An article published in Plos One by researchers from the Spokane Predictive Analytics Group evaluates the efficacy of using utility billing data to predict first-time homelessness. Previous efforts to predict an individual’s risk of experiencing homelessness have relied on data sources like healthcare records or criminal conviction histories, which are not always available for the general population, may vary widely in structure and content, and are not made available in a timely fashion. In contrast, utility billing data are available for most U.S. households, include comparable information on payment history from one jurisdiction to the next, and are typically reported on a monthly basis. The researchers find that utility payment histories hold promise as possible predictors of first-time homelessness.
The authors obtained deidentified utility billing data for residents of Spokane County, Washington, from Avista Utilities (which provided electricity and natural gas billing data) and the City of Spokane (which provided water, sewer, and garbage collection data) for the December 2015-December 2019 period. These data came from over 64,000 premises with roughly 86,300 primary account holders, corresponding to 16.0% of Spokane County residents. The City of Spokane also provided information on individuals experiencing homelessness from its local Homeless Management Information System (HMIS).
The data were used to build two multivariate binary logistic regression models, which were used to estimate whether a utilities account holder was within 12 months of experiencing homelessness for the first time. One model used general utilities information that is typically available in jurisdictions around the country, including the total amount of utilities owed, whether any utilities are in arrears, the number of times a service shutoff arrangement was established for an account, the number of individual account holders associated with the same address (a proxy for resident turnover at a property), and the number of addresses associated with the same individual account holder (a proxy for transiency). The second model incorporated additional data points available in the Spokane County data that may not be readily accessible in other jurisdictions, such as total amount of money owed for specific utilities over a 90-120-day period, and the number of times a utility provider manually contacted an account holder about payment arrearages.
Based on the utilities billing data, the general model correctly predicted that 1,603 account holders would experience first-time homelessness within 12 months of exhibiting signs of financial instability. This represents 86.2% of the 1,845 total account holders documented as experiencing first-time homelessness during the study period as per the HMIS data.
Although the model also had a high number of false positives, the precision of the model can be adjusted as needed to accommodate the differing goals and resources of homelessness prevention programs (HPPs), which aim to prevent homelessness by addressing financial or other crises that put households at greater risk of losing their homes. These programs often provide rental assistance, landlord-tenant mediation, eviction legal services, and other community-based social services that can help households remain stably housed. In addition to helping people avoid the trauma and long-term negative impacts of homelessness on an individual level, previous studies have demonstrated that HPPs play a role in reducing homelessness rates in communities they serve and may be more cost-effective than homeless assistance programs. However, the success of HPPs is highly dependent on their ability to identify households at risk and intervene in a timely manner.
For this reason, the findings of the study may be especially relevant to HPPs. For example, HPPs that use resource-intensive interventions to prevent homelessness can adjust the model to correctly identify individuals who are truly at risk of homelessness more often (with fewer false positives), even if the model does not identify as many of those individuals (giving more false negatives). The authors believe the model “shows promise in facilitating identification of those in need of aid by increasing processing speed and prediction accuracy” and could be used as a first-line screening tool to identify and connect with households facing financial precarity before more detailed screening processes are conducted. The study’s findings may be especially beneficial for cities like Los Angeles that are already exploring ways to use artificial intelligence to prevent homelessness and keep residents stably housed.
Read the article here.