Occupancy estimation and modelling the effects of
covariates:
Studies have demonstrated that spatial replication can serve as a good
surrogate for temporal replication in occupancy studies of sloth bears
(Srivathsa et.al.2017). For our grids we considered 1 km search trail as
one spatial replicate. We employed an equal search effort of 4 km in
each grid, thus maintaining four spatial replicates in each grid.
Detection histories were constructed for each grid, where ‘1’ indicates
detection of species and ‘0’ indicates non-detection and ‘.‘ indicates a
missing observation. For example a detection history of ‘010-‘ indicates
that the sloth bear or its sign was not detected in first and third
search trail, detected in the second search trail and the sampling was
not done in the fourth search trail. For each covariate, data recorded
in each segment along the search trail were pooled to obtain an single
average covariate score. We ran single-species single season occupancy
analysis using a maximum likelihood-based approach in the PRESENCE
software 2.12.31to derive calculated occupancy (Hines, 2006). We
followed a two-step process to estimate probability of detection and
probability of bear occurrence. First, we modelled detection (p) keeping
constant structure for occupancy model as ψ (.). We hypothesized that
three ground based covariates 1) Termite 2) Fruit and 3) Disturbance
Index would affect our probability of detecting sloth bears and its
signs along the search trail, so we used them in the first step for
modelling detection probability. We hypothesized that sloth bear signs
detection will be higher in areas with termite mounds and fruits and
they will be lower in areas with high disturbance. We modelled different
combinations of the detectability (p) covariates and selected the best
model based on minimum AIC, keeping the ψ fixed.
In the second step, we modelled probability of occupancy (ψ) keeping the
top detection model from step one as a constant structure for detection
model (Srivathsa et al. 2017; Panthi et al. 2017; Doherty et al. 2012).
We constructed a set of 27 priori candidate models, each representing a
different ecological hypothesis. These models included either single or
additive effects of two or more covariates to investigate the influence
of covariates on occurrence. Model fit was assessed using the parametric
bootstrap procedure (MacKenzie & Bailey, 2004). The covariate models
were compared and ranked using an information theoretic approach,
relying on Akaike Information Criterion (AIC) for testing relative model
fits. Models with ΔAIC of <2 were strongly supported by the
data. Due to inherent advantage of model averaging (Burnham & Anderson,
1998), the final occupancy estimates and associated standard error were
averaged across the model set. To infer relative influence of covariates
on occurrence, we summed the computed model weights of all the model
containing the particular covariate. Additionally, we used the estimated
β-coefficients to assess the strength of association of each covariate
with occupancy probability.