Homelessness service provision: A data science perspective

We study homeless service provision in the United States from a data science perspective, with the goal of informing homelessness prevention efforts. We use machine learning techniques to predict household reentry into a homeless system using an administrative dataset containing both demographic and service information. This data recorded all publicly funded services provided in a Midwestern US community from 2007 through 2014. We find that several techniques can provide useful lift in the prediction task, with random forests achieving an AUC around 0.7. Prediction improves significantly when conducted within calendar years, compared to across years, suggesting that changing dynamics drive repeated need for homeless services. We also analyze key service usage patterns that are associated with lower probabilities for reentry. Counterintuitively, individuals receiving the least intensive services provided through the homelessness system exhibit significantly lower likelihoods for further system involvement compared to individuals who received more intensive services, even after accounting for initial differences through propensity score and nearest neighbor matching. These result provide intriguing insights into homelessness service delivery that need to be further probed. In particular, it is unclear whether these less intensive services sustainably address housing needs, or whether, in contrast, frustration with inadequate services drives clients away from the homelessness system. Our results provide a proof-of-concept for how data science approaches can drive interesting, socially important research in the provision of public services.