Improving occupancy estimation when two types of observational error occur: non-detection and species misidentification

1. What they did

Occupancy studies estimate the share of sites where a species lives. Standard methods allow for false negatives: a species is missed at a site where it really occurs. They usually assume there are no false positives: a species recorded at a site where it does not occur, most often because it was misidentified. Even trained observers make such mistakes, and false positives push occupancy estimates up. An earlier model that allows for false positives (the Royle-Link model) proved hard to use in practice.

The authors built two new models that handle both kinds of error at once. Both rely on having some detections that can be trusted to be correct:

• Multiple detection state model. One survey method, but each detection is classed as “certain” or “uncertain”. The authors give the example of a sign such as scat or tracks counted as uncertain, and seeing the animal as certain.
• Multiple detection method model. Two survey methods, one that can produce false positives (for example, listening for calls) and one that is assumed not to (for example, catching and handling the animal).

They tested the models with computer simulations (1000 runs for each of 324 combinations of settings) against the standard occupancy model, which assumes no false positives. They then applied them to an experimental study with known occupancy (broadcast calls of the southern leopard frog) and to real surveys of three frog species at 124 sites in and around the Chesapeake and Ohio Canal National Historic Park, Maryland, in 2005. In the frog surveys, observers listened for calls on three or more visits, and calls were graded by intensity. Low-intensity calls (individuals can be counted) were treated as uncertain, and medium and high intensity calls (overlapping calls up to a full chorus) as certain. At 34 nearby wetlands, two independent observers also looked for adults, egg masses and tadpoles and caught them with dip nets, giving certain detections.

2. What they found

Simulations

• When false positives occurred, the standard model consistently overestimated occupancy. Both new models had less bias and more precise estimates.
• The gain was greatest when occupancy was low, when the chance of detecting the species was small, and when there were many survey visits.
• More visits made the bias from false positives worse, not better. Increasing the chance of detection or the number of sites did little to reduce it.
• Both models did better as the share of certain detections rose. The state model did worst when occupancy was low and visits were few; the method model did worst when occupancy was high and detection was low.
• When there were no false positives in the data, model selection preferred the simpler models more than 95% of the time.
• Keeping the uncertain detections in the new model gave better estimates than throwing them out and using the standard model.

Experimental study (broadcast calls)

• The standard model overestimated true occupancy by nearly 80%. When a small share of true detections were classed as certain, both new models were highly accurate.

Three frog species

• Estimated false positive rates were low: 0.030 for the bullfrog, 0.008 for the green frog and 0.027 for the pickerel frog.
• The model allowing for false positives was better supported (lower AIC) for the green frog and pickerel frog; the model without them was slightly better supported for the bullfrog.
• For all three species, occupancy estimates were lower when false positives were allowed for. The biggest difference was for the pickerel frog, where the estimate was less than half that of the standard model (roughly 0.14 versus 0.30, read from Figure 3).

3. What they concluded

• Even small rates of misidentification can badly bias estimates of how many sites a species occupies. Allowing for false positives reduces that bias and increases precision.
• The models are flexible. They can use extra information already in many survey protocols, and they can limit costly certain methods (such as handling animals) to some sites and visits.
• Methods known to produce false positives, such as large volunteer surveys, interviews with local experts, historical records, call surveys, computer call recognition and laboratory tests, should address this bias.
• More survey effort alone does not solve the problem.
• Because it can be hard to tell whether false positives are present, the authors suggest averaging results from models with and without false positives.
• The models could be extended to two species that are mistaken for each other, to use covariates such as distance, temperature or occupancy of nearby sites, and to occupancy that changes over time.

6. Other technologies, skills or innovations

• Two new occupancy models that separate certain from uncertain detections, either within one method or across two methods.
• Example R code supplied with the paper, and the models built into the free software Program PRESENCE.
• Auxiliary information (calling intensity and air temperature) used to help tell true from false detections.

7. Implications for conservation (from the paper)

• The largest bias occurred when occupancy was low. Since wrong estimates of status can have serious consequences for managing rare species, accounting for false positives is especially important for them.
• Analytical methods that account for false positives are necessary for many studies.