Can We Predict (And Possibly Prevent) Homelessness Among Veterans?

In the United States, the number of homeless people has increased every year for the past five years. One group at high risk for homelessness is military veterans. In the US today, it is estimated that there are 40,000 homeless veterans, which makes up about 8% of the country’s homeless population. In 2009, the Obama Administration dedicated extensive resources to the Department of Veterans Affairs (VA) to end homelessness among veterans. More than a decade later, veterans remain over-represented in the homeless population.

Efforts to address homelessness most often focus on providing housing and health interventions to individuals who are already homeless; however, preventing homelessness has emerged as a feasible approach to this problem. In a recent study, Katherine Koh, MD MSc, a psychiatrist at MGH and member of the Street Team at the Boston Health Care for the Homeless Program, and colleagues including researchers from the Veterans Administration have focused on US Army veterans and have used machine learning to identify risk factors most strongly associated with becoming homeless after transitioning to civilian life with the ultimate goal of generating a model which could be used to predict homelessness before discharge from military service.

This prospective cohort study included observations from a total of 16,589 US soldiers who had previously completed a baseline survey of the Army Study to Assess Risk and Resilience in Service members (STARRS) between 2011 and 2014 while in active service. Participants were again surveyed between 2016 and 2018 when they were no longer on active duty, defined as being either separated completely from Army service or deactivated/no longer activated but still in a Reserve or National Guard Component.

The prevalence of homelessness in the 12 months before the survey was 2.9% in the total sample and was higher among the separated (3.4%) versus deactivated (1.7%) veterans. Persistent homelessness, defined as being homeless for more than 3 months, was less common, affecting 1.3% of the population.

Indicators of Mental Health are the Strongest Predictors of Homelessness

The researchers used machine learning to identify the factors most strongly associated with homelessness, screening approximately 2,000 potential predictor variables. The three risk factors most strongly associated with becoming homeless were a history of depression, post-traumatic stress disorder and experiencing the loss of a loved one being murdered.

We have never before had tools to be able to predict homelessness with the degree of accuracy that our model found.

Other important predictors included non-military trauma (e.g., exposure to natural disaster, 4 or more interpersonal losses), and indicators of adverse childhood experiences (childhood homelessness, being on welfare as a child, and physical neglect). Other mental health variables, including lifetime history of generalized anxiety disorder and self-reported indicators of suicidality (lifetime ideation and 2 or more attempts) were also among the significant predictors, all associated with increased risk of homelessness.

The optimal predictive model included 26 predictors (11 survey variables and 15 administrative or geospatial variables). In the test sample, the model was able to identify 61% of the veterans who would experience homelessness in the 4 highest ventiles (top 20%) of predicted risk (area under curve or AUC 0.78, SE=0.02). Model performance varied across important subsamples, performing somewhat better in men (AUC=0.79) than in women (AUC=0.74) and in the separated (AUC=0.78) than in deactivated veterans (AUC=0.67).

Next Steps: Identifying Veterans at Risk for Homelessness

The VA is actively involved in a series of new initiatives to help transitioning veterans adjust to life back in the civilian world. Identifying soldiers at highest risk for homelessness before separation/deactivation could facilitate provision of targeted preventive interventions for homelessness. The results of the current study clearly show that homelessness can be predicted significantly using a flexible machine learning model that captures about 60% of the soldiers who will become homeless.

Our hope is to be able to identify these individuals before leaving so that they already are linked up with support and a person that they can trust prior to leaving.

A relatively small number of survey predictors (n=11) had a significant impact on risk of homelessness. From a practical standpoint, it would be possible to include these variables in a brief self-report questionnaire administered prior to discharge from active service, such as the Department of the Army Career Engagement Survey which all soldiers are required to complete before leaving the Army.

Information in the Department of Defense electronic health records will ultimately be linked with the Veterans Health Administration (VHA) electronic health records. Once these two agencies complete implementation of this shared system, it would be possible to create an automated system to alert clinicians to veterans at increased risk for homelessness.

It is striking that out of the approximately 2,000 potential predictor variables considered, indicators of mental health emerged as the most important predictors of homelessness. While tending to the mental health needs of service members and veterans has become a priority and clearly has implications for improving overall health and well-being, the findings of this study raise the possibility that interventions targeting mental health in veterans may also help to decrease homelessness and mitigate the negative sequelae of homelessness. Looking forward, Dr. Koh states, “The next stage of this study will be to develop and implement a real-world case management intervention designed to target these high-risk soldiers and prevent them from falling into homelessness, which we are working on designing currently.”