Explaining and Forecasting Homelessness in the United States

Abstract: We implement a Bayesian state-space model with the dual objectives of projecting future homelessness rates in each of the 50 US states and investigating the association between homelessness and a collection of demographic and economic factors. Homelessness data is obtained from the US Department of Housing and Urban Development, while covariates are extracted from the US Census Bureau, the US Environmental Protection Agency, and Forbes Media data. We model fluctuations over time via second-order random walks and temporal smoothing with P-splines, finding that the former yields superior prediction performance on held-out data. Our final model identifies median monthly housing cost as having a significant association with state homelessness rate.