of housing that is owner occupied, price-to-rent ratio, and county population share that is married, foreign born, Hispanic, Black, Asian or Pacific Islander, English speaking, and Hispanic speaking. See Table A1 for a complete description of the variables. Robust standard errors clustered by MSA are in parentheses. *|$p$| < .1; **|$p$| < .05; ***|$p$| < .01. Open in new tab Table 4 Proposed ban on dual compensation . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . . Dual . Front . Back . Dual . Front . Back . Dual . Front . Back . Dep. var.: ln(Broker Fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.04*** 0.03*** 0.01 0.04*** 0.05*** 0.04** 0.03*** 0.01 -0.02 (0.01) (0.01) (0.02) (0.00) (0.01) (0.02) (0.01) (0.03) (0.03) Minority borrower |$\times$| Minority broker 0.00 0.02 0.05 -0.04** -0.04 0.05 -0.06* -0.07* -0.09 (0.01) (0.02) (0.05) (0.02) (0.03) (0.08) (0.04) (0.04) (0.07) Observations 106,147 39,697 7,981 97,976 32,998 7,209 70,766 21,108 6,191 Adjusted |$R^2$| .63 .59 .69 .65 .63 .69 .62 .62 .68 log loan amount Y Y Y Y Y Y Y Y Y Broker FE Y Y Y Y Y Y Y Y Y Other controls Y Y Y Y Y Y Y Y Y Year-Quarter FE Y Y Y Y Y Y Y Y Y MSA FE Y Y Y Y Y Y Y Y Y Minority/minority premium 0.04 0.05 0.06 0.00 0.01 0.09 -0.03 -0.06 -0.12 p-value .00 .00 .18 .87 .56 .26 .35 .19 .10 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . (9) . . Dual . Front . Back . Dual . Front . Back . Dual . Front . Back . Dep. var.: ln(Broker Fees) . HW . HW . HW . BW . BW . BW . AW . AW . AW . Minority borrower 0.04*** 0.03*** 0.01 0.04*** 0.05*** 0.04** 0.03*** 0.01 -0.02 (0.01) (0.01) (0.02) (0.00) (0.01) (0.02) (0.01) (0.03) (0.03) Minority borrower |$\times$| Minority broker 0.00 0.02 0.05 -0.04** -0.04 0.05 -0.06* -0.07* -0.09 (0.01) (0.02) (0.05) (0.02) (0.03) (0.08) (0.04) (0.04) (0.07) Observations 106,147 39,697 7,981 97,976 32,998 7,209 70,766 21,108 6,191 Adjusted |$R^2$| .63 .59 .69 .65 .63 .69 .62 .62 .68 log loan amount Y Y Y Y Y Y Y Y Y Broker FE Y Y Y Y Y Y Y Y Y Other controls Y Y Y Y Y Y Y Y Y Year-Quarter FE Y Y Y Y Y Y Y Y Y MSA FE Y Y Y Y Y Y Y Y Y Minority/minority premium 0.04 0.05 0.06 0.00 0.01 0.09 -0.03 -0.06 -0.12 p-value .00 .00 .18 .87 .56 .26 .35 .19 .10 This table reports OLS estimates. The dependent variable is the natural log of broker fees. Columns labeled Dual include only loans that have both front- and back-end fees. Columns labeled Front include loans that only have front-end fees. Columns labeled Back include loans that only have back-end fees. Each reported covariate is a dummy variable that equals one when true, and zero otherwise. Each column restricts the observations to the subgroup specified in the header. Borrower controls include age, gender, marital status, income, credit (FICO) score, a subprime mortgage indicator (FICO|$<$|620), debt-to-income ratio, and employment status. Property-type controls include owner-occupancy versus investor status and single-family versus condominium or multiple unit structure. Loan controls include indicators for broker face-to-face interaction, purchase versus refinance, coborrowing, combined debt loan-to-value, loan type (adjustable rate, fixed rate, interest only), interest rate spread, and borrower documentation. Area controls include distance between borrower and broker, broker competition, area unemployment rate, regulatory environment, education, income, share of housing that is owner occupied, price-to-rent ratio, and county population share that is married, foreign born, Hispanic, Black, Asian or Pacific Islander, English speaking, and Hispanic speaking. See Table A1 for a complete description of the variables. Robust standard errors clustered by MSA are in parentheses. *|$p$| < .1; **|$p$| < .05; ***|$p$| < .01. Open in new tab In the BW subsample, none of the minority coefficients in column 5 or 6 is significantly different from those in column 4. Regardless of the fee structure, Blacks pay a premium with the same white broker, but pay no premium when the broker is Black. In the AW subsample, minorities pay a premium on dual compensation loans when the broker is white (column 7); however, they receive a discount within the same API broker on back-end-fee-only loans (column 9). None of the minority coefficients in column 8 is significantly different from those in column 7, and only the minority borrower coefficient in column 9 is statistically different from the coefficient in column 7. Although we find some evidence that relative to dual compensation loans, Hispanic and API premiums in back-end-fee-only loans arranged by white brokers decrease slightly, overall Table 4 suggests that the elimination of dual compensation to increase price transparency, per se, is unlikely to completely eliminate racial price disparities. 3.2 Broker costs, fee caps, and credit rationing In the postcrisis period, a residential mortgage loan can be categorized as “qualified” or “nonqualified.” Broadly speaking, the CFPB deems a loan as a Qualified Mortgage (QM) if it has features that make it affordable and safe to the typical borrower.35 For the loan to qualify for QM status, the lender must follow certain underwriting criteria and the loan must exclude prohibited contract features that are deemed risky (e.g., negative amortization, interest-only payments, balloon loans, terms greater than 30 years).36 Among the restrictions, loan originator points and fees are capped at 3% of the loan amount (see Section 1026.43(e) of Regulation Z).37 The fee caps are meant to make loans affordable and to reduce broker discretion in pricing. But, the caps may also have an unintended consequence of causing mortgage brokers to withdraw from providing services to loan applicants that require higher levels of effort or service.38 To consider how fees vary by loan applicant, we model the broker’s revenue (fees) as follows: $$\begin{equation}\label{profit} P_{ijmt} = k_{ijmt} + \pi_{ijmt}, \end{equation}$$(2) where |$P_{ijmt}$| is the total dollar amount of revenue generated from loan applicant |$i$| (including fees and yield spread premium), by broker |$j$|⁠, in market |$m$|⁠, at time |$t$|⁠. |$k_{ijmt}$| is the broker’s production cost for originating the loan and |$\pi_{ijmt}$| is the excess profit generated on the loan applicant. The implicit assumption is that brokers charge borrowers fees to cover their production costs. Although there are multiple sources of costs (e.g., marketing, overhead, and so on), a large portion of this cost compensates the mortgage broker for time, effort and search costs. In a perfectly competitive market, |$\pi_{ijmt}$| would be driven to zero. However, the mortgage market is not perfectly competitive.39 Thus, the model allows both cost and excess profit to vary across individuals (e.g., loan and borrower characteristics), brokers, markets, and time to address the heterogeneity in the provision of brokerage services. Empirically, we estimate the production cost for each loan applicant (⁠|$\hat{k}_{ijmt}$|⁠) and then compare it to the fee cap imposed by Regulation Z. If the estimated production cost exceeds the fee cap, then we assume that the borrower would be credit rationed under current regulations because the broker cannot recover the associated costs. We follow an approach similar to Berndt, Hollifield, and Sandås (2017) by fitting a quantile regression model of broker fees:40 $$\begin{equation} q_\alpha(Fees|\mathbf{\Gamma}) = \mathbf{\Gamma}'\mathbf{\beta_\alpha},\label{eq:q_regression} \end{equation}$$(3) where |$\alpha$| is the quantile of interest and |$\mathbf{\Gamma}$| includes the conditioning variables from Equation (1).41 The selected value of |$\alpha$| fits a regression line where |$(1-\alpha)$| of the observations lie above the regression line. The predicted values from this regression provide an estimate of the minimum (conditional) fee required for the broker to originate a loan, which is an estimate of the loan production cost, |$\hat{k}_{ijmt}$| (Liu, Laporte, and Ferguson 2008). If this cost estimate exceeds the fee caps imposed on QM loans, then we assume that the borrower would be unable to obtain a loan. Put differently, the borrower would be credit rationed under current regulations. To generate a baseline of possible credit