Informs: Improving Homelessness Prevention and Client Outcomes through Data Science
In the United States, a Continuum of Care (CoC) Program is a nonprofit umbrella organization that coordinates efforts to aid those who find themselves homeless and help prevent homelessness proactively for a given metropolitan area. Within a CoC, you will find organizations that offer services including emergency shelter, rapid rehousing, permanent supportive housing, street outreach and prevention via shelter diversion. One key feature of a CoC is the coordination of a Homeless Management Information System (HMIS), which acts as a shared database across all partner organizations. Requirements for an HMIS are defined by the U.S. Department of Housing and Urban Development. Because such a database is standardized and sufficiently adopted, it provides a tremendous resource to these organizations for ensuring the efficacy of their efforts. Since 2007, the CoC serving my home city of Cincinnati has been Strategies to End Homelessness (STEH, or “Strategies”). Strategies also manages the Central Access Point (CAP) for Hamilton County, Ohio, which provides a single contact point for those who are at risk of, or currently experiencing, homelessness to receive services.
In 2017, Strategies to End Homelessness reached out to 84.51° – a Cincinnati-based retail data science, insights and media company – for help with their data. Strategies, which has been at the forefront of HMIS adoption in the U.S., hoped to partner with a data scientist to receive guidance on the ways in which this data could be utilized to guide its ongoing assessment of efficacy and inform its decisions on how to best serve the community. We’ve assisted Strategies’ data work in manners ranging from tactical – data cleansing and technical support – to more aspirational diagnostic and prescriptive science solutions. In doing so, I’ve observed firsthand the meaningful improvement that data science can have on outcomes for those served by Strategies and, more broadly, the impact that support from the data science community can make for nonprofits.
Community-facing Performance Insights As an umbrella organization working with many partners, having a meaningful view of data and insights across its full system is a matter of importance for Strategies to End Homelessness. The challenges for such an ask are ensuring the insights visualized are summarized to a level that is meaningful to users and that the data visualization (data viz) remains continually accessible and updated. For a few key initiatives, we worked to build and host data visualizations that met these needs for STEH and its community partners.
The first such exercise involved designing a dashboard to keep a pulse on the KEYS program, which serves youths (age 24 and under) experiencing homelessness. The designed data viz provides a lens over time for key measures of efficacy, including number of entries for services, frequency of recidivism and proportion of exits to a variety of permanent and temporary housing destinations. The dashboard, now embedded within Strategies’ website via Tableau Public, provides a straightforward means for the CoC and its community partners to access updated measures for the program.
A similar request surfaced for a streamlined view of the overall health of the system of care. Especially of interest was understanding how effectively positive outcomes, such as exits to permanent housing and increased income, are being achieved. These measures are important to understand relative to each type of program, as well as benchmarked over time. An effective visualization would therefore be able to answer questions such as “Has rapid rehousing been effective at increasing clients’ income?” and “Are emergency shelters improving clients’ exit destinations over time?” With those goals in mind, we worked to design and embed Strategies’ system performance dashboard in a manner accessible to community partners and continually updated with fresh data.
Data-driven Assessment of Risk Factors For organizations who aim to extend aid to those who are most at need, having a factually driven basis for assessing need is a key aspiration. Accordingly, Strategies faced a need to find the factors that most adversely affect risk of recidivism (return to homelessness). This request brings with it a few pieces of nuance. First, we’re trying to assess not only the risk of an event occurring but also how quickly that event would occur. In other words, if someone is at high risk of recidivism, they’re likely to recidivate in a shorter amount of time as well. Second, observed data for this is what we would call “right-censored,” meaning that for records where recidivism did not occur, there is still some chance that recidivism could happen in the future, and we need a model to account for that. Survival analysis, commonly leveraged in the healthcare space, is a branch of statistics that was designed precisely to handle these nuances. It aims to assess the risk of an event occurring while accounting for a temporally censored training set.
For Strategies, application of such a technique came down to how to best identify and serve families who are most at risk of returning to homelessness, and thereby having a model to assess this risk so that the proper resources could be directed. In part owing to the standardization of HMIS data, such a model was successfully trained and applied to guide the response of the CAP at STEH. The analysis focused on identifying the most differentiating risk factors of recidivism via Cox regression and Kaplan-Meier curves. In doing so, Strategies was able to more prescriptively guide its response to CAP calls by extending shelter diversion and aftercare support for families who had previously been in shelter. Additionally, Strategies shared descriptive data from the analysis to better inform the respective programs of partner organizations that serve families.