Using Machine Learning to Improve Policy Problems; An Econometric Perspective

Can machine learning improve policy outcomes? Can it do so even if the algorithms do not draw causal conclusions? How do we manage the possibility that algorithms might magnify racial and other biases? To examine these questions, Harvard University Economics Prof. Sendhil Mullainathan will talk through one end-to-end example: pretrial detention decisions.

Using a large historical data set, he and his colleagues build and evaluate the potential for a purely predictive algorithm to improve on judges' decisions making. On the one hand, their results suggest room for optimism--we can reduce crime and incarceration rates and simultaneously reduce racial biases. At the same time, there is room for caution. Their application highlights the dangers of "naive" applications. In particular, Prof. Mullainathan highlights two central and thorny econometric problems -- selective labels and omitted payoff biases - that are often ignored. Along the way he will present a simple econometric framework that lays out these problems but also provides a way to understand the role of prediction in policy as well as racial (or other) biases in machine learning.

Please RSVP for this special workshop to Tyler Gabrielski at tgabs@bu.edu. Refreshments provided.