Statistical modeling can pinpoint areas of lending discrimination.

Bank regulators and prosecutors have entered the computer age. Just as the examination procedures for identifying discrimination have evolved from testing for technical compliance to comparative file reviews, so have the examiners' techniques and tools. Statistical modeling is now used by the Fed and the Office of the Comptroller of the Currency in fair-lending examinations and by the Justice Department in investigating and prosecuting lending discrimination.

Statistical modeling, we predict, will also become the standard self-assessment tool for financial institutions with extensive residential mortgage portfolios.

Regulators, prosecutors, and financial institutions use statistical modeling to accomplish two principal objectives.

The first is to identify whether race or some other prohibited basis was a significant factor in the institution's lending decisions.

This is the approach that the Justice Department used in its case against Decatur Federal government officials have called Decatur a model for future fair-lending prosecutions. It is also the approach used in the Boston Fed study, and has apparently been embraced by Fed and OCC examiners.

A second objective of statistical modeling is to provide a more efficient and less judgmental method to identify loans for comparative file review.

The comparative file review techniques used in current examination procedures analyze similarly qualified applicants whose loan applications have different outcomes - sometimes referred to as matched pairs.

Without computer techniques, examiners review a sample of loan files and judgmentally compare approved and denied applicants.

With more efficient computer techniques, examiners, prosecutors and internal investigators can access a wider number of loan files. They can use statistical analysis tools to rank applicants according to their Strengths, weaknesses, and underwriting decisions.

Based on these rankings, applicant files can also be subjected to a more targeted comparative file review to examine whether institutions provided inappropriate consideration or assistance to applicants.

Moreover, because samples are selected to be representative of all loan applicants - not minorities alone - cases of extraordinary assistance provided protected class applicants may be identified.

The first step in building a lending discrimination model is to understand what an underwriter considers when approving or denying an application.

The universe of data that an underwriter can access is vast. Much information is obtained through the application and subsequent communication (e.g., credit bureau reports, employment, income, deposit verifications, and property appraisals).

The model builders must identify the variables that allow underwriters to distinguish acceptable from unacceptable credit risks - "explanatory variables."

Explanatory variables are selected through review of loan policies, loan files, and underwriter interviews.

Once the explanatory variables are identified, statisticians can determine the adequate number of loan files to sample. This sample will include both approvals and denials from protected and nonprotected classes.

The fair-lending investigator then extracts the explanatory variables from each of the sampled loan application files, and enters the statistics into a computer database.

With the assistance of their computer tools, statisticians analyze the application data base and compute the weight and statistical significance of each explanatory variable.

A "dummy variable" representing protected class status is added as a potential explanatory variable to test for discrimination.

For example, if the model is investigating gender discrimination, a dummy variable-is set to a certain value (such as 1) if the applicant is female, and another value (such as 0) if the applicant is male.

If the dummy variable is considered statistically significant - measured by certain standard tests - then the model suggests that the protected class status is a factor in the lending decision, independent of all of the applicant's other qualifications.

If such status hinders an applicant, the model is considered evidence of lending discrimination.

Because modeling is a new element in fair lending examination, it is uncertain what regulatory actions will follow from a statistical finding of discrimination.

Fed examiners suggest that a failure to find statistical discrimination will end the fair-lending review. Alternatively, if statistics indicate discrimination. Fed examiners may use the model to chose applicants for file review, as described above.

As the OCC's modeling program is quite new, less is known about their operating procedures. At this time, it appears as though OCC is using statistical modeling in addition to manual file reviews.

An institution that is prepared for this new generation of fair-lending examinations will have fewer unpleasant surprises. In our practice, we have recommended that all institutions undergo intensive fair-lending self-examination.

At a minimum, this should include a manual comparative file review, based on current examination guidelines.

Prior to beginning such a review, an institution may wish to consult with counsel to discuss any privilege or immunity that may attach to findings.

For midsize and large institutions, we now recommend that self-examination include statistical modeling. In fact for many institutions, modeling is the most cost effective way to structure a self-examination. Manual comparative file reviews require a sizable staff of experienced employees or consultants.

Alternatively. a modeling effort can rely on less experienced personnel to extract applicant data from loan files, and analytical work can be completed by a smaller team experienced statisticians and analysts.

Further. if comparative file review is necessary, the model results will pinpoint those applications with anomalous results for more targeted review.

In addition, the model may assist in compliance monitoring. The model can be converted for use in future periods as a self-testing tool. For example, each quarter a fair-lending reviewer can enter data on closed loans into the model and review those applications whose results were not predicted by the model.

Recent cases and regulatory pronouncements suggest that financial institutions should actively investigate past and current mortgage lending to uncover and resolve possible discrimination.

If discrimination is suspected. in addition to correcting policies, and disciplining personnel. institutions should, with advice of counsel, offer credit to applicants who were inappropriately denied or else compensate them for any damages.

Self-testing efforts - whether by modeling, file review, or both - should be directed at finding any problems before federal fair-lending examiners do.

Mr. Neiman is executive vice president of New York-based Waterhouse Investor Services Inc. (which is not affiliated with Price Waterhouse). Mr. Lavine is a senior consultant with Price Waterhouse Regulatory Advisory Services.

For reprint and licensing requests for this article, click here.
MORE FROM AMERICAN BANKER