The usefulness of statistical analysis as a fair-lending testing tool has been much debated in recent months. Out of this debate a consensus is emerging:
*Denial rates are not very useful for discrimination testing and monitoring.
*Advanced statistical techniques such as regression analysis, even using application file data, are not a testing/monitoring panacea.
*File reviews focused on individual applications may not be well suited to detecting patterns or practices of discrimination.
*The ideal process for discrimination testing and monitoring at this point in time is a combination of statistical analysis and file review.
The reason for this emerging consensus is that statistical analysis and file review techniques are complementary, with offsetting strengths and weaknesses.
Thus, a good fair-lending monitoring strategy for a financial institution is to use statistical analysis to focus its fair-lending review on the possibly troublesome files, and then to review those files to determine why they might be troublesome.
Experience indicates that this approach focuses on a reduced number of select application files. The net benefit to the institution is more effective monitoring at a lower cost.
Despite being easy to compute and readily available, almost everyone agrees that denial rates by themselves provide little insight into whether a pattern of discrimination exists. We wish the world we live in were so simple that single indicators would give us the answer. Clearly, that is not our world.
File comparisons look at accepted and denied applicants who have similar characteristics along several dimensions. The objective of the examiners in comparing the two groups of files is to see if the denied applicants are being given the same access to credit as the similarly situated approved applicants.
This process must be based on similarly situated applicants. There are two disadvantages with this approach. The first problem is that exact or close matches may not exist, and even if they do, it's not clear that the examiner will find them. The second difficulty is that once files are identified, the question is whether the files represent a pattern or practice of discrimination.
One of the advantages of the file review process is that it requires less data entry than statistical testing. However, as automated application tracking systems become the norm, this problem will evaporate.
It is also true that file reviews can be done on small numbers of files, whereas statistical testing requires more files to satisfy the assumptions of the underlying models. Perhaps the most important advantage of file reviews is that they allow the individual circumstances of the applicant to be considered.
The other broad class of tests is a specialized form of regression analysis. Like the file comparison technique, statistical tests have their own set of advantages and disadvantages.
The first advantages of statistical analysis are those associated with determining the existence of exact or close matches, finding those file matches, and deciding whether or not they constitute a pattern or practice of discrimination.
Finding clones is done by statistically weighing an applicant's characteristics. The statistical process will not only identify the clone files that may have received differential treatment, but it will also identify files which may look very different on an individual characteristic basis, but are on average very similar.
These types of offsets are very difficult for individuals doing file reviews to make; however, computers analyze compensating factors very well.
The shortcomings of the file review and statistical analysis techniques individually and the recognition of their complementary virtues has led to the emerging consensus that the statistical analysis and file review processes should be integrated. The role of statistical analysis in this combination is to focus file reviews on particular files.
The approach works because the statistics provide a list of files, which then need manual review to ensure that elements outside the scope of the model are considered. A question that is often asked is, If I have to go to the file anyway, why bother with the statistics? The answer is that statistical analysis focuses the review on potentially troublesome files. To use a forest and trees analogy, the statistics look at the whole forest and tell which trees to inspect.
The files that fall out of the statistical process as outliers can be divided into two groups: those that were rejected by the institution and according to the model should not have been, and those that were accepted by the institution and, according to the model, should have been rejected.
While some bankers view the files that were accepted but should have been rejected as brag loans, these loans might also be viewed as exceptions to the policy, exceptions that potential applicants might also expect to receive. Thus, by accepting these applications, the institution may have created an exception to policy that it will have to honor for future applicants.
The second group of files, those that were not accepted by the institution but should have been according to the statistical model, is usually the larger of the two groups of files. There are two interesting points about these files. First, the applicants in this group of files are typically marginal in that they meet some, but not all, of the institution's acceptance criteria.
The acceptance/rejection decisions from this group of files are not always clear-cut and the underwriter must often make a judgment call. Making such calls requires a mental calculus - counterbalancing several offsetting criteria - that can become quite complex. And, where there is judgment, there is an opportunity for discrimination, albeit unintended.
Further, if there is no offsetting information in the file, these applications represent forgone business. In one case, the files in this group were 10% of the total number of files. Assuming no countervailing file information, these files represent 10% more business for the institution.
Supplementing statistical analysis with file reviews has one final advantage. This integrated process, may encourage financial institutions to aggressively market to the special markets without fear of incurring high rejection rates and, therefore, increase the number of applications going to people in protected classes.
In essence, institutions will be able to support their rejection rates on the basis of economic and special applicant situations, and thereby refute possible accusations of discrimination founded on high rejection rates.
The fair-lending testing and monitoring process continues to evolve. The debate about which testing and monitoring approach is best appears to be resulting in some consensus.
Specifically, a combination of statistical analysis and file reviews seems to be the best way to test and monitor the mortgage credit decision process for disparate treatment. It is also the most cost effective alternative at this point.
Additionally, financial institutions may also find that using the integrated approach enables them to expand their business by marketing more aggressively into lower- and moderate-income and minority census tracts as well as individuals without fear of rising rejection rates, because this integrated process provides strong support for credit decisions.
Mr. Preiss is president and founder of Preiss & Co., a compliance consulting firm in Lake Forest, Ill., specializing in CRA and fair-lending issues.
