Online Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 492 - Application of Nonparametric Methods
Type: Contributed
Date/Time: Thursday, August 6, 2020 : 10:00 AM to 2:00 PM
Sponsor: Section on Nonparametric Statistics
Abstract #313853
Title: Counterfactually Fair Risk Assessment Instruments
Author(s): Alan Mishler* and Edward Kennedy
Companies: Carnegie Mellon University and Carnegie Mellon University
Keywords: causal inference; fairness; risk assessment ; optimization; semiparametric theory; influence function

Risk Assessment Instruments (RAIs) are predictive tools used to aid decision making in domains such as criminal justice and child welfare. RAIs estimate the risk of an adverse outcome, such as child neglect; this estimated risk may then inform a decision such as whether to screen in a child welfare case for investigation. RAIs are naturally concerned with the potential outcomes associated with available decisions, such as what would happen to a child if no investigation were initiated. This means that, contrary to current practice, both the predictive performance and the fairness properties of RAIs should be evaluated with respect to potential outcomes, rather than historical observed outcomes (Coston et al., 2020). Here, we develop novel methods to post-process arbitrary existing predictors to render them (1) loss-optimal with respect to potential outcomes and (2) fair in the sense that they satisfy (approximate) counterfactual equalized odds (Hardt et al., 2016; Coston et al., 2020). We derive estimators and show both theoretically and via simulations that they are consistent at fast rates.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2020 program