Prioritizing Which Homeless People Get Housing Using Predictive Algorithms

We present the methodology for developing a predictive model for identifying homeless persons likely to have high future costs for public services. It was developed by linking administrative records from 2007 through 2012 for seven Santa Clara County agencies and identifying 38 demographic, clinical and service utilization variables with the greatest predictive value. 57,259 records from 2007 to 2009 were modelled, and the algorithm was validated using 2010 and 2011 records to predict high cost status in 2012. The model generated a good area under the ROC curve of 0.83. A business case scenario shows that two-thirds of the top 1,000 high-cost users predicted by the model are true positives with estimated post-housing cost reductions of over $19,000 per person in 2011. The model performed very well in giving low scores to homeless persons with one-time cost spikes, achieving the desired result of excluding cases with single-year rather than ongoing high costs. This triage tool identifies individuals for whom the solution of housing costs less than the problem of homelessness. These individuals are likely to remain homeless and use public services extensively, setting them apart from the majority of homeless individuals who have short stints and use fewer services. The tool is a predictive analytic statistical algorithm that uses 38 pieces of information to calculate the probability that a homeless individual will have ongoing high costs. The information includes age, gender, time spent in jail, medical diagnoses, and use of hospital facilities. The purpose of the tool is to produce accurate and objective ranking scores for targeting scarce housing and service resources in order to make the greatest possible headway in ending chronic homelessness.