Spatial prediction of species distribution: an interface between ecological theory and statistical modelling

Abstract

Neglect of ecological knowledge is a limiting factor in the use ofstatistical modelling to predict species distribution.Three components are needed for statistical modelling, an ecologicalmodel concerning the ecological theory to be usedor tested, a data model concerning the collection and measurementof the data, and a statistical model concerning thestatistical theory and methods used. This component framework is reviewedwith emphasis on ecological theory. Theexpected shape of a species response curve to an environmental gradientis a central assumption on which agreementhas yet to be reached. The nature of the environmental predictorswhether indirect variables, e.g. latitude that have nophysiological impact on plants, or direct variables, e.g. temperaturealso influence the type of response expected.Straight-line relationships between organisms and environment areoften used uncritically. Many users of canonicalcorrelation analysis use linear (straight-line) functions to relateordination axes to variables such as slope and aspectthough this is not a necessary part of the method. Some statisticianshave used straight lines for species/environmentrelationships without testing, when evaluating new statistical procedures.Assumptions used in one component oftenconflict with those in another component. Statistical models can beused to explore ecological theory. Skewed speciesresponse curves predominate contrary to the symmetric unimodal curvesassumed by some statistical methods.Improvements in statistical modelling can be achieved based on ecologicalconcepts. Examples include incorporatinginterspecific competition from dominant species; more proximal predictorsbased on water balance models and spatialautocorrelation procedures to accommodate non-equilibrium vegetation.