Sports

Experimental Tests for Discrimination by Mortgage Loan Originators

Categories
Published
of 28
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Related Documents
Share
Description
Experimental Tests for Discrimination by Mortgage Loan Originators Andrew Hanson Department of Economics, Marquette University P.O. Box 1881 Milwaukee, WI Zackary Hawley
Transcript
Experimental Tests for Discrimination by Mortgage Loan Originators Andrew Hanson Department of Economics, Marquette University P.O. Box 1881 Milwaukee, WI Zackary Hawley Department of Economics, Texas Christian University 2855 Main Drive Fort Worth, TX Hal Martin Department of Economics, Georgia State University P.O. Box 3992 Atlanta, GA30302 Bo Liu Department of Economics, Georgia State University P.O. Box 3992 Atlanta, GA30302 Abstract: We design and implement an experimental test for differential response by Mortgage Loan Originators (MLOs) to requests for information about mortgage loans. Our experiment, based on correspondence with 5,464 MLOs is designed to analyze differential treatment by client race and credit score. Our results show that on net 1.9 percent of MLOs discriminate by not responding to African American clients while responding to white clients. We find that this level of discrimination is small compared to the differential treatment across credit score groups, with 8.3 percent of MLOs responding to inquiries from relatively high credit score clients while not responding to clients that do not report a credit score, and 3.7 percent of MLOs responding to the high credit score group while not responding to inquiries from clients with a relatively low credit score. The effect of being African American on MLO response is roughly equivalent to the effect of having a credit score that is 50 points lower. We also find that the content of response to our inquiries favors whites by offering more details about a loan and using more friendly language. Keywords: Discrimination, Field Experiment, Mortgage Lending, Race JEL: J15, C93 1 I. Introduction Allegations of discriminatory lending practices during the housing boom have resulted in some of the largest cash settlements ever between mortgage lenders and the Department of Justice (DOJ). The two largest settlements, $335 million from Bank of America s Countrywide group and $175 million from Wells Fargo, 1 allege that these institutions steered equally qualified minority applicants into higher interest (sub-prime) loans and charged higher fees than for white borrowers. Interestingly, over this same time period, Home Mortgage Disclosure Act (HMDA) data released by the Federal Financial Institutions Examination Council 2 shows only an 8 basis point difference in the average interest rate charged to white and non-white borrowers (favoring whites). 3 The substantial cost of discriminatory lawsuits and the lack of corroboration in the aggregate data begs the question, do mortgage lenders really discriminate against minority borrowers?, and if so, how is this possible in an age of computerized, nearly automatic, underwriting? We answer each of these questions by testing for racial discrimination using a matched-pair correspondence experiment on Mortgage Loan Originators (MLOs). MLOs are essentially licensed mortgage sales workers who assist customers with loan applications and have the ability 1 The Countrywide and Wells-Fargo lawsuits were settled in December, 2011 and July, 2012, respectively. The DOJ alleges that Countrywide charged higher fees and rates to more than 200,000 African American and Hispanic borrowers and steered more than 10,000 African American and Hispanic borrowers into high interest mortgages (New York Times, December 21 st, 2011). Similarly, the DOJ alleges that Wells-Fargo charged higher fees and rates to approximately 30,000 African American and Hispanic borrowers, and steered more than 4,000 African American and Hispanic borrowers into high interest mortgages (DOJ, July 12 th, 2012). It is typical that lending institutions do not admit guilt when settling a charge of discrimination with the DOJ. 2 The FFIEC maintains summary statistics of HMDA data on its website at: 3 This difference is a raw mean, not conditional on borrower characteristics. The 8 basis point difference between white and non-white borrowers is the difference in the mean interest rate reported for loans where the interest rate is known on conventional, 1-4 family home purchase loans (excluding manufactured homes) between 2004 and Whites have an approximately 27 basis point differential (favoring whites) with African American borrowers over this time period, again not conditional on borrower characteristics. 2 to offer and negotiate the terms of a mortgage with applicants. 4 MLOs are typically the initial and primary contact person for borrowers seeking a mortgage, and, as sales workers whose compensation is tied to performance, have some discretion over how they respond to customer inquiries. MLOs may, for example suggest that a borrower attempt to improve their credit score before completing a loan application, or may encourage a borrower to act quickly to take advantage of low interest rates. They may also present different fees or interest rates to borrowers, offer encouragement or discourage the borrower from moving forward with the loan, or offer other financial advice related to obtaining a mortgage. The role of information provider and advisor in the lending process, and the discretion MLOs have in dealing with customers makes them an integral part of the borrowing process from a client s perspective. Discrimination by MLOs could result in different lending outcomes between minority and majority borrowers, and also influence outcomes as the home purchase process proceeds. For example, a borrower whose pre-approval is delayed or is pre-approved for a smaller loan amount may be treated differently by a real estate agent in terms of search effort, neighborhood choice, or expediency of service. If differences in initial treatment by an MLO are severe (offering different interest rates, fees, or suggesting credit repair services), this could conceivably affect a home buyer in all aspects of the home purchase, even if they are successful in obtaining a loan. Our matched-pair experiment examines the response MLOs offer to initial contact from a potential client interested in obtaining information about a mortgage loan. We design the 4 The Secure and Fair Enforcement for Mortgage Licensing Act (SAFE), part of the larger Housing and Economic Recovery Act of 2008, included several provisions to tighten regulations of MLOs. These provisions included requiring licensing of MLOs, creating a Nationwide Mortgage Licensing System (NMLS), issuing uniform licensing applications and reporting requirements across states, and creating a national clearing house for collecting consumer complaints. 