Filling the gaps: spatial interpolation of residential survey data in the estimation of neighborhood characteristics

Abstract

The measurement of area-level attributes remains a major challenge in studies of neighborhood health effects. Even when neighborhood survey data are collected, they necessarily have incomplete spatial coverage. We investigated whether interpolation of neighborhood survey data was aided by information on spatial dependencies and supplementary data. Neighborhood "availability of healthy foods" was measured in a population-based survey of 5186 persons in Baltimore, New York, and Forsyth County (North Carolina). The following supplementary data were compiled from Census 2000 and InfoUSA, Inc.: distance to supermarkets, density of supermarkets and fruit and vegetable stores, housing density, distance to a high-income area, and percent of households that do not own a vehicle. We compared 4 interpolation models (ordinary least squares, residual kriging, spatial error regression, and thin-plate splines) using error statistics and Pearson correlation coefficients (r) from repeated replications of cross-validations. There was positive spatial autocorrelation in neighborhood availability of healthy foods (by site, Moran coefficient range = 0.10-0.28; all P<0.0001). Prediction performances were generally similar for the evaluated models (r approximately 0.35 for Baltimore and Forsyth; r approximately 0.54 for New York). Supplementary data accounted for much of the spatial autocorrelation and, thus, spatial modeling was only advantageous when spatial correlation was at least moderate. A variety of interpolation techniques will likely need to be utilized in order to increase the data available for examining health effects of residential environments. The most appropriate method will vary depending on the construct of interest, availability of relevant supplementary data, and types of observed spatial patterns.

FIGURE 1

Empirical and model semivariograms for “availability of healthy foods” scale before (“original value”) and after adjustment for supplementary data using ordinary least squares (“residual value”), by study site; Community Survey, 2004. Supplementary data were distance to supermarkets, density of supermarkets and fruit and vegetable stores, housing density, distance to a high-income area, and percent of households that do not own a vehicle. All semivariograms are omnidirectional except for New York’s original value semivariogram which is in the north/south direction.