Exploring the Forest Instead of the Trees: An Innovative Method for Defining Obesogenic Environments

Claudia L. Nau, Johns Hopkins University
Hugh Ellis, Johns Hopkins University
Hongtai Huang, Johns Hopkins University
Annemarie Hirsch, Geisinger Center for Health Research
Lisa Bailey-Davis, Geisinger Center for Health Research
Brian Schwartz, Johns Hopkins University and Geisinger Center for Health Research
Jonathan Pollak, Johns Hopkins University
Ann Liu, Johns Hopkins University
Thomas Glass, Johns Hopkins University

Past research has focused on assessing the association of single neighborhood characteristics with health ignoring spatial co-occurrence of multiple community-level risk factors. We demonstrate the use of random forests (RF), a non-parametric machine learning approach to identify the combination of community features that best predict obesegenic and obesoprotective environments for children. We use data from electronic health records on >160,000 children living in 1289 Pennsylvania communities and include a large number of contextual variables, previously linked to childhood obesity to analyze the joint, spatially co-occurring distribution of features of the food, land use, physical activity and social environments. This analysis allows us to (1) identify the combination of features that render an environment obesogenic, (2) determine their relative importance, and (3) provide evidence regarding the time-lag with which they operate. RF allows consideration of the neighborhood as a system of risk factors, an approach more likely to reflect residents’ experiences.