rationing, we compare the actual fees observed to the fee caps outlined above to determine which borrowers would be credit rationed under current regulations assuming that the actual fees equal loan production costs and that brokers earned zero excess profit. The zero profit condition is, of course, a heroic assumption, which we relax after reporting our baseline results. The top-left panel in Figures 4, 5, and 6 in the Internet Appendix report the results for the HW, BW, and AW loans, respectively. We note that nearly half of Hispanic borrowers that worked with Hispanic brokers and about 44% of Hispanic borrowers that obtained a loan from a white broker would have been credit rationed. Meanwhile, around 40% of white borrowers would be credit rationed regardless of the broker’s race. In addition, more than half of Black borrowers would have been rationed, whereas Asian and Pacific Islander borrowers appear less likely to be exposed to rationing risk, especially Asians who obtain loans from Asian brokers. Overall, these results provide an upper bound on credit rationing created by the current regulations and suggest that racial credit rationing disparities would exist under a zero profit assumption.42 Next, we turn to our quantile regression results reported in Table 5. We present results for different values of |$\alpha$| due to the mechanical relationship between the choice of |$\alpha$| and the fraction of borrowers that are classified as credit rationed in our data; larger values of |$\alpha$| shift the estimated cost function upwards. Columns 1, 2, and 3 provide the point estimates using the 10th, 20th, and 30th quantiles, respectively. Panels A, B, and C, show the quantile estimates for the HW, BW, and AW subsamples. The minority status dummy variables and interaction term are strongly significant for each quantile regression. The estimates in column 1 of panel A, for example, indicate that Hispanic borrowers pay |${\$}$|305 more than whites at the 10th quantile. At higher quantiles, the racial price disparities are greater. Table 5 Quantile regressions of broker fees . (1) . (2) . (3) . Dep. var.: Broker fees|$_\alpha$| . Quantile 10 . Quantile 20 . Quantile 30 . A. HW Minority borrower 304.96*** 359.34*** 411.31*** (21.25) (16.74) (16.64) Minority broker 91.53*** 174.05*** 186.13*** (32.98) (25.97) (25.83) Minority borrower |$\times$| Minority broker 159.99*** 139.61*** 155.54*** (40.32) (31.75) (31.57) Observations 234,413 234,413 234,413 B. BW Minority borrower 236.27*** 292.95*** 352.03*** (15.98) (14.39) (13.89) Minority broker 124.16*** 176.30*** 239.73*** (37.95) (34.17) (32.99) Minority borrower |$\times$| Minority broker -99.08** -116.93*** -177.77*** (44.91) (40.44) (39.04) Observations 213,575 213,575 213,575 C. AW Minority borrower 308.79*** 392.76*** 447.51*** (37.58) (31.82) (30.34) Minority broker 271.86*** 318.00*** 346.93*** (49.62) (42.03) (40.07) Minority borrower |$\times$| Minority broker -458.24*** -471.99*** -454.88*** (70.71) (59.89) (57.10) Observations 160,936 160,936 160,936 Borrower controls Y Y Y Property type N N N Loan controls Y Y Y Area controls N N N Year-quarter fixed effects Y Y Y State fixed effects Y Y Y . (1) . (2) . (3) . Dep. var.: Broker fees|$_\alpha$| . Quantile 10 . Quantile 20 . Quantile 30 . A. HW Minority borrower 304.96*** 359.34*** 411.31*** (21.25) (16.74) (16.64) Minority broker 91.53*** 174.05*** 186.13*** (32.98) (25.97) (25.83) Minority borrower |$\times$| Minority broker 159.99*** 139.61*** 155.54*** (40.32) (31.75) (31.57) Observations 234,413 234,413 234,413 B. BW Minority borrower 236.27*** 292.95*** 352.03*** (15.98) (14.39) (13.89) Minority broker 124.16*** 176.30*** 239.73*** (37.95) (34.17) (32.99) Minority borrower |$\times$| Minority broker -99.08** -116.93*** -177.77*** (44.91) (40.44) (39.04) Observations 213,575 213,575 213,575 C. AW Minority borrower 308.79*** 392.76*** 447.51*** (37.58) (31.82) (30.34) Minority broker 271.86*** 318.00*** 346.93*** (49.62) (42.03) (40.07) Minority borrower |$\times$| Minority broker -458.24*** -471.99*** -454.88*** (70.71) (59.89) (57.10) Observations 160,936 160,936 160,936 Borrower controls Y Y Y Property type N N N Loan controls Y Y Y Area controls N N N Year-quarter fixed effects Y Y Y State fixed effects Y Y Y This table reports the coefficient estimates at the 10th, 20th, and 30th quantiles. The dependent variable is the nominal value of broker fees. Each reported covariate is a dummy variable that equals one when true, and zero otherwise. Borrower controls include age, gender, marital status, income, credit (FICO) score, debt-to-income ratio, and employment status. Loan controls include indicators for broker face-to-face interaction, purchase versus refinance, coborrowing, combined debt loan-to-value, loan type (adjustable rate, fixed rate, interest only), interest rate spread, and borrower documentation. See Table A1 for a complete description of the variables. Standard errors are in parentheses. *|$p$| < .1; **|$p$| < .05; ***|$p$| < .01. Open in new tab Table 5 Quantile regressions of broker fees . (1) . (2) . (3) . Dep. var.: Broker fees|$_\alpha$| . Quantile 10 . Quantile 20 . Quantile 30 . A. HW Minority borrower 304.96*** 359.34*** 411.31*** (21.25) (16.74) (16.64) Minority broker 91.53*** 174.05*** 186.13*** (32.98) (25.97) (25.83) Minority borrower |$\times$| Minority broker 159.99*** 139.61*** 155.54*** (40.32) (31.75) (31.57) Observations 234,413 234,413 234,413 B. BW Minority borrower 236.27*** 292.95*** 352.03*** (15.98) (14.39) (13.89) Minority broker 124.16*** 176.30*** 239.73*** (37.95) (34.17) (32.99) Minority borrower |$\times$| Minority broker -99.08** -116.93*** -177.77*** (44.91) (40.44) (39.04) Observations 213,575 213,575 213,575 C. AW Minority borrower 308.79*** 392.76*** 447.51*** (37.58) (31.82) (30.34) Minority broker 271.86*** 318.00*** 346.93*** (49.62) (42.03) (40.07) Minority borrower |$\times$| Minority broker -458.24*** -471.99*** -454.88*** (70.71) (59.89) (57.10) Observations 160,936 160,936 160,936 Borrower controls Y Y Y Property type N N N Loan controls Y Y Y Area controls N N N Year-quarter fixed effects Y Y Y State fixed effects Y Y Y . (1) . (2) . (3) . Dep. var.: Broker fees|$_\alpha$| . Quantile 10 . Quantile 20 . Quantile 30 . A. HW Minority borrower 304.96*** 359.34*** 411.31*** (21.25) (16.74) (16.64) Minority broker 91.53*** 174.05*** 186.13*** (32.98) (25.97) (25.83) Minority borrower |$\times$| Minority broker 159.99*** 139.61*** 155.54*** (40.32) (31.75) (31.57) Observations 234,413 234,413 234,413 B. BW Minority borrower 236.27*** 292.95*** 352.03*** (15.98) (14.39) (13.89) Minority broker 124.16*** 176.30*** 239.73*** (37.95) (34.17) (32.99) Minority borrower |$\times$| Minority broker -99.08** -116.93*** -177.77*** (44.91) (40.44) (39.04) Observations 213,575 213,575 213,575 C. AW Minority borrower 308.79*** 392.76*** 447.51*** (37.58) (31.82) (30.34) Minority broker 271.86*** 318.00*** 346.93*** (49.62) (42.03) (40.07) Minority borrower |$\times$| Minority broker -458.24*** -471.99*** -454.88*** (70.71) (59.89) (57.10) Observations 160,936 160,936 160,936 Borrower controls Y Y Y Property type N N N Loan controls Y Y Y Area controls N N N Year-quarter fixed effects Y Y Y State fixed effects Y Y Y This table reports the coefficient estimates at the 10th, 20th, and 30th quantiles. The dependent variable is the nominal value of broker fees. Each reported covariate is a dummy variable that equals one when true, and zero otherwise. Borrower controls include age, gender, marital status, income, credit (FICO) score, debt-to-income ratio, and employment status. Loan controls include indicators for broker face-to-face interaction, purchase versus refinance, coborrowing, combined debt loan-to-value, loan type (adjustable rate, fixed rate, interest only), interest rate spread, and borrower documentation. See Table A1 for a complete description of the variables. Standard errors are in parentheses. *|$p$| < .1; **|$p$| < .05; ***|$p$| < .01. Open in new tab Using the point estimates from Table 5, we infer the cost (conditional) for each loan applicant, and the proportion of borrowers that would be at risk of credit rationing under current regulations. For example, Figure 5 