TOC4Fairness Seminar – Alexander Tolbert

Date: Wednesday, November 2nd, 2022
9:00 am – 10:00 am Pacific Time
12:00 pm – 1:00 pm Eastern Time

Location: Weekly Seminar, Zoom

Title: A Capabilities Approach for Measuring Disparities


I develop a capabilities approach to justify variable selection for the causal estimation of discrimination. Capabilities are the practical opportunities that a person is free to pursue to enhance their well-being [Anand et al., 2009, Nussbaum, 2009, Sen, 1992]. The capabilities approach (CA) emphasizes the ends that people achieve (functionings) and the alternative ends that they could achieve (capabilities). In the bioethics literature, many define a health disparity as an avoidable, systematic difference between socially advantaged versus marginalized groups, wherein the marginalized group is further disadvantaged in terms of health. To have a notion of disparity that incorporates causality, Duan et al. [2008] proposed a framework where variables of interest are classified into allowable and non-allowable categories. Allowable variables are morally permissible to cause a difference and are in the conditioning set of variables before measuring disparity. The remaining variables are classified as non-allowable. These variables are morally impermissible to cause a difference and are not in the conditioning set of variables before measuring disparity.

There has been a lack of justification for this demarcation, which is rarely explicitly described in analyses. I claim a capabilities approach lets us determine which sorts of variables are allowable and which are not allowable. I argue that if any variable or combination directly causes your capabilities to fall below some acceptable threshold, those variables are not allowable. Further, if being below the acceptable threshold of capabilities causes the realization of a particular variable or combination of variables, those variables should be considered non-allowable. 

I then defend my normative theory for variable selection. I describe what I call a 3-stage capabilities justification. The first stage identifies the capabilities relevant to a person’s well-being, freedom, and flourishing. The second stage determines the ranking or the weights of the capabilities. Lastly, the third stage determines the threshold for capabilities. I claim that a deliberative approach based on public reasons should be used to determine each stage. I then lay out a few general criteria that must be satisfied at each stage for a proper constructivist deliberation procedure. The first condition requires participation only from agents with proper moral deliberation and reflection ability. The second condition requires an impartiality condition, such as the original position. Finally, I propose a mathematical capabilities model for detecting discrimination for non-manipulable variables. I establish identifiability formulae and causal estimates for this model.


I am 4th year Ph.D. in Philosophy and Masters in Statistics student at the University of Pennsylvania and a member of the Kearns-Roth Research Group in
Machine Learning and Algorithmic Fairness. My fundamental interests are normative questions about equity, just social structure, and discrimination, particularly racial discrimination. I prefer formulating
issues and finding proposed solutions in ways that have clear policy implications. My fundamental methodology seeks to put aspects of those questions into frameworks to which various formal or empirical methods can be applied to assess the results of such applications and, where possible, contribute to improvements. My interest in formal techniques ranges over topics in machine learning, causal inference, algorithmic fairness, computational social science and economics, and algorithmic game theory. I use my training in social and political philosophy to try to inform and criticize the design of algorithms and models to embed social values such as fairness and avoid harm such as discrimination and bias.