Using informative priors to account for identifiability issues in occupancy models with identification errors

Abstract

AbstractCamera traps, autonomous recording units and environmental DNA metabarcoding have become important non-invasive monitoring techniques to collect environmental data with the objective to better understand species distribution. These new data have fueled the development of statistical models to suit with specific sampling designs and get reliable ecological inferences.Among others, site occupancy models enable the estimation of occurrence patterns for one or more target species, considering that their presence may be observed or not. In that context, it is useful to distinguish a) the detection process, which is unobserved and corresponds to the situation where the species of interest leaves signs of its presence (by triggering a camera-trap for example), and b) the identification process, which allows to recognize the species of interest from its signs of presence. Detection and identification are both imperfect processes in which false-negative and false-positive errors may occur, especially when collected data are massive and require automated treatment. Misclassification at both steps can lead to significant biases in ecological estimates, and several extensions have been proposed to correct for these potential errors.The naive occupancy model does not account for false detection since it results in identifiability issues, hence the existence of alternative models that combine several data sources to allow the identification of error parameters. These models however are more data-hungry and require to include a source of data without false-positives. As an alternative, we propose to reuse available data from the identification process in a Bayesian framework through an informative prior, in order to overcome the identifiability issue and reduce estimation bias. We compare through a simulation study these different approaches considering various monitoring designs (e.g., sensors or eDNA data, for species monitoring or disease ecology).Overall, what is at stake is enhancing statistical methods together with sampling non-invasive technologies, in a way to provide ecological outcomes suitable for conservation decision-making.