Abstract:
Knowledge of the range, behavior, and feeding habits of large carnivores is
fundamental to their successful conservation. Traditionally, the best method to obtain feeding
data is through continuous observation, which is not always feasible. Reliable automated
methods are needed to obtain sample sizes sufficient for statistical inference. Identification of
large carnivore kill sites using Global Positioning System (GPS) data is gaining popularity. We
assessed performance of generalized linear regression models (GLM) versus classification trees
(CT) in a multi-predator, multi-prey African savanna ecosystem. We applied GLMs and CTs to
various combinations of distance travelled data, cluster durations, and environmental factors to
predict occurrence of 234 female African lion (Panthera leo) kill sites from 1,477 investigated GPS clusters. Ratio of distance moved 24 hours before versus 24 hours after a cluster was the
most important predictor variable in both GLM and CT analysis. In all cases, GLMs
outperformed our cost-complexity-pruned CTs in their discriminative ability to separate kill from
non-kill sites. Generalized linear models provided a good framework for kill site identification
that incorporates a hierarchal ordering of cluster investigation and measures to assess trade-offs
between classification accuracy and time constraints. Implementation of GLMs within an
adaptive sampling framework can considerably increase efficiency of locating kill sites,
providing a cost-effective method for increasing sample sizes of kill data.
