Discretization of continuous predictor variables in Bayesian networks: an ecological threshold approach

Abstract

Bayesian networks (BNs) are a popular tool in natural resource management but are limited when dealing with ecological assemblage data and when discretizing continuous variables. We present a method that addresses these challenges using a BN model developed for the Upper Murrumbidgee River Catchment (south-eastern Australia). A selection process was conducted to choose the taxa from the whole macroinvertebrate assemblage that were incorporated in the BN as endpoints. Furthermore, two different approaches to the discretization of continuous predictor variables for the BN were compared. One approach used Threshold Indicator Taxa Analysis (TITAN) which estimates the thresholds based on the biological community. The other approach used was the expert opinion. The TITAN-based discretizations provided comparable predictions to expert opinion-based discretizations but in combining statistical rigor and ecological relevance, offer a novel and objective approach to the discretization. The TITAN-based method may be used together with expert opinion.