Measuring and Reducing Algorithmic Discrimination with Quasi-Experimental Data

This research project will develop new tools to measure and reduce algorithmic discrimination in several high-stakes settings. Algorithms guide an increasingly large number of decisions. Alongside this rise is a concern that algorithmic decision-making will entrench or worsen discrimination against legally protected groups.

However, quantifying algorithmic discrimination is often hampered by a selection challenge: an individual's qualification for a decision, which is often used to define discrimination, is typically only available for the group of individuals who were selected for treatment by an existing human or algorithmic decision-maker. This project will overcome this fundamental selection challenge by developing new tools to measure algorithmic discrimination.

The project also will develop alternative algorithms that minimize or reduce discrimination. The researchers will apply these tools in multiple high-stakes settings, including pretrial detention, employment screening, medical testing, and child welfare investigations. The research is of considerable policy interest given the rapid adoption of algorithms in a variety of settings.

The investigators are committed to increasing diversity in the economics research community by recruiting, training, and mentoring women, under-represented minorities, and first-generation college students as undergraduate research assistants and predoctoral fellows. Code produced by this project will be made publicly available.

This research project will develop tools to measure algorithmic discrimination. The project also will develop alternative non-discriminatory algorithms when qualification is unobserved for a subset of individuals. For example, in the employment context, whether an individual would be hired after an interview is not observed for applicants screened out before the interview is held.

The investigators will show that this selection challenge can be overcome with knowledge of average qualification rates across different groups. Further, these average qualification rates can be estimated by utilizing random assignment of decision-makers to individuals. This insight can be used not only to measure algorithmic discrimination, but to develop alternative algorithms that reduce or eliminate discrimination.

The project will consider several extensions. The investigators will utilize experimentation to measure algorithmic discrimination and improve accuracy. The interaction between algorithms and human decision-making also will be explored, as human discretion remains important in most real-world settings.

The results of this research will have implications for more accurately quantifying the trade-offs between algorithmic transparency, accuracy, and fairness.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.