3 experiment to test for differential treatment by client race (white or African American) as well as differential treatment by credit score. We use three credit score categories: no, low, or high credit score offered. We randomly assign pairs of inquiries to MLOs according to our design to test for the effects of race, credit score, and the interaction between these two. Client race is revealed to MLOs by the name associated with each inquiry, where we use names with a high likelihood of being associated with only one race. The source of names in our study is birth certificate counts from male babies born in New York City in We examine both the propensity for MLOs to respond to our inquiries and the content of their response to test for differential treatment. To our knowledge, this is the first experimental test of discrimination by MLOs that uses e- mail correspondence. 5 This is in contrast to earlier studies by Smith and Delair (1999) and Ross et al. (2008), that rely on in-person interaction between MLOs and actors (referred to as an audit). 6 Heckman (1998), and Heckman and Siegelman (1993) critique using actors when testing for discrimination, as they may bias results if they are not identical along all dimensions except race. The severity of actor bias can be diminished by proper choice and training of actors, but actors may also bias an experiment if personal beliefs affect how they act in the experiment. Hanson and Hawley (2011) point out actors may also cause bias if one race reports treatment differently or acts to prompt behavior in subjects in a different manner. While we believe there is value in using in-person studies, and they offer ways to examine discrimination by MLOs that our study cannot, our work provides some advantages over in-person studies. Most importantly, 5 There are several recent studies that use correspondence to test for discrimination in housing markets. See Hanson and Hawley (2011) for a recent example and a review of this literature. 6 See Ross and Yinger (2002) for a particularly lucid explanation of discrimination in the lending process, including an explanation and critique of research methodology. See Ross et al. (2008) for a more recent review of the literature on lending discrimination. 4 we avoid actor bias by relying solely on electronic communication with MLOs, allowing us to dramatically increase the scope of the experiment and the geographic area covered in a costeffective way. Using electronic communication also provides a detailed record of correspondence which allows us to examine the timing and content of MLO response to our inquiries. In addition, the use of the internet in general is becoming a standard part of the home search and borrowing process. Bricker et al. (2010) report that 41.7 percent of borrowers use the internet for information about borrowing, 7 and over 90 percent of home buyers in 2012 reported using the internet in some capacity during their home search (NAR, 2012). Our results show that MLOs discriminate on the basis of race, and treat clients differently by their reported credit score. We find that on net, 1.9 percent of MLOs discriminate by not responding to inquiries from African Americans while responding to inquiries from white clients. 8 We find larger net response differences across credit score types, with 8.3 percent of MLOs responding to the high credit score group while not responding to the no credit score group, and 3.7 percent of MLOs responding to the high credit score group while not responding to the low credit score group. We also find that credit score differences exacerbate differences in differential response between races. Whites are favored by nearly 10 percent of MLOs in audits where the white client reports a higher credit score than the African American client. In cases where African American clients report higher credit scores than white clients, African Americans 7 In addition to the internet, 39.5 percent of borrowers report using sellers of financial services as a method of obtaining information about borrowing. The most commonly used source of information about borrowing is friends, relatives, and associates with 43.9 percent of borrowers using that channel (Bricker et al., 2010). 8 The net level of discrimination measures the difference in the percentage of MLOs that only reply to an inquiry from a white client against the percentage of MLOs that only reply to an inquiry from an African American client. The gross level of discrimination or the percentage of MLOs that only reply to an inquiry from a white client is 17.1 percent of MLOs. The overall difference in response rates is 2.7 percentage points favoring whites- this difference does not match the net discrimination level because some of our experiments involved sending two inquiries to an MLO that had the same race, but different credit score. 5 are favored over whites only 6.6 percent of the time. Examining the content of the response shows that whites are favored even among MLOs that respond to both inquiries. The primary difference in the content of response between whites and African Americans is the inclusion of details about the loan. II. Experiment Design To test for discrimination among MLOs we design a matched pair correspondence experiment using to inquire about assistance with a home mortgage. 9 The matched-pairs are structured to account for race and credit score differences among potential borrowers. Consistent with the matched-pair design, each MLO receives two s in the experiment. We use three credit score groups in our experiment: no credit score, low credit score, and high credit score. The low credit score group reports a randomly assigned credit score between 600 and 650, the high credit score group reports a randomly assigned credit score between 700 and As a precaution against exposing the experiment, we randomly assign a credit score for each (rather than each pair) from a uniform distribution within each category (low or high). Although there is a chance the credit scores within a matched pair are exactly the same, most often the scores will be different within a small range. We use the randomly assigned differences in credit scores within each range to further test how MLOs respond to credit score 9 The experiment attempts to uncover discrimination from differential treatment based on minority status. Scholars (and governments) have also recognized that disparate impact, or having a policy that disproportionately impacts minorities while lacking business purpose is discrimination. See Turner and Skidmore (1999) for a discussion about the difference between differential treatment and disparate impact in mortgage lending. 