Hybridizing Geographically Weighted Regression and Multilevel Models: A New Approach to Capture Contextual Effects in Geographical Analyses

Abstract

Multilevel models are one of the main statistical methods used in modeling contextual effects in social sciences. A common limitation of these methods is the use pre-set boundaries—usually administrative units—to define contexts, when these boundaries do not always match up with the “true” causally relevant contexts that may affect the outcomes of interest. In this study applied to the obesity geography in the Paris area (France), we propose a new spatially explicit two-step procedure to tackle this methodological issue. The first step consists in estimating a geographically weighted regression model, then using it to reveal and delineate relevant nonstationarity-based data-driven spatial contexts, and finally including them as a random effect into a random slope multilevel model. In applying this hybrid methodology for modeling body mass index within a sample of 9,089 French adults, we demonstrate that it outperforms administrative-based multilevel models in terms of decreasing Akaike information criteria, and is better at accounting for contextual effects through intraclass correlation coefficient and increasing slope variance. We suggest that this procedure might be generalized to quantitative geographical analyses involving contextual effects.