COS 95-8 - Ignoring sampling design biases mixed-effects models for resource selection

Wednesday, August 9, 2017: 10:30 AM
C122, Oregon Convention Center
Elizabeth L. Ng, Quantitative Ecology and Resource Management, University of Washington, Seattle, WA and Timothy R. Johnson, Statistical Science, University of Idaho, Moscow, ID
Background/Question/Methods

Resource selection functions are a class of models used to estimate the relative probability of resource use by animals. Matched case-control designs are used in ecology and wildlife management to collect data to estimate resource selection functions. Recent suggestions to incorporate random effects into resource selection models have little statistical justification because they incorrectly assume unconstrained sampling of study sites, ignoring matching in the study design. Matched case-control designs have been used extensively in epidemiology, where conditional likelihood functions are used to account for constrained sampling. Here, we illustrate the discrepancies between the constrained and unconstrained models, and evaluate the bias of parameter estimates using simulation. We evaluated the performance of the mixed-effects logistic model compared to the conditional logistic model, which should produce consistent estimates.

Results/Conclusions

Conditional logistic models had the lowest bias across a wide range of sampling schemes and parameter values. In contrast, mixed-effects logistic models tended to have greater bias and poor confidence interval coverage rates. The bias was greater for small sample sizes and when there was high variability between individuals. As a result, mixed-effects models may underestimate the relative importance of habitat characteristics when estimating resource selection functions. These results illustrate the importance of sampling design considerations when formulating models for resource selection functions.