Twenty years of dynamic occupancy models: a review of applications and look to the future

Since their introduction over 20 years ago, dynamic occupancy models (DOMs) have become a powerful and flexible framework for estimating species occupancy across space and time while accounting for imperfect detection. As their popularity has increased and extensions have further expanded their capabilities, DOMs have been applied to increasingly diverse datasets and research objectives in applied ecology. At the same time, technological advancements have resulted in massive increases in available data, offering both new opportunities and challenges for users of DOMs. Given these developments, it is timely to examine common practices in building these models to understand the breadth of modelling approaches, determine potential vulnerabilities, and identify priorities for future research. We reviewed a sample of articles that have fit DOMs in the past 20 years, examining the contexts of their application and the approaches taken to the model‐building process. We find that these models have been used to pursue diverse objectives, based on datasets with wide‐ranging spatial and temporal scales collected using a variety of survey methods. Our comparisons of modelling approaches indicate that many applications of DOMs considered relatively few covariates on key model parameters, as well as a tendency towards linear responses over more complex non‐linear or interactive forms. Model selection techniques were largely idiosyncratic with little consensus on the best approaches, and model evaluation was rare across reviewed applications. Based on these findings we highlight aspects of the modelling process that merit closer attention, such as the possible impacts of low complexity and missing drivers of heterogeneity on model performance, the uncertainties around robust and appropriate model selection techniques for different contexts, and the need for trusted and reliable tools for model assessment and evaluation.