Predicting gray wolf landscape recolonization: logistic regression models vs. new field data

Recovery of populations of wolves (Canis lupus) and other large, wide‐ranging carnivores challenges conservation biologists and resource managers because these species are not highly habitat specific, move long distances, and require large home ranges to establish populations successfully. Often, it will be necessary to maintain viable populations of these species within mixed‐use landscapes; even the largest parks and reserves are inadequate in area. Spatially delineating suitable habitat for large carnivores within mixed, managed landscapes is beneficial to assessing recovery potentials and managing animals to minimize human conflicts. Here, we test a predictive spatial model of gray wolf habitat suitability. The model is based on logistic regression analysis of regional landscape variables in the upper Midwest, United States, using radiotelemetry data collected on recolonizing wolves in northern Wisconsin since 1979. The model was originally derived from wolf packs radio‐collared from 1979 to 1992 and a small test data set of seven packs. The model provided a 0.5 probability cut level that best classified the landscape into favorable (road density < 0.45 km/km²) and unfavorable habitat (road density > 0.45 km/km²) and was used to map favorable habitat with the northern Great Lake states of Wisconsin, Minnesota, and Michigan. Our purpose here is to provide a better validation test of the model predictions based on data from new packs colonizing northern Wisconsin from 1993 to 1997. In this test, the model correctly classified 18 of 23 newly established packs into favorable areas. We used compositional analysis to assess use of the original habitat probability classes by wolves in relation to habitat class availability. The overall rank of habitat preference classes (P, the percentage favorability from the original model), based on the new packs, was probability class 2 (P = 75–94%) > 3 (P = 50–74%) > 1 (P = 95–100%) > 4 (P = 25–49%) > 5 (P = 10–24%) > 6 (P = 0–9%). As more of the landscape becomes occupied by wolves, classes of lower probability than the 95% class, but above the favorability cut level, are slightly more favored. The 95% class is least abundant on the landscape and is usually associated with larger areas of classes 2 and 3. Wolves may continue to occupy areas of slightly lower habitat probability if adequate population source areas are present to offset the greater mortality in these lower quality areas. The model remains quite robust at predicting areas most likely to be occupied by wolves colonizing new areas based on generally available road network data. The model has also been applied to estimate the amount and spatial configuration of potential habitat in the northeastern United States.