reports the results for the 30th quantile (Figures 4 through 6 in the Internet Appendix report comparisons for the 10th, 20th, and 30th quantiles). In general, the HW and BW subsamples show that Hispanic and Black borrowers represent the groups most at risk of losing access to credit, with 20% of Hispanic and 29% of Black borrowers who obtained financing from white brokers at risk of credit rationing. In contrast, 17% and 14% of white and Asian and Pacific Islander borrowers who obtain financing from white brokers would encounter credit rationing risk, respectively. Thus, our results indicate that racial disparities in credit rationing risk are likely to exist after imposition of the regulation. In particular, Hispanic and Black borrowers—who account for 45% of the loans in our sample—encounter the highest risk of credit rationing regardless of how we estimate loan production costs for each loan applicant. Figure 5 Open in new tabDownload slide Borrowers at risk of credit rationing This figure displays the proportion of borrowers at risk of credit rationing. The underlying sample comprises only Hispanic and white borrowers or brokers (HW), only Black and white borrowers or brokers (BW), or only Asian and white borrowers or brokers (AW). Each graph replaces the actual broker fees with predicted fees from the regression at the 30th quantile in Table 5. L(M)B(M) stands for minority broker, minority borrower; L(M)B(W) stands for minority broker, white borrower; L(W)B(M) stands for white broker, minority borrower; and L(W)B(W) stands for white broker, white borrower. Figure 5 Open in new tabDownload slide Borrowers at risk of credit rationing This figure displays the proportion of borrowers at risk of credit rationing. The underlying sample comprises only Hispanic and white borrowers or brokers (HW), only Black and white borrowers or brokers (BW), or only Asian and white borrowers or brokers (AW). Each graph replaces the actual broker fees with predicted fees from the regression at the 30th quantile in Table 5. L(M)B(M) stands for minority broker, minority borrower; L(M)B(W) stands for minority broker, white borrower; L(W)B(M) stands for white broker, minority borrower; and L(W)B(W) stands for white broker, white borrower. 3.3 Subprime resurgence The qualified mortgage (QM) designation requires that loan origination fees adhere to the caps outlined in the previous section. The benefit of these caps is that they may reduce the scope for brokers to price discriminate; however, they also increase the likelihood of credit rationing. But borrowers that would be credit rationed in the QM market still may be able to obtain mortgage credit in the non-QM market because non-QM loans are not subject to the same fee limitations. In the non-QM market, mortgage brokers retain considerable discretion over mortgage pricing. These mortgages are typically extended to borrowers with a blemished credit history. In other words, non-QM loans are subprime mortgages. But because the term “subprime” carries a negative connotation in the wake of the recent financial crisis, the industry has rebranded these loans as “nonprime” mortgages. Non-QM loans carry significantly higher interest rates and down payment requirements relative to their QM counterparts, and thus, even if high loan production costs do not fully eliminate access to credit, a steep penalty exists for obtaining a mortgage in the nonprime market. Although the current mortgage lending environment is dominated by QM loans, non-QM loans are gaining market share. In the years following the financial crisis, subprime mortgage lending virtually disappeared. However, in 2017, |${\$}$|4.1 billion in securities backed by nonprime mortgages was issued, and the first quarter of 2018 saw |${\$}$|1.3 billion in nonprime issuances, a sum more than double the amount issued in the same quarter a year earlier (McLannahan and Rennison 2018). The growing demand for nonprime securities from secondary mortgage market investors is clear. Existing research reveals that subprime lending, whose primary recipients have blemished credit histories, is concentrated in minority neighborhoods, and subprime mortgages are disproportionately originated to minority borrowers (Mayer and Pence 2009, Faber 2013, Calem, Gillen, and Wachter 2004, Pennington-Cross, Yezer, and Nichols 2000). This is driven, at least in part, by the fact that the two largest minority groups (Hispanics and Blacks) have lower credit scores than whites, on average (Board of Governors of the Federal Reserve System 2007). Moving forward, these differences in credit scores across racial groups suggest that nonprime lending will continue to disproportionately focus on minority borrowers. Given that nonprime lending (a) is more likely to be a source of credit for minorities, (b) is gaining market share, and (c) is not subject to QM fee caps that limit broker pricing discretion, the racial pricing disparities we identified in the precrisis period are likely to persist despite recent regulatory changes.43 3.4 Summary Although the new rules on broker fees limit how mortgage brokers may collect fees on loan originations, potentially reducing pricing disparities, our analysis indicates that these rules alone are unlikely to eradicate minority premiums and instead may place borrowers at risk of credit rationing. However, even if borrowers can avoid credit rationing by obtaining credit in the nonprime market, significant pricing disparities are likely to exist in that market because brokers have considerable pricing discretion on non-QM loans. 4. Conclusion This paper uses a nationwide data set of loans originated between 2003 and mid-2007 by over 124,000 unique mortgage brokers to test for differential treatment in financial contracts by examining both front- and back-end fees that borrowers pay at origination. Our unique data set also allows us to infer the broker’s race, providing the opportunity to observe the race of both sides to the contract. We find that minority pricing premiums exist when the mortgage broker is white after conditioning on an extensive set of borrower, loan, property, and area characteristics. Premiums are smaller, but remain significant, after including individual broker fixed effects, which indicates that a minority borrower pays more than a comparable white borrower when obtaining a loan from the same mortgage broker. The results also suggest that minorities tend to select into high-fee brokers. Importantly, we find that the premium a minority pays depends critically on the race of the mortgage broker. Hispanic brokers charge Hispanic borrowers a premium relative to comparable white borrowers, but observably similar white and Black borrowers pay the same fees when obtaining a loan from the same Black broker. We also find some evidence that Asian borrowers pay lower fees than comparable white borrowers when obtaining loans from Asian brokers. The Dodd-Frank Act of 2010 enacted a number of regulations designed to curb perceived abuses in the mortgage industry. These regulations severely restrict broker discretion in setting mortgage origination fees. For example, the yield spread premium rebate can no longer be paid unless the borrower also reviews a similar loan without it, and mortgage brokers are limited in the ability to collect compensation on the basis of the loan terms other than the loan balance at origination. We document a possible negative or adverse effect of these regulations in the form of potential credit rationing. Assuming that loan fees reflect broker production costs, we estimate that the restrictions on broker fees could result in a large percentage of minority borrowers being at risk of credit rationing. As a result, our study fills the need articulated by Campbell et al. (2011) for a rigorous analysis of the effectiveness of regulatory interventions following the Great Recession, and it outlines the trade-offs policy makers must contend with when designing new regulations. Finally, we offer some observations regarding the interpretation of our results in light of the two leading economic theories of discrimination: statistical and taste based. Since we find that fees charged to borrowers are not independent of broker