You are here:
REGRESSION MODELS THAT RELATE STREAMS TO WATERSHEDS: COPING WITH NUMEROUS, COLLINEAR PEDICTORS
Citation:
Van Sickle, J. REGRESSION MODELS THAT RELATE STREAMS TO WATERSHEDS: COPING WITH NUMEROUS, COLLINEAR PEDICTORS. Presented at American Geophysical Union annual meeting, San Francisco, CA, December 8-12, 2003.
Description:
GIS efforts can produce a very large number of watershed variables (climate, land use/land cover and topography, all defined for multiple areas of influence) that could serve as candidate predictors in a regression model of reach-scale stream features. Invariably, many of these candidate predictors are correlated with each other, leading to ambiguities in regression model choice, prediction and interpretation. Strategies for coping with collinearity and model choice are illustrated, using watershed and stream data collected from the Willamette Basin in Oregon. Collinearity greatly inhibits one's ability to determine the relative importance of individual predictors. Sometimes it is easier and clearer to assess the relative importance of groups of conceptually-similar predictors (for example, "natural-gradient" versus "human disturbance" predictors). In addition, one can estimate the amount of model-explained variance in a response variable that is "shared" by multiple predictors and hence cannot be disentangled. Finally, if multiple, collinear variables are considered as candidate predictors, then model development will likely yield several "best" models, all of them believable and all of them nearly equal in their quality of fit. In this case, one can carry out model averaging of predictions, of relative importance measures, or of the effects of a particular watershed predictor on the stream response variable.