Selecting Spatial Scale of Contextual Covariates to Explain Childhood Obesity

• Activity Number: 516
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 11:15 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Lauren Grant*+ and David Wheeler and Chris Gennings
• Companies: Virginia Commonwealth University and Virginia Commonwealth University and Virginia Commonwealth University
• Keywords: childhood obesity ; spatial scale ; model selection

Abstract

Area-level covariates used as contextual variables in regression models are typically assumed to all be appropriately modeled at the same spatial scale (geographic areal unit). Generally, a smaller spatial unit (i.e., census block group versus a county) is thought to better capture heterogeneity in regression relationships. However, the assumption that all contextual variables should be modeled at the same spatial scale needs exploration. Conveniently, the selection of spatial scale may be viewed as a model selection problem. In this research, we developed a family of algorithms to select the best spatial scale for each neighborhood-level variable entered into a regression model to explain childhood obesity among a large sample of patients of the Virginia Commonwealth University Medical Center. We modified the forward stepwise, forward stagewise, and lasso algorithms and built models considering a large number of socioeconomic covariates at the census block, block group, and tract levels. Our results demonstrate that neighborhood-level socioeconomic effects operated at different spatial scales in regression models to explain childhood obesity.