A COMPARISON OF LINEAR DISCRIMINANT ANALYSIS WITH TREE-BASED CLASSIFICATIONS WITH IMPLICATIONS FOR PREDICTIVE MODELING APPROACHES TO BIOLOGICAL ASSESSMENTS

J. D. Ostermiller^1,2 and C. P. Hawkins^1,2. ¹Department of Fisheries and Wildlife, Utah State University, Logan, Utah, USA 84322-5210, ²Ecology Center, Utah State University, Logan, Utah, USA 84322-5205

RIVPACS-type assessments compare the taxa expected in the absence of pollution (E) to the taxa observed at a site (O). Sensitivities of assessments are inversely related to the magnitude of error associated with predicting E. Calculations of E require estimates of the probabilities that sites belongs to each of several previously defined groups of biologically similar sites (P_q's). In currently used models, P_q's are generated using Linear Discriminant Function Analysis (LDA). Tree-based classification methods (TC) provide a potentially more precise method for estimating P_q's and hence E, because they have less restrictive statistical assumptions and are more adept at capturing non-additive interactions among predictor variables. In this study, we compared RIVPACS-type predictive models based on both LDA and TC. Models were derived from data collected at 200 reference streams in western Oregon and Washington classified into 21 groups. Predictor variables used by LDA included latitude, longitude, elevation, slope, conductivity, %fines, %boulder/cobble, sampling date, and 3 ecoregion dummy variables; whereas the TC model used latitude, longitude, elevation, slope, conductivity, and %large-gravel. Misclassification errors were substantially higher for LDA (60%) than TC (42%). TC procedures show promise as one means of improving predictive modeling approaches (i.e. RIVPACS and BEAST) to biological assessments.