© Marcel Burkhardt
Kéry M., H. Schmid & N. Zbinden (2009)
Grundlagen der Bestandserfassung und Folgerungen für die Datenerfassung und -analyse in großräumigen Monitoringprogrammen.
Vogelwarte 47: 45–53
Large-scale monitoring programs represent a two-level, nested sampling scheme: first, a spatial sample of quadrats or other study sites is selected, within which a second sample, of individuals, occupied quadrats or species, is chosen. To produce meaningful numbers, a monitoring program ought to be based on a spatial probability sample, otherwise the inferences obtained may be biased with respect to the desired statistical population about which one wants to learn something. Moreover, all bird counts and detection-nondetection records (misleadingly also called “presence-absence data”) are binomial random variables, much like the flip of a coin. The binomial distribution is the theoretical basis of all animal or plant surveys and explains and predicts all of their most salient features: 1. repeated counts C vary automatically, even under identical conditions; 2. on average, a count amounts to a proportion p of true population size N , where p is the detection probability, and 3. any comparison between two or more counts represents the simultaneous comparison of the associated true population size N and of the detection probability p. For instance, a temporal trend in counts may be due to a genuine trend in the underlying population size or to a trend in detection probability or to a combination of the two. Any direct interpretation of counts always implies one of two assumptions, either that of p = 1 or that of p < 1 constant. It is useful to think about the genesis of bird counts in a hierarchical way. In a first random process, the true population sizes are generated. In a second random process, the actual counts are generated conditional on these true population sizes and on detection probability. For inference about the underlying true population size free from distorting effects of the observation process, extra information is required, which usually comes as distance information or as repeated observations of a system within a period of closure. Then, distance sampling and capture-recapture methods can be used to estimate true population size or true distributions, corrected for imperfect detection. During the past few years, we have used data from the Swiss breeding bird survey MHB to experiment with, adapt and develop several such methods of the capture-recapture type. Here, we review these briefly, describe some of our key findings and provide pointers to more specific work. These methods correct counts and detection-nondetection data for the binomial observation error inherent in all bird observations. We believe that use of these methods is hard to avoid in a monitoring program if absolute population size or the absolute extent of distributional ranges, corrected for imperfect detection, are required, or if one needs to correct for “dangerous patterns” in detection probability, for instance time trends in p.