Dynamic multistate occupancy modeling to evaluate population dynamics under a scenario of preferential sampling

Abstract

Effective conservation of animal populations depends on the availability of reliable data derived from rigorous monitoring protocols, which allows us to assess trends and understand the processes they are governed by. Nevertheless, population monitoring schemes are hampered by multiple sources of errors resulting from specific logistical and survey constraints. Two common complications are the non‐visitation of some sites in certain years and preferential sampling (PS), that is, the tendency to survey “better” sites disproportionately more often. Both factors can lead to serious biases unless accommodated into models. We used 22 yr of nest‐monitoring data to develop a dynamic multistate occupancy model, including a PS component to investigate occupancy and reproduction dynamics in a peripheral Black Stork (Ciconia nigra) population in Spain. We analyzed the effects of climate and nesting substrate (tree vs. cliff) on population dynamics and accounted for PS and non‐visitation biases using a model that distinguished three territorial states: unoccupied, occupied without, or occupied with successful reproduction. We found strong evidence for positive PS, and when accounting for this bias, lower population size estimates were generated. Black stork nests had a high probability of remaining in the same state from one year to the next, with successful nests more likely to be occupied again and to be successful the following year than occupied but unsuccessful or unoccupied nests. Nesting substrate and spring precipitation did not influence state transition probabilities or the probability of reproductive success; nevertheless, cliff nest occupancy was overall higher than tree nest occupancy. Our results highlight the importance of correcting for non‐visitation and PS in habitat occupancy models. If these potential biasing effects are not accounted for, inferences of population size may be overestimate. Multistate occupancy models with correction for PS offer a powerful analytical framework for data collected as part of population studies of unmarked animals. These models compensate for common methodological biases in biological surveys and can help implement efficient conservation strategies based on robust population dynamics estimates.

Spatio‐temporal variation in post‐recovery dynamics in a large Peregrine Falcon (Falco peregrinus) population in the Jura mountains 2000–2020†

“…2018, Fandos et al . 2021), we must model site occupancy jointly with site visitation. Otherwise, if the likelihood with which a site is surveyed depends in some way on whether it is occupied by a pair, an occupancy model that does not take into account this non‐ignorable missing‐value generation process (Rubin 1976, Gelman et al .…”

“…2018, Fandos et al . 2021, Kéry & Royle 2021: pp. 272–284, 385–397) and trend estimates may be biased.…”

“…However, to correct population size estimates for potentially non-random coverage bias, i.e. for preferential sampling (Diggle et al 2010, Pacifici et al 2012, 2016, Conn et al 2017, Monneret et al 2018, Fandos et al 2021, we must model site occupancy jointly with site visitation. Otherwise, if the likelihood with which a site is surveyed depends in some way on whether it is occupied by a pair, an occupancy model that does not take into account this non-ignorable missing-value generation process (Rubin 1976, Gelman et al 2003 will suffer from bias in the estimates of population size and trends.…”

“…An additional focus of this paper was preferential sampling (PS), which denotes a situation where the probability that a site is surveyed in some way depends on its 'quality' (Diggle et al 2010, Pacifici et al 2012, 2016, Conn et al 2017, Fandos et al 2021. Most commonly it is the better sites, which are more likely to hold a pair or to produce young, that will be surveyed in preference to inferior sites, which are known or believed to be less likely to be occupied or to have successful reproduction.…”

“…Simple, random imputation of the status of sites with missing data will then fail (Monneret et al 2018) and the non-random nature of site visitation must then be modelled jointly with the observed presence/absence patterns at the visited sites to properly impute the missing survey data. When PS is not accounted for in a model that estimates population size by extrapolating results from the surveyed to the unsurveyed sites, population size estimates will become too high (Conn et al 2017, Monneret et al 2018, Fandos et al 2021, Kéry & Royle 2021 and trend estimates may be biased. This is a widespread challenge for population studies such as ours.…”