4 Discussion
We provide an efficient framework for estimating detection parameters
required for SCR studies and validating empirical study designs for
species where baseline detection data is not available. Our results
using seven empirical datasets indicate that our genotyping protocol was
highly successful, our capture and recapture rates were sufficient, and
our study design was appropriate in producing precise and reliable
density estimates. We followed the aerial survey protocol outlined in
Hettinga et al. (2012) to inform our sampling design and obtained
similar recapture rates between sampling occasions. We found that the
detection parameters g0 (detection probability) and (the spatial extent
of an individual’s use of the landscape) varied among our study
populations and between sexes (Table 2, Table S2.1, and Table S2.2). Our
results were robust to reduced sampling intensity (both in frequency and
spatially), with the best study design dependent upon range size, and
not dependent upon estimated population density or the spatial
distribution of individuals.
For multiple species, the SCR model assumption that animals are
independently and uniformly distributed over a study area is often
violated, as is the case for boreal caribou (Després-Einspenner, Howe,
Drapeau, & Kuhl, 2017; López-Bao et al., 2018; Stevenson et al., 2015).
The fission-fusion social structure and dynamics exhibited by boreal
caribou during the winter months leads to frequent exchanges between
groups (Thomas & Gray, 2002). Our simulation results show that SCR
models performed reliably; the grouping and movement patterns of boreal
caribou during our sampling period had minimal impact on the precision
or relative bias of the density estimates. We found a slight
overestimation in density estimates (Appendix 2), but the precision and
relative bias were not impacted. Few studies have looked at the effect
that non-independence of individuals has on SCR methodologies. López-Bao
et al. (2018) simulated scenarios of non-independence and spatial
aggregation of individual wolves (Canis lupus ) with only a slight
underestimation in population abundance estimates of aggregated
individuals, while Després-Einspenner et al. (2017) were unsure to what
extent the measures of uncertainty in their study of a community western
chimpanzees (Pan troglodytes verus ) were underestimated.
Study designs can be inappropriate when poorly matched with the spatial
behaviour of the target species (Williams et al., 2002). Detector arrays
that are significantly smaller than one home range, or extreme detector
spacing that leads to few or no spatial recaptures can result in biased
SCR estimates (Efford, 2011; Efford & Boulanger, 2019; Sollmann et al.,
2012; Tobler & Powell, 2013). Reducing the sampling intensity had a
greater impact on populations with smaller range sizes regardless of
density; reducing the number of transects flown led to extreme detector
spacing with few or no spatial recaptures (Figs S4.4-S4.6). Increasing
the temporal period of sampling can be an effective way of increasing
the number of detected captures and recaptures available for analysis,
which increases precision, however, increasing the temporal sampling
period can also violate the assumption of population closure and lead to
biased estimates (Dupont, Milleret, Gimenez, & Bischof, 2019). We found
that the effects of reducing the number of sampling occasions on density
estimates was influenced by the timing of the survey. If resources were
only available to perform 2, rather than 3, sampling sessions, we
recommend focusing on collecting samples early in the winter, rather
than later in the winter, as we achieved relatively unbiased estimates
(RB <20%) when retaining December, January, or February
sampling occasions. Weather conditions during March surveys were not
always favourable, with poor snow conditions and warm temperatures
creating difficulties for finding animals and identifying fresh tracks
and feeding areas.
Results from our empirical study provides a range of estimates that can
be used for simulating surveys of boreal caribou in other locations. For
poorly studied species, completing an initial empirical study is
critical for obtaining accurate detection probability estimates. Due to
the clustered, nonhomogeneous distribution of boreal caribou, extensive
sampling of the entire population is recommended to ensure that clusters
of caribou are not missed during sampling. Our subsampling scenarios
showed how less extensive sampling in smaller ranges can miss a large
portion of the population, increasing the relative bias and imprecision
of the density estimates. Applying the same sampling design to all seven
of our study populations proved to be suboptimal; detector spacing for
the smaller populations relative to sigma led to imprecise estimates.
Our analytical framework allowed us to examine the results of empirical
surveys in depth, providing confidence in the density estimates. Through
different simulations we were able to explore how relative bias and
precision of estimates vary when assumptions are violated. We showed
that the number of individuals and recaptures of individuals can be used
to predict precision, but that they cannot be used to predict relative
bias. Efford & Boulanger (2019) state that subsampling of data to
emulate different configurations of detectors, or different temporal
sampling can be prohibitively slow, due to model fitting being
computer-intensive; however, we found that even for our largest
population model (24,737 km2, 386 unique individuals,
and 545 recaptures), modelling with time and behaviour effects on bothg0 and \(\sigma\) ran relatively quickly (~7-10
days on a high-performance computer cluster) in a maximum likelihood
framework, where the density model was fitted by maximizing the
conditional likelihood.
We recommend the combination of non-invasive DNA sampling, together with
SCR modeling and distribution simulations, to be an effective, accurate
and precise approach to monitoring
wildlife.