Analysis of Resource Selection Using Utilization Distributions

Abstract

Often resource selection functions (RSFs) are developed by comparing resource attributes of used sites to unused or available ones. We present alternative approaches to the analysis of resource selection based on the utilization distribution (UD). Our objectives are to describe the rationale for estimation of RSFs based on UDs, offer advice about computing UDs and RSFs, and illustrate their use in resource selection studies. We discuss the 3 main factors that should be considered when using kernel UD-based estimates of space use: selection of bandwidth values, sample size versus precision of estimates, and UD shape and complexity. We present 3 case studies that demonstrate use of UDs in resource selection modeling. The first example demonstrates the general case of RSF estimation that uses multiple regression adjusted for spatial autocorrelation to relate UD estimates (i.e., the probability density function) to resource attributes. A second example, involving Poisson regression with an offset term, is presented as an alternative for modeling the relative frequency, or probability of use, within defined habitat units. This procedure uses the relative frequency of locations within a habitat unit as a surrogate of the UD and requires relatively fewer user-defined options in the modeling of resource selection. Last, we illustrate how the UD can also be used to enhance univariate resource selection analyses, such as compositional analysis, in cases where animals use their range nonrandomly. The UD helps overcome several common shortcomings of some other analytical techniques by treating the animal as the primary sampling unit, summarizing use in a continuous and probabilistic manner, and relying on the pattern of animal space use rather than using individual sampling points. However, several drawbacks are apparent when using the UD in resource selection analyses. Choice of UD estimator is important and sensitive to sample size and user-defined options, such as bandwidth and software selection. Extensions to these procedures could consider behavioral-based approaches and alternative techniques to estimate the UD directly.