The problem with HMDA.
In recent years, nothing has been more vexing to the mortgage lending industry than allegations of unfair lending practices that affect racial minorities and center-city neighborhoods. These charges arise from two undisputed facts: 1) the volume of home mortgage loans per mortgageable dwelling in predominantly white areas is two to three times that of minority neighborhoods, and 2) the rejection rate for minority mortgage applicants is roughly twice as high as for white applicants. Regulators, housing advocates and politicians routinely express concern over these disparities. Some even argue they constitute a prima facie case of discrimination by lending institutions.
What is the basis for the assertion that discrimination is widespread in mortgage lending markets? And what data do regulators frequently call on to analyze the issue? In most cases, the answer to both questions is the same: data provided under provisions of the Home Mortgage Disclosure Act (HMDA), originally enacted in 1975.
HMDA fits into an array of related legislation passed during the past 30 years, including the Fair Housing Act of 1968, the Equal Credit Opportunity Act (ECOA) of 1974, the Community Reinvestment Act (CRA) of 1977, and the Financial Institutions Reform, Recovery and Enforcement Act of 1989 (FIRREA), which contained amendments to both HMDA and CRA.
Unlike the prohibitions contained in the Fair Housing Act and ECOA, HMDA simply requires the collection and dissemination of data. The original legislation required institutions to disclose the geographic distribution, by census tract, of home purchase and home improvement loans. HMDA's requirements have been extended to more types of institutions over the years, most recently including independent mortgage companies in 1989. FIRREA amendments to HMDA were even more significant in that they required institutions to begin to collect and report information on all applications for home-mortgage secured-credit, together with the disposition of those applications. The earlier reporting had applied simply to the geographic distribution of loans made. HMDA information is submitted to regulators annually.
Data required includes geographic identifiers (state, county, MSA and census tract), loan amount, purpose of loan, occupancy category, borrower income, race and disposition of the loan application (approved, denied or withdrawn).
Early research using HMDA data
Before HMDA was passed, many studies sought to connect mortgage lending patterns with urban decline. The premise was that reductions in credit usage implied reduced credit availability, which in turn caused neighborhood deterioration, declining owner occupancy and even housing abandonment. But it was never clear whether loan volume was a cause or an effect of the changing urban reality.
During this same time, researchers began studying the causes of mortgage default, often expanding the scope to include neighborhood demographic factors but generally with inconclusive results. It was clear, however, that positive levels of owner equity were significantly and negatively related to default probability. This validated lenders' traditional reliance on loan-to-value (LTV) ratio as a key underwriting guideline.
After the passage of HMDA in 1975, researchers could couple census tract demographics with the original HMDA data, namely, loan volume by census tract. Early studies, however, inevitably encountered the problem of inferring demand and supply of credit based solely on actual extensions of credit. Later studies attempted to control for housing and mortgage demand by adjusting observed levels of lending by measures of mortgageable housing stock. These studies verified the disparity between white and minority urban neighborhoods and showed that neighborhood racial composition was negatively related to the flow of mortgage credit. Moreover, a much larger fraction of transactions in lower-income and minority central-city neighborhoods proved to be financed with cash than is the case in white suburban neighborhoods.
Another line of research during the 1980s focused on the distinction between FHA and conventional mortgage sectors. The premise here was that if lenders discriminate in the conventional mortgage arena, then minorities would be forced into the more costly FHA segment of the market. While research has shown that minorities are relatively more likely to choose FHA loans regardless of their levels of wealth and income, FHA underwriting guidelines have differed from those used in the conventional loan segment in some important ways, especially loan-to-value ratio. Consequently, borrowers with little savings available for down payments may self-select into the FHA category for reasons unrelated to discrimination.
Research since 1990
An alternative line of research seeking to account for persistent racial disparities in origination data is a body of work known as default-based studies. These studies reason that if discrimination exists, loans made to minorities should show lower default rates, or be more profitable compared with loans made to whites. Nobel Laureate Gary Becker, who developed one of the first economic theories of discrimination, has been a proponent of this view.
In an effort to test Becker's theory, researchers Berkovec, Canner, Gabriel and Hannan analyzed several hundred thousand FHA loans. Data included substantial detail on original application characteristics contained in the FHA records, including LTV, income, liquid assets, debt burden and whether the borrower was a first-time homebuyer. Census information was obtained to include property characteristics, including census tract racial composition. Among important variables, only credit history was unavailable.
Results indicated that blacks and predominantly black neighborhoods had significantly higher risk of default and larger losses conditional on default than whites and largely white neighborhoods. Other minority groups were statistically indistinguishable from whites with respect to both measures. Becker's hypothesis held that if discrimination existed, then loans to minorities would be relatively more profitable than loans to whites because minorities would be held to a higher credit standard. But the Berkovec study did not support this hypothesis because loans to blacks among the pool of FHA loans studied actually appeared to be less profitable. A weakness in this result is the absence of credit history information.