race, our results are counter to the hypothesis of statistical discrimination as outlined by Antonovics and Knight (2009). Alternatively, a taste-based theory of discrimination predicts that brokers charge premiums to borrowers that do not share the same race. We find some evidence consistent with this taste-based prediction: white and Asian brokers charge premiums to minority and white borrowers, respectively. We also find that Hispanic brokers charge premiums to Hispanic borrowers, which under certain assumptions, could be consistent with taste-based preferences. However, we find no evidence of taste-based discrimination among Black brokers. When viewed through the lens of law, we note that our results offer some evidence of disparate treatment (using race as a factor in pricing), but we caution that we are unable to entirely rule out disparate impact (application of a race-neutral policy that disproportionately affects minorities.) But we reiterate that regardless of whether one ascribes the results as being associated with disparate treatment or disparate impact, the observed pricing premiums would likely be deemed as illegal discrimination by the courts. Authors have furnished an Internet Appendix, which is available on the Oxford University Press Web site next to the link to the final published paper online. Acknowledgement We thank Philip Strahan (the editor) and three anonymous referees. We also thank Sumit Agarwal, Wayne Archer, Patrick Bayer, Neil Bhuta, Larry Cordell, Ed Coulson, Stuart Gabriel, Kris Gerardi, Andra Ghent, Aurel Hizmo, Lawrence Katz, David Ling, Rich Martin, Gonzalo Maturana, Andy Naranjo, Ruchi Singh, Will Strange, Joe Tracy, Susan Woodward, Tyler Yang, Tony Yezer, and the seminar participants at Baruch College, the Federal Reserve Bank of Philadelphia, University of Florida, Georgia State University, Laval University, the National University of Singapore, the Pennsylvania State University, University College London, University of Georgia, the 2018 Singapore Management University Conference on Urban and Regional Economics, the 2018 AREUEA National Meeting, the 2018 American Real Estate Society Meeting, and the 2020 Allied Social Science Association meeting for their helpful comments and suggestions. We thank the Penn State Institute for Real Estate Studies for providing access to the New Century Mortgage database. James Conklin received research support from the University of Georgia’s Terry-Sanford research award. We received the 2018 ARES Best Paper Award in Real Estate Finance, sponsored by Real Capital Analytics. We affirm that we have no material financial interests related to this research. Supplementary data can be found on The Review of Financial Studies web site. Footnotes 1 Yezer (2006) and Ladd (1998) review the earlier literature about mortgage discrimination, and Courchane and Ross (2019) summarize more recent research and court cases covering discrimination and disparate treatment. However, using loan denial rates and ex post loan performance to measure potential discrimination is controversial since race may be correlated with unobservable borrower credit risk (Horne 1997, Ross and Yinger 2002, Brueckner 1996,Ross 1996, Yinger 1996). 2 Mortgage brokers receive compensation two ways: (1) origination fees paid by borrowers and (2) lender rebates (yield spread premiums). The former refers to the numerous potential expenses, such as points, application fees, and underwriting fees, and other miscellaneous fees borrowers pay the broker at mortgage closing. The latter refers to the rebate the lender pays the broker for negotiating a contract interest rate above the minimum market rate the borrower qualifies to receive. Because borrowers enter the market infrequently and mortgages are heterogeneous products, consumers are at an informational disadvantage relative to market specialists (mortgage brokers) that have considerable discretion over pricing. Thus, mortgage markets are conducive to price dispersion. 3 Although brokers may be exposed to reputational risk for delivering low-quality (high credit risk) loans, our study uses mortgages originated during an economic expansionary period characterized by rising house prices and low early termination events (a key trigger for lender mortgage rescission). As a result, during this period, broker concerns about reputational risks arising from credit risks are most likely minimal. 4 The Dodd-Frank Act is available online at https://www.congress.gov/bill/111th-congress/house-bill/4173/text. See Section 1403, Prohibition on Steering Incentives, which amends Section 129B of the Truth in Lending Act. 5 If Hispanic brokers are relatively high income and well educated compared to their customers, or their customers are more recent immigrants, or substantial heterogeneity in race exists within the Hispanic ethnicity, then Hispanic brokers could exhibit preference-based discrimination against borrowers of the same ethnicity or even race. We thank an anonymous referee for bringing our attention to this possibility. 6 New Century began originating loans in 1997 and stopped in March 2007, when it filed for bankruptcy. We restrict the sample to the period after 2002 because of incomplete data on brokers’ surnames prior to 2003. 7 During the application stage, applicants (or loan officers) fill out a form that asks applicants to identify their race and ethnicity. The ethnicity question allows the applicant to self-identify as either “Hispanic or Latino” or “Not Hispanic or Latino,” while the race question allows the applicant to self-identify as “American Indian or Alaska Native,” “Asian,” “Black,” “Native Hawaiian or Other Pacific Islander,” or “white.” The categories follow the classification standards of federal data on race and ethnicity (62 Fed. Reg. 131, July 9, 1997). Hence, to be consistent with the federal classification standards, we categorize borrowers as American Indian or Alaska Native, Asian or Pacific Islander, Black, Hispanic, or white. The Hispanic category applies to all borrowers who self-identify as Hispanic or Latino. The other categories apply to borrowers who self-identify as the corresponding race, but not Hispanic or Latino. 8 In cases with multiple borrowers on the loan, the income represents the combined income of these borrowers. Since approximately 41% of the loans are low-doc (stated income) loans, the average income reported in the data is likely inflated (Ambrose, Conklin, and Yoshida 2016). 9 Here, we report the exponential of the average log loan amount (⁠|$\exp^{12.06}=172,800$|⁠). Table A2 in the Internet Appendix reports the average loan amount of |${\$}$|206,000. 10 The name matching method employed in the Guardians Association case was devised in a study conducted by the Rand Institute that inferred the racial profile of subjects in the case by comparing their names to 8,000 surnames obtained from the U.S. Census Bureau (Guardians Ass’n of N.Y.C. Police Dep’t v. Civil Serv. Comm’n 1977, ¶ 33). Guardians Ass’n of N.Y.C. Police Dep’t v. Civil Serv. Comm’n (1977) was ultimately upheld on appeal (Guardians Asso. of N.Y.C. Police Dep’t, Inc. v Civil Serv. Com. 1980, Civil Serv. Com. v. Guardians Ass’n. 1983) and has been cited in subsequent rulings (e.g., United States v. Brown 2018). 11 In Section A.2 of the Internet Appendix, we explore other discrete Bayesian classification systems and find that the results are materially unaffected by the choice of classification scheme. We also provide accuracy tests for the BIFSG methodology using publicly available Florida voter data. In an earlier version of the paper, we identified broker race using Bayesian scores calculated using only surnames and first names. Using the Florida voter registration data, we find that using the BIFSG method, which adds ZIP code information, significantly improves race classification accuracy rates. Our main results are not materially affected when we use only surnames and first names to classify race. This provides additional evidence that our results are not driven by our choice of classification technology. 