Modeling Discrimination with Causal Abstraction

The paper introduces a framework for reasoning about racial discrimination, treating race as a high-level abstraction. It addresses the challenge of modeling discrimination causally by allowing precise, explicit statements about social construction, ensuring modularity. The abstraction framework clarifies disagreements in the literature, pinpointing where precise causal accounts of discrimination differ. Causal discrimination theory posits that direct racial discrimination occurs when a person is treated worse due to their race. This concept is supported by audit studies, like Bertrand and Mullainathan's (2004) experiment, which found racial bias in the labor market. The theory applies to other protected attributes and is used in algorithmic fairness criteria and legal contexts, such as the U.S. Title VII, which prohibits actions "because of" sex. The causal notion of discrimination is invoked when the outcome would change if the attribute were different, aligning with but-for causation. This approach promises to address limitations in probabilistic notions of discrimination and is computationally feasible, with methods to mitigate causal discrimination under uncertainty.