Modeling Obesity Rate with Spatial Auto-correlation: A Case Study

The geographical pattern, if any, is an essential factor to consider when modeling obesity rate (areas close to each other with comparable area-specific characteristics often have similar obesity rates); however, non-spatial statistical models assume that the obesity rate is spatially homogeneous, i.e., the rate operates similarly, everywhere. This strong assumption might not always hold in practice. Hence, the objective of this study is to demonstrate how to incorporate spatial auto-correlation from observed data into a statistical model as a case study for modeling obesity rates across 117 health regions in Canada. To achieve this objective, the current research formulates a non-spatial model, a random effect model (unstructured random effect), and several spatial models (structured random effect) using the Bayesian hierarchical formulation. The obesity rate, along with 15 socio-demographic and environmental characteristics at the health region level, was collected and published by Statistics Canada. The model performances were compared using information criteria, cross-validation, and residual analysis. The percentages of the immigrant population and graduates from health regions were negatively associated with the obesity rate. The models identified several obesity clusters across Canada when the estimated rates were mapped. While this study sets an example for applied researchers to develop a parsimonious and robust model, the reported results could aid the public health researchers in the development of a more focused or locally adapted public health policy and planning for obesity prevention.