12 Our goal is to capture broker compensation (not lender compensation). For most of the fee categories in the data, we can straightforwardly determine who collected the fee between New Century (the lender) or the broker. The underwriting fee and the application fee are exceptions. However, our results are materially unaffected whether we include or exclude these fees. In the mortgage literature, points generally refer to fees directly paid by the borrower to the lender to “buy down” the interest rate on the loan. In contrast, in our sample, points represent compensation the broker negotiates from the borrower with one point equivalent to 1% of the loan balance at origination. 13 For refinance loans, origination fees are often rolled into the loan amount. In other words, borrowers do not pay fees out-of-pocket at closing, but rather obtain a larger loan amount to cover the fees. 14 In theory, the broker can use this rebate to offset origination fees. However, evidence suggests that increases in yield spread premiums are associated with relatively small decreases in origination fees (Woodward and Hall 2010, Ambrose and Conklin 2014). 15 Since large differences exist in stated income shares across groups, and these loans were often used to inflate income (Jiang, Nelson, and Vytlacil 2014,Ambrose, Conklin, and Yoshida 2016), in the robustness checks we exclude stated income loans from the analysis. The main results of the paper are materially unaffected. 16 The contract spread reflects market conditions (e.g., market rates, competition) and default risk. A potential concern is that including the contract spread as a covariate in Equation (1) may effectively control for the outcome of interest since the yield spread premium, which is a component of broker compensation, is also a function of the contract rate. However, the unconditional correlation between the contract spread and yield spread premium is only 0.09. As a result, the coefficient estimates on the indicator variables of interest (⁠|$\delta_j$|⁠) remain virtually identical whether or not we exclude the contract spread from the set of controls. 17 The county share of college-educated adults also controls for the broker’s education level, which we do not observe directly. County-level education is likely a noisy proxy for borrower and broker education, especially in large counties. We thank Susan Woodward for bringing this potential issue to our attention. Unfortunately, data on ZIP code (our most granular measure of location) education levels are not available for the time period of our study. However, ZIP-code-education-level data are available from the American Community Survey (ACS) for year 2011. Assuming that educational attainment at the ZIP code level is relatively stable, we estimated the model using the ZIP-code-level share of the adult population with a bachelor’s degree or higher from the ACS 2011 5-year estimates as a proxy for education. Table A11 in the Internet Appendix reports the results and indicates that a 10% increase in educational attainment decreases broker fees by approximately 1.2% to 1.7%, but the main results showing differences in fees paid by minorities remain unchanged. We also estimate our models with ZIP code fixed effects and find that the results are unchanged (see Internet Appendix Table A12). 18 We thank an anonymous referee for this suggestion. The set of demographic control variables mirrors those included in Bayer, Ferreira, and Ross (2017). 19 Exploiting within-broker variation comes at a cost, however, as many individual brokers in the sample originate only a few loans. For example, approximately 60% of the unique loan officers originated only one loan. Forty-three percent of the white loan officers originated loans to both minority and white borrowers, while 36% of the Hispanic, Black, and Asian loan officers originated loans to both white and minority borrowers. Thus, identification in the broker fixed effects regression relies on variation in fees and minority status within the subset of brokers that originated loans to both minorities and whites. We note that over 50% of the mortgages in the sample are originated by brokers that meet these criteria. 20 In the interest of brevity, we do not report coefficient estimates for all controls used in the regression. Tabulated results are available on request, and we note that the results are consistent with those in the extant literature. The default regression model and fee model used to create Figure 2 include the entire sample (HW, BW, and AW). For ease of interpretation, we use credit score bins. However, our primary results do not use credit score bins, but rather let credit (FICO) score and a subprime indicator (FICO|$<$|620) enter directly into the model. The results are materially unaffected by this change. 21 Broker compensation also may be inversely related to risk. For example, lenders may offer greater yield spread premiums to brokers on lower-risk loans. If this is the case, then the coefficient estimates in the right panels should mirror those in the left panels. Again, this is not borne out in the figure. 22 We also estimate a default model that directly controls for broker compensation. Specifically, we create categorical variables of five broker fee quintiles and include them as a control in the default regression (quintile one is the omitted base category). This allows broker compensation to be related to unobserved credit risk factors. We thank an anonymous referee for suggesting this analysis. The coefficient estimates are reported in Internet Appendix Table A7 and presented graphically in Internet Appendix Figure 1. Across the first four quintiles, there is no relationship between broker compensation and credit risk. Default risk, however, is slightly higher (statistically significant at the 5% level) for loans in the highest fee quintile. Thus, some evidence suggests that loans with extreme levels of broker compensation (average broker compensation in this quintile is |${\$}$|11,200) have greater credit risk. As we will describe in more detail below, our main results are materially unaffected when we exclude these high fee loans from our analysis. 23 Coefficient estimates for the full set of controls are reported in Table A9 of the Internet Appendix. Table A10 reports the coefficient estimates from Table 3 transformed to dollar values. 24 In a log-linear model with a dummy variable (⁠|$ln(y)=\alpha + \beta D + \epsilon$|⁠), the percentage increase in |$y$| when the dummy changes from zero to one is |$100 \times (exp(\beta)-1)$|⁠. However, when |$\beta$| is relatively small, as is the case in our study, the percentage change in |$y$| can be approximated by |$100 \times \beta$|⁠. For ease of interpretation, we will use this approximation throughout the paper. 25 The last two rows of Table 3 report the minority premium charged by minority brokers (“Minority/Minority Premium”) along with the corresponding p-value from a test of the null hypothesis that the minority/minority premium is zero. The minority/minority premium is calculated as |$\delta_1 + \delta_2 + \delta_3 - \delta_2$|⁠. 26 We use the reghdfe package (Correia 2019, 2016) to estimate the broker fixed effects models in Stata. This package iteratively eliminates singleton groups (e.g., loans by brokers who originated only one loan), which explains the reduction in sample size. To test for differences in coefficient estimates across models, we follow the procedure outlined in the regdhfe Stata help file. The minority premium coefficients across columns 1 and 2 are statistically significantly different from one another. 27 Broker race drops from the model in the broker fixed effects specifications. Thus, the minority/minority premium is calculated as |$\delta_1 + \delta_3$| in these models. 28 The coefficients (8% vs. 5%) are statistically significantly different from each other. 29 Although our results are from models estimated separately for each subsample (HW, BW, AW), we also estimate a single saturated regression model that includes the full sample. The results, reported in Table A13 of the Internet Appendix, are virtually identical to the results in columns 3, 6, and 9 of Table 3. As discussed above, Figure 1 in the Internet Appendix shows that borrowers with loans with extreme levels of broker compensation are more likely to default. To ensure that the minority pricing premiums in Table 3 do not simply reflect greater default risk on high fee loans, we reestimate our saturated regression models, excluding loans with broker compensation in the highest (fifth) quintile. The results, which are reported in Table A14 of the Internet Appendix, are similar to those from the saturated models in Table 3. 