Looming-Eye buoys do not reduce seabird depredation in poundnets (R script for data analysis)

Materials and methods

The experiment was carried out over a period of 46 days (2021-04-26 to 2021-06-11) at two pound net sites (’test’ and ’control’) spaced 1200 meters apart. The number of birds by species was counted at each site between 4 and 8 times a day on 8 selected dates in the time interval 07:30 to 11:40 with a total of 80 observations. Each count was made by one of three different observers, and a net was never observed by more than one observer on the same day, which precludes estimation of any observer effects. The looming eyes device was introduced at the test site between day 5 and 6 while the control site remained unchanged.

While all species of bird were counted, we focus on cormorant and gulls in this analysis, since these are the ones attracted by the opportunity to feed on garfish in the pound nets. More precisely, the following species/groups were considered: great cormorant, greater black-backed gull, herring gull, lesser black-backed gull, other small gulls, terns, and unidentified gulls. Three analyses were performed using different groupings of the bird counts as response variable:

• 1. All (counts of great cormorants and all gulls).
• 2. Cormorants only
• 3. Gulls only

For each of these, a GAM model was used to test the effectiveness of the device:

log(μ_i) = f1 (time_i) + f2 (deployTime_i) I(Looming_i) + f3(timeOfDay_i)

where f1, f2, and f3 are Duchon splines with first derivative penalization. First derivative penalization implies that splines go toward constant values beyond the data range as opposed to second derivative penalization, where trends are extrapolated. The overall effect of time is described by f1, and f2 describes the deviation from that overall pattern at the treatment site as a function of time since the looming eyes were deployed. This is accomplished by multiplying f2 by and indicator function I(Looming_i), which is zero when there was no looming eyes (i.e. either the control net or before looming eyes were introduced at the treatment net site), and one if the looming eyes device was present. Smoothness selection was carried out with the maximum likelihood (ML) method [1]. The distribution of the response variable is assumed to be either Poisson or negative binomial. Model selection is based on the Akaike information criterion (AIC).

References

[1] Simon N. Wood. Fast stable restricted maximum likelihood and marginal likelihood estimation of semiparametric generalized linear models. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 73(1):3–36, 2011.