The Measure and Mismeasure of Fairness

Authors: Sam Corbett-Davies, Johann D. Gaebler, Hamed Nilforoshan, Ravi Shroff, Sharad Goel

JMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Here we first assemble and categorize these definitions into two broad families: (1) those that constrain the effects of decisions on disparities; and (2) those that constrain the effects of legally protected characteristics, like race and gender, on decisions. We then show, analytically and empirically, that both families of definitions typically result in strongly Pareto dominated decision policies. For example, in the case of college admissions, adhering to popular formal conceptions of fairness would simultaneously result in lower student-body diversity and a less academically prepared class, relative to what one could achieve by explicitly tailoring admissions policies to achieve desired outcomes.
Researcher Affiliation Academia Johann D. Gaebler EMAIL Department of Statistics Harvard University Cambridge, MA 02138, USA; Hamed Nilforoshan EMAIL Department of Computer Science Stanford University Stanford, CA 94305, USA; Ravi Shroff EMAIL Department of Applied Statistics, Social Science, and Humanities New York University New York, NY 10003, USA; Sharad Goel EMAIL Harvard Kennedy School Harvard University Cambridge, MA 02138, USA
Pseudocode Yes Algorithm 1: Path-specific Counterfactuals
Open Source Code Yes Reproduction materials are available at https://github.com/jgaeb/measure-mismeasure.
Open Datasets Yes We base our risk estimates on age, BMI, and race, using a sample of approximately 15,000 U.S. adults aged 18 70 interviewed as part of the National Health and Nutrition Survey (NHANES; Centers for Disease Control and Prevention, 2011-2018)...For our analysis, we use the data released by Obermeyer et al. (2019), which contain demographic variables, cost information, comorbidities, biomarker and medication details, and health outcomes for a population of approximately 43,000 White and 5,600 Black primary care patients at an academic hospital from 2013 2015. Obermeyer et al. released a synthetic data set closely mirroring the real data set, available at: https://gitlab.com/labsysmed/dissecting-bias.
Dataset Splits No The paper describes using a 'simulation study of one million hypothetical applicants' and a 'population of approximately 43,000 White and 5,600 Black primary care patients' from a released dataset. However, it does not specify any training/test/validation splits for models trained or evaluated within this paper.
Hardware Specification No The paper does not provide specific details about the hardware used to run its simulations or analyses.
Software Dependencies No The paper does not specify any software names with version numbers, or specialized packages with versions used for implementation.
Experiment Setup Yes In the example that we consider in Section 4.1, the exogenous variables in the DAG, U = {u A, u D, u E, u M, u T , u Y }, are independently distributed as follows: UA, UD, UY Unif(0, 1), UE, UM, UT N(0, 1). For fixed constants µA, βE,0, βE,A, βM,0, βM,E, βT,0, βT,E, βT,M, βT,u, βT,B, βY,0, βY,D, we define the endogenous variables V = {A, E, M, T, D, Y } in the DAG by the following structural equations:...We use constants µA = 1/3, βE,0 = 1, βE,A = 1, βM,0 = 0, βM,E = 1, βT,0 = 50, βT,E = 4, βT,M = 4, βT,u = 7, βT,B = 1, βY,0 = 1/2, βY,D = 1/2.