Review of Bayesian dynamic occupancy models and the implications of incomplete sampling

One popular tool for modeling long-term wildlife monitoring data is the dynamic occupancy model. Dynamic occupancy models require detection/non-detection observations recorded during repeated visits to the same site across many seasons. However, sampling all sites repeatedly for multi-year studies is not always possible due to limited funding, sites being difficult to access, the addition of sites after monitoring has begun, or even global pandemics. We explore the ramifications of incomplete sampling of sites across seasons on the modeling of dynamic occupancy data. We review four popular choices for modeling dynamic occupancy data, then, with simulation, we explore the implications of incomplete sampling on the inferences gleaned from the different models. These models are also fit to a data set monitoring American Bullfrog breeding populations along the Yellowstone River floodplain in Montana, USA. Results from the bullfrog data analyses are consistent with the results from the simulation, and help support our modeling conclusions. Ideally, if there are no limitations on data collection, we suggest repeatedly visiting all sites in all seasons, however this is often difficult, or impossible for a variety of reasons. For data already collected, regardless of incomplete sampling, we suggest using the dynamic or auto-logistic models.