Vignette Bayesian Site Occupancy Model Bat Acoustic Data

The purpose of this document is to demonstrate how data collected by the Pacific Northwest Bat Hub, a member of the North American Bat Monitoring Program (NABat), may be used to make inferences about species occurrence. As motivation, we consider occurrence data collected on Silver-haired bats (Lasionycteris noctivagans, LANO) in Oregon and Washington in 2019. In particular, we are interested in the following: 1) understanding the relationships between various covariates and the probability of occurrence and detection, and 2) predicting species occurrence probabilities across the study area. Throughout this document, we assume the reader has a cursory understanding of the NABat sampling design and program; see @loeb2015 for a complete description of the NABat protocol. The Pacific NW Bat Hub selected cells for surveys following the NABat probabilistic master sample design [@rodriguez2019]. The analytical unit is the 10km ×\times× 10km grid cell which corresponds to the resolution of the NABat grid [@talbert2018] and we will refer to it as the "NABat grid cell" or just "cell". The NABat grid cell is the unit at which we estimate occurrence (referred to as the "site" in the occupancy literature). Our detection events are the replicated visits to a cell (referred to as a "visit" in the occupancy literature). The revisit design could arise from multiple stationary detector locations (spatial replication within a grid cell) or multiple nightly surveys (temporal replication within a grid cell). We refer to each unique revisit as a "site-night." Because bats may be using a 10-km by 10-km cell for many reasons (e.g., foraging, roosting, etc), we use the term "occurrence" as opposed to "occupancy" throughout for our statistical inferences. While we are using the traditional framework referred to in the literature as "occupancy" modeling, we maintain the language of "occurrence" throughout the vignette to emphasize that we are estimating the probability that at least one individual from a species occurred in a cell.
For our analysis, we assume that all false-positive detections are removed prior to analysis [@banner2018], enabling the use of a single-season, single-species occupancy model [@mackenzie2002]; following @banner2018, we refer to this model as the "remove" model. This document serves as a step-by-step guide to applying the remove model to an example dataset that was a portion of previously analyzed data that contained ecological interpretations (see @wright2021). We provide annotated code to clean and prepare the data for analysis; fit the single-season, single-species Bayesian occupancy model to the cleaned data; assess the convergence of the fitted model and quality of the fitted model through residual diagnostics; and produce summaries of estimated model parameters and predicted occurrence. Throughout this document, considerable emphasis is placed on the assessment of model assumptions, and code is provided to allow researchers to make rigorous assessments of these assumptions for related analyses.
Application of this vignette and the Bayesian occupancy model assumes basic understanding of Bayesian modeling and operational knowledge of Markov chain Monte Carlo (MCMC) sampling. All model fitting and inference is completed in R and we assume base proficiency in R for use of this vignette (e.g., installing packages, familiarity with tidyverse). We use the spOccupancy package developed by @doser2022 to fit the occupancy model; for more information on this package see the R help file or package webpage. The spOccupancy package utilizes the desired Bayesian occupancy model, and includes user-friendly model-fitting syntax and output. This package provides far faster model fitting compared to alternative model fitting algorithms by utilizing Póly-Gamma data augmentation to induce Gibbs updates