30 For example, the consent decree between the New York Attorney General and GreenPoint Financial (New York, Office of the Attorney General, Civil Rights Bureau 2007) establishes a 25 basis point (bp) threshold in fee disparity that would require GreenPoint to “implement appropriate remedial measures to minimize the potential for future pricing disparities by the Broker, including mandatory fair lending training and oral and/or written counseling (p. 6, Section 5.2(a)).” In conducting the analysis, the consent decree stipulates that GreenPoint may only control for “race-and-ethnicity-neutral factors,” such as credit score, loan product type characteristics, and property type and location, which is similar to our regression specification. 31 See New York, Office of the Attorney General, Civil Rights Bureau (2008). Our analysis indicates that 2.8%, 24.1%, and 0.2% of Hispanic, Black, and Asian borrowers, respectively, paid premiums above the cutoff used to indicate possible disparate treatment. 32 We thank Aurel Hizmo for pointing out this possibility in his discussion of our paper at the 2020 Allied Social Science Association meeting. 33 The Board’s 2010 Loan Originator Final Rule, which amended Regulation Z of the Truth-in-Lending Act (TILA), prohibits dual compensation (Consumer Financial Protection Bureau 2012). After the CFPB inherited responsibility for the Regulation Z, the rule was republished at 12 CFR 1026.36(d) (Consumer Financial Protection Bureau 2012). 34 Details are available on the CFPB’s website: www.consumerfinance.gov/policy-compliance/rulemaking/final-rules/. 35 See www.consumerfinance.gov/ask-cfpb/what-is-a-qualified-mortgage-en-1789/ for details. A key incentive for lenders to originate QM loans is that they are afforded certain legal protections against borrower-initiated lawsuits (Bhutta and Ringo 2015). Additionally, lenders (or sponsors) do not face risk retention requirements on securitizations of qualified residential mortgages (QRM), which have the same definition as QM loans (see 24 CFR Part 267, which is available at www.gpo.gov/fdsys/pkg/FR-2014-12-24/pdf/2014-29256.pdf). As a result of the regulatory benefits afforded on QM loans, an overwhelming majority of newly issued mortgages are classified as QM. 36 We focus on regulations directly related to mortgage broker compensation; the potential impacts of other regulations affecting mortgage brokers are outside the scope of our analysis. However, we note that these regulations do cover some of the contract features that are prevalent in our data (e.g., interest-only loans) and thus would preclude them from obtaining QM status. 37 For loans less than |${\$}$|100,000, the fees can exceed 3%. The fee caps for these loans are as follows: |${\$}$|60,000 to |${\$}$|100,000: |${\$}$|3,000 |${\$}$|20,000 to |${\$}$|60,000: 5% of the loan amount |${\$}$|12,500 to |${\$}$|20,000: |${\$}$|1,000 or less |${\$}$|12,500 or less: 8% of the loan amount The definition of “points and fees” for this rule are complicated. For example, fees for third-party services where the lender requires the borrower to use a specific vendor or retains part of these fees are included in this calculation. Unfortunately, we are unable to observe some of these ancillary fees in our data. Thus, our measure of points and fees should be considered a lower-bound estimate of the measure used by the CFPB. 38 For example, a mortgage broker quoted in the New York Times complained, “I will now get paid the same amount to process a plain-vanilla loan as I will a complex loan of equal size that requires more work,” and the director at the National Association of Mortgage Brokers expressed concerns that the new rules will drive small, independent brokerages out of business (Browning 2011). 39 Mortgage markets contain a high degree of information asymmetry (Agarwal, Chang, and Yavas 2012; Albertazzi et al. 2015; Adelino, Gerardi, and Willen 2013; Keys et al. 2009). In the context of our paper, mortgage brokers, who participate in the market frequently, enjoy an informational advantage over borrowers that enter the market infrequently. This information asymmetry means that broker revenue may frequently significantly deviate from loan production costs. 40 An alternative is a stochastic frontier model similar to the approach used in Woodward (2008). In the approach, |$k$| is symmetrically distributed around a mean, and |$\pi$| is distributed nonnegative with an asymmetric distribution. However, this approach is intractable in our context because of the large number of covariates and fixed effects that likely affect costs and profits. 41 The |$\Gamma$| vector excludes property-type controls since they do not directly relate to the loan production costs as well as the log loan amount since we express fees on the left-hand side in nominal terms. Furthermore, to determine whether a loan would have been credit rationed, we divide the predicted broker compensation by the loan amount and compare the ratio to the fee caps. We suppress the subscripts for ease of interpretation. 42 The final rule implementing Regulation Z adopted an amendment allowing the exclusion of up to two “bona fide” discount points from the points and fees calculation for qualified mortgages. However, this exclusion is unlikely to affect our analysis because true discount points—paid to the lender to buy the interest rate down—are virtually nonexistent in our sample of primarily subprime hybrid ARMs. 43 In a contemporaneous paper, Bhutta and Hizmo (2020) provide evidence of post regulation racial pricing disparities in a sample of Federal Housing Administration (FHA) loans originated in 2014. Note that although the Bhutta and Hizmo (2020) paper is similar in spirit to our own, a key difference is that we control for the minority status of the loan officer. References Adelino, M. , Gerardi K., and Willen P. S.. 2013 . Why don’t lenders renegotiate more home mortgages? Redefaults, self-cures and securitization . Journal of Monetary Economics 60 : 835 – 53 . Google Scholar Crossref Search ADS WorldCat Agarwal, S. , Chang Y., and Yavas A.. 2012 . Adverse selection in mortgage securitization . Journal of Financial Economics 105 : 640 – 60 . Google Scholar Crossref Search ADS WorldCat Agarwal, S. , Choi H.-S., He J., and Sing T. F.. 2019 . Matching in housing markets: The role of ethnic social networks . Review of Financial Studies 32 : 3958 – 4004 . Google Scholar Crossref Search ADS WorldCat Albertazzi, U. , Eramo G., Gambacorta L., and Salleo C.. 2015 . Asymmetric information in securitization: An empirical assessment . Journal of Monetary Economics 71 : 33 – 49 . Google Scholar Crossref Search ADS WorldCat Ambrose, B. W. , and Capone C. A.. 1996 . Do lenders discriminate in processing defaults? Cityscape: A Journal of Policy Development and Research 2 : 89 – 98 . Google Scholar OpenURL Placeholder Text WorldCat Ambrose, B. W. , and Conklin J. N.. 2014 . Mortgage brokers, origination fees, price transparency and competition . Real Estate Economics 42 : 363 – 421 . Google Scholar Crossref Search ADS WorldCat Ambrose, B. W. , Conklin J. N., and Yoshida J.. 2016 . Credit rationing, income exaggeration, and adverse selection in the mortgage market . Journal of Finance 71 : 2637 – 86 . Google Scholar Crossref Search ADS WorldCat Antonovics, K. , and Knight B. G.. 2009 . A new look at racial profiling: Evidence from the Boston Police Department . Review of Economics and Statistics 91 : 163 – 77 . Google Scholar Crossref Search ADS WorldCat Anwar, S. , and Fang H. M.. 2006 . An alternative test of racial prejudice in motor vehicle searches: Theory and evidence . American Economic Review 96 : 127 – 51 . Google Scholar Crossref Search ADS WorldCat Arrow, K. 1971 . The theory of discrimination . Discussion Paper , Princeton University . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Avery, R. B. , Brevoort K. P., and Canner G. B.. 2007 . The 2006 HMDA data . Federal Reserve Bulletin A73 : 93 Google Scholar OpenURL Placeholder Text WorldCat Avery, R. B. , Canner G. B., and Cook R. E.. 2005 . New information reported under HMDA and its application in fair lending enforcement . Federal Reserve Bulletin 91 : 344 . Google Scholar OpenURL Placeholder Text WorldCat Ayers, I. , and Siegelman P.. 1995 . Race and gender discrimination in bargaining for a new car . American Economic Review 85 : 304 – 21 . Google Scholar OpenURL Placeholder Text WorldCat Bartlett, R. , Morse A., Stanton R., and Wallace N.. 2019 . Consumer-lending discrimination in the FinTech era . Working Paper , University of California at Berkeley . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Bayer, P. J. , Ferreira F. V., and Ross S. L.. 2016 . The vulnerability of minority homeowners in the housing boom and bust . American Economic Journal: Economic Policy 8 : 1 – 27 . Google Scholar Crossref Search ADS WorldCat Bayer, P. J. , Ferreira F. V., and Ross S. L.. 2017 . What drives racial and ethnic differences in high-cost mortgages? The role of high-risk lenders . Review of Financial Studies 31 : 175 – 205 . Google Scholar Crossref Search ADS WorldCat Bayer, P. , McMillan R., and Rueben K.. 2004 . What drives racial segregation? New evidence using Census microdata . Journal of Urban Economics 56 : 514 – 35 . Google Scholar Crossref Search ADS WorldCat Bayer, P. J. , Casey M. D., Ferreira F. V., and McMillan R.. 2012 . Price discrimination in the housing market . Working Paper , Duke University . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Becker, G. S. 1957 . The economics of discrimination . Chicago : University of Chicago Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Berkovec, J. A. , Canner G. B., Gabriel S. A., and Hannan T. H.. 1994 . Race, redlining, and residential mortgage loan performance . Journal of Real Estate Finance and Economics 9 : 263 – 94 . Google Scholar Crossref Search ADS WorldCat Berkovec, J. A. , Canner G. B., Gabriel S. A., and Hannan T. H.. 1998 . Discrimination, competition, and loan performance in FHA mortgage lending . Review of Economics and Statistics 80 : 241 – 50 . Google Scholar Crossref Search ADS WorldCat Berndt, A. , Hollifield B., and Sandås P.. 2017 . What broker charges reveal about mortgage credit risk . Working Paper , Australian National University . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC Bertrand, M. , Luttmer E., and Mullainathan S.. 2000 . Network effects and welfare cultures . Quarterly Journal of Economics 115 : 1019 – 55 . Google Scholar Crossref Search ADS WorldCat Bhutta, N. , and Hizmo A.. 2020 . Do minorities pay more for mortgages? . Review of Financial Studies , Advance Access published April 29, 2020 , 10.1093/rfs/hhaa047 . Google Scholar OpenURL Placeholder Text WorldCat Crossref Bhutta, N. , and Ringo D.. 2015 . Effects of the ability to repay and qualified mortgage rules on the mortgage market . Report, Board of Governors of the Federal Reserve System , Washington, DC . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC Black, H. A. , Boehm T. P., and DeGennaro R. P.. 2003 . Is there discrimination in mortgage pricing? The case of overages . Journal of Banking & Finance 27 : 1139 – 65 . Google Scholar Crossref Search ADS WorldCat Board of Governors of the Federal Reserve System. 2007 . Report to the congress on credit scoring and its effects on the availability and affordability of credit . Report , Washington, DC . https://www.federalreserve.gov/boarddocs/rptcongress/creditscore/creditscore.pdf. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Boehm, T. P. , and Schlottmann A.. 2007 . Mortgage pricing differentials across Hispanic, African-American, and white households: Evidence from the American Housing Survey . Cityscape 9 : 93 – 136 . Google Scholar OpenURL Placeholder Text WorldCat Browning, L. 2011 . New Fed rule for mortgage brokers . New York Times , February 17 . https://nyti.ms/2jKQm7M. Brueckner, J. K. 1996 . Default rates and mortgage discrimination: A view of the controversy . Cityscape: A Journal of Policy Development and Research 2 : 65 – 68 . Google Scholar OpenURL Placeholder Text WorldCat Calem, P. S. , Gillen K., and Wachter S.. 2004 . The neighborhood distribution of subprime mortgage lending . Journal of Real Estate Finance and Economics 29 : 393 – 410 . Google Scholar Crossref Search ADS WorldCat Campbell, J. Y. , Jackson H. E., Madrian B. C., and Tufano P.. 2011 . Consumer financial protection . Journal of Economic Perspectives 25 : 91 – 113 . Google Scholar Crossref Search ADS PubMed WorldCat Cheng, P. , Lin Z., and Liu Y.. 2015 . Racial discrepancy in mortgage interest rates . Journal of Real Estate Finance and Economics 51 : 101 – 20 . Google Scholar Crossref Search ADS WorldCat Civil Serv. Com. v. Guardians Ass’n. 1983 . 463 U.S. 1228, 103 S. Ct. 3568 . Clarke, K. A. , and Rothenberg L. S.. 2017 . Mortgage pricing and race: Evidence from the Northeast . American Law and Economics Review 20 : 138 – 67 . Google Scholar Crossref Search ADS WorldCat Conklin, J. N. 2017 . Financial literacy, broker–borrower interaction and mortgage default . Real Estate Economics 2 : 376 – 414 . Google Scholar Crossref Search ADS WorldCat Consumer Financial Protection Bureau. 2012 . Loan originator compensation requirements under the Truth in Lending Act . Report , Washington, DC . https://files.consumerfinance.gov/f/201301_cfpb_final-rule_loan-originator-compensation.pdf. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Consumer Financial Protection Bureau. 2014a . 2013 Loan originator rule: Small entity compliance guide . Report , Washington, DC . https://files.consumerfinance.gov/f/201401_cfpb_complaince-guide_loan-originator.pdf. Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Consumer Financial Protection Bureau. 2014b . Using publicly available information to proxy for unidentified race and ethnicity: A methodology and assessment . https://files.consumerfinance.gov/f/201409_cfpb_report_proxy-methodology.pdf. Correia, S. 2016 . Estimating multi-way fixed effect models with REGHDFE . Presentation , July 28-29 . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Correia, S. 2019 . REGHDFE: Stata module to perform linear or instrumental-variable regression absorbing any number of high-dimensional fixed effects . Working Paper , Boston College . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Courchane, M. , and Nickerson D.. 1997 . Discrimination resulting from overage practices . In A special issue of the Journal of Financial Services Research: Discrimination in financial services , vol. 11 , 133 – 51 . New York : Springer . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC Courchane, M. J. , and Ross S. L.. 2019 . Evidence and actions on mortgage market disparities: Research, fair lending enforcement, and consumer protection . Housing Policy Debate 29 : 769 – 94 . Google Scholar Crossref Search ADS WorldCat Courchane, M. J. , and Zorn P. M.. 2012 . Differential access to and pricing of home mortgages: 2004 through 2009 . Real Estate Economics 40 : S115 – S158 . Google Scholar Crossref Search ADS WorldCat Crawford, G. W. , and Rosenblatt E.. 1999 . Differences in the cost of mortgage credit implications for discrimination . Journal of Real Estate Finance and Economics 19 : 147 – 59 . Google Scholar Crossref Search ADS WorldCat Demyanyk, Y. , and Van Hemert O.. 2009 . Understanding the subprime mortgage crisis . Review of Financial Studies 24 : 1848 – 80 . Google Scholar Crossref Search ADS WorldCat Deng, Y. , and Gabriel S.. 2006 . Risk-based pricing and the enhancement of mortgage credit availability among underserved and higher credit-risk populations . Journal of Money Credit and Banking 38 : 1431 – 60 . Google Scholar Crossref Search ADS WorldCat Faber, J. W. 2013 . Racial dynamics of subprime mortgage lending at the peak . Housing Policy Debate 23 : 328 – 49 . Google Scholar Crossref Search ADS WorldCat Ghent, A. C. , Hernández-Murillo R., and Owyang M. T.. 2014 . Differences in subprime loan pricing across races and neighborhoods . Regional Science and Urban Economics 48 : 199 – 215 . Google Scholar Crossref Search ADS WorldCat Guardians Ass’n of N.Y.C. Police Dep’t v. Civil Serv. Comm’n. 1977 . 