Quantifying the Spatial Issues in Human Dimensions Research

Global assessment of human-environment interactions is intrinsically hierarchical, requiring that we infer how local and regional processes will manifest themselves in regional and global conditions, respectively. But the ability to embed local intensive case studies of social and ecological processes within spatially extensive regional or global demographic and environmental data is hindered by the presence of spatial autocorrelation. Spatial autocorrelation wherein nearby points are more alike in social and ecological characteristics than are more remote points can create, as it were, hills and valleys of data similarity that correspond to real world, local or regional human communities and ecosystems that are very different one from another. Many conventional statistical techniques, notably multiple linear regression, fail in these circumstances. First, they require either uniform conditions across the sampling domain or require a priori knowledge of the form of the non-uniformity present. Second, the presence of spatial autocorrelation violates key assumptions made in such conventional statistical analysis. Following our earlier, prototype applications of classification and regression tree (CART) analysis to human-environment data, several members of the IHDP community have obtained promising results on applying CART analysis within their own studies. Yet CART analysis is itself vulnerable to statistical artefacts arising from spatial autocorrelation. We present here a method of extending CART analysis to address spatial autocorrelation issues in spatially extensive human-environment studies, using socioeconomic and ecological datasets for the conterminous United States as examples. We show how semivariogram functions can be incorporated into CART analysis to model the spatial dependence of the data. We show that, by combining conventional with spatially-aware CART analysis, one can determine where (geographically) spatial autocorrelation modifies one’s estimate of the strength of the link between humans and environment (or vice versa). Finally we use CART models to partition the spatial dependence of human-environment responses between 1) spatial patterning in the environmental predictor variables and 2) spatial patterning in human activity that is entirely independent of the environmental predictors. These techniques solve what has previously been acknowledged to be one of the biggest technical hurdles in modeling the human dimensions of the environment.