Can Machine Learning Help Predict the Outcome of Asylum Adjudications?

Abstract

In this study, we analyzed 492,903 asylum hearings from 336 different hearing locations, rendered by 441 unique judges over a thirty-two year period from 1981-2013. We define the problem of asylum adjudication prediction as a binary classification task, and using the random forest method developed by Breiman [2], we predict twenty-seven years of refugee decisions. Using only data available up to the decision date, our model correctly classifies 82 percent of all refugee cases by 2013. Our empirical analysis suggests that decision makers exhibit a fair degree of autocorrelation in their rulings, and extraneous factors such as, news and the local weather may be impacting the fate of an asylum seeker. Surprisingly, granting asylum is predominantly driven by trend features and judicial characteristics- features that may seem unfair- and roughly one third-driven by case information, news events, and court information.

Replaces

Daniel L. Chen, and Jess Eagel, “Can Machine Learning Help Predict the Outcome of Asylum Adjudications?”, TSE Working Paper, n. 17-782, March 2017.

Reference

Daniel L. Chen, and Jess Eagel, “Can Machine Learning Help Predict the Outcome of Asylum Adjudications?”, in Proceedings of the ACM Conference on AI and the Law, 2018.