Applications of Machine Learning Methods to Predict Readmission and Length-of-Stay for Homeless Families: The Case of Win Shelters in New York City

ABSTRACT

New York City faces the challenge of an ever-increasing homeless population with almost 60,000 people currently living in city shelters. In 2015, approximately 25% of families stayed longer than nine months in a shelter, and 17% of families with children that exited a homeless shelter returned to the shelter system within 30 days of leaving. This suggests that “long-term” shelter residents and those that re-enter shelters contribute significantly to the rise of the homeless population living in city shelters and indicate systemic challenges to finding adequate permanent housing. This article focuses on our preliminary work with Win (Women-in-Need) shelters to understand the factors that predict readmission and length-of-stay of homeless families. We create a unified, comprehensive database of the homeless population being served by Win shelters, accounting for more than 6,000 homeless families. We apply logistic regression models and an unsupervised clustering algorithm to identify predictors of re-entry and long-term length-of-stay. Citizenship, age, medical conditions, employment, and history of foster care or shelter stays as a child are found to be significant predictors. The results of the K-means clustering identify three primary groups, consistent with previous typologies characterized by transitionally homeless, episodically homeless, and chronically homeless.

KEYWORDS:

• Clustering
• homelessness
• machine learning
• predictive modeling
• shelter services

Acknowledgments

Our thanks to Win for their partnership in this research, especially to Christine Quinn, Ira M. Bellach, and Meghan Linehan for their support and guidance. We would also like to thank CUSP MS students Xueqi Huang, Kristi Korsberg, Dara Perl, and Avikal Somvanshi for their work on a preliminary version of this analysis presented at the 2017 Bloomberg Data for Good Exchange.

Additional information

Funding

This work is supported, in part, by the John D. and Catherine T. MacArthur Foundation. This research has been conducted in accordance with NYU Institutional Review Board approval IRB-FY2017-1135. Any opinions, findings, and conclusions expressed in this article are those of the authors and do not necessarily reflect the views of any supporting institution. All errors remain the responsibility of the authors.