A multiple detection state occupancy model using autonomous recordings facilitates correction of false positive and false negative observation errors

• Published in: Avian conservation and ecology 2019-12, Vol.14 (2), p.1, Article art1
• Main Authors: Rempel, Robert S., Jackson, Janet M., Van Wilgenburg, Steven L., Rodgers, Jennifer A.
• Format: Article
• Language: English
• Subjects: Acoustic surveying; aru; Bias; Birds; Classification; Computer simulation; Data; Detection; Error analysis; Error correction; Errors; false negative; false positive; Identification; Interpreters; misclassification; Modelling; Occupancy; occupancy modeling; Performance evaluation; Simulation; songbird survey; Species identification; survey design; Surveys
• ISSN: 1712-6568; 1712-6568
• DOI: 10.5751/ACE-01374-140201

Bird surveys have relied upon acoustic cues for species identification for decades; however, errors in detection and identification can lead to misclassification of the site occupancy state. Although significant improvements have been made to correct for false negative (FN) error, less work has been done on identifying and modeling false positive (FP) error. In our online survey we found misidentification can occur even among highly skilled observers, thus methods are required to correct for FP error. In this study we model both FP and FN error in bird surveys using a multiple detection state model (MDSM), and found that modeling both types of error lowered occupancy (ψ) relative to the FN only models in 84% of the observation data sets, and this suggests significant bias in ψ can occur in studies that do not correct for both FN and FP error. In our autonomous recording units (ARU) data we had two detection states, "confirmed" and "unconfirmed," where confirmation was based on agreement of two interpreters, and through simulation evaluated performance of the MDSM using this type of ARU data. We found that MDSM can effectively correct for both FN and FP error across a broad of range of survey observation rates and detection rates (d) and is appropriate for data using "confirmed detections." We developed a binary classification model to assign risk of bias to field observation sets based on survey and model parameters, and found that lower risk of bias cannot be predicted by a single variable or value, but rather occurs under certain combinations of low naïve occupancy rate (< ~0.2), detection rate (< ~0.2), number of confirmed recordings (< ~20) and high FP rate (> ~0.07). Our approach to interpreting ARU data along with our analysis guidelines should help reduce potential inflation of ψ resulting from FP error.