Open AccessDeveloping spatially‐explicit weighting factors to account for bias associated with missed GPS fixes in resource selection studies
Author(s) -
Webb Stephen L.,
Dzialak Matthew R.,
Mudd James P.,
Winstead Jeffrey B.
Publication year - 2013
Publication title -
wildlife biology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.566
H-Index - 52
eISSN - 1903-220X
pISSN - 0909-6396
DOI - 10.2981/12-038
Subject(s) - weighting , statistics , logistic regression , odocoileus , global positioning system , vegetation (pathology) , geography , regression , computer science , mathematics , cartography , ecology , medicine , telecommunications , pathology , biology , radiology
Global positioning system (GPS) collars are prone to locational error and missed fixes caused by vegetation and topography, meaning that locational error may be greater, or fix success lower, in certain habitats. These forms of error can lead to bias associated with data loss or censoring. The goals of this paper were to: 1) estimate resource selection functions using logistic regression to map probability of acquisition (P acq ) of a GPS location and subsequent censoring of locational error in relation to landscape features and 2) develop a spatially‐explicit map of weighting factors across the landscape to avoid over‐ or underestimating resource selection. Female mule deer Odocoileus hemionus were used as a case example and to validate maps. Locational error and P acq were influenced by vegetation and topography, thus necessitating a means to weight the data. Applying logistic regression to quantify P acq allowed an easy and straightforward approach to mapping P acq and subsequently, weighting factors (weight = 1/P acq ). Weighting landscape characteristics improved validation of deer‐occurrence maps compared to using the original, unweighted landscape values. Using the best validating deer‐occurrence map, we found that 87.5‐90.2% of locations (N = 1,043) from an independent sample of deer (N = 4) occurred within the highest probability of use bin (∼ 20% of the landscape); 95.4‐96.9% of independent locations occurred within the two highest probability of use bins (∼ 40% of the landscape). By accounting for, and modeling, missed GPS fixes and locational error, we improved the predictive ability of maps based on an independent sample of deer. Without correction (i.e. weighting) factors, the importance of habitat types and terrain features may be over‐ or underestimated, which could have serious consequences when interpreting resource selection by animals and developing management recommendations.