Advantages of Geographically Weighted Regression for Modeling Substrate in Streams Ken Sheehan West Virginia University Dept. of Wildlife & Fisheries June 9th, 2010 Establishment of Need • Habitat Study and Assessment – Integral to (overall) stream health – Management (present and future) – Fish and aquatic organism health – Needs improvement • Non-spatial analysis typically used • Assessment is an Expensive Endeavor Spatial Data and Streams Commonly Collected Variables – Substrate – Flow – Depth • Spatial autocorrelation (Legendre 1993) • Red herring (Diniz 2003) • Or effective new tool ? • Let’s use it to our advantage… • Geographically Weighted Regression Depth Flow Substrate Traditional Linear Regression… Fitting a line to a stream variable data set – Assumes homoskedacity • Static (flat variance) – Great for predicting relationships – Heavily used, perhaps most dominant type of statistical analysis in environmental and other fields • Classic examination of observed versus expected • Independent variables to predict dependent variables Geographically Weighted Regression • Fotheringham and Brunsden (1998) • Modification of linear regression formula to include spatial attributes of data. Standard regression formula GWR regression formula Depth + Flow + = Substrate? Study Sites ` • Research on Grayling and Wapiti Creeks, Greater Yellowstone ecosystem (Montana) • Elk River and Aaron’s Creek, WV Depth Flow 8,580 x,y coordinates Substrate * Each dot represents an x,y coordinate with depth, flow, and substrate values Results Adjusted Site Location Little Wapiti R-squared Little Wapiti 0.93 Grayling AIC Value 0.98 5742.72 0.69 0.98 0.94 0.99 0.55 Radius AdjustedModel R-Squared 6005.73 6982.44 0.82 8637.95 0.80 0.81 0.52 8947.01 0.75 0.76 9756.02 0.83 0.95 0.63 3226.54 0.85 0.94 4789.01 0.78 0.63 0.9 6668.95 0.86 8444.63 0.84 0.49 0.84 0.8 0.81 9948.32 0.85 1 R-squared R-Squared 0.92 0.82 Grayling Search 8879.35 1 0.69 2 3 0.55 3 1 0.52 2 3 0.63 1 0.63 2 1 0.49 3 2 Kernal Method AICBandwidth Model Type 8 neighbor Adaptive 8 neighbor Adaptive 8 neighbor Adaptive default (30) Fixed default (30) Fixed AICc default (30) Fixed AICc 8 neighbor Adaptive 8 neighbor Adaptive 8 neighbor Adaptive default (30) Fixed Bandwidth Parameter 10637.48 3 Bandwidth Parameter Bandwidth Parameter 11980.46 2 AICc 12214.06 1 Bandwidth Parameter 12924.04 1 Bandwidth Parameter 12925.88 3 Bandwidth Parameter AICc default (30) 17901.21 Fixed default (30) Fixed AICc AICc 2 Visual Comparison Predicted Actual Conclusions • Geographically Weighted Regression models stream substrate more effectively – Supported by AIC, adjusted R2, percent match, and visual comparison • Better assessment of streams • Management • Guides future study and • Economically efficient Acknowledgements • Dr.’s Stuart Welsh, Mike Strager, Steve Kite, Kyle Hartman • WVDNR • West Virginia University • West Virginia Cooperative Fish and Wildlife Research Unit (USGS)