431 F. Supp. 526 (S.D.N.Y.) . Guardians Asso. of N.Y.C. Police Dep’t, Inc. v Civil Serv. Com. 1980 . 633 F.2d 232 . Hanson, A. , Hawley Z., Martin H., and Liu B.. 2016 . Discrimination in mortgage lending: Evidence from a correspondence experiment . Journal of Urban Economics 92 : 48 – 65 . Google Scholar Crossref Search ADS WorldCat Haughwout, A. , Mayer C., and Tracy J.. 2009 . Subprime mortgage pricing: The impact of race, ethnicity, and gender on the cost of borrowing . Brookings-Wharton Papers on Urban Affairs 2009 : 33 – 63 . Google Scholar OpenURL Placeholder Text WorldCat Horne, D. 1997 . Mortgage lending, race, and model specification . Journal of Financial Services Research 11 : 43 – 68 . Google Scholar Crossref Search ADS WorldCat Jiang, W. , Nelson A. A., and Vytlacil E.. 2014 . Liar’s loan? Effects of origination channel and information falsification on mortgage delinquency . Review of Economics and Statistics 96 : 1 – 18 . Google Scholar Crossref Search ADS WorldCat Kau, J. B. , Fang L., and Munneke H. J.. 2019 . An unintended consequence of mortgage financing regulation–a racial disparity . Journal of Real Estate Finance and Economics 59 : 549 – 88 . Google Scholar Crossref Search ADS WorldCat Keys, B. J. , Mukherjee T., Seru A., and Vig V.. 2009 . Financial regulation and securitization: Evidence from subprime loans . Journal of Monetary Economics 56 : 700 – 20 . Google Scholar Crossref Search ADS WorldCat Kider, M. , and Kamensky F.. 2015 . Subtitle A: Loan originator compensation restrictions and enforcement . Mortgage Banker Magazine , https://www.mortgagebankermag.com/featured/subtitle-a-loan-originator-compensation-restrictions-and-enforcement/. Ladd, H. 1998 . Evidence on discrimination in mortgage lending . Journal of Economic Perspectives 12 : 41 – 62 . Google Scholar Crossref Search ADS WorldCat Li, Q. 2014 . Ethnic diversity and neighborhood house prices . Regional Science and Urban Economics 48 : 21 – 38 . Google Scholar Crossref Search ADS WorldCat Liu, C. , Laporte A., and Ferguson B. S.. 2008 . The quantile regression approach to efficiency measurement: insights from Monte Carlo simulations . Health Economics 17 : 1073 – 87 . Google Scholar Crossref Search ADS PubMed WorldCat Mayer, C. J. , and Pence K.. 2009 . Subprime mortgages: What, where, and to whom?. In Housing markets and the economy: Risk, regulation, and policy , eds. Glaeser E. and Quigley J.. Cambridge, MA : Lincoln Land Institute . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC McLannahan, B. , and Rennison J.. 2018 . US subprime mortgage bonds back in fashion with |${\$}$|1.3bn of deals in first quarter of 2018 . Financial Times , March 28 . https://www.ft.com/content/6478a8d6-32c3-11e8-b5bf-23cb17fd1498. Munnell, A. H. , Tootell G. M., Browne L. E., and McEneaney J.. 1996 . Mortgage lending in Boston: Interpreting HMDA data . American Economic Review 86 : 25 – 53 . Google Scholar OpenURL Placeholder Text WorldCat New York, Office of the Attorney General, Civil Rights Bureau. 2006 . Assurance of Discontinuance Pursuant to Executive Law 63(15) in the matter of Countrywide Home Loans, Inc . https://ag.ny.gov/sites/default/files/press-releases/archived/Countrywide%20Assurance%20Final%20Signed%20PDF.pdf. New York, Office of the Attorney General, Civil Rights Bureau. 2007 . Assurance of Discontinuance Pursuant to Executive Law 63(15) in the matter of: Greenpoint Mortgage Funding, Inc. https://ag.ny.gov/sites/default/files/pdfs/bureaus/civil_rights/Greenpoint%20AOD%20Final%20PDF.pdf. New York, Office of the Attorney General, Civil Rights Bureau. 2008 . Assurance of Discontinuance Pursuant to Executive Law 63(15) AOD No. 08-160. in the matter of Home Consultants, Inc. D/B/A HCI Mortgage . https://ag.ny.gov/sites/default/files/pdfs/bureaus/civil_rights/HCI%20Settlement%20Agreement%201-5-08.pdf. Pahl, C. J. 2007 . A compilation of state mortgage broker laws and regulations, 1996-2006 . Discussion Paper , Federal Reserve Bank of Minneapolis . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Pennington-Cross, A. , Yezer A., and Nichols J.. 2000 . Credit risk and mortgage lending: Who uses subprime and why? Working Paper , Research Institute for Housing America . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Phelps, E. S. 1972 . The statistical theory of racism and sexism . American Economic Review 62 : 659 – 61 . Google Scholar OpenURL Placeholder Text WorldCat Ross, S. L. 1996 . Flaws in the use of loan defaults to test for mortgage lending discrimination . Cityscape: A Journal of Policy Development and Research 2 : 41 – 48 . Google Scholar OpenURL Placeholder Text WorldCat Ross, S. L. , and Yinger J.. 2002 . The color of credit: Mortgage discrimination, research methodology, and fair-lending enforcement . Cambridge : MIT Press . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC United States Dept. of Justice. 2010 . Consent order, United States v. PrimeLending . https://www.justice.gov/sites/default/files/crt/legacy/2011/02/08/primelendsettle.pdf. United States Dept. of Justice. 2011 . United States of America v. Countrywide Financial Corporation, U.S. District Court Central District of California . https://www.justice.gov/sites/default/files/crt/legacy/2011/12/21/countrywidecomp.pdf. United States v. Brown. 2018 . United States v. Brown. 2018 . U.S. Dist. LEXIS 40326, 2018 WL 1278577 (N.D. Ill. March 12, 2018) . OpenURL Placeholder Text WorldCat Voicu, I. 2018 . Using first name information to improve race and ethnicity classification . Statistics and Public Policy 5 : 1 – 13 . Google Scholar Crossref Search ADS WorldCat Wong, M. 2013 . Estimating ethnic preferences using ethnic housing quotas in Singapore . Review of Economic Studies 80 : 1178 – 214 . Google Scholar Crossref Search ADS WorldCat Woodward, S. E. 2008 . A study of closing costs for FHA mortgages . Report , Washington, DC : The Urban Institute . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Woodward, S. E. , and Hall R. E.. 2010 . Consumer confusion in the mortgage market: Evidence of less than a perfectly transparent and competitive market . American Economic Review 100 : 511 – 15 . Google Scholar Crossref Search ADS WorldCat Woodward, S. E. , and Hall R. E.. 2012 . Diagnosing consumer confusion and sub-optimal shopping effort: Theory and mortgage-market evidence . American Economic Review 102 : 3249 – 76 . Google Scholar Crossref Search ADS WorldCat Yezer, A. 2017 . Commentary: What can we learn from government attempts to modify the allocation of mortgage and consumer credit in the United States . Cityscape: A Journal of Policy Development and Research 19 : 199 – 205 . Google Scholar OpenURL Placeholder Text WorldCat Yezer, A. , Phillips R., and Trost R.. 1994 . Bias in estimates of discrimination and default in mortgage lending: The effects of simultaneity and self-selection . Proceedings of the Federal Reserve Bank of Philadelphia 197 – 222 . Google Scholar OpenURL Placeholder Text WorldCat Yezer, A. M. 2006 . Discrimination in mortgage lending. In A companion to urban economics , eds. Arnott R. J. and McMillen D. P., 197 – 210 . New York : Blackwell Publishing . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Yinger, J. 1996 . Why default rates cannot shed light on mortgage discrimination . Cityscape: A Journal of Policy Development and Research 2 : 25 – 32 . Google Scholar OpenURL Placeholder Text WorldCat Zhang, J. , and Zheng L.. 2015 . Are people willing to pay for less segregation? Evidence from US internal migration . Regional Science and Urban Economics 53 : 97 – 112 . Google Scholar Crossref Search ADS WorldCat © The Author(s) 2020. Published by Oxford University Press on behalf of The Society for Financial Studies. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/open_access/funder_policies/chorus/standard_publication_model)
TI - Does Borrower and Broker Race Affect the Cost of Mortgage Credit?
JF - The Review of Financial Studies
DO - 10.1093/rfs/hhaa087
DA - 2021-01-24
UR - https://www.deepdyve.com/lp/oxford-university-press/does-borrower-and-broker-race-affect-the-cost-of-mortgage-credit-XK6QPvRqnn
SP - 790
EP - 826
VL - 34
IS - 2
DP - DeepDyve
ER -