Abstract Example 3 Bowen Mince
Machine Learning in the Courtroom: Evaluating the Effectiveness of Algorithms to Predict the
Likelihood of Recidivism
Bowen Mince1 and Wagih Henawi1
1 Department of Statistics, Grinnell College
The United States has both the largest prison population and the highest per-capita incarceration rate in
the world. These metrics have increased in the past half-century, even outpacing population growth. To
reduce the prison population in an unbiased manner, several algorithms have been created and
implemented to predict the likelihood of recidivism of convicted people. In this work, we develop models
using three types of algorithms (classification trees, random forests, and logistic regression) and then
evaluate their accuracy in predicting recidivism using a data set from Broward County, Florida that
tracked 10,372 criminals convicted in 2013 and 2014 over a 5-year period. The accuracy for each model
ranges from 60% - 70%. However, the false-positive rate and false-negative rate varied significantly in
each model regarding race. We then discuss the shortcomings of the current use of these algorithms in
courtrooms throughout the United States and how they perpetuate systemic bias.
Acknowledgment: Shonda Kuiper, Department of Statistics, Grinnell College; Funded by Grinnell
College through the Mentored Advanced Project (MAP).
