Implementing Anti-Discrimination Policies in Statistical Profiling Models*

How should statistical models used for assigning prices or eligibility status in insurance, college enrollment, social-benefit provision, and other settings be implemented when there is a concern about discrimination? In many settings policies are in place that prohibit the direct use of factors such as race, gender, and age. However, there is a large and contentious gray area surrounding the use of other variables (e.g., zip codes and credit scores) that are likely to be highly correlated with these omitted characteristics. We provide a framework to discuss this question and propose a method that serves to eliminate bias while maintaining efficiency relative to alternative models that eliminate key variables. We illustrate the value of our proposed method using data from the federally mandated Worker Profiling and Reemployment Services system. Contact Pope at [email protected] and Sydnor at [email protected]. *We thank Dan Black, David Card, Stefano DellaVigna, Jesse Leary, Jaren Pope, Matthew Rabin, Stephen Ross, Jesse Rothstein, Paige Skiba, Jeff Smith, and seminar participants at UC Berkeley for helpful comments and suggestions. We are also grateful to David Card for generously sharing data used for this project. All errors are our own.