Efficient Homelessness Prevention through Better Targeting: Using Machine Learning to Predict Homelessness Among Unstably Housed Veterans

Background and Purpose: Reliable predictions of imminent homelessness have the potential to transform the efficiency of homelessness prevention services.  Prevention providers seek to allocate resources only to households who will become homeless to ensure that cost savings from reduced shelter use exceed program costs, but predicting which households will become homeless is extremely difficult. Most screening tools either miss large swaths of households that become homeless, or cast too wide a net and serve those who would remain stably housed without any intervention.  This study seeks to improve on current screening models by testing a machine learning algorithm’s predictions of subsequent homelessness among a national sample of Veterans at risk of homelessness.

Methods: The study uses data collected from the Homelessness Screening Clinical Reminder (HSCR), a 2-question screener to identify homelessness and imminent risk of homelessness among all Veterans accessing Veterans’ Health Administration outpatient services.  Among a sample of 104,312 individuals who screened positive for homelessness risk in 2012, we use prior homelessness and housing status, medical and behavioral health records, VA benefits eligibility, and demographic data to predict homelessness status at subsequent rescreening.  Forecasts are made using two methods – logistic regression and random forest, a machine learning classification and forecasting algorithm – and compared.

Results: The random forest algorithm forecasts homelessness with significantly greater accuracy than logistic regression: the algorithm produces a higher rate of true positives and a lower rate of false positives.  Findings have broad implications for homelessness prevention at the Department of Veterans Affairs and beyond.

Conclusions and Implications: By using machine learning forecasts to allocate resources, the VA and community-based providers can substantially improve their targeting, increasing shelter savings while reducing costs for false positives.  Through more efficient allocation, programs can make better use of existing resources and have a stronger argument on which to advocate for additional funds.