Data-Driven Models for Eviction Prevention

Innovation is often driven by trying to do more with less. This is certainly the case in Allegheny County, a leader in using data to both understand the vast landscape of needs and direct funding and care with the greatest efficiency. 

In 2023, leaders at the Allegheny Department of Human Services (DHS) began a project to fine-tune their rental assistance and eviction prevention program, based on the knowledge that existing eviction prevention funding was winding down. For Katy Collins, chief analytics officer at DHS, the driving question was “can we understand how to best target limited rental assistance dollars to those who would most benefit from them?” Initially, this included trying to predict homelessness at the point of eviction. If possible, this would allow DHS to catch and support folks before they lost their homes and became unhoused and accurately direct resources where they would have the most impact.

Allegheny DHS is a data and predictive analytics pioneer, thanks to their large, interconnected Data Warehouse. Pulling together information from across the County and beyond, the Warehouse has data related to DHS touchpoints, child welfare, substance abuse disorder, unemployment, intimate partner violence, and other topics that fall under the broad umbrella of Human Services. It also includes data from the state level as well, including public benefits and unemployment information. The DHS team knew that a tool could be built on this data, creating a model to predict homelessness and then operationalizing that model to provide targeted assistance. With this clear goal, Collins accepted graduate student assistants from the Data Science for Social Good (DSSG) team at Pittsburgh’s Carnegie Mellon University to assist with model development.

There are known indicators that the CMU and DHS team were able to review, including past interactions with the homelessness system as a predictor of future homelessness, substance use disorder, and unstable housing conditions. By identifying individual-level data across the Warehouse, and identifying triggers to homelessness like eviction filings, the team was able to put together a predictive model. Of course, this data is protected and nothing at the individual-level is available to the general public. According to DHS Director Erin Dalton, all models created by DHS have external reviews, with method and fairness assessments. Non-personally identifying information (PII) will also be publicly available with the methodology, in order to provide transparency. While some have expressed reservations about the use of predictive models in the past, Dalton stated that the development of an assessment tool actually allows greater interrogation around decision making. Previously, decision-making — and in some cases funding — relied on case conferencing, advocating, calling, and individual experiences, which could introduce bias or lack a full picture. “We identify places where we think [a model] would help and develop them to implement, not just to create a model,” she said.

After several months of developing and refining the model, the graduate students handed over the model to DSSG post-doctoral students who then progressed to outreach and piloting with DHS. The DHS team knew that many of the folks who don’t proactively call in or ask for eviction assistance are some of the people who need the most help. By going back into the Data Warehouse and identifying these cases, the team could see what touchpoints actually connected some of the most vulnerable residents with their system. For example, child welfare workers may already have a relationship with a family which is then a conduit for sharing information about the assistance. The points where the County was in touch with these folks are now targets for spreading the word about eviction prevention assistance. DHS is also working with local community partners that are already doing prevention work, to help them also be more targeted in their assistance.

There’s another benefit to using the data driven model in Allegheny. This predictive tool not only identifies where financial assistance prevents homelessness, but also reveals where that simply isn’t enough. “Money isn’t the deciding factor for why some folks are homeless,” said Dalton, “so we’re really trying to find out for whom this is most impactful.” There are more appropriate interventions for different groups at risk of eviction and homelessness, as “very high risk folks” likely need support beyond financial assistance, according to Dalton. The model is being fine-tuned through piloting to best identify what successfully helps and where different approaches are needed; for example, families with children are often able to avoid eviction and homelessness with financial aid, as opposed to the highest risk individuals. Allegheny data shows that the highest risk individuals are often those entering the mental health crisis system, who may benefit much more from wraparound and co-located services rather than monetary assistance.  

Initial pilot results show success, with Collins reporting that many of the folks who received assistance are still statically housed. And they are continuing to refine the model to understand who can get the most benefit from funding. For example, the DHS team is also interested in where else this tool could help in directing benefits, like during post-incarceration reentry. According to Dalton, DHS is also working with advocates to pilot a “self-sufficiency” screener that would complement the model.

For other cities looking to emulate this work, Collins and Dalton provided the following recommendations:

• Work with partners. The CMU DSSG students and postdocs were critical to this work, providing capacity to DHS. Local community groups were also foundational to the outreach, as trusted messengers and partners on eviction prevention work.
• Prioritize analytics. “We don’t do anything without analytics driving it,” said Dalton, who maintains that public services should be as invested in data and analytics capacities as the private sector. 
• Be realistic. Both Collins and Dalton are clear that rental/eviction assistance isn’t the sole way to prevent homelessness, as preventing evictions isn’t the same as preventing homelessness. But for the folks that it helps, it’s indispensable, so use data to reach those groups. 
• Share across silos. Without data sharing across health and human service systems, folks in need can be missed, undercounted, or mismatched with assistance. This can also lead to wasted costs. By integrating and sharing data, residents can receive the most appropriate care in the most efficient way.