Accounting for Spatial Dependence in Bayesian Belief Networks

• Activity Number: 81
• Type: Contributed
• Date/Time: Monday, August 4, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics & the Environment
• Author(s): Alix I. Gitelman*+
• Companies: Oregon State University
• Keywords: Bayeian belief network ; spatial models ; ecological data

Abstract

Bayesian belief networks and structure equation models have seen increased use in modeling of ecological processes (e.g., Malaeb, Summers & Pugesek 2000; Lee 2000). The spatial independence assumption used in fitting these models, however, is more often than not violated when data are sampled close together in space, which is often the case with ecological data. Two methods for addressing this shortfall are presented. In the first, when a single response variable can be specified, spatial dependence is modeled in the residuals of that response after fitting the Bayes network to the entire system. In the second method, spatial indices for relevant covariates are introduced as additional nodes in the Bayes network to account for autocorrelation in those variables. Both methods are applied to data regarding the health of macroinvertebrates in Mid-Atlantic region watersheds in an effort to understand the mechanisms by which that health may be compromised. In both cases, the accounting for spatial dependence alters the inferences from the independence case, providing more realistic assessments of the complex system.