Cluster Analyses and Preferred Associations
To determine if the clustering of core units into clans was consistent
over time we used SOCPROG (v.2.9: Whitehead, 2009) and split the data
into two sampling periods to compare metrics by year: August
28th, 2017 – August 22nd, 2018
(sample period 1) and August 29th, 2018 – May
13th, 2019 (sample period 2). We created an
association matrix for each year by calculating the simple association
index of each dyad, where a value of 1 indicates the two core units were
always in association and 0 indicates they were never in association.
The simple association index (AI) was chosen because we were always able
to positively identify all core units in association with the focal unit
(Whitehead, 2008). AI was calculated as AI = N AB/ (N A + N B) or the number
of times that two core units were in association during scans, divided
by the total number of scans where either unit was present. Of four
different clustering methods (average linkage, Ward’s weighted, complete
linkage, and single linkage), the average linkage method had the highest
cophenetic correlation coefficient (CCC = 0.891), so this was the
clustering method that we used for hierarchical cluster analyses (see
Stead & Teichroeb, 2019). We used dendrograms (Fig.2) created through
the average linkage method to compare clustering into clans between
sample periods. Previous work examining the graph of cumulative
bifurcations from the dendrogram from sample period one showed one
significant knot at an AI of 0.05 (Stead & Teichroeb, 2019), so this
was the cut-off that we used to determine clan associations. We then
conducted permutation tests for preferred/avoided associations using
SOCPROG, and permuted association matrices 10,000 times to stabilize
p-values. The network may not be static throughout the sample periods,
and thus the results of this test could not reveal the variability
within each sample period. Further analyses using smaller time windows
was performed to adjust for this.