Statistical Discrimination with Affirmative Action Using Agent-Based Modeling

This research develops a simple agent-based model to implement Arrow’s and Phelps’ theory of statistical discrimination. Discriminating and non-discriminating firms can hire from two categories of workers, those subject to discrimination and others who are not. The statistical discrimination presented takes one of two forms. Firms can have biased beliefs about differences in average productivity between the two groups of workers, or firms can believe the reliability of a skills test for assessing worker productivity varies by demographic group. The former leads to discriminating firms being more likely to exit the market, especially in the presence of affirmative action policies. The latter is much less affected by affirmative action policies. To combat this subtler form of statistical discrimination, affirmative action policies cannot only rely on detecting groupwide demographic differences. They must instead check for both groupwide differences and review employee-level data to check for discrimination.