Reducing Entry into Homelessness through Distribution of Rental Assistance
Partner: Allegheny County Department of Human Services
Team: Kasun Amarasinghe, Alice Lai, Kit Rodolfa, Rayid Ghani (continuation of our DSSG 2022 project – Team: Joachim Baumann, Arun Frey, Abby Smith, Catalina Vajiac
More than half a million Americans experienced homelessness in 2021, facing a wide range of negative impacts on their physical and mental well-being. In Allegheny County, PA, several programs administered by the Department of Human Services (ACDHS) seek to serve the existing homeless population (for instance, with temporary shelters and rapid rehousing programs) as well as prevent entry into homelessness (for instance, with rental assistance for individuals and families facing eviction). However, resources for these preventative services are very limited relative to the level of need: with a budget of approximately $3 million, only a small fraction of the roughly 15,000 evictions filed each year can benefit. In addition, the current process of allocating rental assistance is reactive, requiring individuals who are facing an eviction to reach out to ACDHS in order to be placed on a waitlist, where funding is distributed on a first-come-first-served basis. This system does not prioritize individuals based on their risk of falling into homelessness, while also missing out on those who need urgent help but fail to reach out. ACDHS wants to make this process more proactive and prioritize those who face the greatest need but get left behind.
In this project, we have partnered with ACDHS to better allocate rental assistance to high risk individuals in order to keep them from falling into homelessness. To do so, we used data from ACDHS’ rich data warehouse, including information on previous evictions, homeless spells, interactions with mental, behavioral, and physical health institutions, address changes, demographic information, as well as enrollment in a variety of other ACDHS and state programs, among others. Using this historical data, we trained a series of models that predict an individual’s risk of entering homelessness in the next 12 months among all those with a recent eviction filing.
The current rental assistance selection process used by ACDHS achieves poor precision in terms of preventing homelessness: only 2 out of the 100 individuals selected will fall into homelessness in the 12 months following eviction. By contrast, the machine learning models we trained achieved ten times this precision, a substantial improvement over the existing process. This work was done as a project in the 2022 Data Science for Social Good Fellowship.
We initially restricted our scope to individuals who are facing eviction, but eviction is not the only pathway to homelessness. Our next step is to identify individuals who have interacted with other ACDHS services who are at risk of entering into homelessness.