10 The low credit score range approximates the th percentile of the national distribution of credit scores according to the Fair Isaac Company (FICO). The high credit score range approximates the th percentile of the national FICO score distribution (FRB, 2007). Most MLOs seem to operate using a rule on an acceptable credit score (like a minimum of 620); we noticed that the reported rule in response varied across the responses we received. Our low score sample seems to straddle the rule in all areas. 6 and race differences. For the portion of experiments where a credit score is reported, the average credit score reported is 675; 725 for the high score group and 625 for the low score group. We also divide our experiment into groups by the content of correspondence. We chose this design to guard against exposing the experiment. We divide the experiment into two correspondence categories, with the difference being the type of questions asked of the MLO. Using different questions across MLOs may make our inquiries less suspicious for company spam filters or for co-workers who discuss client s. s in the question set #1 group contain one question about interest rates, and a second question about mortgage fees (all questions for this set are listed in Appendix 1 in the boxes labeled Question 1 and Question 2). s in the question set #2 group contain one question about loan availability and a second question about what information is necessary to proceed in the process of obtaining a loan (all questions for this set are listed in Appendix 1 in the boxes labeled Question #3 and Question #4). To further guard against exposing the experiment to MLOs, we randomly assign the content of inquiries without repeating specific texts for the same MLO. For example, we randomly assign each one of five possible greetings (Hello, Hi, Hi There, Hey, or Dear), and ensure that the other sent to the same MLO does not use the exact greeting. We view the benefits from not matching the text exactly (reducing the risk of exposing the experiment) as exceeding the cost that any of our greetings (or other text options) might influence outcomes in a meaningful way. 11 Appendix 1 details the exact layout of our correspondence and the randomly assigned text that populates each There is almost no difference in the response rate across types of s. The response rate for the question set 1 group is The response rate for the question set 2 group is This difference is not statistically meaningful. 7 Our experiment includes 30 different matched pair types, representing all of the combinations between cells in Figure This allows us to examine the marginal effect of race and credit score (on the extensive and intensive margin), as well as to examine if there is a different marginal effect of credit score across races. We randomly assign each MLO to a matched pair type, and randomly vary the credit score within the range for that type. The matched pair, or within-subjects, design means that each MLO in our experiment receives two e- mail inquiries. We reveal borrower race to MLOs through the name associated with each inquiry. The source of first names is the New York City Department of Health and Human Hygiene (DHHH) records for babies born in The DHHH birth records provide counts of babies born by gender, race, and first name. We start the process of choosing names to represent race by calculating the probability a baby is born either white or African American for each name in the sample. We use only male baby names for this calculation. The DHHH data do not report a count for names with fewer than 10 babies born in a given race-gender match. This makes our probabilities for names that are very likely to be associated with only one race equal to one, when in fact they could be less than one. Because of this censoring, and the primary concern of signaling race, we also consider the raw number of occurrences each name has within a given race. After compiling a list based on probabilities and counts, we eliminate names that have a Muslim or Jewish origin from our list as we want to minimize any confounding effects these characteristics would bring to the experiment. 12 We also randomly vary the order in which s are sent. We do this at the matched pair type, so there are 60 types of matched pairs, one with each type of sent first and an identical pair type with the other sent first. We are careful to randomly assign and maintain an equal number of each type with the order of sending different to ensure that order effects do not drive any results. 8 The source of surnames is Word et al. s analysis of 2000 Census data. This analysis reports counts of surnames for the general population, and by race/ethnicity of respondents to the census. For African American surnames we use the same criteria as first names, choosing those with the highest probability of belonging to African Americans. We choose the surnames with the largest probability of belonging to African Americans regardless of total count, as the data shows a large number of African Americans with these surnames in all cases. We use a slightly different criteria for white surnames, as many of the white names with the highest probabilities all have an strong ethnic component (For example the highest probabilities are Yoder, Mueller, Koch, all are from a German origin). For white surnames, we choose three names (Miller, Nelson, Baker) from the most common (by count) names that have greater than a 0.8 probability of being white and less than a.15 probability of being African American. We choose the other two names (Krueger and Schmitt) from the list with the highest probabilities of being white, regardless of how ethnic sounding they are. Choosing names from each list allows us to test if the choice of ethnic sounding last names affects our results. Table 1 shows the list of names used to signal race in the experiment. The first three columns of Table 1 show the probability a baby is African American or white given they are born with that name, the count of babies born with that name in 1990, and the rank (by count) for each name. The last three columns of Table 1 show the probability a person is African American or white given the surname, the count of persons with that name in 2000, and the rank (by count) for each name. White names chosen with the alternative criteria are not ranked in the top ten for all white names, thus we report their value for rank as NA. MLOs are exposed to the name 9 associated with each inquiry in three ways: the actual address, 13 the signature at the bottom of each (set up to be first name and surname), and the name plate in the MLOs inbox (s
Search
Related Search
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks