How biased are estimates of extinction probability in revisitation studies?

Summary Extinction is a fundamental topic for population ecology and especially for conservation and metapopulation biology. Most empirical studies on extinction resurvey historically occupied sites and estimate extinction probability as the proportion of sites where a species is no longer detected. Possible non‐detection of surviving populations is usually not accounted for, which may result in extinction probabilities that are overestimated. As part of a large revisitation study in north‐east Switzerland, 376 sites with historically known occurrences of a total of 11 plant species 80–100 years ago were visited by two independent observers. Based on typical population size, ramet size and plant architecture, we judged six species as easy to find and five species as hard to find. Using capture–recapture methods to separate non‐detection from true extinction, we gauged the bias of extinction probability estimates that do not account for non‐detection. When non‐detection was not accounted for, a single visit resulted in an average estimate of population extinction probability of 0.49 (range 0.27–0.67). However, the mean detection probability of a surviving population during a single visit had an estimated average of only 0.81 (range 0.57–1). Consequently, accounting for non‐detection resulted in extinction probability estimates ranging between 0.09 and 0.61 (mean 0.36). Based on a single survey, our revisitation study would have overestimated the extinction rate on average by 11 percentage points (range 5–22%) or by 59% (range 0–250%) relative to the estimated true value. A simple binomial argument enables the calculation of the minimum required number of visits to detect a surviving population with high probability (e.g. 95%). For the easy to find species, approximately two visits would be required to find most of the surviving populations, whereas up to four visits would be required for the hard to find species. In revisitation studies, only repeated revisits allow the separation of extinction from simple non‐detection. Unless corrected for possible non‐detection, extinction probability may be strongly overestimated, and hence some control for non‐detection is desirable at least in a subset of species/sites in revisitation studies. These issues are also relevant to the estimation of extinction in metapopulation studies and to the collection of quality data for habitat and distribution models.