Limits of Parity Measures

Topics

• Limits of Parity Measures: Infra-Marginality, Intersectionality, Causality

Reading (Required)

Simoiu, Corbett-Davies, and Goel describe the impact that different distributions of behaviors can have on decision making and how parity conditions sometimes incorrectly point toward non-existent discrimination. This article is accessible, clear, and includes a detailed walk-through of the concept through an analysis of the effects of predictive policing on Black communities.

• The Problem of Infra-Marginality in Outcome Tests for Discrimination

Parity Measures requires classifying individuals into defined groups. The politics of how those groups are defined matters.

• Data Feminism, Ch 4: What Gets Counted Counts

Reading (Optional)

This article describes the phenomenon of ‘Fairness Gerrymandering’, in which a model constrained to be ‘fair’ will naturally move unfairness to carefully divided subgroups of people defined through intersectionality.

• Preventing Fairness Gerrymandering:Auditing and Learning for Subgroup Fairness

This article proposes an approach to measuring fairness for individuals that identify with multiple salient groups (intersectionality):

• Bayesian Modeling of Intersectional Fairness: The Variance of Bias

Reading Responses

• In the Infra-Marginality Paper, carefully re-read the section on Omitted variable bias (page 1211). In the COMPAS dataset, we have the variables here. Which of these variables may have a similar effect on the differences in distributions between COMPAS scores for Black vs. White defendants? Why?

• In Data Feminism:

  1. Summarize the main argument of the reading
  2. Give a supporting argument the author uses
  3. Discuss any objects you may have or subtleties the author must deal with.