Under-speciﬁed Models and Detection of
Discrimination: A Case Study of Mortgage Lending
Risk Analysis Division, Ofﬁce of the Comptroller of the Currency, 250 E Street, S.W., Washington D.C. 20219, USA
This study examines how omitted variables affect underwriting models the OCC estimates during fair lending
examinations. The purpose is to assess the effects of omitted variable bias common to most studies of
discrimination in mortgage lending. The results show omitted variables have an important impact on both the
estimate of the effect of race and on the identiﬁcation of outliers for review. Further, there appears to be no
consistent patterns to the direction of these impacts. This suggests that it is inappropriate to make
generalizations concerning the direction of bias based on assumptions about correlations between omitted
variables and race.
Key Words: mortgage lending, discrimination, omitted variable bias
Omitted variable bias is a common problem afﬂicting empirical studies of racial dis-
crimination. As a result, critics are quick to dismiss ﬁndings of discrimination as simply the
effects of omitted variables that are correlated to race. Given the consequences of ﬁndings
of discrimination, it is important to assess the existence and extent of omitted variable bias
in racial estimates from statistical models. This is difﬁcult to achieve directly, precisely
because limited data and information are the cause of the bias. However, information about
the extent of omitted variable bias can be obtained indirectly by identifying situations
where omitted variables are known to be minimal, creating omitted variable bias by
excluding relevant variables, and measuring the impact.
Fair lending exams conducted by the Ofﬁce of the Comptroller of the Currency
(OCC) present just such a situation. During fair lending exams, the OCC has access to
lenders’ policies, underwriters, and loan applications, which allows it to identify all of
the economic factors a lender considers during the underwriting decision-making pro-
cess. This allows the OCC to minimize problems of omitted variable bias when esti-
mating statistical models to test for discrimination. Access to this information also
creates an opportunity to assess how racial estimates from various fair lending studies
may be affected by variables unavailable to these studies, but available to the OCC.
Although, results from this type of analysis cannot be generalized to other fair lending
or discrimination studies with certainty, compelling arguments can be made, especially
The Journal of Real Estate Finance and Economics, 31:1, 83–105, 2005
2005 Springer Science + Business Media, Inc. Manufactured in The Netherlands.