Dustin E. Brewer | Geology Reports

USGS2025

Predicting secretive species distribution using Bayesian networks with and without expert elicitation: A case study incorporating double-blind peer review

1. Species that are secretive, imperilled and consequently data deficient often re-quire conservation action despite limited available information. In such scenarios, Bayesian networks (BNs) offer a versatile and intuitive approach for utilizing various information sources, including literature reviews, community science data sets and expert knowledge. Although it has been suggested that peer review be incorporated during expert elicitations in a BN modelling context, little information exists about how to implement this approach or about how models constructed using this approach perform. 2. We documented a double-blind peer review approach for expert elicitation in a BN modelling context. Further, we compared BN models that were generated by experts who engaged in this peer-review process (PRBNs) to those that were generated by a single expert whose knowledge was supplemented only by a literature review (LRBNs). These comparisons were based on the ability to predict the occurrence (via community science and satellite telemetry data) of a secretive and data deficient species, the King Rail ( Rallus elegans ), throughout a large region. 3. We found that the LRBNs tended to predict King Rail occurrence as well as, or better than, the PRBNs. The LRBNs that we evaluated provided more consistent predictions across our study area. However, preliminary data suggest that the PRBNs may better distinguish between locations of focal and non-focal species within smaller regions. 4. Practical implication. Our framework for utilizing double-blind peer review could serve as a useful guide and have practical implications for incorporating expert knowledge in BN models. Further, our model comparison case study suggests that, in some contexts, a single expert who uses a literature review to inform the creation of BN models may be able to accurately predict the occurrence of a secretive and data-deficient focal species. Taken together, this information could help ecologists decide when a double-blind peer review approach to expert elicitation is necessary and how to implement this approach in a BN modelling context.

Arkansas, Louisiana, Mississippi, Missouri, Tennes↗

USGS2025

Simple bagged movement models for telemetry data

Determining which statistical methods are appropriate for data is both user and data dependent and prone to change as new methodology becomes available. This process encompasses model ideation, model selection, and determining appropriate use of statistical methods. Literature on models for animal movement emerging in the past two decades has yielded a rich collection of statistical methods garnering much deserved positive attention. Among such efforts, there is limited investigation of the broader place for simple machine learning methodology in animal movement modeling. We propose a bagged (i.e., bootstrap aggregated) animal movement model using simple, off-the-shelf machine learning algorithms. The model is intuitive, retains statistical inference about characteristics of animal movement (i.e., estimated from model-based summary statistics), and only requires knowledge of elementary statistical and machine learning analysis to understand. We show by simulation that our model can provide unbiased estimates of pertinent characteristics of animal movement (e.g., daily displacement) in the presence of large and realistic location error. We believe that increasing accessible literature on simple machine learning animal movement models provides valuable pedagogical and practical support for researchers using statistical models to study animal movement.

Ecology and Evolution↗