Prior to the FIRREA amendments to HMDA, some researchers began using nonpublicly available enhanced data sets to apply statistical probability models to gauge the effect of race and racial or income-level composition of neighborhoods on the likelihood of loan rejection. The availability of the new data (prompted by the FIRREA amendments) beginning in 1991 accelerated this line of research. The most comprehensive of these studies was the Boston Federal Reserve Study originally released in 1992 and published in the American Economic Review last March.
The Boston Fed study
Recognizing that HMDA data does not provide information on the debt burdens, assets nor credit histories of individual applicants, the Boston Federal Reserve Bank study attempted to deal with these limitations by gathering additional information. Added information was pulled together on each of the applications reported by Boston mortgage lenders during 1990. All lenders data was aggregated for a marketwide study. The controversial conclusion of the Boston Fed study was that "race does play a role as lenders consider whether to deny or approve a mortgage loan application."
Many disagreed with this conclusion. Liebowitz and Day criticized the Boston Fed's findings, mainly on the basis of alleged errors in the data set. David Horne, a Federal Deposit Insurance Corporation (FDIC) economist with unique data access, examined a subset of the actual loan files and found little support for the Boston Fed's conclusion.
Paradoxically, Horne found that a large fraction of the minority turn-downs were done by a single minority-owned on average, minority-owned institutions are more conservative in their lending criteria and actually have higher rejection rates than comparable white-owned institutions. And higher rejection rates are often construed as indicators of discrimination.
More fundamental questions have been raised about the statistical approaches used by researchers attempting to measure "discrimination" in this fashion. One of the first such papers, published in 1982 by economists G.S. Maddala and Robert Trost, noted that the methods most researchers use (known as single-equation probability models) are biased whenever there are race differences in loan demand.
More recently, two published research papers done by Rachlis and Yezer, and another by Yezer, Phillips, and Trost, advanced similar arguments, noting that racial groups may differ in their average creditworthiness and that loan applicants choose the terms of the loans they apply for, taking into account the odds of rejection and adjusting their applications appropriately. Thus, a mortgage loan applicant desiring a high-LTV loan may pay off other indebtedness, recognizing that the combination of a high LTV and high ratios increases the probability of rejection. These sorts of choices by borrowers skew the data and make the single-equation probability models inappropriate.
Moreover, all of these methods use a disturbingly indirect approach. Rather than precisely defining the phenomenon to be measured and then setting out methods for doing so, studies control for many - but probably not all - other factors and then attribute any unexplained residual variation to "discrimination."
Many studies note that if discrimination plays a role in the underwriting process it is likely to affect borderline applicants. The "totally clean" application is virtually always approved; but a significant fraction of loan applications are imperfect in one way or another. The greater the number of imperfections, the greater the likelihood of rejection, as various threshold levels are reached (for example, a 40 percent back-end ratio or 95 percent LTV).
Recognizing the difficulties of measuring such a nebulous concept as "discrimination," researchers have used computer simulation to assess the ability of various methods to identify and measure discriminatory underwriting practices. Simulation has several important advantages, including the ability to precisely define the quantities to be measured.
Among the first to use these techniques were economists Bauer and Cromwell at the Cleveland Federal Reserve. Bauer and Cromwell were interested in determining how successful the sorts of statistical procedures available to bank examiners were likely to be. In their work, a loan underwriting process with an element of randomness was simulated and discrimination was defined as the increase in rejection probability for a marginally qualified minority loan applicant. Simulated loan applicants were characterized by income and desired loan amount, loan-to-value ratio, front and back-end ratios, a qualitative measure of credit history, and race. Bauer and Cromwell's results showed that the sorts of procedures examiners use tended toward findings of "false positives," that is, indications of discrimination when none, in fact, exists.
The author's own research expanded on Bauer and Cromwell's method, making a number of enhancements. First, the characteristics of the pools of simulated loan applicants were carefully calibrated to actual national HMDA data and other available sources. Second, a variety of statistical techniques were tested for their ability to identify and measure discrimination, again defined as the increase in rejection probability for the marginally qualified minority applicant, under a range of differing data conditions. Results showed that the typical statistical method used produced biased results (findings of discrimination when none, in fact, existed) in virtually all instances. The only case when this was not true, was when minority loan applicants had the same distribution of characteristics as white applicants, a situation unlikely to be encountered with actual empirical data.
The intuition behind this result is as follows. If two groups differ substantially in their average characteristics, then statistical procedures will tend to confuse those average differences with purely group differences. In the mortgage lending case, minorities' lower average qualifications, in terms of credit score, debt burden and LTV, for instance, produce the application approval differences noted, but statistical procedures cannot tell that it is those differences, and not the race difference itself, that explain the outcomes observed.
Debate between the lending industry and housing advocates about HMDA data has often focused on the variables that are missing from that data. Lenders claim that inferences from HMDA data are flawed because of the omission of variables reflecting underwriting rules. Loan-to-value ratio, applicant credit history and debt-burden levels are generally mentioned as key data that is left out. But as simulation research shows, more data will not solve the problem if the data is used in biased testing procedures.
