Because agroforestry habitat variables exhibited substantial
variation over time, we treated the six sampling periods as separate
replicates for each site. However, repeated sampling may
potentially bias tree building (De’ath and Fabricius, 2000) therefore
we used the package ‘coin’ (Hothorn et al., 2008) to examine the
independence of the response variable and the primary explanatory
variable, given site-level stratification. The null distribution
of the test statistic was calculated using Monte-Carlo resampling
and these ‘site-stratified’ p-values were calculated for the first
node of each tree and presented in the text as the ‘site-stratified
p-values’.