Social Epidemiology and Spatial Epidemiology: An Empirical Comparison of Perspectives

Abstract

Social and spatial epidemiologists each bring a unique perspective to how they examine contextual or neighborhood-level determinants of health. Although both perspectives draw from epidemiology, social epidemiology is additionally grounded in sociology and causal counterfactual frameworks while spatial epidemiology is heavily influenced by medical geography and predictive models. No study to date has compared these two distinct perspectives, along with their corresponding analytical approaches and model results. Yet this comparison may advance contextual effects research in epidemiology by suggesting methodological enhancements, providing insights into the robustness of our conclusions to the perspective taken, and suggesting whether we can truly identify contextual effects from observational data. To facilitate this comparison we used both perspectives to examine a research question: What is the estimated effect of increasing neighborhood education or income on overweight/obesity, type 2 diabetes, and current smoking, independent of individual-level differences? The social epidemiology approach employed propensity score matching while the spatial approach used approximated spatial multilevel models. Data for this study came from the California Health Interview Survey (2005, 2007, 2009) and the American Community Survey (2006-2010). Results revealed minimal to no effect of neighborhood education and income on overweight/obesity, type 2 diabetes, or current smoking, but estimated effects did vary somewhat by approach. This comparison highlighted fundamentally different goals in social and spatial epidemiology: identifying causal factors to intervene compared to predicting potential causal factors to describe reality. Attempts to improve causal inference in observational studies by integrating analytical techniques across subfields will likely be hampered by different objectives and model requirements. This incompatibility for integration, lack of strong evidence of effects, and the overall identification problem cast further doubt on our ability to identify causal contextual effects using observational data. However, this work may help in the design of experiments, which is where we should now focus.