The size of the race effect in single-equation rejection probability models may be reduced in some very small measure by adding other control variables such as LTV or debt-burden ratios, but the mortgage lending industry would incur significant costs and would still appear guilty of discrimination under most likely data conditions. Congress would be ill advised to mandate collection and reporting of additional variables under HMDA based on the belief that the availability of such information would settle the question of lending discrimination once and for all. Not only would it not settle the issue, it would fuel an already contentious debate without adding means to resolve the issue. As the old adage goes, it would add considerable heat without adding any light.
Current examination practice reportedly uses single-equation probability model to pretest lending institutions at the onset of CRA examinations. If the tests conducted on relatively small samples of HMDA data show a positive race effect, examiners undertake a more detailed after-the-fact file-matching procedure.
But if these tests are biased toward finding discrimination when none exists in full-population samples, they cannot accomplish the regulatory objective of prescreening for compliance with fair lending. In this context "passing the test" means obtaining a statistically insignificant estimate of the effect of race on rejection probability. So winning this passing grade is most likely to occur when the lender has a relatively small percentage of minority loan applicants so that small-sample statistical procedures will produce large standard errors and relatively imprecise estimates. Thus, lenders that are sincerely interested in expanding minority lending will be punished by the larger fraction of minority loan applicants in their HMDA data. This is so because that will produce sharper parameter estimates, assured to be biased toward finding positive race effects, indicating discrimination even when none, in fact, exists.
As mentioned earlier, simulation research has shown that the sorts of statistical tests likely to be employed by examiners using actual loan files are also biased toward finding discrimination when none exists, whenever there is an element of randomness to underwriting borderline applicants. Consequently, if a lender "fails" the prescreening test, either as a result of conditions described earlier or because of having a large number of minority loan applicants, the likelihood that additional file-by-file analysis will confirm original suspicions is high.
Accordingly, the best course for lenders to take from a strictly statistical viewpoint, given current regulatory procedures, is to accept only applications from the most qualified minority loan applicants, so that the fraction of minority applicants is small and the approval rate for them is high. Such a course of action, which would involve illegal, but hard to detect, excessive prescreening of minority applicants, is certainly contrary to the spirit and intent of CRA, ECOA and fair lending statutes.
While there are no rigorous empirical studies of prescreening, anecdotal examples reported by researchers Cloud and Galster are consistent with the notion that lenders may "overscreen" minority applicants to avoid possible rejections later on in the process.
If minority qualifications, on average, are lower than white qualifications, then it follows that, on average, minority borrowers will be judged poorer risks than white borrowers. But traditionally this difference in credit risk has not been priced under the Equal Credit Opportunity Act, at least by lenders subject to CRA.
Without the price mechanism available, prime quality mortgage lenders traditionally rationed mortgage credit by rejecting those below some threshold level of credit quality, as indicated by observable qualifications. That group may be disproportionately minority, or low-income, or residents of central cities. Because credit risk is eliminated by FHA insurance, lenders will have ample incentive to make loans to marginally qualified applicants, particularly because rates charged on FHA-insured loans diverge very little from the conventional market.
Hence it is not surprising that we observe higher delinquencies and default rates on FHA loans relative to conventional loans and note that lower-income and minority borrowers are more likely to choose FHA products.
In summary, the regulatory framework facing lenders encourages them to channel marginal credits to the government-insured sector since they are not allowed, by government regulation, to price credit risk appropriately in the conventional market segment. As the market for B and C paper evolves, especially through. development of high-LTV products that compete with FHA, these incentives are likely to change.
I have traced the origins of the Home Mortgage Disclosure Act and the research that HMDA data has made possible. We have seen that HMDA has played a key role in fueling the fair lending legislative initiatives over the past 30 years. Data collected under HMDA provided the rationale for enactment of the Community Reinvestment Act and concern that data availability was still inadequate influenced provisions of FIRREA.
For more than 20 years, HMDA has spawned an array of research efforts. Yet existing data reported under HMDA are generally insufficient to identify discrimination in lending, if it does exist. Augmented HMDA data incorporating key loan underwriting factors are also generally insufficient to identify discrimination, except in limited and unlikely circumstances. And the customary method of measuring discrimination even using augmented HMDA data is biased toward finding discrimination, when none exists, in all but exceptional circumstances. Given these results, we might well conclude that, on balance, HMDA has done far more harm than good.
Michael LaCour-Little, Ph.D., is vice president-research at Citicorp Mortgage, Inc., and was the recipient of the 1996 MBA Dissertation Research Award given by the University of Wisconsin-Madison.
|Printer friendly Cite/link Email Feedback|
|Title Annotation:||Home Mortgage Disclosure Act|
|Date:||Nov 1, 1997|
|Previous Article:||RESPA's latest chapter.|
|Next Article:||The Taxpayer Relief Act.|