A Tale of Two Models: a Non-Spatial and Spatial Regression Analysis of Poverty in Oklahoma

Abstract

When data is associated with a spatial attribute, observed values of a variable at one location may influence values of that variable at neighboring locations. Failure to properly account for spatial dependence between observations in statistical models can produce biased coefficient estimates and compromise the quality of prediction. Spatial regression models incorporate this spatial dependence to produce more appropriate results. This project seeks to identify key determinants of poverty at a county and census tract level in Oklahoma using data from the U.S. Census Bureau’s American Community Survey. I fit both a non-spatial and spatial autologistic models and test for the presence of spatial dependence in the residuals to determine which coefficient estimates are more appropriate. Spatial autocorrelation is tested for in the dependent variable using join-count statistics, which test the extent to which the spatial pattern in binary data are clustered, dispersed or random. I find that poverty rates in Oklahoma are spatially autocorrelated, and while the coefficient estimates for both models are not drastically different, the spatial autologistic model eliminates spatial autocorrelation in the county level model. These findings illustrate the value in examining and incorporating spatial attributes into regression analyses. Specifically, the Oklahoma spatial models will allow policy-makers interested in lowering poverty rates to understand how geography relates to pover