Area variations in health: a spatial multilevel modeling approach

Abstract

Both space and membership in geographically-embedded administrative units can produce variations in health, resulting in geographic clusters of good and poor health. Despite important differences between these two types of dependence, one is easily mistaken for the other, and the possibility that both are at work is commonly ignored. We fit a series of hierarchical and spatially-explicit multilevel models to a U.S. county-level life dataset of life expectancy in 1999 to demonstrate approaches for data analysis and interpretation when multiple sources of area-clustering are present. We demonstrate the methods to detect, interpret, and differentiate evidence of spatial and geographic membership effects and discuss key considerations for analyzing data with spatial or/and membership dimensions. We find evidence that life expectancy is driven by both within-state geographic process, and by spatial processes. We argue that considering spatial and membership processes simultaneously yields valuable insights into the patterning of area variations in health.

Fig. 1

(a) Illustrative map of adjacent counties, (b) illustrative adjacency matrix.

Fig. 2

Model 3 cross-classified structure.

Fig. 3

Model 4 cross-classified structure.

Fig. 4

Model 1 residuals exhibit spatial clustering (Moran’s I=.638).

Fig. 5

Model 2 county-level residuals exhibit spatial clustering after accounting for membership in states (Moran’s I=.289, 95% CI:.281–.298).

Fig. 6

Model 2 state-level residuals exhibit spatial clustering (Moran’s I=.641, 95% CI:.561–.707).

Fig. 7

Model 3 county-level residuals are spatially independent (Moran’s I=.008 95% CI: −.021–.038).

Fig. 8

Model 4 county-level residuals are spatially independent (Moran’s I=.006, 95% CI: −.022–.036).

Fig. 9

Model 4 state-level residuals are spatially independent (Moran’s I = −.028, 95% CI: −.180–.152).