Modeling US county premature mortality

In this study, ordinary least square regression and geographically weighted regression models are developed to predict premature mortality at the US county level.  The dependent variable, from among a family of mortality and morbidity metrics, is the years of potential life lost before age 75.  The OLS model is considered a global approach to predicting YPLL75 producing a single R² value, while the GWR approach produces local R² values for each observation or county in addition to a global R². The principal hypothesis of this study is that one or more independent variables will vary in explanatory power across regions of the US and GWR models provide a means to determine which variables are more salient with respect geographic disparities in mortality burden.

Independent variables are taken from the Robert Wood Johnson Foundation’s and University of Wisconsin’s Population Health Institute’s “County Health Rankings National Data” for 2013.  These variables are grouped into two classes for the purpose of this study:  socio-economic determinants and behavioral risk factors.   Using variables from these two classes, OLS regression and GWR models are compared using global R² values.  Preliminary results show that GWR models produce higher global R² values for variables selected from either class.  For the OLS regression models, residual maps show a high degree of positive spatial autocorrelation delineating statistically significant county clusters of over and under prediction, while residual maps derived from GWR show close to no spatial autocorrelation.  County or local R² maps depict county regions of high and low model explanation.