On the Robustness of Racial Discrimination Findings in Mortgage Lending Studies

The binary logistic regression or logit link model is commonly used to test for racial disparate treatment in fairlending studies undertaken by government agencies, including the Office of the Comptroller of the Currency (OCC) and the Federal Reserve Board (FRB). Ensuring race neutrality in lending remains a concern of regulators and consumer advocates. Improving the understanding of any shortcomings of either bank internal models or regulatory agency models will enable those participants in the mortgage industry to better serve the needs of consumers. We explore this issue using five bank studies undertaken by the OCC. We consider the impact of the logit link assumption, as this determines how race affects the likelihood of loan approval, by moving to three other links: probit, gompit and complementary log log; the latter two are examples of asymmetric links. As our data sets have been obtained using stratified sampling procedures, which has been typical at the OCC, rather than being drawn via simple random sampling, moving away from the logit link complicates estimation; it is no longer possible to use a standard estimation command with an adjustment for stratum effects. Our results reveal that the choice of link function, despite exhibiting similar sample fit, can influence findings of disparate treatment at the nominal level of significance commonly accepted as the legal standard. We also find that the use of a resampling method, which aims to better approximate the finite sample null distribution, for obtaining p-values typically leads to support for discrimination more often than arises from use of the standard normal approximation.