Species occupancy estimation and imperfect detection: shall surveys continue after the first detection?

Abstract

Species occupancy, the proportion of sites occupied by a species, is a state variable of interest in ecology. One challenge in its estimation is that detection is often imperfect in wildlife surveys. As a consequence, occupancy models that explicitly describe the observation process are becoming widely used in the discipline. These models require data that are informative about species detectability. Such information is often obtained by conducting repeat surveys to sampling sites. One strategy is to survey each site a predefined number of times, regardless of whether the species is detected. Alternatively, one can stop surveying a site once the species is detected and reallocate the effort saved to surveying new sites. In this paper we evaluate the merits of these two general design strategies under a range of realistic conditions. We conclude that continuing surveys after detection is beneficial unless the overall probability of detection at occupied sites is close to one, and that the benefits are greater when the sample size is small. Since detectability and sample size tend to be small in ecological applications, our recommendation is to follow a strategy where at least some of the sites continue to be sampled after first detection.

Effect of detection heterogeneity in occupancy‐detection models: an experimental test of time‐to‐first‐detection methods

“…Mixture models, similar to the exponential mixture model we fitted, provide one approach, although these can have problems with parameter identifiability (Link ), while other approaches may require additional data collection (Guillera‐Arroita et al , Miller et al 2015). An extension to current TTD survey methods would be to record the time to second and subsequent detections, which may allow heterogeneity in abundances to be better modelled and accounted for (Guillera‐Arroita et al , Guillera‐Arroita and Lahoz‐Monfort ).…”

“…Imperfect detection needs to be accounted for when interpreting survey data because the number of sites at which a species is recorded will underestimate the true number of sites occupied, introducing a negative bias (Tyre et al ). The severity of this bias depends on the detection probability and will be greater for species that have a lower chance of being detected if present, such as hard‐to‐detect species (Durso et al , Specht et al ) or those that occur in difficult‐to‐survey habitats (Boulinier et al 1998, Gu and Swihart , Royle ).…”

“…Detection probability is derived from the proportion of repeat surveys in which the species was detected at occupied sites, which allows the model to account for imperfect detection and so avoid the negative bias induced if it is disregarded (Kéry 2002, Wintle et al ). This approach is widely advocated for species‐occupancy surveys, and increasingly intricate occupancy‐detection models have been developed to deal with data from a range of survey protocols and sampling designs (Royle et al 2005, Garrard et al , Nichols et al , Sólymos et al 2012, Field et al , Guillera‐Arroita , Specht et al ).…”

“…With any method, care must be taken to evaluate the model assumptions (Guillera‐Arroita and Lahoz‐Monfort ). For occupancy‐detection models, a key assumption is that there is no unexplained heterogeneity in detection probability; failure to meet this assumption can result in severely biased estimates of detection probability and site occupancy (Royle and Dorazio , McClintock et al 2010, Dorazio , Miller et al 2015).…”

“…It is often unclear, however, whether covariate inclusion sufficiently accounts for heterogeneity in detection probability and how to assess if this is the case or not. Goodness‐of‐fit tests, such as Bayesian posterior predictive checks and area under the receiver operating characteristic curve analyses, are commonly used to assess the fit of models to data (MacKenzie and Bailey , Zipkin et al 2012, Rodhouse et al 2015, Wright et al , Warton et al ) and information‐theoretic measures, such as Akaike's information criteria, can be used to compare the fit of models making different assumptions (Nichols et al , Riddle et al 2010, Bornand et al ). While these approaches can identify poor‐fitting models that clearly violate assumptions, they may be less successful at identifying situations where models appear to fit well but parameter estimates are biased.…”