Precision and bias of methods for estimating point survey detection probabilities

Wildlife surveys often seek to determine the presence or absence of species at sites. Such data may be used in population monitoring, impact assessment, and species-habitat analyses. An implicit assumption of presence/absence surveys is that if a species is not detected in one or more visits to a site, it is absent from that site. However, it is rarely if ever possible to be completely sure that a species is absent, and false negative observation errors may arise when detection probabilities are less than 1. The detectability of species in wildlife surveys is one of the most important sources of uncertainty in determining the proportion of a landscape that is occupied by a species. Recent studies emphasize the need to acknowledge and incorporate false negative observation error rates in the analysis of site occupancy data, although a comparative study of the range of available methods for estimating detectability and occupancy is notably absent. The motivation for this study stems from the lack of guidance in the literature about the relative merits of alternative methods for estimating detection probabilities and site occupancy proportions from presence/absence survey data. Six approaches to estimating underlying detection probabilities and the proportion of sites occupied from binary observation data are reviewed. These include three parametric methods based on binomial mixtures, one nonparametric approach based on mark-recapture theory, and two approaches based on simplistic assumptions about occupancy rates. We compare the performance of each method using simulated data for which the true underlying detection rate is known. Simulated data were realized from a beta-binomial distribution, incorporating a realistic level of variation in detection rates. Estimation methods varied in their precision and bias. The binomial-with-added-zeros mixture model, estimated by maximum likelihood, was the least biased estimator of detection probability and, therefore, occupancy rate. We provide an Excel spreadsheet to execute all of the methods reviewed. Stand-alone programs such as PRESENCE may be used to estimate all models including the binomial with added zeros model. Our findings lend support to the use of maximum likelihood methods in estimating site occupancy and detectability rates.