An Evaluation of the Accuracy of Kernel Density Estimators for Home Range Analysis

TLDR

Kernel density estimators are increasingly used as home‑range estimators, yet empirical studies of their performance remain scarce despite extensive theoretical interest. The study aimed to evaluate the accuracy of kernel density estimators for home‑range analysis. The authors used computer simulations to compare the area and shape of kernel density estimates to the true area and shape of multimodal two‑dimensional distributions. When least squares cross‑validation selected the smoothing parameter, the fixed kernel produced area estimates with minimal bias and the lowest surface error, whereas the adaptive kernel overestimated area and yielded higher surface error.

Abstract

Kernel density estimators are becoming more widely used, particularly as home range estimators. Despite extensive interest in their theoretical properties, little empirical research has been done to investigate their performance as home range estimators. We used computer simulations to compare the area and shape of kernel density estimates to the true area and shape of multimodal two—dimensional distributions. The fixed kernel gave area estimates with very little bias when least squares cross validation was used to select the smoothing parameter. The cross—validated fixed kernel also gave surface estimates with the lowest error. The adaptive kernel overestimated the area of the distribution and had higher error associated with its surface estimate.